Intelligent and Fuzzy Techniques: Smart and Innovative Solutions

Advances in Intelligent Systems and Computing 1197
Cengiz Kahraman · Sezi Cevik Onar ·

Basar Oztaysi · Irem Ucal Sari ·
Selcuk Cebi · A. Cagri Tolga Editors
Intelligent and
Fuzzy Techniques:
Smart and
Innovative
Solutions
Proceedings of the INFUS 2020
Conference, Istanbul, Turkey,
July 21–23, 2020
Advances in Intelligent Systems and Computing
Volume 1197
Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
Advisory Editors
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
Rafael Bello Perez, Faculty of Mathematics, Physics and Computing,
Universidad Central de Las Villas, Santa Clara, Cuba
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
Hani Hagras, School of Computer Science and Electronic Engineering,
University of Essex, Colchester, UK
László T. Kóczy, Department of Automation, Széchenyi István University,
Gyor, Hungary
Vladik Kreinovich, Department of Computer Science, University of Texas
at El Paso, El Paso, TX, USA
Chin-Teng Lin, Department of Electrical Engineering, National Chiao
Tung University, Hsinchu, Taiwan
Jie Lu, Faculty of Engineering and Information Technology,
University of Technology Sydney, Sydney, NSW, Australia
Patricia Melin, Graduate Program of Computer Science, Tijuana Institute
of Technology, Tijuana, Mexico
Nadia Nedjah, Department of Electronics Engineering, University of Rio de Janeiro,
Rio de Janeiro, Brazil
Ngoc Thanh Nguyen , Faculty of Computer Science and Management,
Wrocław University of Technology, Wrocław, Poland
Jun Wang, Department of Mechanical and Automation Engineering,
The Chinese University of Hong Kong, Shatin, Hong Kong
The series “Advances in Intelligent Systems and Computing” contains publications
on theory, applications, and design methods of Intelligent Systems and Intelligent
Computing. Virtually all disciplines such as engineering, natural sciences, computer
and information science, ICT, economics, business, e-commerce, environment,
healthcare, life science are covered. The list of topics spans all the areas of modern
intelligent systems and computing such as: computational intelligence, soft comput-
ing including neural networks, fuzzy systems, evolutionary computing and the fusion
of these paradigms, social intelligence, ambient intelligence, computational neuro-
science, artificial life, virtual worlds and society, cognitive science and systems,
Perception and Vision, DNA and immune based systems, self-organizing and
adaptive systems, e-Learning and teaching, human-centered and human-centric
computing, recommender systems, intelligent control, robotics and mechatronics
including human-machine teaming, knowledge-based paradigms, learning para-
digms, machine ethics, intelligent data analysis, knowledge management, intelligent
agents, intelligent decision making and support, intelligent network security, trust
management, interactive entertainment, Web intelligence and multimedia.
The publications within “Advances in Intelligent Systems and Computing” are
primarily proceedings of important conferences, symposia and congresses. They
cover significant recent developments in the field, both of a foundational and
applicable character. An important characteristic feature of the series is the short
publication time and world-wide distribution. This permits a rapid and broad
dissemination of research results.
** Indexing: The books of this series are submitted to ISI Proceedings,
EI-Compendex, DBLP, SCOPUS, Google Scholar and Springerlink **
More information about this series at http://www.springer.com/series/11156

Cengiz Kahraman Sezi Cevik Onar
• •
Basar Oztaysi Irem Ucal Sari

• •
Selcuk Cebi A. Cagri Tolga

•
Editors
Intelligent and Fuzzy

Techniques: Smart
and Innovative Solutions
Proceedings of the INFUS 2020 Conference,
Istanbul, Turkey, July 21-23, 2020
123
Editors
Cengiz Kahraman Sezi Cevik Onar
Department of Industrial Engineering Department of Industrial Engineering
Istanbul Technical University Istanbul Technical University
Istanbul, Turkey Istanbul, Turkey
Basar Oztaysi Irem Ucal Sari

Department of Industrial Engineering Department of Industrial Engineering
Istanbul Technical University Istanbul Technical University
İstanbul, Turkey Istanbul, Turkey
Selcuk Cebi A. Cagri Tolga

Industrial Engineering Department Industrial Engineering Department
Yildiz Technical University Galatasaray University
Istanbul, Turkey Istanbul, Turkey
ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-3-030-51155-5 ISBN 978-3-030-51156-2 (eBook)
https://doi.org/10.1007/978-3-030-51156-2
© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
INFUS is an acronym for Intelligent and Fuzzy Systems. It is a well-established

international research forum to advance the foundations and applications of intel-
ligent and fuzzy systems, computational intelligence, and soft computing for
applied research in general and for complex engineering and decision support
systems.
The principal mission of INFUS is to construct a bridge between fuzzy &
intelligence systems and real complex systems via joint research between univer-
sities and international research institutions, encouraging interdisciplinary research
and bringing multidiscipline researchers together.
INFUS 2019 was organized in July 21–23, 2019, in Istanbul, Turkey. The
number of participants was about 200. The proceedings of INFUS 2020 was
published by Springer publishing house with an excellent quality before the con-
ference began. INFUS 2020 proceedings is again published by Springer under
“Advances in Intelligent Systems and Computing Series” and will be indexed in
Scopus. Its title is Intelligent and Fuzzy Techniques: Smart and Innovative
Solutions. Moreover, special issues of indexed journals will be devoted to a strictly
refereed selection of extended papers presented at INFUS 2020.
Smart technologies, when integrated with data analytics platforms and networks,
offer the potential of expanding the fundamental capacity of any system.
Organizations that invest in smart technologies can see rewards including higher
revenues, an improved customer experience, and increased employee satisfaction.
Smart Solutions let you cost-effectively achieve and manage your objectives for
efficiency, capacity, and availability. Innovative technologies in all business seg-
ments provide significant competitive advantages for successful development.
Innovative decisions should be based on the clearly formulated innovative strate-
gies and fit into business strategies. Innovative decisions start with formation of an
updated information field regarding the innovations being implemented and pos-
sible for the implementation in all operating segments of the business. INFUS 2020
focuses on smart and innovative technologies and solutions.
v
vi Preface
Our invited speakers this year are Prof. Krassimir Atanassov, Prof. Vicenc Torra,
Prof. Janusz Kacprzyk, and Prof. Ahmet Fahri Özok. It is an honor to include them
in our conference program. We appreciate their voluntary contributions to INFUS
2020, and we hope to see them at INFUS conferences for many years.
In the beginning of the planning process, we had planned to organize INFUS
2020 in Izmir at Izmir Katip Celebi University as the host. Unfortunately, coron-
avirus pandemic prevented it as all of you know. We hope to organize an interactive
conference in 2021 with your participation in Izmir. Our social program of INFUS
2020 in Izmir will be exactly realized at INFUS 2021. We thank all of you very
much since you did not give up your participation to INFUS 2020. We appreciate
your sincerity and fidelity.
This year, the number of submitted papers became 345. After the review process,
about 40% of these papers have been rejected. The distribution of the remaining
papers is as follows from the most to the least: Turkey, Russia, China, Iran, Poland,
India, Azerbaijan, Bulgaria, Morocco, Spain, Algeria, Serbia, Ukraine, Pakistan,
Canada, South Korea, UK, Indonesia, USA, Vietnam, Finland, Romania, France,
Uzbekistan, Italy, and Austria. We again thank all the representatives of their
countries for selecting INFUS 2020 as an international scientific arena.
We also thank the anonymous reviewers for their hard works in selecting
high-quality papers of INFUS 2020. Each of the organizing committee members
provided invaluable contributions to INFUS 2020. INFUS conferences would be
impossible without them. We hope meeting you all next year in Turkey.
Cengiz Kahraman
Selcuk Cebi
Basar Oztaysi
Sezi Cevik Onar
Irem Ucal Sari
A. Cagri Tolga
Contents
Invited Speakers’ Papers

Fuzzy Meets Privacy: A Short Overview . . . . . . . . . . . . . . . . . . . . . . . . 3
Vicenç Torra and Guillermo Navarro-Arribas
Intelligent Planning of Spatial Analysis Process
Based on Contexts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Stanislav Belyakov, Alexander Bozhenyuk, Janusz Kacprzyk,
and Igor Rozenberg
The Method of Finding the Base Set of Intuitionistic Fuzzy Graph . . . . 18
Alexander Bozhenyuk, Stanislav Belyakov, Janusz Kacprzyk,
and Margarita Knyazeva
Intuitionistic Fuzzy Assessments of the Abdominal Aorta
and Its Branches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Valetin Vasilev, Evdokia Sotirova, Krassimir Atanassov, and Sotir Sotirov
Mathematical Philosophy and Fuzzy Logic . . . . . . . . . . . . . . . . . . . . . . 32
Ahmet Fahri Özok
Clustering
Segmentation of Retail Consumers with Soft Clustering Approach . . . . 39
Onur Dogan, Abdulkadir Hiziroglu, and Omer Faruk Seymen
Segmentation Analysis of Companies’ Natural Gas Consumption
by Soft Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Onur Dogan
Comparison of Fuzzy C-Means and K-Means Clustering
Performance: An Application on Household Budget Survey Data . . . . . 54
Songul Cinaroglu
vii
viii Contents
A Hybrid Approach for Clustering and Selecting of Cloud Services

Based on User Preferences Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Ouassila Hioual, Ouided Hioual, and Sofiane Mounine Hemam
Basket Patterns in Turkey: A Clustering of FMCG Baskets Using
Consumer Panel Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Tolga Kaya, Ahmet Talha Yiğit, and Utku Doğruak
Journey Segmentation of Turkish Tobacco Users Using Sequence
Clustering Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Ahmet Talha Yiğit, Tolga Kaya, and Utku Doğruak
A Survey on Spherical Fuzzy Sets and Clustering the Literature . . . . . . 87
Eren Ozceylan, Baris Ozkan, Mehmet Kabak, and Metin Dagdeviren
Picture Fuzzy Sets and Spherical Fuzzy Sets

Picture Fuzzy Linear Assignment Method and Its Application
to Selection of Pest House Location . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Fatma Kutlu Gundogdu
Simple Additive Weighting and Weighted Product Methods Using
Picture Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Fatma Kutlu Gundoğdu and Eda Bolturk
Evaluating Sustainable Vehicle Technologies for Freight
Transportation Using Spherical Fuzzy AHP and TOPSIS . . . . . . . . . . . 118
Miguel Jaller and Irem Otay
Evaluating Strategic Entry Decisions Using Spherical Fuzzy Sets . . . . . 127
Sezi Cevik Onar, Cengiz Kahraman, and Basar Oztaysi
Spherical Fuzzy Cost/Benefit Analysis of Wind Energy Investments . . . 134
Sezi Cevik Onar, Basar Oztaysi, and Cengiz Kahraman
A Fuzzy Pricing Model for Mobile Advertisements by Using Spherical
Fuzzy AHP Scoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Basar Oztaysi, Sezi Cevik Onar, and Cengiz Kahraman
A Framework for Selection of the Best Food Waste Management
Alternative by a Spherical Fuzzy AHP Based Approach . . . . . . . . . . . . 151
Aysu Melis Buyuk and Gul Tekin Temur
Using Spherical Fuzzy AHP Based Approach for Prioritization
of Criteria Affecting Sustainable Supplier Selection . . . . . . . . . . . . . . . . 160
Yagmur Unal and Gul T. Temur
Contents ix
Estimation and Prediction

Estimating Shopping Center Visitor Numbers Based on Various
Environmental Indicators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Cagatay Ozdemir, Sezi Cevik Onar, and Selami Bagriyanik
Predicting Customers’ Next-to-Be Purchased Products . . . . . . . . . . . . . 180
Buse Mert, Defne İdil Eskiocak, and Işıl Öztürk
Estimating Breathing Levels of Asthma Patients with Artificial
Intelligence Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Asım Sinan Yüksel and Fatma Gülşah Tan
The Problem of Selection with the Fuzzy Axiomatic Design of MEMS
Based Sensors in Industry 4.0 Predictive Maintenance Process . . . . . . . 195
Arif Sercan Erez, Mehmet Cakmakci, and Rasit Imrenci
Fuzzy Based Noise Removal, Age Group and Gender Prediction
with CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Ali Tunç, Sakir Taşdemir, and Murat Köklü
Click Prediction in Digital Advertisements: A Fuzzy Approach
to Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Ahmet Tezcan Tekin, Tolga Kaya, and Ferhan Çebi
Neural Fuzzy Inference Hybrid System with Support Vector Machine
for Identification of False Singling in Stock Market Prediction
for Profit Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Bhupinder Singh and Santosh Kumar Henge
Estimation of Potential Locations of Trade Objects on the Basis
of Fuzzy Set Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
Alekperov Ramiz Balashirin Oglu and Salahli Vuqar Mamadali Oglu
Surface Roughness Prediction Using ANFIS and Validation
with Advanced Regression Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 238
M. A. Vinod Kumar
Predicting Likelihood to Purchase of Users for E-commerce . . . . . . . . . 246
Çağlar İçer, Deniz Parmaksız, and Altan Çakır
Hesitant Fuzzy Sets

Webpage Design Based on Generation Differences Using Hesitant
Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
M. Çağrı Budak and Sezi Çevik Onar
A Hesitant Fuzzy Linguistic Group Decision Making Model
for Energy Storage Unit Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Ahmet Aktas and Mehmet Kabak
x Contents
Solar Energy Power Plant Investment Selection with Unbalanced

Hesitant Fuzzy Linguistic MULTIMOORA Method Based
Score-HeDLiSF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
Veysel Çoban and Sezi Çevik Onar
Application of Linear Programming Model in Multiple Criteria
Decision Making Under the Framework of Interval-Valued Hesitant
Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
Tabasam Rashid and M. Sarwar Sindhu
Neutrosophic Sets
Evaluation of the Fourth Party Logistics Provider Under
Neutrosophic Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
Serhat Aydın, Mehmet Yörükoğlu, and Mehmet Kabak
A* Algorithm Under Single-Valued Neutrosophic
Fuzzy Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
Esra Çakır and Ziya Ulukan
Bipolar Neutrosophic Fuzzy Dijkstra Algorithm and Its Application . . . 311
Esra Çakır and Ziya Ulukan
Physician Selection with a Neutrosophic Multi-criteria Decision
Making Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
Ahmet Sarucan, M. Emin Baysal, and Orhan Engin
Intuitionistic Fuzzy Sets

An Intuitionist Fuzzy Method for Discovering Organizational
Structures that Support Digital Transformation . . . . . . . . . . . . . . . . . . . 331
Zineb Besri and Azedine Boulmakoul
Modeling Humanoid Robots Mimics Using Intuitionistic Fuzzy Sets . . . 339
Cengiz Kahraman, Sezi Cevik Onar, Basar Oztaysi, and Irem Otay
Understanding the Blockchain Technology Adoption from
Procurement Professionals’ Perspective - An Analysis of the
Technology Acceptance Model Using Intuitionistic Fuzzy
Cognitive Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Ayça Maden and Emre Alptekin
Distance and Similarity Measures of Generalized Intuitionistic Fuzzy
Soft Set and Its Applications in Decision Support System . . . . . . . . . . . 355
Muhammad Jabir Khan and Poom Kumam
Intuitionistic Fuzzy Analysis of Variance of Movie Ticket Sales . . . . . . . 363
Velichka Traneva and Stoyan Tranev
Contents xi
Interval Valued Intuitionistic Fuzzy Gaussian Membership Function:

A Novel Extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
Janani Bharatraj
An Integrated IVIF-DEMATEL and IVIF-TOPSIS Methodology
for Hotel Information System Selection . . . . . . . . . . . . . . . . . . . . . . . . . 381
Gizem Erkal, Huseyin Selcuk Kilic, Zeynep Tugce Kalender,
Ahmet Selcuk Yalcin, and Gulfem Tuzkaya
Decision Making Using New Distances of Intuitionistic Fuzzy Sets
and Study Their Application in the Universities . . . . . . . . . . . . . . . . . . . 390
Shuker Mahmood Khalil and Mohanad Abdulkareem Hasan Hasab
Some New Ordered Semi-linear Spaces of Intuitionistic Fuzzy
Processes and the Pair of Closely Related States . . . . . . . . . . . . . . . . . . 397
Nguyen Dinh Phu, Nguyen Nhut Hung, and Le Thi Ngoc Quynh
Pythagorean Fuzzy Sets and Q-Rung Orthopair Fuzzy Sets

Emergency Decision Making Problem of Power Cut in Turkey
Using Pythagorean Fuzzy Thermodynamic Approach
with Prospect Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
Veysel Çoban and Sezi Çevik Onar
Multi-criteria Cloud Computing Service Provider Selection
Employing Pythagorean Fuzzy AHP and VIKOR . . . . . . . . . . . . . . . . . 423
Irem Otay and Tuğba Yıldız
Extension of Classical Analytic Hierarchy Process Using q-Rung
Orthopair Fuzzy Sets and Its Application to Disaster Logistics
Location Center Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432
Seyed Amin Seyfi Shishavan, Yaser Donyatalab,
and Elmira Farrokhizadeh
Interval Valued q-Rung Orthopair Fuzzy Prioritized Dual Muirhead
Mean Operator and Their Application in Group Decision Making . . . . 440
Salih Berkan Aydemir and Sevcan Yilmaz Gündüz
Fuzzy Capital Budgeting Using Fermatean Fuzzy Sets . . . . . . . . . . . . . . 448
Duygu Sergi and Irem Ucal Sari
Interval Valued q-Rung Orthopair Fuzzy EDAS Method
and Its Application to Supplier Selection . . . . . . . . . . . . . . . . . . . . . . . . 457
Elmira Farrokhizadeh, Seyed Amin Seyfi Shishavan, Yaser Donyatalab,
and Seyyed Hadi Seifi
xii Contents
Hamacher Aggregation Operators Based on Interval-Valued q-Rung

Orthopair Fuzzy Sets and Their Applications to Decision
Making Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
Yaser Donyatalab, Elmira Farrokhizadeh, Seyed Amin Seyfi Shishavan,
and Seyyed Hadi Seifi
Similarity Measures of q-Rung Orthopair Fuzzy Sets Based on Square
Root Cosine Similarity Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
Yaser Donyatalab, Elmira Farrokhizadeh,
and Seyed Amin Seyfi Shishavan
The Dice (Sorensen) Similarity Measures for Optimal Selection
with q-Rung Orthopair Fuzzy Information . . . . . . . . . . . . . . . . . . . . . . 484
Elmira Farrokhizadeh, Seyed Amin Seyfi Shishavan, Yaser Donyatalab,
and Sohrab Abdollahzadeh
Technology Selection of Indoor Location Systems Using Interval
Valued Type-2 Intuitionistic Fuzzy WASPAS . . . . . . . . . . . . . . . . . . . . . 494
Basar Oztaysi, Sezi Cevik Onar, and Cengiz Kahraman
Evaluating the Influencing Factors on Adoption of Self-driving
Vehicles by Using Interval-Valued Pythagorean Fuzzy AHP . . . . . . . . . 503
Gozde Bakioglu and Ali Osman Atahan
Evaluation of Maintenance Strategies Using Pythagorean Fuzzy
TOPSIS Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512
Tolga Gedikli, Beyzanur Cayir Ervural, and Durmus Tayyar Sen
Multicriteria Decision Making-Applications

Process Robot Automation Selection with MADM in Airline
Cargo Sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
Gul Durak and A. Cagri Tolga
Smart System Evaluation in Vertical Farming via Fuzzy
WEDBA Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
Murat Basar and A. Cagri Tolga
Blockchain Software Selection for a Maritime Organization
with MCDM Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543
Melis Kaska and A. Cagri Tolga
Prioritization of Influence Factors for Selecting E–Learning Systems . . . 550
Ali Karasan and Melike Erdogan
Special Agriculture Production Selection Using Interval
Type-2 Fuzzy AHP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
Sarper Alem
Contents xiii
Investigating Enablers to Improve Transparency in Sustainable Food

Supply Chain Using F-BWM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
Anish Kumar, Sachin Kumar Mangla, Pradeep Kumar,
and Yasanur Kayikci
Resilience Measurement System: A Fuzzy Approach . . . . . . . . . . . . . . . 576
Merve Kadan, Gökhan Özkan, and Mehmet Hilmi Özdemir
Road Selection for Autonomous Trucks in Turkey with Fuzzy AHP . . . 582
Zeynep Hasgul and Can Aytore
Multi-criteria Oil Station Location Evaluation Using Spherical
AHP&WASPAS: A Real-Life Case Study . . . . . . . . . . . . . . . . . . . . . . . 591
Irem Otay and Serhat Atik
Managing Cultural Built Heritage in Smart Cities Using Fuzzy
and Interval Multi-criteria Decision Making . . . . . . . . . . . . . . . . . . . . . 599
Mimica Milošević, Dušan Milošević, and Ana Stanojević
Criteria Weighting for Blockchain Software Selection Using
Fuzzy AHP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 608
Ferhat Karayazi and Ilke Bereketli
A Proposed Decision Making Methodology to Select IT Suppliers
for Software Development Outsourcing in Banking Sector Using
EDAS Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616
Murat Levent Demircan and Algı Acarbay
Analysis of Traffic Accidents to Identify Factors Affecting Injury
Severity with Fuzzy and Crisp Techniques . . . . . . . . . . . . . . . . . . . . . . . 625
Tutku Tuncalı Yaman, Emrah Bilgiç, and M. Fevzi Esen
Logistics Performance Index (LPI) Analysis with Using AHP
and Fuzzy AHP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634
Emir Şahin, İrem Pala, and Berrin Denizhan
Sustainable Transportation Service Provider Evaluation Utilizing
Fuzzy MCDM Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 642
Mehtap Dursun and Ergun Ari
Development of Assessment Model for Selection of Sustainable Energy
Source in India: Hybrid Fuzzy MCDM Approach . . . . . . . . . . . . . . . . . 649
S. K. Saraswat, Abhijeet Digalwar, and S. S. Yadav
Evaluation of Criteria that Affect the Sustainability of Smart Supply
Chain in a Textile Firm by Fuzzy SWARA Method . . . . . . . . . . . . . . . . 658
Arzu Organ, Kevser Arman, and Ali Katrancı
Green Supplier Selection Using Intuitionistic Fuzzy AHP and TOPSIS
Methods: A Case Study from the Paper Mills . . . . . . . . . . . . . . . . . . . . 666
Ezgi Demir and Gözde Koca
xiv Contents
InterCriteria Analysis Applied to Emerging Europe and Central Asia

University Rankings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674
Veselina Bureva
Determining Significant Factors Affecting Vaccine Demand
and Factor Relationships Using Fuzzy DEMATEL Method . . . . . . . . . . 682
İkbal Ece Dizbay and Ömer Öztürkoğlu
Benchmarking and Visualization of the Ergonomic Risks
in a Manufacturing Plant: Joint Application of Fuzzy AHP
and Treemap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 690
Samira Keivanpour
Multicriteria Decision Making-Theory

An Integrated Fuzzy DEMATEL and Fuzzy Cognitive Mapping
Methodology for Prioritizing Smart Campus Investments . . . . . . . . . . . 701
Ali Karasan and Cengiz Kahraman
Different Approaches to Fuzzy Extension of an MCDA Method
and Their Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 709
Boris Yatsalo and Alexander Korobov
Multi-objective Task Scheduling in Cloud Computing Environment
by Hybridized Bat Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 718
Timea Bezdan, Miodrag Zivkovic, Eva Tuba, Ivana Strumberger,
Nebojsa Bacanin, and Milan Tuba
Qualitative Factors in the Fuzzy Earned Value Method . . . . . . . . . . . . . 726
Dorota Kuchta
The Novel Integrated Model of Plithogenic Sets and MAIRCA Method
for MCDM Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733
Abdullah Özçil, Ayşegül Tuş, Gülin Zeynep Öztaş, Esra Aytaç Adalı,
and Tayfun Öztaş
An Alternative Approach for Performance Evaluation:
Plithogenic Sets and DEA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742
Gülin Zeynep Öztaş, Esra Aytaç Adalı, Ayşegül Tuş, Tayfun Öztaş,
and Abdullah Özçil
Multi-criteria Decision Making Problem with Intuitionistic Fuzzy
Sets: A Novel Ranking Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 750
Hatice Ercan-Tekşen
Preference-Oriented Fuzzy TOPSIS Method . . . . . . . . . . . . . . . . . . . . . 758
Alicja Mieszkowicz-Rolka and Leszek Rolka
Contents xv
Intelligent Statistical Analyses

A New Approach to Analyze Perceived Design Quality of Mobile
Phone Using Fuzzy Hierarchical Conjoint Analysis . . . . . . . . . . . . . . . . 769
Yusra Erdem, Selcuk Cebi, and Esra Ilbahar
Congestion Trajectories Using Fuzzy Gaussian Travel Time Based
on Mesoscopic and Cellular Automata Traffic Model . . . . . . . . . . . . . . . 779
A. Boulmakoul, L. Karim, M. Nahri, and A. Lbath
Data Mining Algorithms for Classification Model
of Engineering Grade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 788
Ching-Lung Fan
A Fuzzy Strategy Analysis Simulator for Exploring the Potential
of Industry 4.0 in End of Life Aircraft Recycling . . . . . . . . . . . . . . . . . . 797
Samira Keivanpour
Knowledge Data Discovery (Frequent Pattern Growth):
The Association Rules for Evergreen Activities
on Computer Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 807
Fauzan Asrin, Saide Saide, Silvia Ratna, and Alex Wenda
Enabling Big Data Analytics at Manufacturing Fields
of Farplas Automotive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 817
Özgün Akın, Halil Faruk Deniz, Doğukan Nefis, Alp Kızıltan,
and Altan Çakır
Big Data Analytics Framework for Smart City Real-Time Fuzzy
Complex Event Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825
Basmi Wadii, Azedine Boulmakoul, Lamia Karim, and Ahmed Lbath
Cloud Based Big Data DNS Analytics at Turknet . . . . . . . . . . . . . . . . . 833
Altan Çakır, Yousef Alkhanafseh, Esra Karabıyık, Erhan Kurubaş,
Rabia Burcu Bunyak, and Cenk Anıl Bahçevan
A Combined Bioinspired Algorithm for Big Data Processing . . . . . . . . . 842
Elmar Kuliev, Dmitry Zaporozhets, Yury Kravchenko, and Ilona Kursitys
Exploration of the Waves of Feminism Using Sentiment Based
Text Mining Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 850
H. Umutcan Ay, S. Nazlı Günesen, and Tolga Kaya
SVM-Based Hybrid Robust PIO Fault Diagnosis for Bearing . . . . . . . . 858
Farzin Piltan and Jong-Myon Kim
Modeling Urban Traffic by Means of Traffic Density Data:
Istanbul Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 867
Tutku Tuncalı Yaman, Hülya Başeğmez Sezer, and Emrah Sezer
xvi Contents
Realising Newspaper Sales by Using Statistic Methods . . . . . . . . . . . . . 875

Onur Dogan and Omer Faruk Gurcan
Extensions of Fuzzy Sets in Big Data Applications:
A Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 884
Nurşah Alkan and Cengiz Kahraman
A Big Data Semantic Driven Context Aware
Recommendation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 894
Manuel J. Barranco, Pedro J. Sanchez, Jorge Castro, and Raciel Yera
Blended Environment of Naive Bayes and Support Vector Machine
(SVM) for Designing Simulation Based E-Learning
Respiratory System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 903
Anuradha Verma Babbar and Santosh Kumar Henge
InterCriteria Analysis of Public Health Data in Bulgaria . . . . . . . . . . . . 910
Evdokia Sotirova, Valentin Vasilev, Sotir Sotirov, and Hristo Bozov
A Case Study on Vehicle Battery Manufacturing Using Fuzzy Analysis
of Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 916
Abbas Parchami, Mashaallah Mashinchi, and Cengiz Kahraman
New Suggestion for Fuzzy Random Variable
and Its Statistical Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 924
Abbas Parchami
Networks
Fuzzy Centrality Analysis for Smart City Trajectories . . . . . . . . . . . . . . 933
Lamia Karim, Azedine Boulmakoul, Ghyzlane Cherradi, and Ahmed Lbath
Improving Customer Experience for an Internet Service Provider:
A Neural Networks Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 941
Özge H. Namlı, Seda Yanık, Faranak Nouri, N. Serap Şengör,
Yusuf Mertkan Koyuncu, and İrem Küçükali
Segmentation of Larynx Histopathology Images via Convolutional
Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 949
Ahmet Haşim Yurttakal and Hasan Erbay
Glioma Brain Tumor Grade Classification from MRI Using
Convolutional Neural Networks Designed by Modified FA . . . . . . . . . . 955
Timea Bezdan, Miodrag Zivkovic, Eva Tuba, Ivana Strumberger,
Nebojsa Bacanin, and Milan Tuba
Dynamic System Control Using Z-number Based Fuzzy
Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 964
Rahib H. Abiyev
Contents xvii
Comprehensive Research of the Commodity Market: Theoretical,

Methodological and Modern Approaches Using Neural Networks
with Fuzzy Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 973
Mikayilova Rena Nuru Kizi
The Maximum Lexicographic Contraflow Finding in a Fuzzy
Dynamic Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 981
Evgeniya Gerasimenko and Vladimir Kureichik
An Enhanced Particle Swarm Optimization with Levy Flight
for RBF Neural Network in Typical Karst Area, South China . . . . . . . 990
Zhangjun Cao, Dong Wang, and Lachun Wang
A Method for Forecasting the Demand for Pharmaceutical Products in
a Distributed Pharmacy Network Based on an Integrated Approach
Using Fuzzy Logic and Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . 998
Alekperov Ramiz Balashirin Oglu and Isgandarova Ilhama Tarlan Kizi
Marketing Strategy Selection Using Fuzzy Analytic Network Process . . . 1008
Serhan Hamal, Bahar Sennaroglu, and Mahmure Övül Arıoğlu
A Deep Learning Model for Skin Lesion Analysis Using Gaussian
Adversarial Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1015
Fatih Ergin and Ismail Burak Parlak
Learning Channel-Wise Ordered Aggregations
in Deep Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1023
Iris Dominguez-Catena, Daniel Paternain, and Mikel Galar
Design of Nano-scale Synaptic Neural Network Using RRAM . . . . . . . . 1031
Ali Mohamed and Osama Rayis
Neural Network-Based Control Framework for SISO Uncertain
System: Passive Fault Tolerant Approach . . . . . . . . . . . . . . . . . . . . . . . 1039
Sejal Raval, Himanshukumar R. Patel, and Vipul A. Shah
Forecasting the Day-Ahead Prices in Electricity Spot Market
of Turkey by Using Artificial Neural Networks . . . . . . . . . . . . . . . . . . . 1048
Berna Tektaş, Aygülen Kayahan Karakul, and Rozi Mizrahi
Intelligent Learning
Using Formal Concept Analysis Tools in Road
Environment-Type Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1059
A. Boulmakoul, Z. Fazekas, L. Karim, G. Cherradi, and P. Gáspár
An Intelligent Overtaking Assistant System
for Autonomous Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1068
Ersin Armağan and Tufan Kumbasar
xviii Contents
Predicting Movie Ratings with Machine Learning Algorithms . . . . . . . . 1077

Sandy Çağlıyor and Başar Öztayşi
Trajectory Tracking of a Quadcopter Using Adaptive Neuro-Fuzzy
Controller with Sliding Mode Learning Algorithm . . . . . . . . . . . . . . . . 1084
Hasan Kemik, Mehmet Berkin Dal, and Yesim Oniz
Influencer Identification System Design Using Machine
Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1092
Elvira Israfilova, Armagan Arslan, Nihan Yildirim, and Tolga Kaya
Efficient and Real-Time Face Recognition Based on Single Shot
Multibox Detector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1100
Youngshin Ahn, Sumi Kim, Fei Chen, and Jaeho Choi
Machine Learning-Based Robust Feedback Observer for Fault
Diagnosis in Bearing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1107
Farzin Piltan and Jong-Myon Kim
Malfunction Detection on Production Line Using Machine Learning:
Case Study in Wood Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1116
Kağan Özgün, Sami Can Aklan, Ahmet Tezcan Tekin, and Ferhan Çebi
Classification of Breast DCE-MRI Images via Boosting
and Deep Learning Based Stacking Ensemble Approach . . . . . . . . . . . . 1125
Ahmet Haşim Yurttakal, Hasan Erbay, Türkan İkizceli, Seyhan Karaçavuş,
and Cenker Biçer
Machine Learning Approach for Automatic Categorization
of Service Support Requests on University Information
Management System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1133
Aytuğ Onan, Erdem Atik, and Adnan Yalçın
Comparative Study of Different Machine Learning Models for Remote
Sensing Bathymetry Inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1140
Shen Wei, Ji Qian, Rao Yali, and Meng Ran
Container Terminal Workload Modeling Using Machine
Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1149
Üstün Atak, Tolga Kaya, and Yasin Arslanoğlu
Usage of Machine Learning Algorithms for Flow Based Anomaly
Detection System in Software Defined Networks . . . . . . . . . . . . . . . . . . 1156
Muhammet Fatih Akbaş, Cengiz Güngör, and Enis Karaarslan
Speech Analysis with Deep Learning to Determine Speech
Therapy for Learning Difficulties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1164
Nogol Memari, Saranaz Abdollahi, Sonia Khodabakhsh, Saeideh Rezaei,
and Mehrdad Moghbel
Contents xix
A Compromise-Based New Approach to Learning Fuzzy

Cognitive Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1172
Miraç Murat and Umut Asan
Fuzzy Analytics
Fuzzy Analytics of Spaces’ Dynamic Use in Smart Buildings . . . . . . . . . 1183
Azedine Boulmakoul, Abdellah Daissaoui, Ahmed Lbath,
and Lamia Karim
Object Monitoring Under Z-Information Based on Rating Points . . . . . 1191
Olga M. Poleshchuk
An Intelligent Decision Support System: Application of Fuzzy Tools
and System Dynamics Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1199
Merve Kadan, Gökhan Özkan, and Mehmet Hilmi Özdemir
Fuzzy Uncertainty Modelling in Cost and Cash Flow Forecasting
in Project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1206
Dorota Kuchta and Adam Zabor
Fuzzy System for Building Energy Flows Management . . . . . . . . . . . . . 1216
Penka V. Georgieva
Fuzzy RFM Analysis: An Application in E-Commerce . . . . . . . . . . . . . 1225
Basar Oztaysi and Mert Kavi
A New Structure of Nullnorms on Bounded Lattices . . . . . . . . . . . . . . . 1233
Gül Deniz Çaylı
A New Perspective to Design Phase of Water Supply Systems
from Aspect of Water Demand Using Fuzzy Automation . . . . . . . . . . . . 1242
Halid Akdemir, Ayşegül Alaybeyoğlu, and Ali Danandeh Mehr
Fuzzy Cognitive Map Based PI Controller Design . . . . . . . . . . . . . . . . . 1250
Aykut Denizci, Sinan Karadeniz, and Cenk Ulu
On 2-Fuzzy Metric Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1258
Elif Güner and Halis Aygün
Some Properties of Partial Fuzzy Metric Topology . . . . . . . . . . . . . . . . 1267
Ebru Aydogdu, Başak Aldemir, Elif Güner, and Halis Aygün
Fuzzy Classifier Based Colorimetric Quantification
Using a Smartphone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1276
Öykü Berfin Mercan and Volkan Kılıç
LabHub: A New Generation Architecture Proposal
for Intelligent Healthcare Medical Laboratories . . . . . . . . . . . . . . . . . . . 1284
Bengi Tugcu Idemen, Emine Sezer, and Murat Osman Unalir
xx Contents
Center-of-Gravity Real Options Method Based on Interval-Valued

Fuzzy Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1292
Jani Kinnunen, Irina Georgescu, and Mikael Collan
Fusing Fuzzy Preferences in Contexts of Social Choice:
Towards a Topological Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1301
María Jesús Campión, Esteban Induráin, and Armajac Raventós-Pujol
Understanding the Behavior of Zadeh’s Extension Principle
for One-to-One Functions by R Programming Language . . . . . . . . . . . . 1309
Abbas Parchami and Parisa Khalilpoor
Intuitionistic Fuzzy Z-numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1316
Irem Ucal Sari and Cengiz Kahraman
Risk Assessment
Risk Assessment of R&D Activities Using Intuitionistic Fuzzy AHP
and FIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1327
Esra Ilbahar, Selcuk Cebi, and Cengiz Kahraman
Risk Matrix Technique Based Fuzzy MCDM Risk Assessment Method
and Application in Glass Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1336
Sukran Seker
A New Risk Assessment Approach for Occupational Health
and Safety Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1345
Selcuk Cebi
Intuitionistic Fuzzy Group Decision Making on Risk Assessment
and Control Efficiency of Accounting Information Systems
with TOPSIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1355
Aygülen Kayahan Karakul, Rozi Mizrahi, and Berna Tektaş
Evaluation of Occupational Safety Risk in Underground Mining
Using Fuzzy Bayesian Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1363
Fatma Yaşlı and Bersam Bolat
Risk Analysis for Digitalization Oriented Sustainable Supply Chain
Using Interval-Valued Pythagorean Fuzzy AHP . . . . . . . . . . . . . . . . . . . 1373
Nurşah Alkan
Intelligent Quality
Design for Six Sigma and Process Capability Using Penthagorean
Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1385
Elif Haktanır and Cengiz Kahraman
Contents xxi
Defects Control Charts Using Interval-Valued Penthagorean

Fuzzy Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1396
Elif Haktanır and Cengiz Kahraman
Fuzzy AHP Based Prioritization and Taxonomy of Information
Quality Factors in Accounting Information Systems . . . . . . . . . . . . . . . 1407
Rozi Mizrahi, Berna Tektaş Aygülen, and Kayahan Karakul
Analytical Techniques to Compute Cp and Cpm Capability
Indices by R Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1415
Abbas Parchami
Heuristics
Proposal of Genetic Algorithm Approach for Solving Single Machine
Scheduling Problem Under Learning Effect . . . . . . . . . . . . . . . . . . . . . . 1423
Derya Deliktas and Mustafa Urhan
Hybrid Krill Herd Algorithm with Particle Swarm Optimization
for Image Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1431
Elnaz Pashaei, Elham Pashaei, and Nizamettin Aydin
Distributed Fuzzy Permutation Flow Shop Scheduling Problem:
A Bee Colony Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1440
M. Emin Baysal, Ahmet Sarucan, Kadir Büyüközkan, and Orhan Engin
Fuzzy Metaheuristics: A State-of-the-Art Review . . . . . . . . . . . . . . . . . . 1447
Nurşah Alkan and Cengiz Kahraman
Fuzzy and Evolutionary Algorithms for Transport Logistics
Under Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1456
Yuriy Kondratenko, Galyna Kondratenko, Ievgen Sidenko,
and Mykyta Taranov
Using GA for Locating Post-disaster Field Hospitals:
The Case of Istanbul . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1464
Yesim Komurcu and Seda Yanik
Optimization
Optimization of Spatial-Time Planning Resource Allocation
Under Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1475
Olesiya Kosenko, Alexander Bozhenyuk, Stanislav Belyakov,
and Margarita Knyazeva
Solving Fuzzy Multi-objective Linear Programming Problems
Using Multi-player Zero-Sum Game . . . . . . . . . . . . . . . . . . . . . . . . . . . 1483
Gizem Temelcan, Inci Albayrak, Hale Kocken, and Mustafa Sivri
xxii Contents
A Bioinspired Algorithm for Improving the Effectiveness

of Knowledge Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1491
Ilona Kursitys, Yury Kravchenko, Elmar Kuliev, and Alexandr Natskevich
Spatial Coherence Radius of Plane Wave Propagating in Anisotropic
Maritime Atmospheric Turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1499
Bing Guan, Fei Chen, and Jaeho Choi
A Comparative Performance Analysis of Consensus Models
Based on a Minimum Cost Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1506
Á. Labella, R. M. Rodríguez, and L. Martínez
Pedestrian Detection Based on Improved Mask R-CNN Algorithm . . . . 1515
Wenjun Yu, Sumi Kim, Fei Chen, and Jaeho Choi
A Revised Weighted Fuzzy C-Means and Center of Gravity Algorithm
for Probabilistic Demand and Customer Positions . . . . . . . . . . . . . . . . . 1523
Engin Bayturk, Sakir Esnaf, and Tarik Kucukdeniz
Mobile Application Based Automatic Caption Generation
for Visually Impaired . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1532
Özkan Çaylı, Burak Makav, Volkan Kılıç, and Aytuğ Onan
Chaos Control of BLDC Motor via Fuzzy Based PID Controller . . . . . . 1540
Alkım Gökçen, Mehmet Uğur Soydemir, and Savaş Şahin
Sustainability Oriented Scheduling Procedure for Public Projects . . . . . 1548
Dorota Kuchta and Ewa Marchwicka
Optimization of an Oil Refinery Valuation System Through
the Intuitionistic Fuzzy InterCriteria Analysis . . . . . . . . . . . . . . . . . . . . 1555
Velichka Traneva and Stoyan Tranev
Double Edge–Vertex Domination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1564
Bünyamin Şahin and Abdulgani Şahin
Order Acceptance Scheduling on a Single Machine
with Energy Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1573
Mariam Bouzid, Oussama Masmoudi, and Alice Yalaoui
The Analysis of Human Oriented System of Weighted Fuzzy Petri
Nets for the Passenger Transport Logistics Problem . . . . . . . . . . . . . . . 1581
Zbigniew Suraj, Oksana Olar, and Yurii Bloshko
Concepts and Methods of “Digital Twins” Models Creation
in Industrial Asset Performance Management Systems . . . . . . . . . . . . . 1589
Nodirbek Yusupbekov, Fakhritdin Abdurasulov, Farukh Adilov,
and Arsen Ivanyan
Contents xxiii
Finding the Optimal Features Reduct, a Hybrid Model of Rough Set

and Polar Bear Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1596
Amer Mirkhan and Numan Çelebi
Design and Manufacture a Vehicle Auxiliary System Model
Controlled by Using Smartphone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1604
Nguyen Phu Thuong Luu
A Study on Vehicle Air Condition System Model Controlled
by Using Smartphone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1612
Nguyen Phu Thuong Luu
Multi-shift Single-Vehicle Routing Problem Under Fuzzy Uncertainty . . . 1620
F. Nucci
On Merrifield-Simmons Index of Trees . . . . . . . . . . . . . . . . . . . . . . . . . 1628
Bünyamin Şahin
A Fuzzy Goal Programming Model to Evaluate Green
Energy Alternatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1635
Beyzanur Cayir Ervural
Application of Fuzzy Logic Model for Correct Lighting in Computer
Aided Interior Design Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1644
Rahib Imamguluyev
Review and Discussion Papers

Extensions of Ordinary Fuzzy Sets: A Comparative
Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1655
Cengiz Kahraman, Basar Oztaysi, Irem Otay, and Sezi Cevik Onar
Systematic Literature Review for Work Sampling with IOT
Technology in Retail Store Operations . . . . . . . . . . . . . . . . . . . . . . . . . . 1666
Gizem Omeroglu, Abdullah Ertug, and Aysegul Ozkavukcu
Engineering Economics Using Fuzzy Sets: A Literature Review . . . . . . 1675
Eda Boltürk
Scientific Cooperation in the Field of Economics in Selected
European Countries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1683
Paweł Lula, Urszula Cieraszewska, and Monika Hamerska
Sentiment Analysis on Students’ Evaluation of Higher
Educational Institutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1693
Mansur Alp Toçoğlu and Aytuğ Onan
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1701

Invited Speakers’ Papers
Fuzzy Meets Privacy: A Short Overview
Vicenç Torra1,2(B) and Guillermo Navarro-Arribas3

1
Maynooth University, Maynooth, Ireland
[email protected]
2
School of Informatics, University of Skövde, Skövde, Sweden
3
Department of Information and Communication Engineering,
Universitat Autònoma de Barcelona, Catalonia, Spain
[email protected]
Abstract. The amount of information currently available is a threat to

individual privacy. Data privacy is the area that studies how to process
data and propose methods so that disclosure does not take place.
Fuzzy sets has been extensively used in all kind of applications. In
this paper we review the areas in which fuzzy sets and systems can
have an impact in the field of data privacy. We will review some of our
contributions in the field and some open problems.
Keywords: Data privacy · Fuzzy sets and systems · Fuzzy meets

privacy
1 Introduction
Data privacy [21] is the field that studies approaches for ensuring that disclosure
does not take place when data is processed and used in computations.
Privacy and disclosure are terms not clearly defined, and, therefore, in order
to develop methods and tools to ensure that disclosure does not take place, we
need to establish concrete formulations. Privacy models can be seen as compu-
tational definitions of privacy. Given a privacy model, then we can establish a
method that is compliant with the privacy model.
The literature provides different privacy models. For example, the most rele-
vant ones are privacy from reidentification, k-anonymity, and differential privacy.
Each of them has led to a plethora of methods that provide implementations of
the privacy model. Methods compete with respect to some properties (e.g., com-
putational cost, data quality).
When a company or an individual has a database (standard SQL database
or even a non-SQL one as representing textual data, social networks) and wants
to release it to a third party, k-anonymity and privacy from reidentification are
the most common privacy models. Both models focus on avoiding reidentifica-
tions and identity disclosure. This is achieved modifying the database. That is,
instead of providing the original database, we modify it and reduce its quality
so that any intruder trying to find an acquaintance in the database will (prob-
ably) not succeed. The quality reduction is expected to be low so data is still
c The Editor(s) (if applicable) and The Author(s), under exclusive license
C. Kahraman et al. (Eds.): INFUS 2020, AISC 1197, pp. 3–9, 2021.
https://doi.org/10.1007/978-3-030-51156-2_1
4 V. Torra and G. Navarro-Arribas
useful. Naturally, quality reduction implies some loss of information. Then, from
a research perspective, the goal is to find methods that avoid reidentification but
have minimal impact to any data use.
Data protection mechanisms for databases are usually known as masking
methods. A simple masking method is noise addition. This is to add noise
following a certain Gaussian distribution to the original data. Unfortunately,
this approach is not very effective with respect to the trade-off information
loss/disclosure risk. In other words, to have low disclosure risk we need to per-
turb the data significantly and this can make the data useless for most analysis.
Alternative methods that are more effective include microaggregation and rank
swapping.
When masking methods are applied, data quality is compromised to reduce
risk. Because of that there are measures that evaluate in what extent information
is lost in the transformation. They are known as information loss measures. In
general, information loss depends on the intended use. When the use is not know
we usually use generic information loss measures. They are measures based on
some statistics. As modifying the data does not imply, in general, that disclosure
risk vanishes completely. Disclosure risk measures have also been proposed in
the literature. These measures depend on the privacy model and the type of
disclosure we want to avoid.
In this paper we review the use of fuzzy sets in data privacy, focusing on
database releases. We will discuss, the role of fuzzy sets in masking, in measur-
ing information loss and in measuring disclosure risk. This paper extends and
updates our previous discussion in [1,23].
2 Data Protection Methods Based on Fuzzy Techniques
Given a database X, a masking method is a function ρ that applied to X returns

the database X . X is a database based on X but with less quality. In the
literature there are different families of functions ρ defined for this objective.
The main three classes are perturbative methods, non-perturbative methods
and synthetic data generators. Perturbative methods modify the original data
introducing some noise (some kind of error is introduced to the data), the non-
perturbative methods modify the original data changing the level of detail but
there are no erroneous data (some change of granularity is applied to the data),
and synthetic data generators construct models of the data and then replace the
original data by the one generated with the models.
We have contributed in perturbative methods and on synthetic data genera-
tors using techniques inspired on fuzzy sets.
– Microaggregation is a very effective method for data protection. Effective

in the sense that it achieves a good trade-off between data utility/information
loss and disclosure risk. With a good parameterization, it is often possible to
achieve low information loss and low disclosure risk.
Fuzzy Meets Privacy: A Short Overview 5
Microaggregation proceeds with the construction of small clusters of data,

and then replaces each record by cluster representatives. If all clusters are
required to be of a certain minimum size, privacy is ensured. If all clusters
are required to be small, the perturbation of the original data is small (i.e.,
information loss is small and data utility is kept maximal). We have con-
tributed with methods based on fuzzy clustering. Fuzzy c-means permits us
to solve one of the inconveniences of applying standard crisp clustering in
data privacy: an intruder can easily guess which is the centroid assigned to
a particular record. This information can be used by the intruder to attack
the database. This problem as well as a possible solution is explained in
detail in [29]. Recent results on fuzzy microaggregation are given in [26,27].
In the fuzzy community the problem of microaggregation has been considered
recently (see e.g. [9,10,14], fuzzy extensions of this work my be of interest.
Microaggregation is a data protection mechanism that is specially well suited
when edit constraints (i.e., constraints on the values of variables) are present.
An example of a constraint follows: retail price is defined in terms of net price
plus taxes. All records in the database need to satisfy this constraint. Data
editing [11,31] is the field in official statistics that focuses on the construction
of a database satisfying the constraints. We have developed an approach for
this problem based on fuzzy c-means (see [19,22] for details). Our variation
of fuzzy c-means is such that the resulting centroids are required to satisfy
the constraints on the variables even when these constraints are not satisfied
in the original data.
– Synthetic data generators (see e.g. [5,18]) are based on constructing a
model from the data. Then, the model is used to generate artificial data.
We introduced an approach [6] based on fuzzy c-regression models. Fuzzy
c-regression models use a parameter c that corresponds to the number of
clusters (i.e., regression models). Then, the larger the parameter c, the more
clusters we have, and the more regression models we have. This parameter
permits the user to find a better trade-off between information loss and dis-
closure risk. That is, large c has low information loss but high risk, and small
c has large information loss and low risk. The selection of the parameter c
is to permit to have a good balance between information loss and disclosure
risk.
3 Information Loss Measures Based on Fuzzy Techniques
Any masking method needs to be evaluated with respect to the extent that it
modifies the original data. It is important to know how much information is lost
in the process. Information loss depends on the data uses. That is, it is different
if data will be used to compute means, variances and correlations or it will be
used to compute a regression model. Information loss measures are to quantify
the information loss.
We have studied information loss when the intended use is clustering and,
more specifically, fuzzy clustering. We have considered information loss when
data is to be clustered using fuzzy c-means, noise clustering, possibilistic c-means
and fuzzy possibilistic c-means. We have studied information loss for some stan-
dard methods for data protection as e.g. microaggregation but also for some
synthetic data generators [16] as IPSO [5] and our method based on fuzzy c-
regression models [7]. We have also compared [15] the measures of information
loss obtained with our approach and the ones obtained with some generic infor-
mation loss measures (e.g., measures based on standard statistics as the mean,
variances, covariances and correlations of the database).
The comparison of fuzzy clusters is not an easy task. Two problems arise, one
is about the comparison of two fuzzy partitions. Another problem is that fuzzy
clustering methods are typically implemented with methods that only ensure
the convergence to a local optima. We have addressed the two problems. In
addition, the need to compare fuzzy clusters taking into account the uncertainty
of the fuzzy clustering methods have lead us to the definition of interval-valued
or intuitionistic fuzzy partitions [28,30]. That is, fuzzy partitions in which the
membership value of an element to a cluster is an interval instead of a number
in [0, 1].
4 Disclosure Risk Measures Based on Aggregation

Functions: WM, OWA and the Choquet Integral
Fuzzy techniques have also been applied to measure disclosure risk. As stated
above, one of the approaches for computing disclosure risk [33] is to count the
number of records of an intruder that can be linked to the protected file. The
most standard approach for linking the two files is to use an Euclidean distance to
measure the dissimilarity between pairs of records. Nevertheless, other distances
can be used. The literature includes results using e.g. Mahalanobis distance and
also kernel-based distance. From a formal point of view, given a certain distance
d with a given parameter p, we have that the disclosure risk is defined by
|{y = x∗ (y)|y ∈ Y }|
DR(Y, X ) = ,
|Y |
where
x∗ (y) = arg min dp (x, y).
x∈X
In this expression we can use the Euclidean distance. That is, dp = AM , where
AM stands for the arithmetic mean. Nevertheless, any other distance is possible.
We have considered weighted distances building them from weighted means,
OWA operators and the Choquet integrals. When the distance is based on the
weighted mean and the OWA operator, the parameter p is a weighting vector;
in the case of the Choquet integral the parameter p is a fuzzy measure.
The consideration of a parameterized aggregation operators for computing
the distance permits us to consider the corresponding optimization problem.
That is, we can study which is the parameter p which maximizes the disclosure
risk. Then, given a pair of files Y and X , and a distance function d defined in
terms of a parameter p, we can determine the p that maximizes the number of
reidentifications. This p is the best parameter that an intruder might have, and,
thus, an upper bound of the disclosure risk.
We have applied this optimization approach to the weighted mean, the
OWA [25] and the Choquet integral [2] (see also [3]).
We have seen that the weighted mean gives better results than the OWA,
and that the Choquet integral leads to better results than the weighted mean.
We have also seen that for most of the cases in which all variables are protected
with the same data protection method, the use of a Choquet integral is not
significantly better than using the weighted mean and the OWA operator. In
contrast, when different data protections are applied to different variables of a
file, the use of a Choquet integral can be useful.
5 Conclusion and Future Work

In this paper we have considered the use of fuzzy techniques in data privacy.
The focus was on aspects related to masking databases.
Most of the results reported are for files of regular size, and corresponding
to standard SQL databases. Some of these results can be applied or extended to
other types of data sets (as e.g., categorical data and time-series). Information
loss measures based on fuzzy clustering can also be applied to other types of
data (as e.g., non-numerical data). Extensions that can be efficient for big data
(data of large dimension, streaming data, etc.) will be useful. Research lines for
data privacy for big data are described in [8,24].
Fuzzy techniques can be applied in other contexts. For example, in line with
secure multiparty computation, we consider several parties that want to build
together a set of clusters. Up to our knowledge, this research line has not been
explored. We need to mention that there is a term used in the literature of
cryptography that has nothing to do with fuzzy sets theory, although one may
think so. It is fuzzy cryptography. For more, the readers can apply to [4,12,13,
17,20,32].
Acknowledgements. Partially supported by Vetenskapsrådet project: “Disclosure

risk and transparency in big data privacy” (VR 2016-03346); and Spanish and Catalan
governments with projects TIN2014-55243-P and 2014SGR-691 respectively.
References
1. Torra, V., Abril, D., Navarro-Arribas, G.: Fuzzy methods for database protection.
In: Proceedings of EUSFLAT-LFA, Aix-Les-Bains, France, 18–22 July 2011, vol.
1, no. 1, pp. 439–443. Atlantis Press (2011)
2. Abril, D., Navarro-Arribas, G., Torra, V.: Choquet integral for record linkage. Ann.
Oper. Res. 195, 97–110 (2012)
3. Abril, D., Navarro-Arribas, G., Torra, V.: Supervised learning using a symmetric
bilinear form for record linkage. Inf. Fusion 26, 144–153 (2016)
4. Aggarwal, C.C., Yu, P.S. (eds.): Privacy-Preserving Data Mining. Springer, Hei-
delberg (2007)
5. Burridge, J.: Information preserving statistical obfuscation. Stat. Comput. 13, 321–
327 (2003)
6. Cano, I., Torra, V.: Generation of synthetic data by means of fuzzy c-regression. In:
Proceedings of FUZZ-IEEE 2009, DVD-ROM, Jeju Island, Korea, pp. 1145–1150.
IEEE (2009). ISBN 978-1-4244-3597-5
7. Cano, I., Ladra, S., Torra, V.: Evaluation of information loss for privacy preserving
data mining through comparison of fuzzy partitions. In: Proceedings of FUZZ-
IEEE 2010/WCCI, pp.1–8. IEEE (2010). (ISBN in press)
8. D’Acquisto, G., Domingo-Ferrer, J., Kikiras, P., Torra, V., de Montjoye, Y.-A.,
Bourka, A.: Privacy by design in big data: an overview of privacy enhancing tech-
nologies in the era of big data analytics. ENISA Report (2015)
9. Endo, Y., Hamasuna, Y., Hirano, T., Kinoshita, N.: Even-sized clustering based
on optimization and its variants. JACIII 22(1), 62–69 (2018)
10. Endo, Y., Hirano, T., Kinoshita, N., Hamasuna, Y.: On various types of even-
sized clustering based on optimization. In: International Conference on Modeling
Decisions for Artificial Intelligence, pp. 165–177. Springer, Cham (2016)
11. Granquist, L.: The new view on editing. Int. Stat. Rev. 65(3), 381–387 (1997)
12. Hay, M., Miklau, G., Jensen, D.: Anonymizing social networks. In: Proceedings of
VLDB (2008)
13. Jones, R., Kumar, R., Pang, B., Tomkins, A.: “I know what you did last summer”:
query logs and user privacy. In: Proceedings of the Sixteenth ACM Conference on
Conference on Information and Knowledge Management, pp. 909–914 (2007)
14. Kitajima, K., Endo, Y.: Even-sized clustering with noise clustering method. In:
2018 Joint 10th International Conference on Soft Computing and Intelligent Sys-
tems (SCIS) and 19th International Symposium on Advanced Intelligent Systems
(ISIS), pp. 837–842. IEEE (2018)
15. Ladra, S., Torra, V.: On the comparison of generic information loss measures and
cluster-specific ones. Int. J. Uncertain. Fuzziness Knowl. Based Syst. 16(1), 107–
120 (2008)
16. Ladra, S., Torra, V.: Information loss for synthetic data through fuzzy clustering.
Int. J. Uncertain. Fuzziness Knowl. Based Syst. 18(1), 25–37 (2010)
17. Liu, K., Terzi, E.: Towards identity anonymization on graphs. In: Proceedings of
the 2008 ACM SIGMOD International Conference on Management of Data, pp.
93–106 (2008)
18. Muralidhar, K., Sarathy, R.: Generating sufficiency-based non-synthetic perturbed
data. Trans. Data Priv. 1(1), 17–33 (2008)
19. Torra, V.: On the definition of linear constrained fuzzy c-means. In: Proceedings of
the EUROFUSE 2009, Pamplona, Spain, September 2009, pp. 61–66 (2009). ISBN
978-84-9769-242-7
20. Torra, V.: Privacy in data mining. In: Data Mining and Knowledge Discovery
Handbook, 2nd edn (invited chapter), pp. 687–716. Springer, Boston (2010)
21. Torra, V.: Data Privacy: Foundations, New Developments and the Big Data Chal-
lenge. Springer, Heidelberg (2017)
22. Torra, V.: Fuzzy clustering-based microaggregation to achieve probabilistic k-
anonymity for data with constraints (2020, to be published)
23. Torra, V., Abril, D., Navarro-Arribas, G.: Fuzzy meets privacy. In: Proceedings of
USB MDAI 2017, pp. 108–117. Springer, Cham (2017)
24. Torra, V., Navarro-Arribas, G.: Big data privacy and anonymization. In: Lehmann,
A., Whitehouse, D., Fischer-Hübner, S., Fritsch, L. (eds.) Privacy and Identity
Management – Facing up to Next Steps, pp. 16–26. Springer, Heidelberg (2016)
25. Torra, V., Navarro-Arribas, G., Abril, D.: Supervised learning for record linkage
through weighted means and OWA operators. Control Cybern. 39, 1011–1026
(2010)
26. Torra, V.: A fuzzy microaggregation algorithm using fuzzy c-means. In: Artifi-
cial Intelligence Research and Development: Proceedings of the 18th International
Conference of the Catalan Association for Artificial Intelligence, pp. 214–223. IOS
Press (2015)
27. Torra, V.: Fuzzy microaggregation for the transparency principle. J. Appl. Logics
23, 70–80 (2017)
28. Torra, V., Min, J.-H.: I-fuzzy partitions for representing clustering uncertainties.
In: International Conference on Artificial Intelligence and Soft Computing, pp.
240–247. Springer, Berlin (2010)
29. Torra, V., Miyamoto, S.: Evaluating fuzzy clustering algorithms for microdata
protection. In: International Workshop on Privacy in Statistical Databases, pp.
30. Torra, V., Miyamoto, S.: A definition for I-fuzzy partitions. Soft. Comput. 15(2),
363–369 (2011)
31. De Waal, T.: An overview of statistical data editing. Statistics Netherlands (2008)
32. Willenborg, L., de Waal, T.: Elements of Statistical Disclosure Control. Lecture
Notes in Statistics. Springer, Heidelberg (2001)
33. Winkler, W.E.: Re-identification methods for masked microdata. In: International
Workshop on Privacy in Statistical Databases, pp. 216–230. Springer, Berlin (2004)
Intelligent Planning of Spatial Analysis Process
Based on Contexts
Stanislav Belyakov1 , Alexander Bozhenyuk1(&) ,

Janusz Kacprzyk2 , and Igor Rozenberg3
1
Southern Federal University, Nekrasovsky 44, 347922 Taganrog, Russia
[email protected], [email protected]
2
Systems Research Institute, Polish Academy of Sciences,
Newelska 6, 01-447 Warsaw, Poland
[email protected]
3
Public Corporation “Research and Development Institute of Railway
Engineers”, 27/1, Nizhegorodskaya Str., 109029 Moscow, Russia
[email protected]
Abstract. Spatial situations are the subject of analysis in many life support
systems, manufacturing and business. Finding a solution of applied problem in
conditions of lack of information gives rise to an interactive process of studying
the situation, the participants of which are the user-analyst and geoinformation
service. In the process of interaction, the analyst seeks to achieve situation
awareness sufficient to generate, evaluate and make rational decisions. This
paper explores the functioning of an intelligent recommender system designed
to reduce the cognitive load on the user. At the heart of its work is planning a
sequence of analysis contexts. The context refers to the information structure
necessary for the presentation of cartographic information to the analyst. It is
assumed that the geoservice has a set of contexts, and each analysis session is a
time series of contexts. The problem of minimizing the number of contexts
providing the necessary situational awareness of the analyst is considered. The
presentation of the experience of analysis by the components of level, trend and
rhythm is proposed. The results of an experimental study of the proposed
method are presented too.
Keywords: Intellectual geo-information systems Context Analysis

management
1 Introduction
Interactive study of geographical maps, schemes and plans is used for decision-making
in various fields of business, production and planning. Geographic Information Sys-
tems (GIS) provide the necessary tools for visual analysis. User analyst and GIS form a
system that solves difficult formalized tasks. Both sides play an active role in the
analysis process. The analyst is trying to comprehend the spatial image of the situation
and develop a strategy for generating, evaluating and selecting alternative solutions. To
do this, he requests useful information from big data GIS. The difficulty of selection is
Intelligent Planning of Spatial Analysis Process 11
the uncertainty, ambiguity and incompleteness of the idea of the problem being solved.
Because of this, the analysis process becomes cyclical. The initial statement of the
problem and the search for solutions are replaced by a new formulation of the purpose
of the search and the repeated generation of solutions. The cycle is accompanied by the
accumulation of a certain potential of knowledge by the analyst, called situation
awareness. However, there is a danger of cognitive overload, which blocks the
development of effective solutions. Protection against cognitive overload rests with the
GIS. Its purpose is to select the most meaningful data in terms of meaning, based on an
assessment of user behavior. The difficulty in solving this problem lies in the uncer-
tainty and ambiguity in assessing the state by the user-analyst.
Both described processes together determine the course of visual analysis. GIS is
considered as a recommendation system, which forms a stream of the most useful
cartographic images. Sequence management does not guarantee finding the best
solution to an applied problem, but it increases the likelihood of this event by
improving situational awareness and reducing the cognitive load on the analyst.
This paper discusses the planning of the analysis based on the use of contexts. The
context refers to the information structure that describes the classes of objects and
relations of a particular spatial domain. Every context reflects some semantic aspect of
the map perception. For example, the simplest context of route search in a road network
consists of highway sections, road junctions, buildings and structures. GIS can include
a wide range of contexts of various kinds. Contexts play the role of “smart filters” of
cartographic information. Changing contexts in an analysis session not only regulates
the flow of information, but also characterizes the behavior of the analyst. The selected
sequences of contexts reflect the intuitive meaning of the analysis. The stabilization of
meaning is important, since the study of useless and insignificant objects and rela-
tionships leads to cognitive overload.
A feature of the proposed approach to analysis planning is the use of expert
knowledge on the rational use of contexts. It is assumed that the intuitive meaning of
the context can be explicitly defined by valid conversions of the context instance.
Accordingly, the use of the experience of analysis will be reduced to the manipulation
of permissible transformations.
The presentation of the article is organized as follows: the results of published
studies of the problem posed are analyzed in Sect. 2; the principle of operation of the
GIS advisory subsystem is described in Sect. 3; a conceptual model of knowledge is
presented in Sect. 4; the optimal number of contexts is defined in Sect. 5; concluding
remarks and further research pathways are given in Sect. 6.
2 Review of Related Publications
Modern GIS implement the basic principles of cartographic visualization [1] in

accordance with which the control of the analysis is entirely left to the user. A char-
acteristic feature of this control option is the explicit choice by the user of the
appropriate goal of analyzing a thematic map and the manual selection of objects and
relationships that are not contained in it, but are useful for making decisions. There is
no control over the image perception by the user, although the number of objects and
12 S. Belyakov et al.
links in the image significantly affects the level of cognitive loading. Thus, the lack of
cartographic visualization method should be considered the lack of control over the
usefulness of the information provided and the user’s reaction to its content. The
analysis process merely follows the analysis pattern, which is implicitly embedded in
the thematic map. This creates the risk of inadequate perception of the spatial situation
and making the wrong decision.
The problem of limiting cognitive load has been studied for a long time. The well-
known Hick’s law [2] indicates an increase in decision-making time with an increase in
the number and complexity of the options considered. This is manifested in a decrease
in the dynamism of viewing a map or a situation scheme when analyzing cartographic
images. At the same time, it is known [3] that viewing from different angles, zooming
in, panning a cartographic fragment is a fundamental mental process necessary for an
analyst to understand the meaning of what he saw. The idea that perception plays a
crucial role in the process of interactive visual analysis is consistent with modern ideas
about the impact of visualization on creativity [4]. The results of these studies indicate
the need to minimize cognitive load, but do not provide a real way to solve the
problem.
The main means of finding effective solutions to the applied problem in the analyst-
GIS system is the situational awareness of the analyst [5]. Awareness of the goal, the
generation of subtasks corresponding to the set goal, the search for their solutions
through the perception of information about the real world situation, forecasting the
development of the situation are components of situational awareness [6]. Achieving
maximum situational awareness is a scientific and practical problem that has been
solved for many years in engineering, psychology, engineering and design [7]. For the
problem considered in this paper, these results should be adapted taking into account
the specificity of cartographic analysis.
Modern research in the field of visual informatics has led to the idea of intelligent
control (guidance) of the analysis process. In the paper [8] describes the general
principles of guidance and their implementation in various information systems. The
essence of guidance is the joint work of the recommendation system and the user in
solving the general problem of improving the quality of the result. The article notes that
the greatest difficulty is the search and use of knowledge to intellectualize the joint
work of the user and the recommendation system. It is not possible to solve this
problem by an universal way. The problem considered in this paper is specific, since
maps and schemes are both a goal and a means. By studying the map formed by the
GIS, the analyst achieves the goal of making the most effective solution to the applied
problem. In this case, the GIS uses the current state of the map as a means of assessing
the state of the analysis process. This feature of guidance has not been investigated.
The article [9] discusses the classes of data and knowledge necessary for guidance,
and also discusses a conceptual model of the process of visual analysis. The concept is
formulated as a model building framework (MBF). From our point of view, this
concept is fully consistent with the process of analyzing cartographic materials. At the
same time, the authors did not consider the problems associated with the presentation
and use of knowledge for guidance, taking into account indicators of situational
awareness and cognitive load.
3 Management of Analysis Process
Consider the implementation of the management of the analysis. We assume that the
GIS is equipped with an intelligent recommender subsystem that performs the fol-
lowing functions.
– The function of selecting useful cartographic objects in the analysis workspace. The
meaning of the function is to supplement any user request to the GIS cartographic
database with information useful for analysis. This is due to the fact that every
response to a request has semantic redundancy. Knowledge of the usefulness of
cartographic images allows to build an “intelligent filter”. The semantic orientation
is determined by the context used. Therefore, the same query in different contexts
has a different view. The selection function returns a lot of objects and relationships,
which are advisable to add to the workspace.
– The function of controlling the perception of the analysis workspace by the user.
The flow of visual information is limited by the subjective psychophysical abilities
of a person for natural reasons. Therefore, difficult-to-understand images increase
the cognitive load on the user. The complexity of the workspace is understood as
the number of cartographic objects and relationships included in it. This function
performs a change in the complexity of the workspace based on knowledge about
the significance of objects and relationships in the established context.
– The stabilization function of the semantic direction of analysis plans a sequence of
contexts corresponding to requests from the analyst. Any context makes sense. At
the same time, not every sequence of contexts can be meaningful. By stabilization
of the semantic direction of analysis, it is necessary to understand the selection of
contexts, the sequence of which can be evaluated by the user as rational. The choice
of the most promising context for continuing the dialogue is complicated by the
uncertainty in assessing the current level of situational awareness and the degree of
cognitive load. In these conditions, it seems advisable to make a choice of context
based on the knowledge of experts who have experience in conducting analysis
using GIS. The presentation and use of knowledge is the basis for the function in
question.
The joint work of the analytic-GIS system is carried out according by the following
algorithm:
1. The analyst is registered in the session with GIS, determining the context of the
upcoming analysis.
2. The analyst forms a request to the GIS cartographic database through the user
interface.
3. The recommender subsystem calls the function of stabilization of the semantic
direction of analysis. If the response to the request does not contradict the meaning
of the current context, then it is planned for future use. Otherwise, among the
contexts described in the system, there is one that retains the most meaning of the
current context. This context is planned for future use.
4. The recommender subsystem invokes the function of selecting useful cartographic
objects in the planned context. The response is added to the workspace.
5. The recommender subsystem calls the function of perception control.

6. GIS visualizes the workspace.
7. If the session is not completed, go to step 2.
8. End the session.
As can be seen from the algorithm, steps 3–5 implement joint control over the
analysis process. Increasing situational awareness is provided in steps 3 and 4, the
limitation of cognitive load is ensured by completing step 5.
4 Conceptual Model of Knowledge
The context in context-sensitive systems includes any information that allows you to
identify the current situation and take adequate actions to solve the problem [10]. The
conceptual and logical structure of the context depends on the application area. Sta-
bilization of the semantic direction of analysis requires a special conceptual model of
knowledge about sequences of contexts in an analysis session. The concept of
“meaning” is proposed to be formally reflected in the description of permissible
changes in the sequence of contexts, within the framework of which the intuitively
understood essence of the analysis process is preserved. If the boundaries of permis-
sible changes are violated, then the meaning of the analysis in the context used is
considered lost.
We denote by C ¼ fck g many contexts, in each of which the local semantic content
of images is supported. Locality means that the data and knowledge of the context are
applicable in the field with a certain spatial, temporal and semantic boundary, which is
explicitly specified in the description of the context. An analysis session generates a
sequence of contexts C of arbitrary length with repeating elements. Any C sequence is
a precedent for the collaboration of GIS and analyst. Consider the model of the image
of the sequence of contexts of C m , which describes the semantic direction of analysis,
in the form:

C m ¼ C; H C ; ð1Þ
Here, C is the center of the image, which should be a precedent, on the basis of which
the expert determines the subjective understanding of the semantic direction of anal-
ysis. The HðCÞ component is a set of permissible transformations of the center that
preserve its meaning. Essentially, Cm model represents a generalization made by an
expert based on a reflection of a precedent.
The image model has the following advantages:
1. Case analysis [11, 12] identifies the similarity of precedents with their semantic
proximity. However, far from always small deviations of the precedent parameters
mean the preservation of its meaning. This conclusion is not reliable. Therefore, it is
advisable to expand the comparison of precedents with additional information;
2. Precedents do not reflect the in-depth knowledge of expert analysts about trans-
ferring the existing experience of analysis to other situations; do not allow making
reliable decisions in new, previously unstudied conditions. Indication of permissible

transformations eliminates the incorrect transfer of experience.
The adequacy of the mapping of the semantic direction by the set of permissible
HðCÞ transformations in (1) is determined by the organization of the procedure for
obtaining expert knowledge. Note that permissible transformations are visualized
cartographic objects that allow the analyst to transfer knowledge in a figurative form in
the familiar user interface.
Knowledge of the sequence of contexts in a session needs to be structured to
increase the reliability of the conclusions. We propose to present C with the following
characteristics:
– by level (LðCÞ), which is the set of contexts used in the session, i.e. LðCÞC. The
level characterizes the use of contexts in a session without regard to their sequence
and possible repetitions;
– by trend (RðCÞ), which sets the direction of the analysis. The direction RðCÞ is
defined by a subset of contexts RðCÞ LðCÞ that substantially determine the
beginning and end of the process, i.e. jRðCÞj [ 1;
– by rhythm (BðCÞ), which reflects the temporal nature of the sequence of contexts.
The rhythm is described by a sequence of arbitrary length BðCÞ ¼ ðc0 ; t0 ; c1 ;
t1 ; ::ck Þ; ci 2 C. Here tj are time intervals. The sequence BðCÞ reflects periodic
returns to information sources, a slowdown or acceleration of the analysis process,
etc.
Knowledge of each characteristic can be represented in the form of images showing
acceptable transformations.
5 Optimal Number of Contexts
Consider the issue of the applicability of context-based planning related to their

number. The feasibility of switching GIS from one context to another is determined by
the ratio of the two components of the complexity of the analysis. The first component
(U1 ) is determined by the costs of the analyst to study the set of available contexts.
Without such preliminary work, it is impossible to make a reasonable choice of the
desired context. As the number of contexts (Q) increases, these costs increase. For
small values Q, growth can be considered linear - U1 ¼ u1 Q.
Here u1 is the unit cost per context.
The study of the workspace in the established context generates the second com-
ponent of the complexity (U2 ). With an increase in the number of contexts, these costs
decrease due to the fact that the analyst receives less redundant data. It can be assumed,
that U2 ¼ u2 =Q. Here u2 are the unit costs of the context. The expression for total costs
takes the form:
U1 þ U2 ¼ u1 Q þ u2 =Q:
It is easy to verify that the minimum cost is achieved when

pffiffiffiffiffiffiffiffiffiffiffiffi
Q ¼ u2 =u1 : ð2Þ
This expression (2) allows us to draw the following conclusions:

– the analyst has the ability to manipulate the workspace with minimal effort
regardless of the context chosen. This follows from the fact that the complexity is of
minimal importance;
– the effectiveness of the analysis process is significantly affected by the number of
contexts. There is a certain value of Q, providing a compromise between the
growth of the complexity of studying the system of contexts and the increasing
complexity of the analysis of the workspace in the established context;
– with significant individual differences in the ability of users to study the system of
contexts, it is advisable to limit the number of available contexts. The values u1 and
u2 can become individual tuning parameters of the user profile.
6 Conclusion
Planning the analysis process in dialogue with the GIS recommendation subsystem
plays an important role in solving difficult formalized spatial problems. The proposed
conceptual model of knowledge representation allows stabilizing the semantic direction
of analysis and increasing the efficiency of the search for solutions to the problem.
Further research is supposed to be carried out in the direction of improving the
mechanisms for extracting expert knowledge in the form of images of a sequence of
contexts.
Acknowledgments. The reported study was funded by the Russian Foundation for Basic
Research according to the research projects N 18-01-00023, N 19-07-00074.
References
1. Huang, B.: Comprehensive Geographic Information Systems. Elsevier, Amsterdam (2017)
2. Hick, W.E.: On the rate of gain of information. Q. J. Exp. Psychol. 4(1), 11–26 (1952)
3. Gibson, J.J.: A theory of direct visual perception. In: Royce, J., Rozenboom, W. (eds.) The
Psychology of Knowing. Gordon & Breach, New York (1972)
4. Palmiero, M., Nori, R., Piccardi, L.: Visualizer cognitive style enhances visual creativity.
Neurosci. Lett. 615, 98–101 (2016)
5. Endsley, M.R.: Design and evaluation for situation awareness enhancement. In: Proceedings
of the human Factors Society 32nd Annual Meeting, Santa Monica, pp. 97–101. Human
Factors Society, CA (1988)
6. Endsley, M.R., Bolte, B., Jones, D.G.: Designing for Situation Awareness: An Approach to
Human-Centered Design. Taylor & Francis, London (2003)
7. Ziemke, T., Schaefer, K.E., Endsley, M.: Situation awareness in human-machine interactive
systems. Cogn. Syst. Res. 46, 1–2 (2017)
8. Ceneda, D., Gschwandtner, T., Miksch, S.: You get by with a little help: the effects of
variable guidance degrees on performance and mental state. Vis. Inform. 3, 177–191 (2019)
9. Collins, C., Andrienko, N., Schreck, T., Yang, T., Choo, J., Engelke, U., Jena, A., Dwyer,
T.: Guidance in the human–machine analytics process. Vis. Inform. 2, 166–180 (2018)
10. Dey, A., Abowd, G.: Towards a better understanding of context and context-awareness. In:
CHI 2000 Workshop on the What, Who, Where, When, and How of Context-Awareness,
pp. 304–307 (2000)
11. Aamodt, A., Plaza, E.: Case-based reasoning: foundational issues, methodological
variations, and system approaches. AI Commun. 7(1), 39–59 (1994)
12. Yan, A., Yu, H., Wang, D.: Case-based reasoning classifier based on learning pseudo metric
retrieval. Expert Syst. Appl. 89, 91–98 (2017)
The Method of Finding the Base Set
of Intuitionistic Fuzzy Graph
Alexander Bozhenyuk1(&) , Stanislav Belyakov1 ,

Janusz Kacprzyk2 , and Margarita Knyazeva1
1
Southern Federal University, Nekrasovsky 44, 347922 Taganrog, Russia
[email protected], [email protected],
[email protected]
2
Systems Research Institute, Polish Academy of Sciences,
Newelska 6, 01-447 Warsaw, Poland
[email protected]
Abstract. In this paper, we consider the concepts of intuitionistic fuzzy path and
minimal intuitionistic fuzzy base of an intuitionistic fuzzy graph. Based on them,
we introduce the concept of a basic set of the intuitionistic fuzzy graph. The base
set is an invariant of the intuitionistic fuzzy graph, that is, it does not change
during its structural transformations. The properties of minimal bases and basic
set of the intuitionistic fuzzy graph are considered too. An approach to finding all
the minimal intuitionistic fuzzy bases is proposed. Finding all the minimal bases
allows us to find the basic set of an intuitionistic fuzzy graph. The proposed
algorithm is a generalization of the algorithm for clear and fuzzy graphs. An
example of finding the basic set of an intuitionistic fuzzy graph is considered.
Keywords: Intuitionistic fuzzy set Intuitionistic fuzzy graph Intuitionistic

fuzzy path Minimal intuitionistic fuzzy base Base set
1 Introduction
Currently, science and technology are characterized by complex processes and phe-
nomena for which complete information is not always available. For such cases, mathe-
matical models of various types of systems containing elements of uncertainty have been
developed. A large number of these models are based on the expansion of the usual set
theory, namely, fuzzy sets. The concept of fuzzy sets was introduced by L. Zadeh [1] as a
method of representing uncertainty and fuzziness. Since then, the theory of fuzzy sets has
become an area of research in various disciplines. In 1983, K. Atanassov [2] introduced
the concept of intuitionistic fuzzy sets as a generalization of fuzzy sets. He added a new
component to the definition of a fuzzy set, which determines the degree of non -
memership. Intuitionistic fuzzy sets are fuzzy sets of a higher order. Their application
makes the solution procedure more complicated, but if the complexity of calculations in
time, volume of calculations, or memory can be neglected, then a better result can be
achieved.
The theory of fuzzy graphs is finding an increasing number of applications for
modeling real-time systems, where the level of information inherent in the system
depends on different levels of accuracy. The original definition of a fuzzy graph [3] was
The Method of Finding the Base Set of Intuitionistic Fuzzy Graph 19
based on fuzzy relations by L. Zadeh [4]. In [5], fuzzy analogs of several basic
graphical concepts were presented. In [6, 7], the notion of fuzzy graph complement was
defined and some operations on fuzzy graphs were studied. The concepts of intu-
itionistic fuzzy relations and intuitionistic fuzzy graphs were introduced in [8, 9] and
some of their properties were investigated. In [10–12], the concepts of a dominating
set, a regular independent set, a domination edge number of edges in intuitionistic
fuzzy graphs were considered.
In this paper we introduce the concepts of minimal intuitionistic base subset and
base set in intuitionistic fuzzy graphs. These concepts are a generalization of the
minimal dominating vertex subsets of a crisp graph [13] and the domination set of a
fuzzy graph [14], respectively.
The contents of this paper are as follows: In Sect. 2, we review some concepts and
definitions of intuitionistic fuzzy graphs. In Sect. 3, we introduce a minimal intu-
itionistic fuzzy base, a base set of intuitionistic fuzzy graph, and consider some of their
properties. In Sect. 4, we propose and justify the method for finding base set, and
consider an example of its finding. The last section deals with conclusions and future
directions.
2 Basic Concepts and Definitions of Intuitionistic Fuzzy

Graphs
Definition 1 [1]. Let X be a nonempty set. A fuzzy set drawn A from X is defined as
A ¼ fhlA ðxÞ; xijx 2 Xg; where lA : X ! ½0; 1 is the membership function of the fuzzy
set A.
Definition 2 [2]. Let X be a nonempty set. An intuitionistic fuzzy set A in X is an
object having the form: A ¼ fhx; lA ðxÞ; vA ðxÞijx 2 Xg; where the functions
lA ðxÞ; vA ðxÞ : X ! ½0; 1 define respectively, the degree of membership and degree of
non-membership of the element x 2 X to the set A, which is a subset of X, and
8x 2 X½lðxÞ þ vA ðxÞ 1:
Let p and q be intuitionistic fuzzy variables that have the form: p ¼
ðlðpÞ; mðqÞÞ; q ¼ ðlðqÞ; mðqÞÞ; here lð pÞ þ mð pÞ 1 and lðqÞ þ mðqÞ 1: Then the
operations “&” and “˅”are defined as [15]:
p & q ¼ ðminðlðpÞ; lðqÞÞ; maxðmðpÞ; mðqÞÞÞ; ð1Þ
p _ q ¼ ðmaxðlðpÞ; lðqÞÞ; minðmðpÞ; mðqÞÞÞ: ð2Þ
We assume that p < q if l(p) < l(q) and m(p) > m(q).
Definition 3 [3]. A fuzzy graph is a triplet G e ¼ ðV; r; lÞ; where V is finite and non-
empty vertex set, r : V ! ½0; 1 is a fuzzy subset of V, and l : V V ! ½0; 1 is fuzzy
relation on V V such that: ð8x; y 2 VÞ½lðx; yÞ minðrðxÞ; rðyÞ:
20 A. Bozhenyuk et al.
This definition considers a fuzzy graph as a collection of fuzzy vertices and fuzzy
edges. Another version of a fuzzy graph was proposed in [7, 16] as a set of crisp
vertices and fuzzy edges:
Definition 4 [16]. A fuzzy graph is a pair Ge ¼ ðV; RÞ; where V is a crisp set of vertices
and R is a fuzzy relation on V, in which the elements (edges) connecting the vertices V,
have the membership function lR: V V ! ½0; 1:
Such fuzzy graph in [16] was called a fuzzy graph of the first kind.
Definition 5 [8, 9]. An intuitionistic fuzzy graph is a pair G e = (A, B), where A ¼
hV; lA ; mR i is an intuitionistic fuzzy set on the set of vertices V, and B ¼ hV
V; lB ; mB i is an intuitionistic fuzzy relation such that:
lB ðxyÞ minðlA ðxÞ; lA ðyÞÞ and mB ðxyÞ maxðmA ðxÞ; mA ðyÞÞ;
and the following condition is fulfilled: ð8x; y 2 VÞ ½0 lB ðxyÞ þ mB ðxyÞ 1:

Definition 6 [17]. An intuitionistic fuzzy graph of the first kind is a pair Ge ¼ ðV; U Þ;
where V is a crisp set of vertices, U ¼ hV V; l; mi is intuitionistic fuzzy relation
(intuitionistic fuzzy edges) such that: ð8x; y 2 VÞ ½0 lðx; yÞ þ m ðx; yÞ 1:
3 Base Set of Intuitionistic Fuzzy Graph
Let G e ¼ ðV; U Þ be an intuitionistic fuzzy graph of the first kind. Let pðx; yÞ ¼
ðlðx; yÞ; mðx; yÞÞ be an intuitionistic fuzzy variable that determines the degree of
adjacency and the degree of non-adjacency from vertex x to vertex y.
Let us give a definition of a fuzzy intuitionistic path from vertex xi to vertex xj in an
e xi ; xj 2 V:
intuitionistic fuzzy oriented graph G;
Definition 7. The oriented intuitionistic fuzzy path of the graph G e is a directed

e
sequence of intuitionistic fuzzy edges Lðxi ; xj Þ leading from the vertex xi to the vertex xj
in which the final vertex of any edge other than the last is the initial vertex of the next
edge.

The strength q e Lðxi ; xj Þ of the path e Lðxi ; xj Þ is determined by the expression:

e
q Lðxi ; xj Þ ¼ & p xa ; xb , where operation & is determined according to
e
xa ;xb 2 L ðxi ;xj
(1) and (2).

Obviously, this definition coincides with the same definition for crisp and fuzzy
graphs.
Definition 8. A vertex xj is reachable from a vertex xi in a graph if there is an

intuitionistic fuzzy path e
Lðxi ; xj Þ with the strength q eLðxi ; xj Þ 6¼ ð0; 1Þ:
Definition 9. The degree of intuitionistic fuzzy reachability of the vertex xj from the
vertex xi is the quantity:

c xi ; xj ¼ max q eLðxi ; xj Þ ;
k21;p
where p is the number of different oriented paths from the vertex xi to xj.
Let us agree that each vertex xi2V in an intuitionistic fuzzy graph is intuitionistic
fuzzy reachable on its own with reachability degree (1,0).
Example 1. For the intuitionistic fuzzy graph shown in Fig. 1, the vertex x1 is not
reachable from the vertex x5, and the vertex x5 is intuitionistic fuzzy reachable from the
vertex x1 with degree:
cðx1 ; x5 Þ ¼ ðð0:2; 0:4Þ & ð0:3; 0:1Þ & ð0:7; 0:2ÞÞ _ ð0:2; 0:6Þ _
ðð0:5; 0::3Þ & ð0:8; 0:1ÞÞ ¼ ð0:5; 0:3Þ:
x2 (0.3,0.1) x4
(0.2,0.4) (0.7,0.2)
(0.2,0.6)
x1 x5
(0.5,0.3) (0.8,0.1)
x3
Fig. 1. Example intuitionistic fuzzy graph with degree c(x1, x5) = (0.5, 0.3).
e ¼ ðV; U Þ with a number of vertices jV j ¼ n be

Let an intuitionistic fuzzy graph G
given.
Definition 10. A minimal intuitionistic fuzzy base with degree a is a subset of vertices
B V from which any vertex of the graph is reachable with a degree a, and which is
minimal in the sense that there is no subset B0 V with the same reachability.
e
We denote by RðBÞ the intuitionistic fuzzy set of vertices that are reachable from an
arbitrary subset B V. Then the subset B is a minimal intuitionistic fuzzy base with
degree a if and only if the conditions are satisfied:

e ðBÞ ¼ cj =xj xj 2 X& 8j ¼ 1; n cj a ;
R ð3Þ
e 0 Þ ¼ f\c0j =xj [ jxj 2 X & ð9j ¼ 1; nÞðc0j \aÞg

ð8B0 BÞ ½ RðB ð4Þ
Condition (3) means that any vertex either enters the subset B, or it is reachable from
some vertex of the same set with a degree of reachability of at least a. Condition (4)
indicates that any subset B0 B does not possess property (3).
The following property follows from the definition of a minimal intuitionistic fuzzy
base:
Property 1. Let cðxi ; xj Þ be the degree of reachability of vertex xj from a vertex xi. Then
the statement is true:
ð8xi ; xj 2 BÞ ½cðxi ; xj Þ\a: ð5Þ
In other words, the reachability of some vertex of a minimal base from any other
vertex of the same base is less than the value a.
We denote by sXk ¼ fXk1 ; Xk2 ; . . .; Xkl g the family of minimal intuitionistic fuzzy
bases consisting of k vertices with degrees ak1 ; ak2 ; . . .; akl . Denote by
ak ¼ maxfak1 ; ak2 ; . . .; akl g. The value ak means that there is a subset consisting of
k vertices from which all the remaining vertices are reachable with a degree ak and
there is no subset consisting of k vertices from which all remaining vertices would be
reachable with a degree more then ak . If the family sXk ¼ ;, then we define ak ¼ ak 1 .
Definition 11. We call set B e ¼ f\a1 =1 [ ; \a2 =2 [ ; . . .; \an =n [ g base set of
e
intuitionistic fuzzy graph G.
The base set is an invariant of an intuitionistic fuzzy graph, that is, it does not
change during its structural transformations.
Property 2. The following expression is true:
ð0; 1Þ a1 a2 . . . an ¼ ð1; 0Þ: ð6Þ
4 The Method for Finding Base Set
We will consider the method of finding a family of all minimal intuitionistic fuzzy base.
Let us assume that set Ba is a minimal intuitionistic base of the graph Ge with the
intuitionistic degree a = (µa,ma). Then for an arbitrary vertex xi2V, one of the fol-
lowing conditions must be true.
ð8xi 2 VÞ½xi 2 Ba _ ð9xj 2 Ba jlðxj ; xi Þ la & mðxj ; xi Þ ma Þ: ð7Þ
To each vertex xi 2 V we assign Boolean variable pi that takes value 1, if xi 2 Ba

and 0 otherwise. We assign the intuitionistic variable nji = a = (µa,ma) for the
proposition ((xj,xi),m(xj,xi)) a. Passing from the quantifier form of proposition (7) to

the form in terms of logical operations, we obtain a true logical proposition:
UB ¼ & ðpi _ _ ðpj & nji ÞÞ:

i¼1;n j¼1;n
Here, n = |V|. Supposing nii ¼ ð1; 0Þ and considering that the equality pi _
_ pj & nji ¼ _ pj & nji is true for any vertex xj, we finally obtain:
j j
UB ¼ & ð _ ðpj & nji ÞÞ: ð8Þ

i¼1;n j¼1;n
We open the parentheses in the expression (8) and reduce the similar terms by
following rules:
a _ a & b ¼ a; a & b _ a & b ¼ a; n1 & a _ n2 & a & b ¼ n1 & aif n1 n2 : ð9Þ
Here, a, b 2 {0,1} and n1,n2 2 [(0,1), (1,0)].

Then the expression (8) may be presented as
UB ¼ _ ðp1i & p2i &. . .& pki & ai Þ: ð10Þ

i¼1;l
We may prove next property:

Property 3. Each disjunctive member in the expression (8) gives a minimum intu-
itionistic base subset with the degree ai.
This property allows you to find all minimum intuitionistic base subsets that make it
possible to define a base set.
Let’s consider an example of finding domination set of graph G e shown in Fig. 2.
x2 (0.2,0.6) x3
(0.3,0.1)
(0.5,0.2)
(0.3,0.1)
(0
.9
,0.
0
)
(0.2,0.4)
x1 x4
e
Fig. 2. Intuitionistic fuzzy graph G.
e has a view:
The adjacent matrix of this graph G
The reachability matrix of this graph will have the form:
The expression (8) for this reachability matrix has the following form:
UB ¼ ½ð1:0; 0:0Þp1 & ½ð0:5; 0:2Þp1 _ ð1:0; 0:0Þp2 & ½ð0:2; 0:6Þp1 _ ð0:2; 0:6Þp2
_ ð1:0; 0:0Þp3 & ½ð0:5; 0:2Þp1 _ ð0:9; 0:0Þp2 _ ð0:3; 0:1Þp3 _ ð1:0; 0:0Þp4 :
Multiplying parenthesis 1 and 2, parenthesis 3 and 4, and using rules (7) we obtain:
UB ¼ ½ð1:0; 0:0Þp1 p2 _ ð0:5; 0:2Þp1 & ½ð0:2; 0:6Þp1 _ ð0:2; 0:6Þp2 _ ð0:5; 0:2Þp1 p3
_ ð0:9; 0:0Þp2 p3 _ ð0:3; 0:1Þp3 _ ð1:0; 0:0Þp3 p4 :
Multiplying the parentheses, we finally obtain:
UB ¼ ð0:2; 0:6Þp1 _ ð0:5; 0:2Þp1 p3 _ ð0:9; 0:0Þp1 p2 p3 _ ð1:1; 0:0Þp1 p2 p3 p4 :
~ has 5 minimal intuitionistic bases, and base set is defined as:

It follows that graph G
e ¼ f\ ð0:2; 0:6Þ=1 [ ; \ð0:5; 0:2Þ=2 [ ; \ð0:9; 0Þ=3 [ ; \ð1; 0Þ=4 [ g:

B
This base set shows, in particular, that there is a subset in the graph (B = {x1}),
consisting of 1 vertex such that all other vertices of the graph (D\B = {x2, x3, x4}) are
reachable with a degree of at least (0.2, 0.6).
5 Conclusion
In this paper we considered the concepts of fuzzy minimal intuitionistic base subsets
and base set of intuitionistic fuzzy graph. The method of finding families of all fuzzy
minimal base subsets and base set were considered. This method is the generalization
of Maghout’s method for fuzzy graphs. The base set definition allows to estimate the
any intuitionistic fuzzy graph with the position of existence of fuzzy invariants.
It should be noted that the considered method is a method of complete ordered search,
since such tasks are reduced to coverage problems, i.e. these tasks are NP-compete
tasks. However, the proposed method can be effective for intuitionistic fuzzy graphs
with inhomogeneous structure and not large dimensionality. In the future, the con-
sidered method and algorithm is supposed to be used to search for other invariants, in
particular, antibases and various centers of intuitive fuzzy graphs. More similar results
and applications will be reported in upcoming papers.
Research according to the research projects N 18-01-00023, N 20-01-00197.
References
1. Zadeh, L.A.: Fuzzy sets. Inf. Control 8, 338–353 (1965)
2. Atanassov, K.T.: Intuitionistic Fuzzy Sets. In: Proceedings of VII ITKR’s Session, Central
Science and Technical Library, vol. 1697/84, pp. 6–24. Bulgarian Academy of Sciences,
Sofia, Bulgarian (1983)
3. Christofides, N.: Graph theory An algorithmic approach. Academic press, London (1976)
4. Zadeh, L.A.: Similarity relations and fuzzy orderings. Inf. Sci. 3(2), 177–200 (1971)
5. Rosenfeld, A.: Fuzzy graphs. In: Zadeh, L.A., Fu, K.S., Shimura, M. (eds.) Fuzzy sets and
their applications, pp. 77–95. Academic Press, New York (1975)
6. Mordeson, J.N., Nair, P.S.: Fuzzy Graphs and Fuzzy Hypergraphs. Springer, Heidelberg
(2000)
7. Sunitha, M.S., Kumar, A.V.: Complement of a fuzzy graph. Indian J. Pure Appl. Math. 33
(9), 1451–1464 (2002)
8. Shannon, A., Atanassov, K.T.: A first step to a theory of the intuitionistic fuzzy graphs. In:
Lakov, D. (ed.) Proceeding of the FUBEST, pp. 59–61. Sofia, Bulgaria (1994)
9. Shannon, A., Atanassov, K.T.: Intuitionistic fuzzy graphs from -, - and (, b)-levels. Notes
Intuitionistic Fuzzy Sets 1(1), 32–35 (1995)
10. Karunambigai, M.G., Sivasankar, S., Palanivel, K.: Different types of domination in
intuitionistic fuzzy graph. Ann. Pure Appl. Math. 14(1), 87–101 (2017)
11. Parvathi, R., Thamizhendhi, G.: Domination in intuitionistic fuzzy graphs. Notes Intuition-
istic Fuzzy Sets 16, 39–49 (2010)
12. Velammal, S.: Edge domination in intuitionistic fuzzy graphs. Int. J. Comput. Sci. Math. 4
(2), 159–165 (2012)
13. Ore, O.: Theory of Graphs. American Mathematical Society Colloquium Publications, vol.
38, p. 270 (1962)
14. Bershtein, L.S., Bozhenuk, A.V.: Maghout method for determination of fuzzy independent,
dominating vertex sets and fuzzy graph kernels. Int. J. Gen Syst 30(1), 45–52 (2001)
15. Atanassov, K., Gargov, G.: Elements of intuitionistic fuzzy logic. Part I Fuzzy Sets Syst. 95,
39–52 (1998)
16. Bershtein, L., Bozhenyuk, A.: Fuzzy graphs and fuzzy hypergraphs. In: Dopico, J., de la
Calle, J., Sierra, A. (eds.) Encyclopedia of Artificial Intelligence, pp. 704–709. Informa-
tion SCI, Hershey (2008)
17. Bozhenyuk, A., Knyazeva, M., Rozenberg, I.: Algorithm for finding domination set in
intuitionistic fuzzy graph. In: Atlantis Studies in Uncertainty Modelling, vol. 1, pp. 72–76
(2019)
Intuitionistic Fuzzy Assessments
of the Abdominal Aorta and Its Branches
Valetin Vasilev1, Evdokia Sotirova1(&), Krassimir Atanassov1,2,

and Sotir Sotirov1
1
Prof. Dr. Assen Zlatarov University, “Prof. Yakimov” Blvd.,
8010 Burgas, Bulgaria
[email protected], {esotorova,ssotorov}@btu.bg
2
Department of Bioinformatics and Mathematical Modelling,
Institute of Biophysics and Biomedical Engineering, Bulgarian Academy
of Sciences, Acad. G. Bonchev Str., Bl. 105, 1113 Sofia, Bulgaria
[email protected]
Abstract. The human body is an open stationary system and has different
adaptation mechanisms in changing conditions of the external and internal
environment. In this aspect there is no exception to the circulatory system,
respectively the arteries of the lower limbs. The narrowed or blocked arteries
reduce blood flow to the lower limbs. In this case, the blood flows through the
smaller blood vessels and the anastomoses (connections) between them above
and below the thrombosis of the main vessel. In this research, an approach using
an intelligent tool for analyzing data through fuzzy logic is presented. The
intuitionistic fuzzy assessments for the blood supply evaluation of the lower
limbs with peripheral artery disease are suggested. The permeability of the artery
is within a certain range, limited by the minimum and maximum thresholds. The
intuitionistic fuzzy evaluations determine the degree of permeability of the
artery. They are formed on the basis of a set of intuitionistic fuzzy estimations
hl; mi of real numbers from the set [0, 1] [0, 1].
Keywords: Intelligent systems Intuitionistic fuzzy sets Modelling

Abdominal aorta Vascular system
1 Introduction
The human body is an open stationary system and as such possesses numerous
adaptation mechanisms in the changing conditions of the external and internal envi-
ronment. In this respect, the circulatory system, including the arteries of the lower
extremities, is no exception [8, 11, 13]. In acute and chronic blockages (thrombosis)
blood begins to move on alternate pathways. It would be of particular interest to
evaluate the possibility of permeability of the arteries. The combination of intelligent
systems and fuzzy instruments for analysis the data is an object of many active
investigations in the recent years.
In acute trunk thrombosis, adaptation mechanisms are severely restricted, especially
in younger individuals [1]. This is because in the elderly, and especially in those over
Intuitionistic Fuzzy Assessments of the Abdominal Aorta and Its Branches 27
65 years of age, more advanced atherosclerosis of the vessels is observed leading to

single or floor narrowing (stenoses) and even to chronic thrombosis, which slowly
generate compensatory mechanisms over time, consisting primarily of the development
of collateral (lateral) pathways along which the blood moves through the smaller blood
vessels and the anastomoses (connections) between them above and below the trunk
thrombosis [12].
In chronic abdominal aortic thrombosis in place of separation of both pelvic arteries
or thrombosis of the pelvic arteries itself, blood moves along collateral of the small
pelvis, deep iliac artery branches, which anastomose with branches of the common
femoral artery and thus circumvent the occlusion [10]. This blood supply to the limbs is
not effective, but their vitality is temporarily preserved and this allows for diagnosis
and planned treatment of the disease [9].
In chronic occlusion of the superficial femoral artery (SFA), the blood finds a
collateral path through the branches of the profunda femoral artery (PFA), which
anastomoses with the so-called geniculate arteries [7]. They are branches of both PFA
and SFA, and this makes it possible to create an alternative blood pathway bypassing
thrombosis.
In the popliteal arteries - anterior tibial artery, posterior tibial artery and peroneal
artery, in the conditions of chronic occlusion of any of them, collaterals develop to those
who are passable and this way they compensate for the lack of blood in the blood supply
regions of the thrombosed artery [14]. In the case of thrombosis of all three or direct
occlusion of the popliteal artery, located above the site of the division of the tibial arteries,
some of the genicular (pericardial) arteries find anastomoses with small vessels of the
knee and develop alternative blood pathways. In any case, the body looks for ways and
means to compensate for the disease process, in this case the effects of atherosclerosis, and
to preserve as much as possible the integrity of all organs and systems.
Subject of the study is to introduce an intelligent method for evaluation the blood
supply of the lower limbs with peripheral artery disease. As an evaluation tool the
assessments from the theory of Intuitionistic fuzzy sets (IFSs, see [2–4]) were used.
IFSs represent an extension of the concept of fuzzy sets, as defined by Zadeh [15],
exhibiting function lA ð xÞ defining the membership of an element x to the set A, eval-
uated in the [0; 1]-interval. The difference between fuzzy sets and IFSs is in the presence
of a second function mA ð xÞ defining the non-membership of the element x to the set A,
where lA ð xÞ 2 ½0; 1; mA ð xÞ 2 ½0; 1, under the condition of lA ð xÞ þ mA ð xÞ 2 ½0; 1.
The theory of IFS contains many developed techniques not only for registering but
also quantification of the so inherent to the real processes and objects imprecision and
uncertainty.
The paper is organized as follows. Section 2 presents the intuitionistic fuzzy
evaluation model. The analysed data are connected with patients with reduced blood
flow to the lower limbs because of narrowed or blocked arteries. Section 2.1 show the
intuitionistic fuzzy assessments of the arteries’ permeability and arteries’ non-
permeability of the data for 202 (126 men and 76 women) patients with diabetes
mellitus treated with endovascular procedures at the University Hospital in Burgas for
2016–2019. Section 2.2 presents the calculation the Intuitionistic fuzzy threshold
values. Finally, Sect. 3 presents our conclusions and future work. Section 4 and Sect. 5
are acknowledgments and references.
28 V. Vasilev et al.
2 Proposed Assessment Model
In [6] a model of the abdominal aorta and its branches as a part of the vascular system
is presented. Objective criteria and parameters can be found which through the model
would give us a clearer understanding about the state of the Cardiovascular System,
presence and weight of the pathology, the risk degree and a relatively correct prognosis
for the development of one disease or another.
The approach proposed here for the evaluation of the abdominal aorta and its
branch-es extends the capabilities of the model developed in [6].
2.1 Intuitionistic Fuzzy Estimations

The presented here intuitionistic fuzzy estimations reflect the degree of each arteries’
permeability l, or arteries’ non-permeability m of the lower limbs, and for them is valid
that l þ m 1 and the ordered pairs hl; mi of real numbers were defined in the sense of
intuitionistic fuzzy sets.
Thus, the evaluation of the arteries’ permeability and arteries’ non-permeability is
obtained according to the following formulas:
b
l¼ ð1Þ
300
and
a
m¼1 ; ð2Þ
300
where:
• a is a higher measured value of the permeability of the arteries (during the systolic
pressure);
• b is a lower measured value of the permeability of the arteries (during the diastolic
pressure).
The degree of uncertainty p ¼ 1 l m represents such cases wherein there is no
information for the blood flow.
The actual data for 202 (126 men and 76 women) patients with diabetes mellitus
treated with endovascular procedures at the University Hospital in Burgas, Bulgaria for
2016–2019 are analyzed. The average age of patients is 66.5 years (from 41 to 92
years).
A significant number of comorbidities were found in all patients, with their total
number exceeding 100%, as more than one comorbidity was detected in some patients.
In the observed data there are:
• 89 patients with hypertension (44.06%);
• 102 patients with ischemic heart disease (50.5%);
• 73 patients with obesity (body mass index > 30) (36.14%);
• 32 patients with cerebrovascular disease (15.84%);
• 26 patients with chronic renal failure (12.87%).
To calculate the intuitionistic fuzzy estimations hl; mi according to formulas (1) and
(2), the averaged values of the blood pressure of the patients group are used:
• Group of patients with hypertension: hlh ; mh i ¼ h0:320; 0:520i, averaged blood
pressure 144/96;
• Group of patients with ischemic heart disease: hlihd ; mihd i ¼ h0:310; 0:543i, aver-
aged blood pressure 137/91;
• Group of patients with obesity (BMI > 30): hlo ; mo i ¼ h0:290; 0:570i, averaged
blood pressure 129/88;
• Group of patients with cerebrovascular disease (15.84%): hlcd ; mcd i ¼ h0:283;
0:547i, averaged blood pressure 136/85;

• Group of patients with chronic renal failure (12.87%): lcrf ; mcrf ¼ h0:287; 0:560i,
averaged blood pressure 132/86.
2.2 Intuitionistic Fuzzy Threshold Values

The permeability of the artery is within a certain range, limited by the minimum and
maximum thresholds, as a follows:
• q – minimal threshold of the value of the permeability of the current place of the
abdominal aorta and its branches;
• p – maximal threshold of the value of the permeability of the current place of the
abdominal aorta and its branches.
The normal condition is
q \ l \ 1 m \ p:
The ordered pair ha; bi is called an Intuitionistic Fuzzy Pair (IFP, see [5]) if and
only if a, b 2 [0, 1] and a + b 1. For two IFPs ha; bi and hc; d i,
ha; bi hc; d i if and only if a c and b d:
Let us have n observations hl1 ; m1 i; hl2 ; m2 i; . . .; hln ; mn i of evaluation of the

arteries’ permeability from the top downwards of a healthy person. To every IFP, each
pair is also associated with a bi number, representing the amount of blood passed. Then
the inequalities
hl1 ; m1 i hl2 ; m2 i . . . hln ; mn i and b1 b2 . . . bn
hold. When there exists i, 1 i n 1 for which

hli ; mi i \ li þ 1 ; mi þ 1 and bi [ [ bi þ 1 ;
then, there is blockage of a blood vessel between i-th and (i+1)-st regions.
30 V. Vasilev et al.
3 Conclusion
This paper suggests an intelligent method for evaluation the blood supply of lower
limbs with peripheral artery disease by introducing intuitionistic fuzzy assessments. In
this way, the blockage of the abdominal aorta and its branches can be determined
because, as the blood vessel becomes obstructed, the permeability below sharply
decreases and increases in places above the obstruction. This is reflected in the system
by exceeding the relevant minimum or maximum thresholds and specifying the exact
location of the obstruction.
In our next work, a more detailed description of the intuitionistic fuzzy assessments
and local (location-related) thresholds will be provided.
Acknowledgment. The authors are grateful for the support provided by the Bulgarian Ministry
of Education and Science under the National Research Programme “Information and Commu-
nication Technologies for a Digital Single Market in Science, Education and Security” approved
by “DCM # 577/17.08.2018”.
The authors declare that there is no conflict of interest regarding the publication of this paper.
References
1. Adelman, R.: Abdominal aortic aneurysm 18 years after apparent resolution of an umbilical
catheter-associated aortic thrombosis. J. Pediatr. 132(5), 874–875 (1998)
2. Atanassov, K.: Intuitionistic Fuzzy Sets. Springer, Heidelberg (1999)
3. Atanassov, K.: On Intuitionistic Fuzzy Sets Theory. Springer, Berlin (2012)
4. Atanassov, K.: Intuitionistic fuzzy logics as tools for evaluation of data mining processes.
Knowl. Based Syst. 80, 122–130 (2015)
5. Atanassov, K., Szmidt, E., Kacprzyk, J.: On intuitionistic fuzzy pairs. Notes Intuit. Fuzzy
Sets 19(3), 1–13 (2013)
6. Atanassov, K., Vassilev, V., Andonov, V., Sotirova, E.: A generalized net model of the
abdominal aorta and its branches as a part of the vascular system. In: IWIFSGN’2019,
Warsaw, Poland, 24–25 October (2019)
7. Beard, D.: Chronic lower limb ischemia. West. J. Med. 173(1), 60 (2000)
8. Standring, S. (ed.): Gray’s Anatomy E-Book: The Anatomical Basis of Clinical Practice.
Elsevier Health Sciences, Amsterdam (2015)
9. Hans, S., Jareunpoon, O., Balasubramaniam, M., Zelenock, G.: Size and location of
thrombus in intact and ruptured abdominal aortic aneurysms. J. Vasc. Surg. 41(4), 584–588
(2005)
10. Kazi, M., Thyberg, J., Religa, P., Roy, J., Eriksson, P., Hedin, U., Swedenborg, J.: Influence
of intraluminal thrombus on structural and cellular composition of abdominal aortic
aneurysm wall. J. Vasc. Surg. 38(6), 1283–1292 (2003)
11. Noble, A., Johnson, R., Thomas, A., Bass, P.: Systems of the Body Series: The
Cardiovascular System. Elsevier eBook, Churchill Livingstone (2010). ISBN
9780702050824
12. Torsney, E., Mandal, K., Halliday, A., Jahangiri, M., Xu, Q.: Characterisation of progenitor
cells in human atherosclerotic vessels. Atherosclerosis 191(2), 259–264 (2007)
13. Tortora, G., Derrickson, B.: Essentials of Anatomy and Physiology. Wiley, Hoboken (2013)
14. Wright, L., Matchett, W., Cruz, C., James, C., Culp, W.: Popliteal artery disease: diagnosis
and treatment. Radiographics 24(2), 467–479 (2004)
15. Zadeh, L.: Fuzzy sets. Inf. Control 8(3), 338–353 (1965)
Mathematical Philosophy and Fuzzy Logic
Ahmet Fahri Özok(&)
Industrial Engineering Department, Istanbul Okan University Tuzla Campus,

Akfırat, 34959 Istanbul, Turkey
[email protected]
Abstract. Mathematics in general is an abstraction. Of course, if it is not an easy

subject, but in the same time it is rewarding one. We have to keep in mind that
Mathematics is a part of our culture. In developed, developing and underdevel-
oped countries nobody can count himself truly good educated without Mathe-
matical Concepts. Perhaps one can say that all individuals have to understand in
some degree what Mathematics is and what it does. Unfortunately it has always
been a great problem in Mathematics, despite the publication of hundreds of
Mathematical books, to understand the essential meaning of it. Above all it is a
human subject with its own concepts, axioms, postulates etc. [1, 2, 4].
1 Mathematical Philosophy and Mathematics in General
Mathematics in general is an abstraction. Of course, if it is not an easy subject, but in

the same time it is rewarding one. We have to keep in mind that Mathematics is a part
of our culture. In developed, developing and underdeveloped countries nobody can
count himself truly good educated without Mathematical Concepts. Perhaps one can
say that all individuals have to understand in some degree what Mathematics is and
what it does. Unfortunately it has always been a great problem in Mathematics, despite
the publication of hundreds of Mathematical books, to understand the essential
meaning of it. Above all it is a human subject with its own concepts, axioms, postulates
etc. [1, 2, 4].
Mathematics has many important practical applications and it is not important only
for scientist and engineers but also for everyone. Because in mathematical thought one
can find many interesting points which should be interesting for everybody.
We know that in former days mankind has always been fascinated by the ordinary
integers, the natural, numbers (excluding 0) 1; 2; . . .; 9. In Mathematical History at a
very early age human being noticed that he possesses 5 fingers on each hand and 5 toes
on each foot. After some years he believes that he himself is unique individual. He
understands that he represents the number 1. I think it is very important in the meaning
of number concept. The number 2 has biological significance and mystic significance
develops around the number 3, etc.
The beginnings of Mathematical Logic are of philosophical interest. Some basic
Mathematical Concepts can be included in Mathematical Philosophy if we don’t have
them satisfactory in other sciences. Two examples can be given; namely the nature of
infinity and continuity. They were belonging philosophy but belong now to
Mathematical Philosophy and Fuzzy Logic 33
Mathematics. As a common language in science Mathematics makes it possible to

express clearly all of important concepts [10].
Mathematical philosophy, because of its content, can not include definite scientific
results. it deals with some important questions on the frontier of science. In the region
of unsolved problems in Mathematics, generally Mathematical Logic is relevant to
philosophy. (In this point one claim that in the future Fuzzy Logic could be helpful
from the point of view of further research).
Constructive and abstract content of Mathematics helps us to understand integers,
fractions, real numbers, complex numbers, differentiation, integration and finally
abstraction [8, 9, 11].
The following subjects are some of essentials to understand Mathematical Phi-
losophy and its background: Symbolic Logic, Propositional Functions, Relations, The
Meaning of Magnitude, The Range of Quantity, Limits and Irrational Numbers Ordinal
Continuity, Dimensions and Complex Numbers, Projective and Descriptive Geometry,
Continuity of Space and Causality. Mathematics in general can be assumed as one of
the oldest test and richest of the disciplines. Unfortunately, it is not commonly
understood that Mathematics is also one of the most rapidly expanding of the disci-
plines among others [5–7].
We can assume that everybody possesses some degree of Mathematical intuition.
Logic is a tool to solve problems. The power of Mathematics is a combination of
intuition and logic. The Art of problem solving, especially in Pure Mathematics
resembles in some sense the fine arts. An intuitive proof in Mathematics, allows us to
understand why the theorem must be true, on the other side logic provides us basic
ground to show that it is true. We have to understand the essential meaning of concepts
and a genuine desire to increase our knowledge steadily.
The abstractions in Mathematics leads us to some consequences: Group Theory has
some applications to rigid motion in spaces, symmetries in geometry, the deformation
of curves in a topological space, geometry in general motivates a proof and makes it
possible to understand the qualitative aspects of a theorem.
Mathematics is not a manipulation of numbers and symbols. It is a comprehensive
and precise body of abstraction but not in any case a static body of knowledge. The
basic elements of it are logic and intuition.
2 Probability and Fuzzy Logic in Industrial Engineering
Probability Concept is somehow different from Definitive Mathematical Concepts.

A thorough grounding in Probability Theory and related topics, makes it possible to
have a better idea and understanding of Statistical Inference. To use probability as a
Mathematical tool it is essential to have some formalism of Probability Theory. To be
able to make a true interpretation from data analysis it is important to know that
probability increases the strength of our confidence.
We cannot have a sound idea if we don’t understand the philosophy of Probability
Theory. Fundamental mathematical concepts of Probability form a major component to
have a better conclusion in statistical inference methods. In fact, the discipline of
34 A. F. Özok
probabilistic methods makes it possible to establish a bridge between descriptive

statistics and statistical inference.
Generally, in Fuzzy Logic, differently from Probability Theory each problem can
be solved differently according to main parameters. In Industrial engineering we have
some problems which cannot be solved using Definitive or Probabilistic Methods. In
many cases it is proved that the importance of Fuzzy Logic is far ahead if we compare
it with other methods. In different kind of problems in science and engineering Fuzzy
Logic seems to be very flexible to solve them. If we have a sound knowledge in Fuzzy
Logic it gives us the opportunity to have marvelous discoveries. Fuzzy Logic enriched
the common language among different sciences. It is possible imagination and inge-
nuity to the utmost in different kind of Engineering Problems. As an Engineering Area
Industrial Engineering really has a plenty of diversity: From Man-Machine System to
Production Planning, from Logistics to Man Power Planning [3].
3 Conclusion
We have to be capable the smooth passing from Definitive to from Probabilistic to

Fuzzy Mathematical Modelling.
If we don’t understand the deep relation among these three Mathematical areas, we
can’t be actively engaged in the method of science. Fuzzy logic is not a subject throw
out all of the traditional ideas and replaced them by weird and strange creations of no
possible use in Engineering and Science.
In History of Mathematics, new ideas have developed naturally out of older ones
and have been incorporated steadily with the passing of years. In Engineering Prob-
lems, our aim has to be directed toward to understand the problem and main parameters
of it. Instead of blind manipulation, we have to decide; which Mathematical Model
yields the best result.
References
1. Exner, R.M., Rosskopf, M.F.: Logic in Elementary Mathematics. Dover Publications, Inc.,
New York (2011)
2. Henle, J.M., Kleinberg, E.M.: Infinitesimal Calculus. Dover Publications, Inc., New York
(2003)
3. Kahraman, C. (ed.): Fuzzy Applications in Industrial Engineering. Springer, Heidelberg
(2006)
4. Kline, M.: Mathematics for the Nonmathematician. Dover Publications, Inc., New York
(1967)
5. Kitcher, P.: The Nature of Mathematical Knowledge. Oxford University Press, Oxford
(1984)
6. Von Mises, R.: Probability, Statistics and Truth. Dover Publications, Inc., New York (1957)
7. Russell, B.: Introduction to Mathematical Philosophy. Dover Publications, Inc., New York
(1993)
8. Russell, B.: The Principles of Mathematics. Georg Allen & Unwin Ltd., New York (1964)
9. Stewart, I.: Concepts of Modern Mathematics. Penguin Books, London (1975)
Mathematical Philosophy and Fuzzy Logic 35
10. Todhunter, I.: A History of the Mathematical Theory of Probability from the Time of Pascal
to that of Laplace. Chelsea Publishing Company, New York (1949)
11. Whitehead, A.N.: An Introduction to Mathematics. Oxford University Press, New York
(1958)
Clustering
Segmentation of Retail Consumers with
Soft Clustering Approach
Onur Dogan1(B) , Abdulkadir Hiziroglu2 , and Omer Faruk Seymen3

1
Department of Industrial Engineering, Izmir Bakircay University,
35665 Izmir, Turkey
[email protected]
2
Department of Management Information Systems, Izmir Bakircay University,
35665 Izmir, Turkey
[email protected]
3
Department of Information Systems Engineering, Sakarya University,
54050 Sakarya, Turkey
[email protected]
Abstract. Defining customer requirements in a huge amount of data of

the digital era is crucial for companies in a competitive business envi-
ronment. Customer segmentation has been attracted to a great deal of
attention and has widely been performed in marketing studies. However,
boundary data which are close to more than one segment may be assigned
incorrect classes, which affects to make the right decisions and evalua-
tions. Therefore, segmentation analysis is still needed to develop efficient
models using advanced techniques such as soft computing methods. In
this study, an intuitionistic fuzzy clustering algorithm were applied to
customer data in a supermarket according to the amount spent in some
product groups. The data represent 33-month customer shopping data
in a supermarket for eight product groups. The results indicate the intu-
itionistic fuzzy c-means based customer segmentation approach produces
more reliable and applicable marketing campaigns than conditional fuzzy
c-means and k-means segmentation method.
Keywords: Customer segmentation · Fuzzy clustering · Intuitionistic

fuzzy c-means · Statistical methods · Marketing perspective
1 Introduction
Understanding customer demands is a crucial task for all companies in their mar-
keting strategies [7]. Many studies examine the spending habits of customers
concerning the transactional data with segmentation methods [1,6,20]. Smith
[19] first presented segmentation to marketing research. Then, segmentation was
introduced as an alternative idea in place of a commodity differentiation app-
roach [2]. The principal opinion of segmentation is to classify similar customers.
A segment or cluster can be explained as a set of customers who have similar
features of demography, spending habits, values, etc.
40 O. Dogan et al.
The choosing segmentation method is another essential task in understanding

customer demands because the improvements in data-oriented techniques have
modified the marketing tactics. The huge amount of data and the inadequate
quality of conventional segmentation tools based on statistical methods on over-
flowing data have prompt scientists and specialists to develop powerful segmen-
tation methods to create valuable information [10]. A broad diversity of customer
segmentation techniques such as statistical clustering [16,22,23], AID/CHAID
[15,18], discrimination analysis [21], inductive learning [14] and soft computing
[8,17] have been used in the previous studies. Many of these techniques use
various segmentation variables, such as demographic, socio-demographic and
behavioural including recency, frequency and monetary (RFM) to group cus-
tomers [10].
As a class of data-based methods, soft computing approaches have begun to
be used in the segmentation area. Moreover, soft computing methods are consid-
ered to be one of the potential areas that can develop the segmentation future
[9,13]. The notable practice of soft computing systems in business difficulties,
especially in segmentation, makes segmentation tasks more engaging, as these
techniques are applicable and efficient.
Fuzzy c-means (FCM) is one of the common soft computing methods of clus-
tering that relies on the fuzzy sets theory [3]. FCM algorithm takes into account
soft boundaries among the clusters, by enabling each data point in the dataset
to be clustered to belong to various groups by a particular degree. The member-
ship degree of a data point to a cluster is commonly calculated according to the
distance between the data point and the cluster center. Various modifications of
the FCM have been introduced through the year, such as modifications applied
to image segmentation [4,24], computer vision [11], customer path segmentation
[8] or generalizations of different fields [12]. In this study, FCM is adapted to
create another extension, intuitionistic fuzzy c-means (IFCM). The main differ-
ence of IFCM is to consider not only membership and non-membership values
but also hesitancy. Moreover, this study includes the entropy due to fuzziness.
The rest of the study is designed as follow. Section 2 presents the main
methodology used in the study. Section 3 gives clustering results and compares
the solutions. Section 4 shows evaluation of segmentation results considering
some demographic variables that are not included in the clustering steps. Finally,
Sect. 5 discuss the conclusions and limitations of the study.
2 Methodology
In this study, the intuitionistic fuzzy c-means clustering (IFCM) algorithm pro-
posed by Chaira [5] were adapted. To convert conditional FCM to IFCM, the
cluster centers are updated by considering hesitance values like in Eq. 1.
μ∗ki = μki + πki (1)
μ∗ki and μki are the intuitionistic and conventional fuzzy membership of the data
i in cluster k, respectively.
Segmentation of Retail Consumers with Soft Clustering Approach 41
Then, new cluster centers are computed as:

N ∗
∗ μki xi
vk = i=1 N
∀k, k = 1, 2, · · · , c (2)
∗
i=1 μki
Using Eq. 1 and Eq. 2, the membership values and cluster centers are updated
in each iteration of the fuzzy c-means algorithm that aims to minimize Eq. 3.
The algorithm stops when the difference between the updated cluster centers
and previous cluster centers is less than , which is a predetermined parameter.
c
N
J1 = (μ∗ki )m d2 (xi , vk ) (3)
k=1 i=1
The intuitionistic fuzzy entropy (IFE) as the second dimension of the objective
function also was considered. IFE aims to maximize good data in the cluster.
The second dimension can be written as [5]:
c
∗
J2 = πk∗ e1−πk (4)
k=1
N
where πk∗ = N1 i=1 πki , k ∈ [1, N ]. πki shows the hesitation value of data i in
cluster k. As a result, the final objective function of the algorithm contains two
terms that should be minimized:
c
N c
∗
J= (μ∗ki )m d2 (xi , vk ) + πk∗ e1−πk (5)
k=1 i=1 k=1
Equation 5 is a general form for a clustering algorithm. For k-means clustering,

μ∗ki is zero or one. The second part (J2 ) is zero for both k-means and FCM
clustering because the hesitancy is ignored. IFCM advances the clustering per-
formance by considering hesitancy and entropy due to fuzziness. This study also
validates that the IFCM results surpass both k-means and FCM results.
3 Results
The research data were obtained from one of the major supermarkets in Turkey.
The transaction data include the spending amount of 4674 customers in eight
main product categories, defined by the supermarket, for 33 months. Some addi-
tional data such as age, gender, income, internet, and households were included
in the transaction data.
One of the challenging tasks in a clustering study is to define the number
of clusters. Figure 1 presents some data pre-processing steps. In this study, a
silhouette measure of cohesion and separation was calculated to find the opti-
mum number of clusters in k-means clustering. Although the number of clusters
was defined 3, three out of nine variables (cerez, konserve and promosyon) were
42 O. Dogan et al.
Fig. 1. Data pre-processing steps (baharat: spice, cerez: snack, yaglar: oils, kahvalti:
breakfast, konserve: canned goods, sarkuteri: delicatessen, sos: sauce, temizlik: cleaning,
promosyon: promotion)
detected statistically non-significant variables by considering eta-squared values,

which is less than 0.3. The remaining variables (baharat, yaglar, kahvalti, sarku-
teri, sos and temizlik) were re-tested by applying both again two-step clustering
and eta-squared to verify whether there is another statistically non-significant
variables for the defined number of clusters. Finally, the optimum number of clus-
ters was defined as two for the remaining six variables. Similar pre-processing
steps were executed for fuzzy clustering but the Xie-Beni index defines the num-
ber of clusters. Besides, translations to English were given after the Fig. 1 caption
since data were collected in Turkish.
Fig. 2. Clustering results
K-means and FCM were applied to cluster customer data into 2 groups. IFCM
clustering was converted from conditional FCM, which was proposed by Chaira
[5]. Figure 2 gives the clustering results by comparing the three algorithms. IFCM
produced better solutions in Fig. 2a by reaching minimum objective function val-
ues. The data split is higher in the k-means algorithm. Although FCM divides
data almost equally, a readjustment considering fuzzy entropy changes the num-
ber of customer data in the clusters.
For validation of the clustering results, one way ANOVA testing was applied
for six variables (spice, oils, breakfast, delicatessen, sauce, and cleaning) by defin-
ing clusters produced by IFCM as a factor variable. Analysis of variance showed
that means for each variable are not equal. Additionally, data should be dis-
tributed non-normally to be able to group data into the distinguishable clus-
ters. The test of normality was rejected for all variables considering skewness
and kurtosis. These results indicated IFCM clustering can be used for customer
segmentation.
4 Managerial Implications
Since considered IFCM shares hesitancy to membership degrees, new and more
reasonable solutions can be revealed. Moreover, IFCM takes into account fuzzy
entropy due to uncertainty conditions. Therefore, IFCM can assign data points
into the different clusters. Table 1 presents a sample of the different clustering
solutions in order to show that IFCM varies from other methods. The clustering
results were given in Table 1 with segment number and membership value of the
related segment in parenthesis. Since the created clusters represent customer
segments, offering effective marketing promotions highly depends on clustering
results. Although customer 1101261 and 1102640 were assigned in segment 1
with a membership degree of 0.5 approximately in FCM and k-means methods,
IFCM produced better results with assignment them into segment 2. It means
IFCM enables to offer better marketing campaigns for the customers.
Table 1. Differences of IFCM clustering
Customer Spice Oils Breakfast Delicatessen Sauce Cleaning IFCM FCM KM

1101261 2,62 2,55 6,34 4,43 1,77 8,77 2 (0,6237) 1 (0,5101) 1 (1,0000)
1102640 5,09 3,70 6,25 4,17 2,66 5,47 2 (0,6417) 1 (0,529) 1 (1,0000)
1157511 2,05 3,30 10,02 0,00 0,00 10,10 1 (0,6162) 2 (0,5023) 1 (1,0000)
Although six demographic variables (age, gender, income, internet, house-

hold, and house type) were selected to characterize customers based on defined
clusters, three of them (age, gender, income) were found statistically significant.
Figure 3 presents the distinguishing features of the demographic variables. When
the average spending of customers by age is analyzed, it is seen that it does not
change much in the same segments, but there is a notable difference between
the two clusters. Although the gender variable is not very different, the rate of
genders in the same segment can be interpreted. In C1 and C2 clusters, the ratio
of males and females are 0.88 and 0.76, respectively. Income distribution has 15
categories defined by the company. The income level increases from category 0
to category 14. It can be easily seen that while the income level increased, the
average spending amount resemble a bell-shaped curve, which firstly increases
then decreases. However, the average spending amount of C1 is higher than C2.
44 O. Dogan et al.
Fig. 3. Statistically significant demographic variables
5 Conclusion and Limitations

The study focuses on the segmentation of supermarket customers considering
the average spending amount on eight product groups and participant rates to
promotions in 33 months. To do that, first, the number of clusters was defined
according to the silhouette of cohesion and separation for k-means and Xie-
Beni index for fuzzy c-means. Then, statistically non-significant variables were
defined by eta-squared value. For the remaining six variables, three methods,
k-means, FCM, and IFCM by converting from FCM, were applied. Since IFCM
gave the minimum objective function value, six demographic variables are used
to distinguish defined clusters. Only three of them, age, gender, and income,
were statistically significant. Some managerial implications were presented for
these variables to distinguish customer clusters. This study contributes to the
literature by giving managerial insights and presenting the impact of some demo-
graphic information.
Although customer value is essential for marketing decisions, clustering meth-
ods do not provide the significance of customer segments. Therefore, some other
techniques such as multi-criteria decision-making methods can support the seg-
mentation results in the future direction.
Three demographic variables, which are internet, household, and house type,
were statistically insignificant. However, these variables could be used for several
purposes. Internet usage is important to investigate online shopping behavior.
The household may affect the average spending amount. Similarly, the type
of house where customers live may be an interesting demographic variable to
investigate customers’ spending behavior.
References
1. Alizadeh Zoeram, A., Karimi Mazidi, A.R.: New approach for customer clustering
by integrating the LRFM model and fuzzy inference system. Iran. J. Manag. Stud.
11(2), 351–378 (2018)
2. Beane, T., Ennis, D.: Market segmentation: a review. Eur. J. Mark. 21(5), 20–42
(1987)
3. Bezdek, J.C., Ehrlich, R., Full, W.: FCM: the fuzzy c-means clustering algorithm.
Comput. Geosci. 10(2–3), 191–203 (1984)
4. Cai, W., Chen, S., Zhang, D.: Fast and robust fuzzy c-means clustering algorithms
incorporating local information for image segmentation. Pattern Recogn. 40(3),
825–838 (2007)
5. Chaira, T.: A novel intuitionistic fuzzy c means clustering algorithm and its appli-
cation to medical images. Appl. Soft Comput. 11(2), 1711–1717 (2011)
6. Chiang, L.L.L., Yang, C.S.: Does country-of-origin brand personality generate
retail customer lifetime value? A Big Data analytics approach. Technol. Forecast.
Soc. Change 130, 177–187 (2018)
7. Dogan, O., Bayo-Monton, J.L., Fernandez-Llatas, C., Oztaysi, B.: Analyzing of
gender behaviors from paths using process mining: a shopping mall application.
Sensors 19(3), 557 (2019)
8. Dogan, O., Oztaysi, B., Fernandez-Llatas, C.: Segmentation of indoor customer
paths using intuitionistic fuzzy clustering: process mining visualization. J. Intell.
Fuzzy Syst. 38(1), 675–684 (2020)
9. Hiziroglu, A.: A neuro-fuzzy two-stage clustering approach to customer segmenta-
tion. J. Mark. Anal. 1(4), 202–221 (2013)
10. Hiziroglu, A.: Soft computing applications in customer segmentation: state-of-art
review and critique. Expert Syst. Appl. 40(16), 6491–6507 (2013)
11. Iakovidis, D.K., Pelekis, N., Kotsifakos, E., Kopanakis, I.: Intuitionistic fuzzy
clustering with applications in computer vision. In: International Conference on
Advanced Concepts for Intelligent Vision Systems, pp. 764–774. Springer, Heidel-
berg (2008)
12. Karayiannis, N.B.: Generalized fuzzy c-means algorithms. J. Intell. Fuzzy Syst.
8(1), 63–81 (2000)
13. Kordon, A.K.: Future trends in soft computing industrial applications. In: 2006
IEEE International Conference on Fuzzy Systems, pp. 1663–1670. IEEE (2006)
14. Leung, C.H.: An inductive learning approach to market segmentation based on
customer profile attributes. Asian J. Mark. 3(3), 65–81 (2009)
15. Levin, N., Zahavi, J.: Predictive modeling using segmentation. J. Interact. Mark.
15(2), 2–22 (2001)
16. Li, J., Wang, K., Xu, L.: Chameleon based on clustering feature tree and its appli-
cation in customer segmentation. Ann. Oper. Res. 168(1), 225–245 (2009)
17. Munusamy, S., Murugesan, P.: Modified dynamic fuzzy c-means clustering
algorithm-application in dynamic customer segmentation. Appl. Intell. 50(6),
1922–1942 (2020)
18. Sargeant, A., McKenzie, J.: The lifetime value of donors: gaining insight through
chaid. Fund Raising Manag. 30(1), 22–27 (1999)
19. Smith, W.R.: Product differentiation and market segmentation as alternative mar-
keting strategies. J. Mark. 21(1), 3–8 (1956)
20. Tarokh, M.J., EsmaeiliGookeh, M.: Modeling patient’s value using a stochastic
approach: an empirical study in the medical industry. Comput. Meth. Prog. Bio.
176, 51–59 (2019)
46 O. Dogan et al.
21. Tsiotsou, R.: Using visit frequency to segment ski resorts customers. J. Vacat.
Mark. 12(1), 15–26 (2006)
22. Wang, C.H.: Outlier identification and market segmentation using kernel-based
clustering techniques. Expert Syst. Appl. 36(2), 3744–3750 (2009)
23. Xia, J.C., Evans, F.H., Spilsbury, K., Ciesielski, V., Arrowsmith, C., Wright, G.:
Market segments based on the dominant movement patterns of tourists. Tour.
Manag. 31(4), 464–469 (2010)
24. Yong, Y., Chongxun, Z., Pan, L.: A novel fuzzy c-means clustering algorithm for
image thresholding. Measur. Sci. Rev. 4(1), 11–19 (2004)
Segmentation Analysis of Companies’
Natural Gas Consumption by Soft
Clustering
Onur Dogan(B)
Department of Industrial Engineering, İzmir Bakırçay University,

35665 Izmir, Turkey
[email protected]
Abstract. Although renewable energy sources have more significance

than traditional sources, limited energy sources such as natural gas have
been consumed in daily life. Natural gas energy is an environmental-
friendly and easy-to-use energy source. Furthermore, it is the cleanest
burning of all fossil fuels. Since the amount of gas consumed by companies
is much more, the companies’ consumption amount was investigated. The
companies concerning the consumption amount were clustered by fuzzy
c-means algorithm in order to enhance the understandability thanks to
the collection of 1049 companies’ data. This study, at the same time,
presents the effects of weather conditions on the consumption amount
in Istanbul. Hence, a calendar view visualization is created to show the
relationship between consumption and weather conditions.
Keywords: Company consumption behavior · Natural gas ·

Segmentation · Fuzzy c-means · Weather temperature · Smart meters
1 Introduction
The energy field manages challenges including energy sources diversity, the var-
ious forms of energy productivity explications and environmental responsibili-
ties [8]. Natural gas demand will rise 30% by 2030 [12]. A growing concern in
the declining consumption of all kinds of energy has occurred in the economy
[18]. Smart grids are vital for energy savings efforts since they facilitate users
to implement environment-friendly choices on their energy consumption. The
notable improvements about combining smart meters to the natural gas system
are the opportunity to provide not static tariffs for consumers and intensified
load profiling [19].
The amount of natural gas usage varies in domestic and industrial zones by
reducing gas charges resulting from overflowing energy resources or increasing
investments and long term contracts [2]. In 2015, natural gas was enough for
just 25% of global energy needs. Natural gas was the second common energy
source in 2016. Technological advances were constantly expanding the ability
48 O. Dogan
of its extraction, carrier and storage methods [1]. Moreover, the natural gas
requirement will rise by 30% in 2030. Households prefer natural gas on heating
and hot water systems due to ease of use [8].
Data-based methodologies have a significant point in numerous smart grid
cases. Nevertheless, previous researches have discussed the electrical energy field
[7]. Data-oriented analysis of natural gas consumption has still a discussion gap
about soft clustering and company-oriented profiling. The data potentially pro-
vide opinions for an unusual consequence for monopolies and policy-makers.
Significant deductions can be obtained from the information on common usage
amounts of various organizations. Information about company gas consumption
is remarkable in examining the impact of weather temperature on various compa-
nies. The research outcomes can support authorities in the gas industry to make
demand-side programs, user-oriented strategies, demand forecasting in terms of
predicted weather conditions.
Data can include a relationship that cannot be obvious in real-world problems
[5]. Clustering is a data-based technique that intends to discover a relationship
by separating clusters considering similarity [4]. The aim is to divide the data
points in such a way that objects are assigned to the same cluster as similar as
possible. Numerous researches studied the clustering of energy users. K-means is
frequently applied in the literature [3,9,10,16,17]. Although studies have widely
concentrated on electricity energy [13,20], natural gas is also another important
study topic thanks to an environment-friendly energy type. Previous studies
focused on residential natural gas consumption [8,21]. However company-based
consumption is more important than residential consumption because the con-
sumption amount is very higher than the residential and rising energy costs [15].
Besides, the weather temperature has more influences on companies than the
residential.
This study aims to produce advantages from the data about company clus-
tering and examine the usage amounts considering weather temperatures. Since
1049 company data are included in this study, many data points are close to
each other, which creates more boundary data. Hence, a fuzzy c-means cluster-
ing algorithm was chosen. The remainder of this paper is organized as follows.
Firstly, the fuzzy c-means clustering is presented. Then, the application of the
algorithm is given. Finally, the discussion and conclusion section is shown.
2 Fuzzy C-Means Clustering
Fuzzy c-means (FCM) is one of the most common clustering algorithms in fuzzi-
ness [14]. It is a soft version of K-means clustering. It pursues the same actions
by taking into account one data can refer to more than one cluster. In many
real-life examples, some data objects are nearly equal distances to more than
one cluster, which is called boundary data point. Fuzzy clustering approaches
allow progressing the probability of the correct classification by assigning the
boundary data objects into the right cluster [6]. Conventional FCM tries to
minimize Eq. (1).
Segmentation Analysis of Companies’ Natural Gas Consumption 49
c
N
J(U, V : X) = (µki )m d2 (vk , xi ) (1)
k=1 i=1
where U is the fuzzy membership matrix; V is the cluster centers, and X is

the data set to be clustered. µki is the membership value of data xi fuzzy cluster
ck . The fuzzifier m must be larger than 1. If m equals to 1, then the clusters are
formed in crisp format. An appropriate U (membership matrix) and V must be
computed like Eq. (2) and Eq. (3) to minimize Jm , for all k = 1, 2, · · · , c.
⎛ ⎞−1
c

µki = ⎝ [(d2ki )/(d2ji )]1/(m−1) ⎠ , (2)
j=1
N m
µki xi
vk = i=1
n m
(3)
i=1 µki
Equation (2) and Eq. (3) are used until the termination criterion is satisfied.
3 Case Study
The Istanbul Natural Gas Network is kept under control continuously for 7
days and 24 h with SCADA (Central Control and Data Acquisition System)
and fully equipped smart devices. First, the collected data from SCADA is used
to compute XBI to determine the optimum number of clusters. The optimum
number of clusters is found four. Table 1 shows gas usage amounts for some
companies on 25th January.
Company names are provided with an ID thanks to privacy issues. The usage
amount of the companies is given in m3 column. As mentioned before, fuzzy
clustering methods have the flexibility for assignment of boundary data points.
In this example, company MRBP207 with 3924.36 m3 consumption is assigned
into cluster 1 (C1) in the K-Means algorithm, whereas it is in cluster 2 (C2) in
FCM. The membership values in FCM are computed using Eq. (2).
Figure 1 shows the temperatures of Istanbul for the analysis period that we
obtain data from 25th January to 16th February. Red and blue lines show the
maximum and minimum temperatures, respectively. While the maximum tem-
perature level varies from 8 ◦ C to 20 ◦ C, the minimum temperature level changes
from 1 ◦ C to 10 ◦ C. The dotted lines present the percentage of the consumption
by companies in the clusters. We evaluate natural gas consumption within clus-
ters. Therefore, the percentages are calculated for the analysis period for four
clusters with respect to the average amount of consumed natural gas in the
cluster.
Although the amount of consumption in C1, C2, and C3 is changed over
the daily temperature, companies in C4 are not directly affected. When the
maximum temperature is the lowest level, 8 ◦ C on 9th February, 789 companies
consumed 925 m3 natural gas on average in C1, which is shown in Table 2. When
the temperature increases 12 ◦ C, from 8 ◦ C to 20 ◦ C, the decrease in C1 is 3.8%.
50 O. Dogan
Table 1. Comparison of two clustering methods
25th January K-Means Fuzzy C-Means

ID m3 C1 C2 C3 C4 C1 C2 C3 C4
MRIT131 930 1 0 0 0 0.981 0.0133 0.0038 0.0017
MRIT348 3084 1 0 0 0 0.826 0.1071 0.0589 0.0082
MRIS079 3139 1 0 0 0 0.778 0.1444 0.0685 0.0091
MRIT029 3152 1 0 0 0 0.764 0.1460 0.0714 0.0186
MRBP207 3924 0 1 0 0 0.4814 0.4465 0.0536 0.0184
MRBP016 6232 0 1 0 0 0.1796 0.7700 0.0399 0.0105
MRBS143 9586 0 1 0 0 0.0709 0.8605 0.0423 0.0262
MRBS068 9820 0 1 0 0 0.0609 0.8105 0.0927 0.0359
MRIS278 12111 0 0 1 0 0.0309 0.1005 0.8523 0.0162
MRIS053 12287 0 0 1 0 0.0294 0.0860 0.8663 0.0183
MRIP037 13491 0 0 1 0 0.0214 0.0654 0.8924 0.0208
MRBS040 16759 0 0 1 0 0.0134 0.0547 0.7924 0.1395
MRIS284 24060 0 0 0 1 0.0001 0.0045 0.0944 0.9009
MRIT014 29535 0 0 0 1 0.0001 0.0037 0.0853 0.9109
MRIS240 30463 0 0 0 1 0.0001 0.0025 0.0064 0.9910
MRBT131 35349 0 0 0 1 0.0000 0.0043 0.0021 0.9936
Fig. 1. Temperature (◦ C) and Changes in Gas Consumption (%) in Istanbul

C2 and C3 are more vulnerable than others against to the temperature changes
concerning natural gas consumption. As expected, when the temperature is the
highest level, the consumption amount reaches the minimum amount in C3.
However, in other clusters, the minimum consumption is not exactly the same,
but very close, with the amount consumed at the highest temperature. This
situation explains that low changes in the temperature do not directly provide
a fluctuation. For example, companies in C2 consumed the minimum amount of
gas on 3rd February that is before one day of the maximum temperature.
Table 2. Consumption amounts at different levels (m3 )
C1 C2 C3 C4
◦
8 C (9th Feb.) 925 5846 18387 36385
20 ◦ C (4th Feb.) 890 3921 11204 32434
Change 3.80% 32.90% 39.10% 10.90%
Average 1096 5709 16479 35042
Minimum 525 3917 11204 31466
Maximum 1498 7039 20500 38608
The average consumption amounts near the amount consumed at the mini-
mum temperature level. At the lowest temperature level, whereas C1 companies
consumed lower gas than the average amount, other companies have a consump-
tion amount above average. This can be an indicator that companies that have
a consumption lower than 1500 m3 take care to consume a lower amount of gas
when the temperature level decreases.
4 Conclusions and Future Directions
In this study, we evaluated companies’ natural gas consumption in Istanbul. We

extracted 23-day gas consumption data from SCADA (Supervisory Control and
Data Acquisition) system which is called the gas control center in Istanbul Gas
Distribution Corporation (IGDAS). We used a k-means clustering method to
group companies concerning the amount of gas consumption. Four groups are
created for each day by calculating the Xie-Beni Index (XBI). We compared
the changes in the consumption amount within each group over different actual
weather temperatures, which is taken from www.accuweather.com. We focused
on the minimum and maximum temperature level, 8 ◦ C to 20 ◦ C to illustrate an
example evaluation. When the temperature is the highest level, the consump-
tion amount reaches the minimum amount in a general manner. However, low
changes such as 1 ◦ C increase, in the temperature do not directly affect the
gas consumption amount. For example, C2 companies consumed the minimum
amount of gas before one day (3rd February) of the maximum temperature (4th
52 O. Dogan
February). Although the average consumption amounts near to the amount con-
sumed at the minimum temperature level, C1 has below average, and others
have above average consumption amounts. Companies in C2 and C3 are more
affected by changes in the temperature. It is interesting that C4 companies are
not directly affected by temperature changes.
The study can be expanded by improving the clustering algorithm. For exam-
ple, other kind of fuzzy extensions such as hesitant fuzzy c-means, intuitionistic
fuzzy c-means and spherical fuzzy clustering methods can yield more sensitive
results. The consumption behaviors of the clustered companies can be inspected
over time. For this purpose, a calendar view can be created. At the same time,
not only temperature but also other kinds of variables such as wind and humidity
that affect the natural gas consumption can be considered. Although clustering
gives meaningful results for the evaluation of companies gas consumption, it
has a drawback because it is sensitive to the selected clustering algorithms and
variables. Hence, the results can lead to misunderstanding of clusters in decision
making [11].
References
1. Abid, M., Alimi, M.: Stochastic convergence in US disaggregated gas consumption
at the sector level. J. Nat. Gas Sci. Eng. 61, 357–368 (2019)
2. Baldacci, L., Golfarelli, M., Lombardi, D., Sami, F.: Natural gas consumption
forecasting for anomaly detection. Expert Syst. Appl. 62, 190–201 (2016)
3. Benı́tez, I., Quijano, A., Dı́ez, J.L., Delgado, I.: Dynamic clustering segmentation
applied to load profiles of energy consumption from Spanish customers. Int. J.
Elec. Power Energy Syst. 55, 437–448 (2014)
4. De Oliveira, J.V., Pedrycz, W.: Advances in Fuzzy Clustering and its Applications.
Wiley, Hoboken (2007)
5. Dogan, O.: Heuristic approaches in clustering problems. In: Handbook of Research
on Applied Optimization Methodologies in Manufacturing Systems, pp. 107–124.
IGI Global (2018)
6. Dogan, O., Oztaysi, B.: Genders prediction from indoor customer paths by
Levenshtein-based fuzzy kNN. Expert Syst. Appl. 136, 42–47 (2019)
7. Fagiani, M., Squartini, S., Gabrielli, L., Pizzichini, M., Spinsante, S.: Computa-
tional intelligence in smart water and gas grids: an up-to-date overview. In: 2014
International Joint Conference on Neural Networks (IJCNN), pp. 921–926. IEEE
(2014)
8. Fernandes, M.P., Viegas, J.L., Vieira, S.M., Sousa, J.M.: Analysis of residential
natural gas consumers using fuzzy c-means clustering. In: 2016 IEEE International
Conference on Fuzzy Systems (FUZZ-IEEE), pp. 1484–1491. IEEE (2016)
9. Figueiredo, V., Rodrigues, F., Vale, Z., Gouveia, J.B.: An electric energy consumer
characterization framework based on data mining techniques. IEEE Trans. Power
Syst. 20(2), 596–602 (2005)
10. Hernández, L., Baladrón, C., Aguiar, J., Carro, B., Sánchez-Esguevillas, A.: Classi-
fication and clustering of electricity demand patterns in industrial parks. Energies
5(12), 5215–5228 (2012)
11. Hsu, D.: Comparison of integrated clustering methods for accurate and stable
prediction of building energy consumption data. Appl. Energy 160, 153–163 (2015)
12. International Energy Agency: World energy outlook special report 2015 (2015).
https://webstore.iea.org/world-energy-outlook-2015. Accessed 25 Feb 2019
13. Motlagh, O., Berry, A., O’Neil, L.: Clustering of residential electricity customers
using load time series. Appl. Energy 237, 11–24 (2019)
14. Oztaysi, B., Isik, M.: Supplier evaluation using fuzzy clustering. In: Supply Chain
Management Under Fuzziness, pp. 61–79. Springer, Heidelberg (2014)
15. Rackow, T., Javied, T., Donhauser, T., Martin, C., Schuderer, P., Franke, J.: Green
Cockpit: transparency on energy consumption in manufacturing companies. Pro-
cedia CIRP 26, 498–503 (2015)
16. Räsänen, T., Voukantsis, D., Niska, H., Karatzas, K., Kolehmainen, M.: Data-based
method for creating electricity use load profiles using large amount of customer-
specific hourly measured electricity use data. Appl. Energy 87(11), 3538–3545
(2010)
17. Rodrigues, F., Duarte, J., Figueiredo, V., Vale, Z., Cordeiro, M.: A comparative
analysis of clustering algorithms applied to load profiling. In: International Work-
shop on Machine Learning and Data Mining in Pattern Recognition, pp. 73–85.
Springer, Heidelberg (2003)
18. Swan, L.G., Ugursal, V.I.: Modeling of end-use energy consumption in the residen-
tial sector: a review of modeling techniques. Renew. Sustain. Energy Rev. 13(8),
1819–1835 (2009)
19. USmartConsumer: European smart metering landscape report 2016 (2016). http://
www.escansa.es/usmartconsumer/documentos/USmartConsumer European
Landscape Report 2016 web.pdf. Accessed 27 Feb 2017
20. Wen, L., Zhou, K., Yang, S.: A shape-based clustering method for pattern recog-
nition of residential electricity consumption. J. Cleaner Prod. 212, 475–488 (2019)
21. Zhu, S.B., Li, Z.L., Zhang, S.M., Liang, L.L., Zhang, H.F.: Natural gas pipeline
valve leakage rate estimation via factor and cluster analysis of acoustic emissions.
Measurement 125, 48–55 (2018)
Comparison of Fuzzy C-Means and K-Means
Clustering Performance: An Application
on Household Budget Survey Data
Songul Cinaroglu(&)
Department of Health Care Management, FEAS,

Hacettepe University, Ankara, Turkey
[email protected]
Abstract. National Household Budget Survey (HBS) data includes sociode-

mographic and financial indicators that are the elements of government public
policy actions. Finding the optimal grouping of households in a given, suffi-
ciently large data is a challenging task for policymakers. Soft classification
techniques such as Fuzzy C-means (FCM) provide a deep understanding of
hidden patterns in the variable set. This study aims to compare FCM and
k-means (KM) classification performance for the grouping of households in
terms of sociodemographic and out-of-pocket (OOP) health expenditure vari-
ables. Health expenditure variables have heavily skewed distributions and that
the shape of the variable distribution has a measurable effect on classifiers.
Incorporating Bayesian data generation procedures into the variable transfor-
mation process will increase the ability to deal with skewness and improve
model performance. However, there is a scarcity of knowledge about the
embedded strategy performance of the Bayesian data generation approach with
unsupervised learning with the application on health expenditures. This study
applied the aforementioned strategy to Turkish HBS data for the year 2015
while comparing FCM and KM classification performance. Normality test
results for the distribution of logarithmic (KS = 0.006; p > 0.05) and Box-Cox
transformed (KS = 0.006; p > 0.05) health expenditure variables, which were
generated using lognormal distributions from a Bayesian viewpoint, are next to
normal. Moreover, KM clustering (Sil = 0.48) results are better than FCM
(Sil = 0.4198) for classifying households. The optimal number of household
groups is 20. Further studies will compare the cluster-seeking performance of
other unsupervised learning algorithms while incorporating arbitrary health
expenditure variables into the study model.
Keywords: Fuzzy C-means K-means Classification Health expenditure

Household Budget Survey Bayesian data generation Unsupervised learning
1 Introduction
Choosing the right clustering technique and finding hidden patterns is a challenging task,
especially for high-dimensional data obtained from official statistics, including arbitrary
variables such as expenditures [1]. There exist several clustering techniques based on
Comparison of Fuzzy C-Means and K-Means Clustering Performance 55
dissimilarity or distance measures between objects and clusters [2]. Among these
methods, k-means (KM) is a type of hard clustering technique [3] that finds “k” clusters
from the given dataset by iteratively computing the distance from each point to “k” cluster
centers [4]. On the contrary, Fuzzy C-means (FCM) is a type of soft clustering technique
that is able to deal more effectively with outliers and perform membership grading [3]. In
the existing literature comparing KM and FCM clustering performance, it has been stated
that KM outperforms FCM in terms of speed when using medical data [5] and various
datasets [6]. Additionally, it has been demonstrated that FCM has clear advantages over
the crisp and probabilistic clustering methods [3]. Moreover, FCM is a good alternative
classification technique for financial time series because the cluster centers express the
representative financial states of the companies, while the varying membership degrees
represent fluctuations in their states over time [7].
On the other side of the coin, it is critical to note that the shape of the variable
distribution has a measurable effect on classifiers that are considered to be “robust” and
that both parametric and non-parametric classifiers may be sensitive to these effects [8].
When the variable is skewed and/or heavily tailed, we can model the costs using
appropriate alternative distributions instead of assuming normality. Incorporating
Bayesian processes into the model will enhance the shape of the distribution [9].
Lognormal distribution is often appropriate, in the case of not normal distribution,
when the results might be non-robust to outliers in the data. To suggest a particular
form of distribution, then, analysis using this distribution should be recommended, but
sensitivity to alternative choices of distribution should be assessed [10]. The literature
also offers transformations to combat highly skewed health expenditure variables.
Using Box-Cox transformation and taking the natural logarithms of the variables [11]
are some of the advisable strategies for dealing with issues of skewed or not normal
distributed expenditure variables [12].
With the existence of the “big data” concept in official statistics, finding hidden
patterns and gathering information is valuable to better understand socioeconomic
behaviors and the grouping of decision-making units [1]. The application of soft
computing techniques such as FCM on official data has become more common than
ever before [13, 14]. Moreover, the incorporation of feature selection into clustering
procedures [15], variable transformations, and standardization before clustering are a
few advisable strategies [16]. However, there is a scarcity of knowledge about the
optimal clustering technique to find hidden patterns in Household Budget Survey
(HBS) data.
This study aims to compare KM and FCM classification performance for grouping
households in terms of their sociodemographic and health expenditure variables.
Before the comparison of clustering techniques, the Bayesian data generation proce-
dure was embedded into the unsupervised learning process for health expenditure
variables. Study algorithms and working principles of FCM and KM algorithms are
explained in Sect. 2. Study results presented in Sect. 3. Finally, Sect. 4 includes
conclusions and advices for future work.
56 S. Cinaroglu
2 Methods
Data representing the sociodemographic and health expenditure variables of house-

holds are derived from Turkish Statistical Institute (TurkStat) HBS data for the year
2015 [17]. The study variables included gender, age (being 65 years of age and over),
insurance, education, marital status, number of household members, and monthly out-
of-pocket (OOP) health expenditures. In this study, FCM and KM algorithms were
applied to examine the classification patterns of households, and the Silhouette
(Sil) index was used to compare classification performance. The working principles of
the FCM and KM algorithms are explained next.
Working principle of FCM algorithm as follows [6]:

1. Initialize U ¼ uij matrix; U ð0Þ

2. At k-step: calculate the centroid CðkÞ ¼ cj with U ðkÞ
PN
i¼1 ij xi
um
c j ¼ PN ð1Þ
i¼1 um
ij
3. Update U ðkÞ , U ðk þ 1Þ
1
uij ¼ 2 ð2Þ
Pc kxi cj k m1
k¼1 kxi ck k

4. If U ðk þ 1Þ U ðkÞ \ h then STOP; otherwise return to 2:
Here, “m” is the fuzziness parameter.
Working principle of KM algorithm as follows [6]:

For “M” sample vector fx1 ; x2 ; . . .; xM g falling into “k” compact clusters (k < M)
Let “mi ” be the mean of the vectors in the cluster “i”
If kx mi k is the minimum of all “k” distances
insert “x” into the respective cluster until there is no change in any “m”.
The classification performance of these two methods is compared using Sil index
values. The Sil index is measured based on the distance between all points in the same
group and separation is based on the distance to the nearest neighbor. The Sil cluster
validity index is calculated as follows [18]:
X X bðxi ; ck Þ aðxi ; ck Þ
SilðCÞ ¼ 1=N ð3Þ
ck 2CXi 2Ck maxfaðxi ; ck Þ; bðxi ; ck Þg
3 Results
3.1 Descriptive Statistics
Table 1 presents descriptive statistics for the categoric and continuous study variables
for the year 2015. It can be seen that 86.8% of households have males; 81.4% are under
65 years old; 96% have health insurance; and 87.6% of them graduated from primary,
secondary, and high school. In addition, 85.3% percent were married and 93.2% have
less than 7 household members. Additionally, the median value for the household’s
monthly OOP health expenditure is 22.70 TL.
Table 1. Descriptive statistics for study variables.
Categoric Categoric
N % N %
variables variables
Gender Education
Male 5903 86.8 Uneducated 842 12.4
Primary & secondary
Female 898 13.2 5959 87.6
& high
Age_65 Marital
Under 65 5537 81.4 Married 5804 85.3
65 and over 1264 18.6 Not married 997 14.7
Insurance Household Size
Yes 6532 96.0 Lower than 7 6339 93.2
No 269 4.0 Equal or higher than 7 462 6.8
Total 6801 100 Total 6801 100
Continuous
N Min. Max. Mean Median Std. Dev.
variable
OOP.h.exp* 6801 0.09 3746 52.93 22.70 139.96
*
: Turkish liras
Distribution analysis and histograms for OOP health expenditure variable before
and after transformations are presented, followingly.
3.2 Distribution Analysis

Figure 1 presents a histogram for the OOP health expenditure variable before trans-
formation. It can be seen that the distribution of the OOP health expenditure variable is
highly positively skewed, which is consistent with previous knowledge [9, 11]. Figs. 2
(a) and (b) show the distributions of the OOP health expenditure variable generated
using lognormal distributions from the Bayesian viewpoint. It is clear that the
embedded Bayesian data generation procedure using lognormal and Box-Cox trans-
formations is a useful strategy to deal with the heavily skewed distribution of the OOP
58 S. Cinaroglu
health expenditure variable. Normality tests are performed for the next step to
understand the normality of the distribution of the original OOP health expenditure
variable and the shape of the distribution after application of the Bayesian data gen-
eration procedure.
Frequency
Households monthly OOP health expenditure
Fig. 1. Histogram for OOP health expenditure variable before transformation.
(a) (b)
Frequency
Frequency
Logarithmic transformed version of OOP Box-Cox transformed version of OOP health

health expenditure by using lognormal expenditure by using lognormal distribution from
distribution from Bayesian viewpoint Bayesian viewpoint
Fig. 2. Histograms for logarithmic and Box-Cox transformed versions of heavily skewed OOP
health expenditure variables generated by using lognormal distribution from Bayesian viewpoint.
3.3 Normality Tests

Normality test results for the distribution of the original OOP health expenditure
variable verifies that the OOP health expenditure variable is heavily skewed
(K-S = 0.353, p < 0.001). In addition to that, logarithmic transformed (K-S = 0.043;
p < 0.001) and Box-Cox (K-S = 0.039, p < 0.001) transformed versions of the original
OOP health expenditure variable indicate not normal distributions, too. Interestingly,
the incorporation of Bayesian data generation with logarithmic and Box-Cox trans-
formation processes are able to effectively deal with the skewness problem and cause
the distribution to become normal (see Table 2).
Table 2. Normality test results.
Original OOP.h.exp* variable and transformed versions K-S** p

OOP.h.exp 0.353 <0.001
Logarithmic transformation of OOP.h.exp variable 0.043 <0.001
Box-Cox transformation of OOP.h.exp variable 0.039 <0.001
LOGNORMAL&_LOG_transformed.OOP.h.exp 0.006 0.2
LOGNORMAL§_Box_Cox_transformed.OOP.h.exp 0.006 0.2
*:Out-of-pocket health expenditure variable; **:Kolmogorov-Smirnov test; &:Logarithmic
transformed version of OOP health expenditure by using lognormal distribution from
Bayesian viewpoint; §: Box-Cox transformed version of OOP health expenditure by using
lognormal distribution from Bayesian viewpoint.
3.4 Classification Results

Classification results based on KM and FCM clustering performance are presented in
Figs. 3 and 4, respectively. Sil index values of KM clustering results obtained by
changing the number of clusters and generated by incorporating the original OOP
health expenditure and Bayesian data generation embedded transformed versions of
this variable into the dataset are presented in Figs. 3 and 4, respectively. It can be seen
that the clustering performance obtained from the dataset that includes the original
OOP health expenditure variable is better than the dataset that includes the Bayesian
embedded and transformed OOP health expenditure variable. KM and FCM classifi-
cation performances are comparatively visualized in further experiments that are per-
formed by incorporating the original OOP health expenditure variable into the dataset
(see Fig. 5).
60 S. Cinaroglu
k-means
1
Silhouette Index
0
5 10 15 20 25 30
OOP.h.exp Number of clusters
LOGNORMAL_LOG_transformed
LOGNORMAL_Box_Cox_transformed
Fig. 3. Silhouette index values of KM clustering results for three datasets includes original and
Bayesian data generation embedded transformed OOP health expenditure variables.
Figure 5 presents the Sil index values obtained from KM and FCM clustering
results for the dataset, including the original OOP health expenditure variable and other
study variables. KM clustering performance outperforms (Sil20 = 0.6633) FCM clus-
tering (Sil20 = 0.4198). In addition to that, better Sil index values are obtained when
the number of clusters determined is 20.
fuzzy c-means
1.0
Silhouette Index
0.0
5 10 15 20 25 30
OOP.h.exp Number of clusters
LOGNORMAL_LOG_transformed
LOGNORMAL_Box_Cox_transformed
Fig. 4. Silhouette index values of FCM clustering results for three datasets includes original and
Bayesian data generation embedded transformed OOP health expenditure variables.
0.8
Silhouette Index
0.6
0.4
0.2
kmeans fuzzy_cmeans
0
5 10 15 20 25 30
Number of clusters
Fig. 5. Silhouette index values of k-means and fuzzy c-means clustering performances by
incorporating original OOP health expenditure variable into the study variables.
4 Conclusions and Future Work
The results of this study demonstrate that KM clustering performs better than FCM for
classifying households in terms of sociodemographic and OOP health expenditure
variables. Moreover, the optimal number of household groups obtained from KM
clustering is 20. Additionally, the incorporation of Bayesian data generation with
logarithmic and Box-Cox transformation processes can effectively deal with the
skewness problem of the OOP health expenditure variable. However, study results
verify that embedding Bayesian data generation and transformation process for the
OOP health expenditure variable did not enhance its classification performance. In light
of these study results, it is evident that KM clustering better fits official HBS data,
including the arbitrary OOP health expenditure variable. Future studies should compare
other unsupervised learning algorithms to examine the grouping of hidden patterns in
official datasets.
References
1. Degirmenci, T., Ozbakır, L.: Differentiating households to analyze consumption patterns: a
data mining study on official household budget data. WIREs Data Mining Knowl. Discov.
8(1), 1–15 (2017)
2. Palarea-Albaladejo, J., Fernández, J.A.M., Soto, J.: Dealing with distances and transforma-
tions for fuzz c-means clustering of compositional data. J. Classif. 29(2), 144–169 (2012)
3. Sun, H., Wang, S., Jiang, Q.: FCM-based model selection algorithms for determining the
number of clusters. Pattern Recogn. 37(10), 2027–2037 (2004)
4. Kumar, K.M., Reddy, A.R.M.: An efficient k-means clustering filtering algorithm using
density based initial cluster centers. Inf. Sci. 418–419, 286–301 (2017)
5. Husein, A.M., Harahap, M., Aisyah, S., Purba, W., Muhazir, A.: The implementation of two
stages clustering (k-means clustering and adaptive neuro fuzzy inference system) for
prediction of medicine need based on medical data. J. Phys: Conf. Ser. 978(1), 012019
(2017)
6. Panda, S., Sahu, S., Jena, P., Chattopadhyay, S.: Comparing Fuzzy-C Means and K-Means
clustering techniques: a comprehensive study. In: Wyld, D., Zizka, J., Nagamalai, D. (eds.)
Advances in Computer Science, Engineering & Applications. Advances in Intelligent and
Soft Computing. Springer, Berlin (2012)
7. Sarlin, P., Eklund, T.: Fuzzy clustering of the self-organizing map: some applications on
financial time series. In: Laaksonen J., Honkela T. (eds.) Advances in Self-Organizing Maps.
WSOM 2011. Lecture Notes in Computer Science, vol. 6731. Springer, Berlin (2011)
8. Marko, N.F., Weil, R.J.: Non-Gaussian distributions affect identification of expression
patterns, functional annotation, and prospective classification in human cancer genomes.
PLoS ONE 7(10), 1–15 (2012)
9. Manning, W.G., Mullahy, J.: Estimating log models: to transform or not to transform?
J. Health Econ. 20(4), 461–494 (2001)
10. Nixon, R.M., Thompson, S.G.: Parametric modelling of cost data in medical studies. Stat.
Med. 23(8), 1311–1331 (2004)
11. Manning, W.: Dealing with skewed data on costs and expenditures. In: Jones, A.M. (ed.)
The Elgar Companion to Health Economics, 2nd edn. Edward Elgar Publishing (2012).
(chapter 44)
62 S. Cinaroglu
12. Box, G.E.P., Cox, D.R.: An analysis of transformations. J. Roy. Stat. Soc.: Ser.
B (Methodol.) 26(2), 211–243 (1964)
13. Yapıcı-Pehlivan, N., Gursoy, Z.: Determination of individuals’ life satisfaction levels living
in Turkey by FMCDM methods. Kybernetes 48(8), 1871–1893 (2019)
14. Gao, S., Zhou, C.: Differential privacy data publishing in the big data platform of precise
poverty alleviation. Soft Comput. 67 (2019). https://doi.org/10.1007/s00500-019-04352-1
15. Dash, M., Choi, K., Scheuermann, P., Liu, H.: Feature selection for clustering - a filter
solution. In: International Conference on Data Mining Proceedings, Maebashi City, Japan,
pp. 115–122. IEEE (2002)
16. Mohamad, I.B., Usman, D.: Standardization and its effects on k-means clustering algorithm.
Res. J. Appl. Sci. Eng. Technol. 6(17), 3299–3303 (2013)
17. Turkish Statistical Institute (TurkStat) Statistical Indicators. Household budget survey data
(2015). http://www.turkstat.gov.tr/UstMenu.do?metod=temelist. Accessed 01 Jan 2020
18. Rousseeuw, P.J.: Silhouettes: a graphical aid to the interpretation and validity of cluster
analysis. J. Comput. Appl. Math. 20, 53–65 (1987)
A Hybrid Approach for Clustering
and Selecting of Cloud Services Based on User
Preferences Evaluation
Ouassila Hioual1(&), Ouided Hioual2, and Sofiane Mounine Hemam3

1
LIRE Laboratory of Constantine, Abbes Laghrour University,
Khenchela, Algeria
[email protected]
2
Abbes Laghrour University, Khenchela, Algeria
[email protected]
3
ICOSI Laboratory, Abbes Laghrour University, Khenchela, Algeria
[email protected]
Abstract. With the increasing use of cloud computing, it is very important for
the Cloud users to analyze and compare performance of the Cloud services.
Since Cloud services selection problem contains several conflicting criteria, it is
considered as a multi-criteria decision making (MCDM) problem. On another
side, one of the most popular unsupervised data mining methods is Clustering
which is used for grouping set of objects. The contribution of this paper is to
propose an approach based on clustering, Pareto Optimal and MCDM methods.
Our approach allows users to specify the quality requirements of the cloud
services they want to use. It consists of three steps: in the first step, we use the
clustering, more precisely the artificial neural network, to minimize the very
large number of cloud services on the Net. In the second step, we apply Pareto
Optimal algorithm to select non-dominated services. Finally, in the third step,
we use the weights provided by the user to select the most appropriate cloud
service for these requirements. To demonstrate the effectiveness of the proposed
approach, a case study is presented.
Keywords: Cloud services User requirements Pareto optimal Clustering

Neural network MCDM
1 Introduction
Today, cloud computing has grown in popularity in the research community and the
business world. Many end users and businesses use cloud services to save their data or
to gain more computing power.
Using a cloud service provides many advantages for the end users. First, it allows a
significant reduction in costs, since users buy only the resources according to their
needs, without investing in infrastructure or maintenance. It allows, also, the guarantee
of instant and uninterrupted access to computing and storage resources for any user
having a connected machine to the Internet. In addition, users can easily adapt
resources to their specific needs and can add resources on demand.
64 O. Hioual et al.
All these advantages have led to an increase in the number of the cloud computing
end-users. Consequently, many new needs have emerged, among which: the need to
have a system that allows searching and selecting Cloud services corresponding to the
needs of the end-users. Our contribution is in this research axis consists to propose a
hybrid approach based on machine learning, Pareto Optimal and TOPSIS method
(Technique for Order Preference by Similarity to Ideal Solution) [1]. Our approach
allows users to specify and optimize the quality of the required Cloud services they
wish to use.
To do this, we propose an approach that allows selecting the Cloud services that
correspond to user requirements by giving to the users the possibility of optimizing the
value of some of these requirements, namely costing and response time.
The proposed approach consists of three stages: in the first stage, we use clustering,
more precisely the Artificial Neural Network (ANN) [2], in order to minimize the very
large number of Cloud services on the Net. In the second step, we apply the Pareto
Optimal algorithm [3] in order to select the non-dominated services. Finally, in the
third step, we apply the TOPSIS technique. The latter uses the weights provided by the
user in order to select the Cloud service most suited to these requirements.
The rest of this paper is organized as follows. Related work relative to the problem
of the Cloud services selection is discussed in the Sect. 2. Then, in the Sect. 3, we
explain the proposed approach. In Sect. 4, a case study is presented to demonstrate
effectiveness of the proposed approach. Finally, we conclude the paper by a conclusion
and some perspectives.
2 Related Work
The increased use of Cloud Computing has led to the emergence of new needs, such as
that to have a system that allows searching and selecting Cloud services that meet user
requirements.
In order to help users to choose the adequate Cloud services that meet their needs,
many research work have been proposed to offer new solutions. As already mentioned
above, our main goal is to find the Cloud services that best suit the needs of users. In
this section, we will cite some work that treat the problem of selecting Cloud services.
These works are classified into two main classes: those based on similarity and those
based on MCDA methods (Multi-Criteria Decision Analysis).
In [4], the authors presented an algorithm for selecting cloud services. Their
algorithm allows to determine the gains and the cost of the available Cloud services that
can be achieved by Proxy and, consequently returns those that optimize both the gains
and the cost. This algorithm is performed in two steps. In the first step, the proxy selects
the cloud services available following the request sent by the user. The second step of
this algorithm is devoted to calculate the gains and the cost of the pre-selected Cloud
services in order to select those, which their above-cited criteria are optimized. In [5],
the Cloud portal with a Cloud service search has been presented. This system uses the
concept of similarity [6] and consults the Cloud ontology which is adopted to select the
Cloud services that correspond to the requirements specified by the user. These authors
also proposed Cloudle in [7] which is a search engine for Cloud services and its main
A Hybrid Approach for Clustering 65
functionalities are: request processing, similarity reasoning and scoring. As in [5], to

calculate the similarity between Cloud services the Cloudle consults the Cloud
ontology in order to return a list of results sorted by aggregated similarity.
The authors in [8] presented a resource selection service based on the Cloud
ontology. This generates virtual ontologies (Vons) based on virtual resources and
combines them into new resources. Then, it calculates the similarity between these new
resources to determine those that best meet the requirements and needs of the user.
In [9], the authors presented a service-matching algorithm and a service compo-
sition algorithm. The goal of these algorithms is to calculate the semantic similarity [6]
between cloud services to determine if two cloud services are interoperable or not.
As it can be seen, these different works are mainly based on similarity to determine
which of the services best meet the requirements of a user. Thus, they would be better
suited to users who want to find similar Cloud services.
There are many studies that have used MCDA methods to address the problem of
cloud services selection. The authors in [10] conduct an in-depth study of Cloud service
selection techniques, including MCDA-based techniques such as AHP (Analytic Hier-
archy Process) [11], ANP (Analytic Network) Process) [12], MAUT (Multiple Attribute
valUe Theory) [13] and outranking methods such as the ELECTRE method [14].
The authors in [15] proposed an AHP-based approach for the cloud services
selection of SaaS. The used criteria are: functionality, architecture, usability, supplier
reputation and cost. Among the limitations of this method: it has only been tested using
three Cloud services only SaaS. In addition, it is used only to compare SaaS, leaving
aside the other categories of cloud services.
The authors in [16] presented a Quality of Service (QoS) based mapping approach
to combine SaaS and IaaS, then classify the combined Cloud services for end users.
The authors used the AHP method to classify the various Cloud services. The analysis
of literature shows that there is a little work done that considers using MCDA, opti-
mization and machine learning methods together to select adequate services in cloud
computing systems. Thus, there is a need for an approach that considers both end user
preferences and profile to select the needed Cloud service in an adaptive manner. In the
following section, we will present the problem and the proposed approach.
3 Proposed Approach
Several approaches based on Multi Criteria Decision Analysis (MCDA) have been
proposed to solve the problem of service selection. The major problem that arises at the
time of selection is the significant time necessary to meet the needs of the user. In order
to optimize the time to search and select a service, it is necessary to organize, in a first
step, Cloud services in the form of classes using one of the classification methods to
have a pre-selection of a sub set of services, namely an ANN. Then, in the second step,
we apply Pareto Optimal algorithm to select non-dominated classes. Finally, in the
third step, we use the weights provided by the user to select the most appropriate cloud
service class for these requirements.
66 O. Hioual et al.
3.1 Cloud Architecture

In order to provide its services by meeting users’ requirements, the Cloud environment
should be able to receive, perform, monitor and control users’ requests. This sub
section describes the proposed architecture that enables the selection of cloud services.
As shown in Fig. 1, the architecture of our system assumes that the cloud is composed
of two main layers: selection and Cloud services layers. One of the functions of the
selection layer is to allow end users to interact with the cloud. In addition, it selects the
best service according to their profiles and requirements. In order to fulfill these
functions, the structure of the selection layer is composed of five components: manager
component, extractor component, CS-Classifier, CS-Filter and CS-Selector.
Fig. 1. Services selection layer for the Cloud Computing Architecture
3.2 Selection Layer Functionality

In this sub-section, we describe the functionality and the interaction between the dif-
ferent components of the selection layer.
Manager Component. The role of this component is to build a cloud services data-
base. This later contains all information relative to cloud services such as, financial,
location, agility, availability, reliability and so one [17]. So, this component is in,
continuously, interaction with Cloud service layer in order to update the cloud services
data base belonging to the selection layer.
Extractor Component. The extractor component is an intermediary component
between the cloud end user and the others components of the selection layer. The main
functionalities of this component is to receipt the request from the cloud end user, to
extract the main information from this request such as: the cloud end user profile, the
needed cloud service and its requirements; and, finally, it sends the selected cloud
service to the cloud end user. When the extractor component receives the request it:
• Extracts the requested cloud service from the end user request and, then it sends the
requested cloud service to the Manager component in order to allow, to this latter, to
query the cloud services database and to send the features of cloud services, having
the same functionalities, to the CS-classifier component.
• Sends the cloud end user profile to the CS-classifier component.

• Sends to the CS-Selector component the requirements of the cloud end user.
CS-Classifier. As shown in the Fig. 2, the CS-classifier has two inputs: (1) the cloud
end user profile and, (2) a set of cloud services, according to the need of the end user.
This component is an Artificial Neural Network (ANN), it produces as output two
classes: (1) acceptable Cloud services and, (2) non acceptable Cloud services, i.e. this
component classify the cloud services having the same functionalities according to the
profile of the end user. So, the acceptable class will contain the services which cor-
respond to the end user profile.
Fig. 2. Classification of Cloud services according to user profile
CS-Filter. Once the first step is achieved, the cloud services will be divided into two
classes as shown in Fig. 2. Some cloud services are retained and the others are rejected
according to the end user profile.
In this second step, the non-dominate sorting algorithm [3] is used to estimate the
Pareto optimal solutions. The pseudo-code of the function of assignment of the Pareto
rank is given below. In this algorithm, the variable N is the number of the Cloud
services belonging to the acceptable class. At the end of this algorithm, the resources
having a rank equals to 1 form a Pareto optimal.
Pseudo-algorithm Rank Pareto

Currant-Rank ← 1
m←N
while N 6¼ 0 do
For i = 1 to m do
If Xi is not dominated then
Rank(Xi, t) ← Currant-Rank
Rank Xi in a temporary population
N←N−1
End For
Currant-Rank ← Currant-Rank + 1
m←N
End While
End
68 O. Hioual et al.
CS-Selector. In the third step, TOPSIS (Technique for Order of Preferences by

Similarity to Ideal Solution) proposed by [18] is used to select the best compromise
solution from the set of the Pareto front solutions (Rank = 1). In this last step, the user
must provide the weights of the cost and response time criteria. These latter are
transmitted to CS-Selector by the Extractor component (Fig. 1 arrow 3). Finally, once
the best cloud service is selected (according to the end user profile and its require-
ments), the CS-Selector transmit this cloud service to the Extractor component.
4 Case Study
In this section, we evaluate the feasibility of our proposed approach throw a case study.
So, the code is written entirely in the Python programming language [19] and will run
on any platform that supports Python (i.e. almost all platforms).
In order to show how our approach work, we take as example, in this case study,
the cloud service (CS1) which is requested by the end user. The set (S) of cloud
services having the same functionalities as CS1 is as follow: S = {csi1 =i 2 ½1::12 }, i.e.
that there are, totally, 12 cloud services. The cost and response time of these cloud
services csi1 are done in Table 1.
Table 1. Values of the cost and response time of the cloud services csi1
csi1 Cost Response time csi1 Cost Response time
cs11 25 38 cs71 21 31
cs21 19 11 cs81 19 27
cs31 17 17 cs91 26 33
cs41 22 27 cs10
1
21 29
cs51 16 81 cs11
1
21 43
cs61 17 47 cs12
1
33 21
In the first step, the CS-Classifier (the ANN component) produces as output two
categories of the cloud services of the Table 1. To do this, it takes into consideration
the profile of the cloud end-user on one hand and, in addition of the two criteria (cost
and response time) it takes also others features of the cloud services, such as local-
ization, agility, availability, reliability and so one, on the second hand. Therefore, the
two obtained categories are as follow:
Acceptable Coud services = {cs21 , cs31 , cs41 , cs91 , cs10
1 }
Unacceptable Cloud services = {cs1 , cs1 , cs1 , cs1 , cs81 , cs11
1 5 6 7
1 , cs1 }
12
In the second step, the Pareto rank algorithm is applied on the category of the
acceptable cloud services in order to determine the non-dominated cloud services.
As shown in Fig. 3, the cloud services cs21 and cs31 are non-dominated because their
rank is equal to one.
Finally, in the third step, we apply the TOPSIS method in order to select the cloud
service according to the weight of the cost and response time criteria provided by the
cloud end-user. In this case study, they are 0,65 and 0,35 respectively. At the end of the
sixth phase of the TOPSIS method, we obtain the following alternatives ranking:
cs21 > cs31 . So, the CS-Selector selects the Cloud Service cs21 .
Fig. 3. The non-dominated cloud services
5 Conclusion
In this paper, we proposed an approach composed of three steps. In the first, we

proposed an Artificial Neural Network (ANN) in order to classify the set of the
requested cloud services (cloud services having the same functionalities) into two
categories (Acceptable and unacceptable cloud services). This classification is based on
the profile of the cloud end user, and its aim is to minimize the very large number of
Cloud services. In the second step, we applied the Pareto front algorithm on the set of
the cloud services belonging to the acceptable class, to select the non-dominated cloud
services. The selected cloud services are considered to be the best cloud services that
meet the cloud end user requirements. So, this step allows to minimize the search space
since it allowed to select, only, a few cloud services, which allowed to optimize even
more the search and selection time. And, in the third step, we applied the TOPSIS
technique. The latter uses the weights, of the cost and respond time, provided by the
cloud end user in order to select the Cloud service best suited to these requirements.
In this work, we have based on two user requirements, namely the optimization of
the cost and response time. Thus, this work can be improved by taking into account
other criteria (requirements) of the user on one side and improving the performance of
the system on the other side.
70 O. Hioual et al.
References
1. Hioual, O., Boufaïda, Z., Hemam, S.M.: Load balancing, cost and response time
minimisation issues in agent-based multi cloud service composition. Int. J. Internet Protoc.
Technol. 10, 73–88 (2017). https://doi.org/10.1504/IJIPT.2017.085187
2. Lloyd, S.: Least squares quantization in PCM. IEEE Trans. Inf. Theory 28, 129–137 (1982)
3. Hemam, S.M., Hioual, O.: A hybrid load balancing algorithm for P2P-cloud system aware of
constraints optimisation of cost and reliability criteria. Int. J. Internet Protoc. Technol. 10,
99–114 (2017). https://doi.org/10.1504/IJIPT.2017.085189
4. Zeng, W., Zhao, Y., Zeng, J.: Cloud service and service selection algorithm research. In:
Proceedings of the first ACM/SIGEVO Summit on Genetic and Evolutionary Computation.
pp. 1045–1048 (2009)
5. Kang, J., Sim, K.M.: Cloudle: a multi-criteria cloud service search engine. In: 2010 IEEE
Asia-Pacific Services Computing Conference, pp. 339–346. IEEE (2010)
6. Resnik, P.: Semantic similarity in a taxonomy: an information-based measure and its
application to problems of ambiguity in natural language. J. Artif. Intell. Res. 11, 95–130
(1999)
7. Kang, J., Sim, K.M.: Cloudle: An agent-based cloud search engine that consults a cloud
ontology. In: Cloud Computing and Virtualization Conference, pp. 312–318. Citeseer (2010)
8. Yoo, H., Hur, C., Kim, S., Kim, Y.: An Ontology-Based Resource Selection Service on
Science Cloud. In: Ślęzak, D., Kim, T., Yau, S., Gervasi, O., Kang, B. (eds.) GDC 2009.
CCIS, vol. 63, pp. 221–228. Springer, Heidelberg (2009). https://doi.org/10.1007/978-3-
642-10549-4_26
9. Zeng, C., Guo, X., Ou, W., Han, D.: Cloud Computing Service Composition and Search
Based on Semantic. In: Jaatun, M., Zhao, G., Rong, C. (eds.) CloudCom 2009. LNCS, vol.
5931, pp. 290–300. Springer, Heidelberg (2009). https://doi.org/10.1007/978-3-642-10665-
1_26
10. Sun, L., Dong, H., Hussain, F.K., Hussain, O.K., Chang, E.: Cloud service selection: state-
of-the-art and future research directions. J. Netw. Comput. Appl. 45, 134–150 (2014)
11. Saaty, T.L.: The analytic hierarchy process for decision in a complex world. Pittsburgh RWS
Publ. (1980)
12. Satty, T.L.: Decisions with the analytic network process (ANP). Univ. Pittsburgh (USA),
ISAHP (1996)
13. Churchman, C.W., Ackoff, R.L., Arnoff, E.L.: Introduction to operations research (1957)
14. Roy, B.: The Outranking Approach and the Foundations of the ELECTRE Methods. Theory
and Decision (1991)
15. Godse, M., Mulik, S.: An approach for selecting software-as-a-service (SaaS) product. In:
2009 IEEE International Conference on Cloud Computing, pp. 155–158. IEEE (2009)
16. Karim, R., Ding, C., Miri, A.: An end-to-end QoS mapping approach for cloud service
selection. In: 2013 IEEE Ninth World Congress On Services, pp. 341–348. IEEE (2013)
17. Tripathi, A., Pathak, I., Vidyarthi, D.P.: Integration of analytic network process with service
measurement index framework for cloud service provider selection. Concurr. Comput. Pract.
Exp. 29, e4144 (2017)
18. Tzeng, G.-H., Huang, J.-J.: Multiple Attribute Decision Making: Methods And Applications.
CRC Press, USA (2011)
19. Van Rossum, G., Drake, F.L.: The Python Language Reference Manual. Network Theory
Limited, UK (2011)
Basket Patterns in Turkey: A Clustering
of FMCG Baskets Using Consumer Panel Data
Tolga Kaya, Ahmet Talha Yiğit(&), and Utku Doğruak
Department of Management Engineering, Istanbul Technical University,

34367 Istanbul, Turkey
{kayatolga,yigitahm}@itu.edu.tr,
[email protected]
Abstract. The purpose of this study is to suggest a clustering approach to

define the main groups of baskets in Turkish fast-moving consumer goods
(FMCG) industry based on the sectoral decomposition, the total value and the
size of the baskets. To do this, based on the information regarding the 2,965,837
baskets (8,147,233 transactions) of 14293 households which took place in the
calendar year 2018, alternative unsupervised learning methods such as K-means,
and Gaussian mixtures are implemented to obtain and define the basket patterns
in Turkey. A supervised ensembling approach based on XG-Boost method is
also suggested to assign the new baskets into the existing clusters. Results show
that, “SaveTheDay”, “CareTrip”, “Breakfast”, “SuperMain” and “MeatWalk”
are among the most important basket types in Turkish FMCG sector.
Keywords: Basket analysis Cluster analysis K-means FMCG Supervised

learning Consumer panel
1 Introduction
Shopping is one of the most essential actions for the contemporary individuals and
households of the society. With the constantly increasing access to all kinds of prod-
ucts, people’s shopping behaviors are transforming both at the individual and social
levels. Today, this transformation is still being continued by many factors one which is
the increasing frequency of shopping enabled by the increased numbers of both con-
ventional stores such as retailers, and groceries, and modern stores such as online
stores. A significant portion of people have a chance to reach and purchase the product
they need whenever they want; consequently, the shopping lists are almost diminished
to single items [1]. As a consequence, the shopping action of people is becoming more
and more dynamic; moreover, it is essential for companies to analyze and understand
this behavior to create more profitable strategies especially in industries like fast-
moving consumer goods (FMCG) industry which is more dynamic in nature because it
provides inexpensive products which are purchased in daily basis.
Shopping mission term can be defined as the reason why the customers visit the
stores. For example, a customer may visit a store to buy some daily food for breakfast
or make a quick trip to the local grocery to buy some milk which is forgotten to be
bought. There can be various types of shopping missions among a group of customers
72 T. Kaya et al.
depending on their needs and habits. Hence, as the shopping environments are getting
more and more dynamic, the shopping missions of customers are getting more versatile
and important in many sectors such as FMCG [2]. Companies should be aware of these
missions and make decisions about their products such as the in-store locations and
bundles accordingly. Moreover, having a deeper understanding of shopping missions
of customers, companies will have a significant competency over their competitors in
their sectors.
The aim of this study is to provide a clustering model to segment and detect the
different shopping missions in Turkish FMCG industry by using Ipsos Household
Panel data. The data used in the study is containing 2,965,837 baskets of 14293
households which have attended the panel in 2018. Considering, the data is statistically
representative of Turkey, the results found from this sample of Turkish households are
generalisable and beneficial to understand the main basket patterns in Turkish FMCG
sector. After finding the basket patterns exist in Turkish FMCG sector, a classification
model has been proposed to assign future baskets to the existing segments. Conse-
quently, the resulting model is providing a guide for Turkish FMCG companies to
communicate and understand their customers much better and effectively.
The rest of the paper is organized as follows: the literature review of market basket
analysis and shopping missions, the methodology used in the study, data collection and
manipulation, model building, examination of the findings, discussion and suggestions
for the future studies.
2 Literature Review
Like customer segmentation, market basket analysis is a powerful tool for companies to
develop a better understanding of customers by revealing purchasing patterns besides
exploring associations or co-occurrences [3]. Market basket data is utilized in many
researches in the literature by the application of various data mining methods for
different purposes such as apriori analysis [4], eclat [5], and clustering [6] to analyze
the customer behaviors. Applying clustering analysis to market basket data, companies
can gain insights to propose better planned product development, and positioning,
advertisement, and campaign strategies [7, 8].
Shopping missions can be defined as the main motivation of the customers before
they take the shopping action [9]. Even though, there can be some missions that are
common for many societies, it is shown that the motivations behind shopping varies
among different cultures [10]. Consequently, it is a changing and dynamic phenomenon
which should be analyzed accordingly in different cultures and even in different years.
There are many different studies in the literature which are aiming to explore basket
patterns in different countries [11, 12] with different approaches [13, 14].
There are several studies in the literature which aims to explore shopping missions
in different sectors. Griva et al. [15] introduced a data mining-based framework to
discover patterns in customers’ visits in a supermarket and identify their shopping
missions. The authors demonstrated the performance of their model in real data of eight
representative stores of a Greek retailer. Sarantpoulos et al. [16] developed an ana-
lytical method to identify several shopping missions at the store level and demonstrate
Basket Patterns in Turkey 73
the utility, validity, and replicability of this method using a data set with 4 million
baskets from a multinational supermarket chain. In a similar study, Griva et al. [8]
categorized a customer visit by the purchased product categories in the basket and
identified the shopping intention behind the visit e.g. a ‘breakfast’ visit to purchase
cereal, milk, bread, cheese etc. They also suggested a semi-supervised feature selection
approach that uses the product taxonomy as input and suggests customized categories
as the output.
3 Methodology
In the study, different clustering methods, which refers to a very broad set of unsu-
pervised machine learning techniques for finding subgroups, or clusters, in a data set,
have been applied to get the best basket segments. Clustering algorithms mainly aims
to partition the data set into distinct groups so that, the observations within each group
are similar to each other, while in different groups they are different from each other.
K-means clustering is a simple machine learning approach for partitioning a data
set into K distinct, non-overlapping clusters. The idea behind K-means clustering is that
a good clustering is one for which the within-cluster variation is as small as possible.
The within-cluster variation for cluster Ck is a measure W ðCk Þ of the amount by which
the observations within a cluster differ from each other. As shown in Eq. (1), the aim of
the K-means algorithm is to minimize the total within-cluster variation [17].
( )
X
K
minimize W ðCk Þ ð1Þ
C1 ;...;CK
k¼1
Extreme gradient boosting method has been introduced to use gradient boosting
algorithm in a more scalable and computational way [18]. It is one of the most accurate
ensemble learning methods which can be applied both on regression and classification
problems [19].
4 Data
The data used in the study contains 2,965,837 FMCG baskets formed by 8,147,233
transactions of 14293 Turkish households which have been actively participating to
Ipsos Household Panel in the calender year 2018. A full year data is obtained to avoid
the possible misleading of seasonal effects of some certain products. All of the
2,965,837 baskets have certain characteristics as the products purchased in the baskets
belongs to 127 different FMCG categories ranging from toothpaste to legumes. To
extract the best information, firstly all the baskets are resembled with their spending
decompositions among the FMCG categories. However, using a similar approach to
Grive et al. [6], through an iterative process, these initial 127 macro categories are
aggregated into first 82, then 52, and finally 8 micro categories as shown in Table 1.
74 T. Kaya et al.
Table 1. Some example micro categories and their assigned macro categories
Example micro categories Macro category
Chocolates, Ice creams, Nuts Confectionery
Garbage bags, Bath and Kitchen Cleaners House Care
Fabric Conditioners, Laundry and Dish Detergents LaunDish
Flour, Spices, Legumes Lunch and Dinner
Red and White meat Meat
Cheese, Eggs, Milks Milk and Breakfast
Diapers, Skin care, Hair care Personal and Baby Care
Tea, Ayran, Carbonated soft drinks Refreshment
After creating the macro categories, the baskets are represented with the distribution
of the money spent on each macro category. Additionally, to have a grasp of each
basket’s general identity, the total value of each baskets and the total number of items
purchased in each basket are added as the final features. As a result of this stage of data
preparation, every individual basket are represented by 10 main features which are
spendings on “Confectionery”, “House Care”, “LaunDish”, “Lunch and Dinner”,
“Meat”, “Milk and Breakfast”, “Personal and Baby Care”, “Refreshment” categories,
the total value of basket (Totval), which is equal to the sum of category spendings, and
the total number of items purchased (Totitem) in the basket.
In the second part of data preparation, a new category called “Snacks and Drinks” is
formed by merging Confectionery and Refreshment categories to use in the second
stage of clustering analysis which is aiming to cluster the relatively smaller baskets
(Save the Day cluster which is obtained as a main cluster) with respect their total value
and total number of purchased items. Crafting this new feature, enabled the model to
give more distinct clusters.
As the final step of data manipulation, all the featıres obtained are standardized to
have the ready to use data; thus, the problems which can be faced while working with
Euclidean distance because of the possible differences in the ranges of values of
features such as the Totval and Totitem features are avoided. The standardization for
each feature is done by the method shown in Eq. (2).
xi x
x0i ¼ ð2Þ
r
Completing the standardization, the data has become ready to use in the further
clustering analysis, classification model building for the future basket assignments.
5 Model Building and Comparison
As the first step of the model building process, the several clustering techniques
including K-means and Gaussian mixtures clustering methods have been applied to the
data to have a proper shopping mission segmentation by exploring the patterns
emerging in baskets. As the second step, a supervised model based on XG-Boost
method is built to allocate future baskets to the existing segments that created in the
first stage. Both unsupervised and supervised modelling stages are applied for both
main clusters and sub-clusters.
In the first step, K-means clustering method has been applied to the data. By
examining the elbow plot, and many trial and error steps, the optimal number of
clusters to define the best clusters is determined as 9. To compare alternative methods,
including Gaussian mixtures, 9 cluster results of them are also computed. After the
comparison of the results with respect to the compliance with the aim of the project,
and computation costs, K-means clustering is determined as the best model because of
its reproducible and computationally less expensive nature, and the more interpretable
results. The number of baskets within each main cluster found by K-means algorithm
are distributed as shown in Table 2.
Table 2. Basket decomposition among main clusters created by K-means model

1 2 3 4 5 6 7 8 9
2278212 32777 315186 60913 30374 50645 121371 49819 26540
One of the 9 main clusters obtained by the result of K-means model, which is
cluster 1 shown in Table 2, is formed by the majority of the baskets. This cluster is
representing the single or few item and low value baskets which are quite common in
FMCG sector. Consequently, another K-means clustering is applied to this particular
cluster with the aim of partitioning it into smaller sub-clusters which may give further
interpretable missions. After the K-means application, 4 sub-clusters of the main cluster
1 have been detected. As a result, the unsupervised part of the model building and
comparison stage is finalized by obtaining 7 main clusters and 4 sub-clusters.
Having all the main and sub-clusters, the supervised stage of model building has
started. XG-Boost classification models are built to allocate future baskets to the
existing clusters. As in the clustering approach, two different models for both main and
sub clusters have been built in the classification approach. Because of high accuracies
achieved by XG-Boost models, no alternative method has been applied.
For the main cluster allocation, at first the total baskets are split into a 75 percent
training data set and 25 percent test data set. A tree based, 9 class soft probability XG-
Boost model is trained on training data, with the defined parameters. The parameters
are determined after some trials to optimize the convergence and accuracy of the
model. The eta parameter which is equal to learning rate in other machine learning
methods is defined as 0.001, max_depth which means the maximum depth of trees
defined as 5, gamma which means minimum loss reduction is set to 1 and other
parameters are used as default values. After around 10000 iterations of training, the
model has converged, and the model has become ready to test. All the predictions of
the created model on the test data set have been belonging probabilities of each baskets
to each class. As these soft assignments are not evaluable, all baskets are assigned to
classes in which the have the highest probability of being a member. After, assigning
classes for every basket, the predictions and the true labels are compared. The model
76 T. Kaya et al.
achieved a 0.995 accuracy on test set. The confusion matrix comparing the predictions
and true labels is shown Fig. 1.
Fig. 1. The confusion matrix of the XG-Boost model for main clusters
After having a sufficient accuracy on main clusters, the sub-clusters obtained by

clustering the main cluster 1 are also needed to be allocated for any basket belonging to
main cluster 1. To achieve this, a new tree based 4 class soft probability XG-Boost
model with the exact same parameters is trained on the training data which is a 75
percent sample of cluster 1 baskets. This model also converged to achieve a high
accuracy of 0.997 which is even higher than the model for the main clusters.
As a result, the clustering and allocation tasks are finalized. The allocation of new
baskets can also be done with low error with respect to the high accuracies of the
classification models.
6 Findings
The 9 main clusters obtained are named after their average feature values which are
shown in Table 3.
After examining the table, clusters are named according to their characteristic
features. For example, because of very high value of Meat category in the baskets of
cluster 2, it is names as “MeatWalk” which implementing a shopping basket oriented to
meat related products. The names assigned to each cluster are listed in Table 4.
Using the same method, the sub-clusters of main cluster 1, which is called
“SaveTheDay” are also named as, “MeatRun”, “MilkRun”, “SnacksNDrinks”, “Sin-
gleShots”. As the result, the basket patterns existing in the Turkish FMCG sector in
2018 are defined by using a two stage K-means clustering approach. Besides, with the
high accuracy classification models proposed, the future baskets are assured to be
assigned to the patterns they belong to.
Table 3. The average feature values of main clusters

Cluster TVal TItm CNF HC LND L&D Meat M&B P&B Ref
1 11 2 1 0 0 2 2 3 1 2
2 113 3 2 0 0 3 101 4 1 1
3 46 7 12 1 1 16 4 6 3 3
4 57 4 3 0 1 6 2 3 2 40
5 56 7 4 20 5 8 3 6 7 4
6 68 6 4 2 36 8 3 5 7 4
7 58 5 2 0 0 8 3 40 1 2
8 67 5 3 1 1 4 2 3 52 2
9 210 19 32 5 10 80 16 33 18 16
Table 4. The main clusters and their assigned pattern names

Cluster Pattern name
1 SaveTheDay
2 MeatWalk
3 MainShop
4 Refreshers
5 HouseCare
6 LaunDish
7 Breakfast
8 CareTrip
9 SuperMain
7 Conclusion and Discussion
The output of this study proposed two models one of which has determined the existing
basket pattern in Turkish FMCG sector in 2018, while the other has allocated new
baskets in the following years. Even though, both models are very successful at
achieving their goals, many different approaches could be used to create both of them.
For example, the macro categories in the data part can be designed in different ways
that may result in different missions. A fully connected neural network can be trained
for classification part to achieve even higher accuracies. Moreover, further analysis
such apriori and eclat analysis can be applied to find relationships and cliques between
products in each basket pattern.
Acknowledgement. This study is an outcome of the collaboration project conducted by IPSOS

and ITUNOVA (2019) entitled “Clustering FMCG Consumers and Shopping Baskets in Turkey
using Basket Size, Product Decomposition and Household Profiles”.
78 T. Kaya et al.
References
1. Schmidt, M.: Retail shopping lists: reassessment and new insights. J. Retail. Consum. Serv.
19(1), 36–44 (2012)
2. Severin, V., Louviere, J.J., Finn, A.: The stability of retail shopping choices over time and
across countries. J. Retail. 77(2), 185–202 (2001)
3. Charlet, L., Ashok, D.: Market basket analysis for a supermarket based on frequent itemset
mining. IJCSI Int. J. Comput. Sci. Issues 9 (2012). 1694-0814. www.IJCSI.org
4. Rathod, A., Dhabariya, A., Thacker, C.: A survey on association rule mining for market
basket analysis and apriori algorithm. Int. J. Res. Adv. Technol. 2(3), 2321–9637 (2014)
5. Pachunkar, D., Kothari, A.: A survey on ECLAT based algorithm. Int. J. Sci. Res. Dev.
5(10) (2017). 2321-0613
6. Yun, C., Chuang, K., Chen, M.: Using category-based adherence to cluster market-basket
data. In: Proceedings of the 2002 IEEE International Conference on Data Mining, 2002,
Maebashi City, Japan, pp. 546–553 (2002)
7. Liu, R., Lee, Y., Mu, H.: Customer classification and market basket analysis using k-means
clustering and association rules: evidence from distribution. In: Big Data of Korean Retailing
Company (2018)
8. Griva, A., Bardaki, C., Pramatari, K., Papakiriakopoulos, D.: Retail business analytics:
customer visit segmentation using market basket data. Expert Syst. Appl. 100, 1–16 (2018)
9. Berkhout, C.: Shopping Missions. In: Assortment and Merchandising Strategy. Palgrave
Macmillan, Cham (2019)
10. Evanschitzky, H., Emrich, O., Sangtani, V., Ackfeldt, A.-L., Reynolds, K.E., Arnold, M.J.:
Hedonic shopping motivations in collectivistic and individualistic consumer cultures. Int.
J. Res. Mark. 31(3), 335–338 (2014)
11. Jamal, A., Davies, F., Chudry, F., Al-Marri, M.: Profiling consumers: a study of Qatari
consumers’ shopping motivations. J. Retail. Consum. Serv. 13(1), 67–80 (2006)
12. Mehta, R., Sharma, N., Swami, S.: A typology of Indian hypermarket shoppers based on
shopping motivation. Int. J. Retail Distrib. Manag. 42(1), 40–55 (2014)
13. Yun, C., Chuang, K., Chen, M.: An efficient clustering algorithm for market basket data
based on small large ratios. In: 25th Annual International Computer Software and
Applications Conference (2001)
14. Gil, J., Tobari, E., Lemlij, M., Rose, A., Penn, A.: The differentiating behaviour of shoppers:
clustering of individual movement traces in a supermarket (2009)
15. Griva, A., Bardaki, C., Panagiotis, S., Papakiriakopoulos, D.A.: A data mining-based
framework to identify shopping missions. In: MCIS, p. 20 (2014)
16. Sarantopoulos, P., Theotokis, A., Pramatari, K., Doukidis, G.: Shopping missions: an
analytical method for the identification of shopper need states. J. Bus. Res. 69(3), 1043–1052
(2016)
17. Hastie, T., Tibshirani, R., Friedman, J.: The Elements of Statistical Learning: Data Mining,
Inference, and Prediction, 2nd edn. Springer, New York (2017)
18. Chen, T., Guestrin, C.: XGBoost: a scalable tree boosting system. In: 22nd ACM SIGKDD
Conference (2016)
19. Gradient Boosting and XGBoost. https://medium.com/@ODSC/gradient-boosting-and-
xgboost-9b4a23b84944. Accessed 13 Apr 2020
Journey Segmentation of Turkish Tobacco
Users Using Sequence Clustering Techniques
Ahmet Talha Yiğit(&), Tolga Kaya, and Utku Doğruak

{yigitahm,kayatolga}@itu.edu.tr,
[email protected]
Abstract. In this paper, using individual tobacco panel data, a novel and
behavioral approach based on sequence clustering techniques is proposed to
have a deeper understanding of different behavior types of Turkish tobacco users
during the consecutive price markups of 6th April and 2nd May in 2019. To
achieve this, main brands before markups are determined for each of the 5052
individuals. Then, having some prior assumptions, their purchase behaviors are
obtained as time-stamped event sequences. Finally, while a portion of the
obtained sequences which are less complex are segmented with empirical
analyses, the rest of them are segmented using hierarchical clustering with
optimal matching event (OME) distance. Results suggested seven main type of
behavior among the tobacco users in Turkey during the markup period.
Keywords: Sequence analysis Individual tobacco panel Hierarchical

clustering Sequence clustering Unsupervised learning
1 Introduction
Understanding the consumer behavior changes with respect to occurrences which are
significantly affecting the circumstances of a sector is a common concern for all players of
all markets. The change in the behaviors of consumers of a product in any sector can be
perceived and analyzed in many different dimensions and aspects. As an example, in
Turkish tobacco sector, price changes are having more and more impact on the prefer-
ences of tobacco users with the ongoing regulations some which are preventing tobacco
companies to position their brands. Consequently, the annual and occasional markup
periods of tobacco products create a significant area of concern for tobacco companies to
create policies to take the most profitable actions by having a deeper understanding of the
tendencies and behaviors of their customers during these periods. However, because of
the high purchase frequency and the highly competitive nature of tobacco sector, the
static representations may not provide the most advantageous insights about the ten-
dencies and preferences of the costumers during these crucial periods. In this highly
active environment, it is essential for tobacco companies to regard and solve this problem
in a more comprehensive and dynamic way by using novel approaches.
The aim of this study is to provide a novel approach to segment the Turkish tobacco
users by their purchase preferences during markup periods to enable tobacco
80 A. T. Yiğit et al.
companies to have sufficient information about the market dynamics. Considering

individual purchasing journeys as sequences, sequence clustering techniques are
applied with appropriate usage of sequence distance metrics such as optimal matching
(OM) and optimal matching event (OME). Applying these methods, the problem of
understanding the customer journeys of Turkish tobacco users is inspected from a
different angle; moreover, a promising model using OME distance in hierarchical
clustering is proposed as the output of this paper.
The rest of the paper is organized as follows: the literature review of customer
segmentation, clustering and sequence clustering, the methodology used in the study,
data collection and preparation, model building, examination of the findings, discussion
and suggestions for the future studies.
2 Literature Review
Customer segmentation is one of the most essential topics of marketing; thus, there has
been many scientific studies which has created a variety of different approached to this
crucial issue over years [1]. Traditional segmentation techniques, which are still widely
used in many sectors, mostly considers non-behavioral features of customers such as
demographics and socioeconomic status [2]. On the other hand, with the increase in the
availability of data in recent years, data-driven behavioral applications of market
segmentation are becoming more and more promising to create more reliable seg-
mentations [3].
Machine learning techniques are categorized in three main topics, which are
supervised, unsupervised, and semi-supervised machine learning techniques, in the
literature [4]. Clustering is one of the most common application fields of unsupervised
learning techniques. Because of the nature of market segmentation is to create sub-
groups of the customer population without having any pre-determined labels, it can be
considered as a clustering problem. By using clustering techniques in market seg-
mentation applications, the more similar customers with respect to the regarded features
are kept in the same cluster while the less similar ones are distributed to other clusters
[5]. This natural harmony of clustering techniques and market segmentation problem
creates many different available approaches to the possible beneficial solutions [6–8].
In social sciences, sequence analysis is a tool to examine individuals with respect to
their biographical data such as employment histories, and family life cycles [9]. One of
the methods to analyze sequence data is clustering them using distance metrics suitable
to the type of the sequence [10]. The sequence data are categorized as state sequences
and event sequences with respect to the way they are created. However, this nuance
difference between the two types of data, creates a need for different distance metrics to
cluster them efficiently [11]. There are many studies about the alternative distance
metrics for both types of sequence data in the literature; however, optimal matching
distance for state sequences and optimal matching event distance for event sequences
are generally preferred. On the other hand, having appropriate purchase data of cus-
tomers, they can be clustered using their purchase history as state or event sequences to
create novel and behavioral approaches to market segmentation and related fields.
There are several examples of these possible novel approaches in the market seg-
mentation, and marketing literature ranging from creating a recommendation system for
Journey Segmentation of Turkish Tobacco Users 81
customers by analyzing their journeys [12] to analyze market baskets by sequences of

buying order of products [13] and using sequence analysis in market research to
analyze market behavior [14].
This paper proposes a novel behavioral approach for customer segmentation by
utilizing the sequence analysis and clustering techniques using the purchase journey
data obtained from individual tobacco panel. Besides, by examining the markup
periods, which are periods of interest for tobacco companies, an innovative method-
ological approach is also introduced.
3 Methodology
In this paper hierarchical clustering and partitioning around medoids techniques are
used to cluster the event sequence data created by the purchase histories of Turkish
tobacco users during markup periods. The dissimilarity matrix is created using OME
distances between each pair of event sequences.
OM analysis is used to calculate pairwise distances between state sequences with
respect to their dissimilarities. By using the obtained pairwise distances, patterns in the
data are explored with the application of clustering techniques. OME is a modified
version of OM for the application of the same principle on event sequences [15]. OME
is basically explained as the total cost of transforming one event sequence into another.
The cost is computed by the amount of insertion-deletion, substitution, and time
synchronization operations needed to complete the sequence transformation. The cost
for each operation can be defined by the user in TraMineR R package [16]. As in OM,
OME cost is computed for all possible pairs of observations to create a distance matrix
for the application of clustering algorithms.
Hierarchical clustering is a clustering technique which takes a dissimilarity matrix and
a linkage criterion as inputs and gives clusters as an output. While creating the clusters,
Ward’s minimum variance method has been used to compute the inter-cluster distances,
which are calculated as shown in Eq. (1), to decide which clusters to merge [17].
1 XX 2
d ðC1 ; C2 Þ ¼ d xi ; xj ð1Þ
NC1 NC2 i2C1 j2C2
Partitioning around medoids (k-medoids) is a divisive clustering algorithm which is

highly similar to k-means clustering method. However, in k-medoids the cluster centers
are chosen from the observations while in k-means it is not necessary. Another dif-
ference between is that arbitrary distance metrics such as optimal matching event
distance can be used with partitioning around medoids technique, while the k-means
algorithm uses Euclidian distance. The cost function of partitioning around medoids
method is shown in Eq. (2) where there are k clusters [18].
Xk X 2
Cost ¼ j¼1 i2Cj
d xi ; zj ð2Þ
4 Data
The analysis in this paper, the aim of which is obtaining insight and knowledge about
the effect of consecutive price markups of 6th April and 2nd May in 2019 on brand
preferences in tobacco sector, has been done using the journeys of individuals in IPSOS
Individual Tobacco Panel data. IPSOS Individual Tobacco Panel provides a reliable
data source which gathers tobacco purchase information from around 6000 people
periodically, and because of the high sample size, the data provides a high statistical
representativeness of Turkish tobacco users. The data originally includes various fea-
tures of every purchase event such as the date and place of the purchase, the brand of
the purchased cigarette, and many socioeconomic and demographic features of the
customer. Because the aim of this paper is to analyze the brand journeys of customers,
only the date, brand and price information about the purchases have been used to create
the sufficient event sequences to analyze and cluster.
Before creating the event sequence data, a comprehensive preliminary study has
done to determine the scope of the project, and to provide a more robust dataset to
clustering algorithms. As the first step, the period that has been analyzed is determined
and divided into three sub-periods as before, between, and after the two markup dates.
Before markups period has been used to determine a main brand of cigarette for each of
the individuals in the data. While determining a main brand for an individual, following
two constraints are used for every customer: the number of main brand purchases
should be at least half of the number of total purchases, and the number of main brand
purchases should be at least 1.5 times more than the number of purchases of the second
most frequently bought brand. The individuals who actively give data during the
determined period and meet the main brand determination constraints are selected to be
analyzed in the study.
The next step of data preparation is to create event sequences by chronologically
ordering the purchases of each individual during the between and after markups
periods. After creation of the event sequences, prices of each purchased brand between
and after markup periods have been compared with the main brand price for each
sequence to diminish the variety of possible events caused by the high number of
different brands into four main events which are main brand purchase (MAINB),
purchase of a different brand which has the same price with the main brand (SAMEP),
purchase of a different brand which has a higher price than the main brand (HIGH-
ERP), purchase of a different brand which has a lower price than the main brand
(LOWERP). As the result, the time-stamped event sequences which are created with
four different possible events have become ready for further analysis.
As the final step of data preparation for clustering algorithms, a portion of the
obtained sequences which are less complex are segmented with empirical analyses. The
individuals who have bought only their main brand during the determined period has
been segmented as “OnlyMB” group. On the other hand, the individuals who started to
purchase a different brand are segmented into three groups of “NonMBEnd-High-
erPrice”, “NonMBEnd-SamePrice”, and “NonMBEnd-LowerPrice” with respect to the
comparison made between the prices of the brands, that they switched to at the end of
the markup period, and the main brand prices.
As the result, 1162 individuals who have not assigned to any segment by the
empirical analysis, whose sequences are shown in Fig. 1, have become ready for
clustering analysis.
Fig. 1. The unassigned sequences after empirical analysis.
5 Model Building and Comparison
Having the ready to use event sequence data, sequence clustering techniques have been
applied to determine different patterns among the purchase behaviors of the unassigned
1162 individuals. As the event sequence data is represented by categorical variables,
OME distance measure is used to form the difference matrix by computing all pairwise
distances between the observations. Because the aim of this paper is to explore different
patterns regardless of time, the unit time change cost is taken as 0. Moreover, the
insertion and deletion costs for all features in sequences are determined 1.
After creating the dissimilarity matrix, hierarchical clustering and partitioning
around medoids methods, which are suitable to use with dissimilarity matrices instead
of data itself, have been applied to find clusters. The resulting clusters from both
methods were very similar to each other, having the value of Rand Index as 0.847.
To make a choice between two alternative clustering results, the performance
measures have been compared as seen on Table 1. Examining these measures, it is
decided that both models have their advantages and drawbacks with respect to different
comparison criteria. As most of the measures are very close to each other, silhouette
width is used as the decision criteria. As a result of this comparison, clustering results
from the hierarchical clustering method is chosen as it provides two better separated
clusters than the partitioning around medoids method [19].
Resulting clusters obtained by hierarchical clustering method are visualized in
Fig. 2. As can be seen, the model has mainly separated the journeys with respect to the
price segments that they switch to at the beginning or in the middle of the period. After
this stage, all the journeys are assigned to a journey segment.
Table 1. Comparison of hierarchical clustering and partitioning around medoids results

Clustering method Dunn index Clusterwise 1 Average 2 Silhouette widths 3
Hierarchical clustering 0.214 0.608 0.694 0.486
Partitioning around medoids 0.214 0.550 0.635 0.739
Fig. 2. Resulting clusters of hierarchical clustering application
6 Findings
The results obtained by the paper, proposes that there exist 7 different patterns of
tobacco user journeys during the markup period examined.
Table 2 shows the proposed journey types and their percentage distribution among
all Turkish tobacco users. As shown, a big majority of total tobacco users have never
switched to any other brand during the period after the markup. The rest of them either
changed their brands at the end of the period, or at the beginning or in the middle.
Table 2. Journey types, their descriptions, and percentages with respect to total size
Journey type Description %Size
OnlyMB Purchase only main brand 61.4
NonMBEnd-HigherPrice Switch to a higher price brand at the end 5.2
NonMBEnd-SamePrice Switch to a same price brand at the end 4.3
NonMBEnd-LowerPrice Switch to a lower price brand at the end 7.5
MBEnd-HigherPrice Tries a higher price brand yet ends with main brand 7.6
MBEnd-SamePrice Tries a same price brand yet ends with main brand 5.3
MBEnd-LowerPrice Tries a lower price brand yet ends with main brand 8.7
In conclusion, it is shown that, using panel data, households and individuals can be
examined according to their purchase journeys by converting their daily purchases into
sequences of events. The model and methodology proposed as the output of this study
has a high potential to be applied in many different product categories in FMCG and in
many different sectors, and to explore beneficial patterns in different scenarios. How-
ever, the methodology can be improved with further developments to diminish the
amount of assumptions made in the data preparation process. A possible alternative
method can be using state sequences instead of event sequences while creating the data.
Acknowledgement. This study is an outcome of the collaboration project conducted by IPSOS

and ITUNOVA (2020) entitled “Exploration of the Shopping Journeys and Customer Churn in
FMCG Sector in Turkey Using Data Analytics Techniques”.
References
1. Ngai, E.W.T., Xiu, L., Chau, D.C.K.: Application of data mining techniques in customer
relationship management: a literature review and classification. Expert Syst. Appl. 36(2),
2592–2602 (2009)
2. Teichert, T., Shehu, E., von Wartburg, I.: Customer segmentation revisited: the case of the
airline industry. Transp. Res. Part A: Policy Pract. 42(1), 227–242 (2008)
3. Cooil, B., Aksoy, L., Keiningham, T.L.: Approaches to customer segmentation. J. Relat.
Mark. 6(3–4), 9–39 (2008)
4. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. The MIT Press, Cambridge
(2016)
5. Saunders, J.A.: Cluster analysis for market segmentation. Eur. J. Mark. 14(7), 422–435
(1980)
6. Tsai, C.-Y., Chiu, C.-C.: A purchase-based market segmentation methodology. Expert Syst.
Appl. 27(2), 265–276 (2004)
7. Kim, K., Ahn, H.: Using a clustering genetic algorithm to support customer segmentation for
personalized recommender systems. In: Artificial Intelligence and Simulation, pp. 409–415
(2005)
8. Namvar, M., Gholamian, M.R., KhakAbi, S.: A two phase clustering method for intelligent
customer segmentation. In: 2010 International Conference on Intelligent Systems, Modelling
and Simulation (2010)
9. Piccarreta, R., Studer, M.: Holistic analysis of the life course: methodological challenges and
new perspectives. Adv. Life Course Res. 41, 100251 (2018)
10. Ritschard, G., Burgin, R., Studer, M.: Exploratory mining of life event histories. In:
McArdle, J.J., Ritschard, G. (eds.) Contemporary Issues in Exploratory Data Mining in
Behavioral Sciences, pp. 221–253. Routeledge, New York (2018)
11. Studer, M., Ritschard, G.: What matters in differences between life trajectories: a
comparative review of sequence dissimilarity measures. J. R. Stat. Soc. Ser. A (Stat. Soc.)
179(2), 481–511 (2015)
12. Terragni, A., Hassani, M.: Analyzing customer journey with process mining: from discovery
to recommendations. In: 2018 IEEE 6th International Conference on Future Internet of
Things and Cloud (FiCloud), Barcelona, pp. 224–229 (2018)
13. Kamakura, W.A.: Sequential market basket analysis. Mark. Lett. 23(3), 505–516 (2012)
14. Silberer, G.: Analyzing sequences in marketing research. In: Diamantopoulos, A., Fritz, W.,
Hildebrandt, L. (eds.) Quantitative Marketing and Marketing Management. Gabler Verlag,
Wiesbaden (2012)
15. Studer, M., Müller, N., Ritschard, G., Gabadinho, A.: Classer, discriminer et visualiser des
séquences d’événements. In: Revue des nouvelles technologies de l’information (RNTI),
E-19, pp. 37–48 (2010)
16. Gabadinho, A., Ritschard, G., Müller, N.S., Studer, M.: Analyzing and visualizing state
sequences in R with TraMineR. J. Stat. Softw. 40(4), 1–37 (2011)
17. Ward, J.H.: Hierarchical grouping to optimize an objective function. J. Am. Stat. Assoc. 58
(301), 236 (1963)
18. Hastie, T., Tibshirani, R., Friedman, J.: The Elements of Statistical Learning: Data Mining,
Inference, and Prediction, 2nd edn. Springer, New York (2017)
19. Rousseeuw, P.J.: Silhouettes: a graphical aid to the interpretation and validation of cluster
analysis. J. Comput. Appl. Math. 20, 53–65 (1987)
A Survey on Spherical Fuzzy Sets
and Clustering the Literature
Eren Ozceylan1(&) , Baris Ozkan2 , Mehmet Kabak3 ,

and Metin Dagdeviren3
1
Industrial Engineering Department, Gaziantep University,
27100 Gaziantep, Turkey
[email protected]
2
Industrial Engineering Department, Ondokuz Mayıs University,
55270 Samsun, Turkey
[email protected]
3
Industrial Engineering Department, Gazi University, 06560 Ankara, Turkey
{mkabak,metindag}@gazi.edu.tr
Abstract. In addition to the well-known fuzzy sets, a novel type of fuzzy sets
called as spherical fuzzy set (SFS) is recently introduced in the literature. SFS is
the generalized structure over existing structures of fuzzy sets (intuitionistic
fuzzy sets, Pythagorean fuzzy sets, and neutrosophic sets) based on three
dimensions (truth, falsehood, and indeterminacy) to provide a wider choice for
decision makers (DMs). Although the SFS is introduced recently, the topic
attracts the attention of academicians at a remarkable rate. Hence, a compre-
hensive literature review of recent and state-of-the-art papers is vital to draw a
framework of the past, and to shed light on future directions. To do so, a
systematic review methodology is followed and 43 academic papers include
SFS are selected and reviewed. The papers are then analyzed and categorized
based on the titles and abstracts to construct a useful foundation of past research.
Finally, trends and gaps in the literature are identified to clarify and to suggest
future research opportunities in fuzzy logic area.
Keywords: Spherical fuzzy sets Survey Fuzzy logic Literature review

Clustering
1 Introduction
After the introduction of fuzzy logic proposed by Zadeh [1] which can take values in
the range [0, 1], soft computing has gained a remarkable acceleration in the literature.
However, the concept of an ordinary fuzzy set (henceforth called a type-1 fuzzy set) is
criticized by most of the researchers due to its disadvantages [2]. These disadvantages
can be attributed to the problem that membership functions are mapped to exact real
numbers [3]. To cope with the aforementioned drawback, the boundaries of the
membership areas themselves are considered fuzzy which results to second order fuzzy
sets (type-2) [4]. Later on, different fuzzy set extensions are proposed in literature. For
instance, intuitionistic fuzzy sets (IFS) and type-2 fuzzy sets are introduced by
88 E. Ozceylan et al.
Atanassov [5, 6] who stated that there may be a situation of determining the degree of
non-membership, emphasized that the degree of indecision of DMs should also be
taken into consideration. At the same year, Yager [7] added the concept of fuzzy multi-
sets to the literature. In these multi-sets, one member can belong to the same cluster
with the same or different degrees of membership. In 1999, Smarandache [8] intro-
duced neutrosophy to deal with the problems involves indeterminate and inconsistent
information. Garibaldi and Ozen [9] think that membership functions may change over
time and include a dynamic environment. Therefore, the nonstationary fuzzy sets is
developed by Garibaldi and Ozen [9] to add uncertainty to the membership functions of
a fuzzy system to model variability caused by experts. Afterwards, the concept of
hesitant fuzzy sets is first introduced to the literature by Torra [10]. Hesitant fuzzy
clusters give the DM the ability to assign more than one membership value for a
particular element. As an extension of IFSs, Pythagorean fuzzy sets (PFSs) are intro-
duced by Yager [11]. The sum of squares of membership and non-membership degrees
in PFSs should be less than or equal to 1. PFSs also provide that l2 þ t2 þ p2 ¼ 1
where l is the membership function, t is the non-membership function, and p is the
hesitation margin (that is, neither membership nor non-membership functions). Then,
picture fuzzy set which is a direct generalization of the fuzzy set introduced by Zadeh
[1] and the IFS proposed by Atanassov [5] is developed by Cuong [12]. In addition to
the aforementioned three parameters (l, t; and p), the picture fuzzy set defines a new
parameter namely refusal membership equals the following condition:
ð1 ðl þ m þ pÞÞ. Finally, Yager [13] introduce qRung orthopair fuzzy sets as an
extension of the IFSs and the PFSs to express uncertain information. Their eminent
characteristic is that the sum of the qth power of the membership degree and the qth
power of the degrees of non-membership is equal to or less than 1.
Considering the fuzzy extensions above, a novel concept of SFSs is developed by
Kutlu Gundogdu and Kahraman [14] to provide a larger preference domain for DMs.
DMs can assign their hesitancy information about an alternative with respect to a
criterion independently from membership and non-membership degrees. SFS meet the
condition that the squared sum of its membership degree, non-membership degree and
hesitancy degree is less than or equal to one. In SFSs, while the squared sum of
membership, non-membership and hesitancy parameters can be between 0 and 1, each
of them can be defined between 0 and 1 independently. For detail information about the
theoretical background of SFS, the reader is referred to Kutlu Gundogdu and Kahraman
[14, 15]. Although the SFS is introduced recently, the topic attracts the attention of
academicians. However, no paper which reviews existing studies about SFS is
observed. To do so, a systematic review methodology is followed in this study. After
the introduction section, review methodology with the detailed analysis of reviewed
papers is clarified in Sect. 2. The current gaps and future research opportunities are
presented in Sect. 3.
A Survey on Spherical Fuzzy Sets and Clustering the Literature 89
2 Review Methodology and Results
The study was conducted by covering the accepted journal papers (available online)
and proceedings in scientific English language. The term of “spherical fuzzy set” is
searched using the “title, abstract, keywords” search bar in Scopus database, and sorted
by relevance. It should be mentioned that the search engine is updated periodically due
to the acquisition of new publications, relevance, citations, and so forth, so the process
of collecting papers is terminated on March 1, 2020. According to the search, 43 papers
are investigated (Table 1).
Table 1. Reviewed papers.

Paper Explanation
[16] By discussing IFSs and picture fuzzy sets, a numerical and geometric comparison is
made between them
[17] New aggregation operators have been proposed for T-SFSs. A multi-criteria
decision-making approach based on these operators is presented
[18] By introducing SFSs, spherical fuzzy numbers based weighted averaging and
weighted geometric aggregation operators are presented
[19] The family of novel similarity measures between SFSs based on cosine function is
studied considering the positive, neutral, negative and refusal grades in SFSs
[20] The classical VIKOR is extended using SFSs. Its applicability and validity is showed
through in warehouse location selection problem
[21] Ten similarity measures are defined between SFSs based on the cosine function by
considering the degree of positive-membership, degree of neutral-membership
membership, degree of non-membership membership and degree of refusal-
membership in SFSs
[15] The classical WASPAS method has been extended to the spherical fuzzy WASPAS
method and its application with the industrial robot selection problem has been
shown
[22] Operations such as spherical fuzzy t’-norms and spherical fuzzy T-Conorms are
introduced. A food circulation center evaluation problem is given as a practical
application
[14] A new theory on SFSs is proposed together with its fundamentals. The classical
TOPSIS method is extended to spherical fuzzy TOPSIS and it has been compared on
illustrative example with intuitionistic fuzzy TOPSIS
[23] By extending the classical CODAS method to the spherical fuzzy CODAS method,
shown its application with an illustrative example. It also performs comparative and
sensitivity analysis between intuitionistic fuzzy TOPSIS, intuitionistic fuzzy CODAS
and spherical fuzzy CODAS
[24] Interval-valued T-SFSs are introduced, the basic operations are proposed for
IVTSFSs and their properties are investigated
[25] Aggregation operator for the SFS which is based on harmonic mean operator by
using algebraic and Einstein strict Archimedean T-norm and T- Conorm is proposed.
Based on this developed aggregation operators, the new multi attribute group
decision making method has been established for decision making problems and this
method’s effectiveness and reliability examined by illustrative example
(continued)
Table 1. (continued)
Paper Explanation
[26] The spherical linguistic sets and a set of spherical linguistic aggregation operators to
deal with the interrelationship among spherical linguistic numbers are proposed. The
proposed method has been applied in the investment selection problem
[27] T-spherical fuzzy graph are introduced along with the operations. T-spherical fuzzy
graphs have been presented for solving problem of service centers
[28] New operators are proposed for SFSs and T-SFSs. Using on these proposed
operators a novel method to multiple-attribute group decision-making are developed
[29] The definition of spherical fuzzy soft sets and some of its properties has been
introduced. An algorithm is proposed to solve the decision-making problem based on
adjustable soft discernibility matrix
[30] Dombi aggregation operators have been developed for SFSs. New approach to
decision-making problem based on the proposed Dombi aggregation operators
introduced
[31] By developing Z-fuzzy hypothesis tests, illustrative examples are presented with
sensitivity analysis
[32] They proposed distance aggregation operators for SFSs
[33] Spherical aggregation operators for spherical fuzzy numbers utilizing these strict
Archimedean t‐norm and T‐Conorm are proposed. Using these aggregation proposed
operators, a decision-making method is proposed for ranking alternatives
[34] The TOPSIS approach is proposed through covering-based spherical fuzzy rough set
models by means of SF b-neighborhoods
[35] The notion of linguistic SFS, which is the combination of linguistic fuzzy sets and
SFSs are proposed. New algorithm for the decision-making based on the defined
linguistic spherical fuzzy aggregation operators is proposed
[36] Some novel logarithmic operations of SFSs and the spherical fuzzy entropy method
are proposed
[37] Operators named the generalized -spherical fuzzy -weighted geometric and
arithmetic interaction functions are introduced
[38] SFSs are extended to the interval-valued SFSs. Corresponding operators, arithmetic
operations, aggregation operators are proposed for interval-valued SFSs. Original
score and accuracy functions are introduced. Interval-valued spherical fuzzy TOPSIS
is introduced
[39] T-spherical fuzzy generalized Maclaurin symmetric mean (GMSM) and T-spherical
fuzzy weighted GMSM operators have been proposed for T-SFSs. The proposed
method is used for R&D project selection problem of a company producing
toothpaste
[40] The notion of qRung picture fuzzy graphs and discuss regularity of these graphs are
introduced
[41] The concept of SFS and T-SFS is introduced as a generalization of fuzzy set, IFS and
PFS. Some operations of SFSs and T-SFSs along with spherical fuzzy relations are
defined
[42] New interactive averaging aggregation operators by assigning associate probabilities
for T-SFSs are proposed
(continued)
Paper Explanation
[43] A novel enhanced TOPSIS-based procedure for tackling multi attribute group
decision making issues under spherical fuzzy setting are proposed
[44] Spherical fuzzy TOPSIS method is applied to the site selection of photovoltaic power
station
[45] Spherical fuzzy CODAS method is applied to the warehouse location selection
problem
[46] A new divergence measure under T-SFS structure is proposed by utilizing the
advantages of Jensen-Shannon divergence, which is called TSFSJS distance. TSFSJS
distance not only satisfies the distance measurement axiom, but also can better
distinguish the difference between TSFSs than other distance measures
[47] Linguistic spherical fuzzy weighted averaging operator is proposed. Combination of
the operator with MABAC is tested on the evaluation of shared bicycles in China
[48] Multi attribute decision making method is proposed, using aggregation operators for
SFSs and score functions. Proposed method has been applied to debt collection
problem
[49] Spherical fuzzy VIKOR is proposed. The proposed method has been used to solve
the waste disposal site selection problem problem
[50] Spherical fuzzy MULTIMOORA is proposed. The proposed method is compared
with neutrosophic MULTIMOORA and intuitionistic fuzzy TOPSIS methods
[51] Spherical fuzzy (SF) QFD has been proposed. The proposed method has been used
in a linear delta robot technology design. A comparative analysis using SF-TOPSIS
is applied for competitive firms.
[52] Multi-attribute decision making method based on spherical fuzzy aggregation
operators is proposed. Spherical fuzzy aggregation operators have been successfully
applied to debt collection problem
[53] Spherical fuzzy AHP method is used to solve the problem of industrial robot
selection
[54] 2-tuple spherical fuzzy linguistic aggregation operators: 2-tuple spherical fuzzy
linguistic weighted geometric, 2-tuple spherical fuzzy linguistic weighted geometric
average, 2-tuple spherical fuzzy linguistic ordered weighted geometric, and 2-tuple
spherical fuzzy linguistic hybrid geometric operators are proposed. Then, these
operators are used to solve the 2-tuple spherical fuzzy linguistic multi attribute
decision making problem
[55] New correlation coefficients for TSFSs are proposed, and their generalizations are
proved using some examples and remarks
[56] Fuzzy AHP is extended to spherical fuzzy AHP. It is compared with neutrosophic
fuzzy AHP by using spherical fuzzy AHP in solving site selection of wind power
farm problem
Fig. 1. Distribution of publications based on different journals/proceedings (43 papers).
In total, the number of articles is 6 times greater than the number of proceeding
papers. While 37 of 43 papers are articles, rest of them is proceedings papers. The
distribution of the journals is presented in Fig. 1. 6 of 43 papers are conference studies
and all of them are published in Advances in Intelligent Systems and Computing.
18 journals have contributed to the publication of 43 articles. Among the journals,
“Journal of Intelligent and Fuzzy Systems” seems dominant, representing 21% of all
published articles. Considered papers are also investigated according to their subject
area. 9 different subject areas with the number of contributing papers are shown in
Fig. 2. As expected, the majority of the papers (81.4%) belong to the computer science
area, followed by mathematics and engineering. Table 2 gives information about top 10
contributors in SFS topic with their affiliation addresses. While Kahraman, C. and
Mahmood, T. are at the top with 12 papers, they are followed by Abdullah, S. and
Ashraf, S. with 10 papers. The distribution of affiliation addresses of the authors are
also illustrated in Fig. 3. As it can be seen, majority of the studies belongs to Turkey
and Pakistan. Rest of them spreads to east of the world.
Fig. 2. Distribution of publications based on subject area.

Table 2. Top 10 authors with their affiliations.

Author Affiliation Country # of
papers
Kahraman, C. Istanbul Technical University Turkey 12
Mahmood, T. International Islamic University Islamabad Pakistan 12
Abdullah, S. Abdul Wali Khan University Mardan Pakistan 10
Ashraf, S. Abdul Wali Khan University Mardan Pakistan 10
Kutlu Gundogdu, Istanbul Technical University Turkey 6
F.
Ullah, K. International Islamic University Islamabad Pakistan 6
Jan, N. International Islamic University Islamabad Pakistan 5
Garg, H. Thapar Institute of Engineering & India 4
Technology
Munir, M. International Islamic University Islamabad Pakistan 4
Liu, P. Shandong University of Finance and China 3
Economics
In addition to descriptive analysis above, the titles, abstract and keywords of 43

reviewed papers are analzyed. 160 different keywords are observed in 43 papers, and
they are used 427 times in total. According to Table 3, the most frequently used 15
keywords share 52.46% of total keywords. While “spheres” term is the most used
keyword, “fuzzy sets”, “decision making” and “spherical fuzzy sets” comes later.
Fig. 3. Spatial distribution of affiliation addresses.
Table 3. Most frequently used keywords.

# Keywords % # Keywords %
1 Spheres 7.03 9 Membership functions 1.87
2 Fuzzy sets 6.79 10 Neutrosophic sets 1.64
3 Decision making 6.56 11 Arithmetic operations 1.41
4 Spherical fuzzy sets 6.56 12 Comparative analysis 1.41
5 Aggregation operator 4.92 13 Multi criteria decision making 1.41
6 Intuitionistic fuzzy sets 3.98 14 Decision-making problem 1.17
7 Mathematical operators 3.75 15 T-spherical fuzzy sets 1.17
8 Multi attribute decision making 2.81 Total 52.46
In addition to the keywords analysis above, 43 papers are clustered in terms of the
titles and abstracts. To do so, the sklearn library of Python is used. The results are given
in Table 4. The algorithm classifies 43 papers within 5 clusters. To determine the
number of clusters, Silhouette analysis is applied. While the second and third columns
give the classification based on titles, the rest of columns provide the same data based
on abstracts. Due to the differences on the scanned words in titles and abstracts, the
assigned papers to each cluster are different. According to Table 4, same or similar
authors take part in the same clusters (e.g. [15, 23, 49, 50]). When the titles of men-
tioned studies are investigated, it is clear to see that all of them are related with the
extension of multi-criteria decision making with SFSs.
Table 4. Clustering of papers based on titles and abstracts.

Cluster Clustering based on titles Clustering based on abstracts
#1 [15, 23, 49, 50] [51]
#2 [17, 18, 24, 26, 28, 34, 35, 37, 39, 42, [14, 15, 20, 23, 31, 38, 44, 45, 48–50,
47, 54, 55] 52, 53, 56]
#3 [25, 29–31, 33, 36, 41, 48, 56] [16, 19, 21, 27, 40, 41, 46, 55]
#4 [14, 20, 27, 38, 43, 51–53] [22, 25, 30, 32, 39, 43, 47]
#5 [16, 19, 21, 22, 32, 40, 45, 46] [17, 18, 24, 29, 33–37, 42, 54]
3 Conclusion
In this study, a systematic review methodology was followed and academic articles
containing SFS were reviewed. 43 papers were taken into consideration. Results may
be subject to limitations of the sources used in this paper. Following gaps are sum-
marized as: (i) aggregation operators and distance measurement equations for interval-
valued SFSs should be developed; (ii) existing MCDM methods such as CODAS,
ARAS should be extended using SFS; (iii) interval-valued spherical fuzzy – MCDM
methods (e.g. VIKOR) should be proposed; (iv) SFSs based approaches should be
applied to engineering problems; (v) spherical fuzzy preference relations for decision-
making should be developed for intuitionistic and Pythagorean fuzzy sets; (vi) interval-
valued spherical fuzzy QFD should be developed; (vii) spherical fuzzy preference
relations and inter-valued SFSs should be used in SF-AHP.
References
1. Zadeh, L.A.: Fuzzy sets. Inf. Control 8(3), 338–353 (1965)
2. Mendel, J.M.: Advances in type-2 fuzzy sets and systems. Inf. Sci. 177(1), 84–110 (2007)
3. Karnik, N.N., Mendel, J.M.: Operations on type-2 fuzzy sets. Fuzzy Sets Syst. 122, 327–348
(2001)
4. Zadeh, L.A.: The concept of a linguistic variable and its applications to approximate
reasoning. Part I. Inf. Sci. 8, 199–249 (1975)
5. Atanassov, K.T.: Intuitionistic fuzzy sets. Fuzzy Sets Syst. 20(1), 87–96 (1986)
6. Atanassov, K.T.: More on intuitionistic fuzzy sets. Fuzzy Sets Syst. 33(1), 37–45 (1989)
7. Yager, R.R.: On the theory of bags. Int. J. Gen. Syst. 13(1), 23–37 (1986)
8. Smarandache, F.: A Unifying Field in Logics: Neutrosophy, Neutrosophic Probability, Set
and Logic. American Research Press, Rehoboth (1999)
9. Garibaldi, J.M., Ozen, T.: Uncertain fuzzy reasoning: a case study in modeling expert
decision making. IEEE Trans. Fuzzy Syst. 15(1), 16–30 (2007)
10. Torra, V.: Hesitant fuzzy sets. Int. J. Intell. Syst. 25(6), 529–539 (2010)
11. Yager, R.R.: Pythagorean fuzzy subsets. In: Proceedings of the Joint IFSA World
Congress NAFIPS Annual Meeting, pp. 57–61 (2013)
12. Cuong, B.C.: Picture fuzzy sets. J. Comput. Sci. Cybern. 30(4), 409–420 (2014)
13. Yager, R.R.: Generalized orthopair fuzzy sets. IEEE Trans. Fuzzy Syst. 25(5), 1222–1230
(2017)
14. Kutlu Gundogdu, F., Kahraman, C.: Spherical fuzzy sets and spherical fuzzy TOPSIS
method. J. Intell. Fuzzy Syst. 36(1), 337–352 (2019)
15. Kutlu Gundogdu, F., Kahraman, C.: Extension of WASPAS with spherical fuzzy sets.
Informatica 30(2), 269–292 (2019)
16. Ullah, K., Mahmood, T., Jan, N.: Similarity measures for T-spherical fuzzy sets with
applications in pattern recognition. Symmetry 10(6), 193 (2018)
17. Garg, H., Munir, M., Ullah, K., Mahmood, T., Jan, N.: Algorithm for T-spherical fuzzy
multi-attribute decision making based on improved interactive aggregation operators.
Symmetry 10(12), 670 (2018)
18. Ashraf, S., Abdullah, S., Mahmood, T., Ghani, F., Mahmood, T.: Spherical fuzzy sets and
their applications in multi-attribute decision making problems. J. Intell. Fuzzy Syst. 36(3),
2829–2844 (2019)
19. Rafiq, M., Ashraf, S., Abdullah, S., Mahmood, T., Muhammad, S.: The cosine similarity
measures of spherical fuzzy sets and their applications in decision making. J. Intell. Fuzzy
Syst. 36(6), 6059–6073 (2019)
20. Kutlu Gundogdu, F., Kahraman, C.: A novel VIKOR method using spherical fuzzy sets and
its application to warehouse site selection. J. Intell. Fuzzy Syst. 37(1), 1197–1211 (2019)
21. Wei, G., Wang, J., Lu, M., Wu, J., Wei, C.: Similarity measures of spherical fuzzy sets based
on cosine function and their applications. IEEE Access 7(8882324), 159069–159080 (2019)
22. Ashraf, S., Abdullah, S., Aslam, M., Qiyas, M., Kutbi, M.A.: Spherical fuzzy sets and its
representation of spherical fuzzy t-norms and t-conorms. J. Intell. Fuzzy Syst. 36(6), 6089–
6102 (2019)
23. Kutlu Gundogdu, F., Kahraman, C.: Extension of codas with spherical fuzzy sets. J. Mult.
Valued Logic Soft Comput. 33(4–5), 481–505 (2019)
24. Ullah, K., Hassan, N., Mahmood, T., Jan, N., Hassan, M.: Evaluation of investment policy
based on multi-attribute decision-making using interval valued T-spherical fuzzy aggrega-
tion operators. Symmetry 11(3), 357 (2019)
25. Donyatalab, Y., Farrokhizadeh, E., Garmroodi, S.D.S., Shishavan, S.A.S.: Harmonic mean
aggregation operators in spherical fuzzy environment and their group decision making
applications. J. Mult. Valued Logic Soft Comput. 33(6), 565–592 (2019)
26. Cao, H., Zhang, R., Wang, J.: Some spherical linguistic Muirhead mean operators with their
application to multi-attribute decision making. J. Intell. Fuzzy Syst. 37(6), 8097–8111
(2019)
27. Guleria, A., Bajaj, R.K.: T-spherical fuzzy graphs: operations and applications in various
selection processes. Arab. J. Sci. Eng. 45, 2177–2193 (2020)
28. Liu, P., Khan, Q., Mahmood, T., Hassan, N.: T-spherical fuzzy power Muirhead mean
operator based on novel operational laws and their application in multi-attribute group
decision making. IEEE Access 7(8631030), 22613–22632 (2019)
29. Perveen, F., Sunil, J.J., Babitha, K.V., Garg, H.: Spherical fuzzy soft sets and its applications
in decision-making problems. J. Intell. Fuzzy Syst. 37(6), 8237–8250 (2019)
30. Ashraf, S., Abdullah, S., Mahmood, T.: Spherical fuzzy Dombi aggregation operators and
their application in group decision making problems. J. Ambient Intell. Humanized Comput.
19, 1–19 (2019). https://doi.org/10.1007/s12652-019-01333-y
31. Haktanir, E., Kahraman, C.: Z-fuzzy hypothesis testing in statistical decision making.
J. Intell. Fuzzy Syst. 37(5), 6545–6555 (2019)
32. Ashraf, S., Abdullah, S., Abdullah, L.: Child development influence environmental factors
determined using spherical fuzzy distance measures. Mathematics 7(8), 661 (2019)
33. Ashraf, S., Abdullah, S.: Spherical aggregation operators and their application in multi-
attribute group decision-making. Int. J. Intell. Syst. 34(3), 493–523 (2019)
34. Zeng, S., Hussain, A., Mahmood, T., Ashraf, S., Munir, M.: Covering-based spherical fuzzy
rough set model hybrid with TOPSIS for multi-attribute decision-making. Symmetry 11(4),
547 (2019)
35. Jin, H., Ashraf, S., Abdullah, S., Bano, M., Zeng, S.: Linguistic spherical fuzzy aggregation
operators and their applications in multi-attribute decision making problems. Mathematics
7(5), 413 (2019)
36. Jin, Y., Ashraf, S., Abdullah, S.: Spherical fuzzy logarithmic aggregation operators based on
entropy and their application in decision support systems. Entropy 21(7), 628 (2019)
37. Quek, S.G., Selvachandran, G., Munir, M., Kumar, R., Priyadarshini, I.: Multi-attribute
multi-perception decision-making based on generalized T-spherical fuzzy weighted aggre-
gation operators on neutrosophic sets. Mathematics 7(9), 780 (2019)
38. Kutlu Gundogdu, F., Kahraman, C.: A novel fuzzy TOPSIS method using emerging interval-
valued spherical fuzzy sets. Eng. Appl. Artif. Intell. 85, 307–323 (2019)
39. Liu, P., Zhu, B., Wang, P.: A multi-attribute decision-making approach based on spherical
fuzzy sets for Yunnan Baiyao’s R&D project selection problem. Int. J. Fuzzy Syst. 21(7),
2168–2191 (2019)
40. Akram, M., Habib, A.: q-rung picture fuzzy graphs: a creative view on regularity with
applications. J. Appl. Math. Comput. 61(1–2), 235–280 (2019)
41. Mahmood, T., Ullah, K., Khan, Q., Jan, N.: An approach toward decision-making and
medical diagnosis problems using the concept of spherical fuzzy sets. Neural Comput. Appl.
31(11), 7041–7053 (2019)
42. Zeng, S., Garg, H., Munir, M., Mahmood, T., Hussain, A.: A multi-attribute decision making
process with immediate probabilistic interactive averaging aggregation operators of T-
spherical fuzzy sets and its application in the selection of solar cells. Energies 12(23), 4436
(2019)
43. Barukab, O., Abdullah, S., Ashraf, S., Arif, M., Khan, S.A.: A new approach to fuzzy
TOPSIS method based on entropy measure under spherical fuzzy information. Entropy
21(12), 1231 (2019)
44. Kutlu Gundogdu, F.: Principals of spherical fuzzy sets. Adv. Intell. Syst. Comput. 1029, 15–
23 (2020)
45. Kutlu Gundogdu, F., Kahraman, C.: Spherical fuzzy sets and decision making applications.
Adv. Intell. Syst. Comput. 1029, 979–987 (2020)
46. Wu, M.-Q., Chen, T.-Y., Fan, J.-P.: Divergence measure of t-spherical fuzzy sets and its
applications in pattern recognition. IEEE Access 8(8946628), 10208–10221 (2020)
47. Liu, P., Zhu, B., Wang, P., Shen, M.: An approach based on linguistic spherical fuzzy sets
for public evaluation of shared bicycles in China. Eng. Appl. Artif. Intell. 87, 103295 (2020)
48. Kahraman, C., Onar, S.C., Oztaysi, B.: Performance measurement of debt collection firms
using spherical fuzzy aggregation operators. Adv. Intell. Syst. Comput. 1029, 506–514
(2020)
49. Kutlu Gundogdu, F., Kahraman, C., Karasan, A.: Spherical fuzzy VIKOR method and its
application to waste management. Adv. Intell. Syst. Comput. 1029, 997–1005 (2020)
50. Kutlu Gundogdu, F.: A spherical fuzzy extension of MULTIMOORA method. J. Intell.
Fuzzy Syst. 38(1), 963–978 (2018)
51. Kutlu Gundogdu, F., Kahraman, C.: A novel spherical fuzzy QFD method and its application
to the linear delta robot technology development. Eng. Appl. Artif. Intell. 87, 103348 (2020)
52. Bolturk, E.: AS/RS technology selection using spherical fuzzy TOPSIS and neutrosophic
TOPSIS. Adv. Intell. Syst. Comput. 1029, 969–976 (2020)
53. Kutlu Gundogdu, F., Kahraman, C.: Spherical fuzzy analytic hierarchy process (AHP) and
its application to industrial robot selection. Adv. Intell. Syst. Comput. 1029, 988–996 (2020)
54. Abdullah, S., Barukab, O., Qiyas, M., Arif, M., Khan, S.A.: Analysis of decision support
system based on 2-tuple spherical fuzzy linguistic aggregation information. Appl. Sci.
(Switzerland) 10(1), 276 (2020)
55. Ullah, K., Garg, H., Mahmood, T., Jan, N., Ali, Z.: Correlation coefficients for T-spherical
fuzzy sets and their applications in clustering and multi-attribute decision making. Soft
Comput. 24(3), 1647–1659 (2020)
56. Kutlu Gundogdu, F., Kahraman, C.: A novel spherical fuzzy analytic hierarchy process and
its renewable energy application. Soft Comput. 24(6), 4607–4621 (2020)
Picture Fuzzy Sets and Spherical Fuzzy
Sets
Picture Fuzzy Linear Assignment Method
and Its Application to Selection of Pest House
Location
Fatma Kutlu Gundogdu(&)
Industrial Engineering Department, Turkish Air Force Academy,

National Defence University, 34149 Istanbul, Turkey
[email protected]
Abstract. The theory of picture fuzzy sets is useful for handling uncertainty in
multiple attribute decision making problems by considering membership, non-
membership and indeterminacy degrees independently for each element. In this
paper, by extending the classical linear assignment method, we propose a novel
method which is called picture fuzzy linear assignment method (PF-LAM) for
solving multiple criteria group decision-making problems with picture fuzzy
sets. A ranking procedure consisting of aggregation functions, score functions,
accuracy functions, and weighted rank frequency and a binary mathematical
model are presented to determine the priority order of various alternatives. The
applicability and validity of the proposed method is shown through the selection
of pest house locations. The proposed method helps managers to find the best
location to construct the pest house based on the determined criteria.
Keywords: Picture fuzzy sets Linear assignment model Multiple criteria

decision-making model Optimization Pest house location selection
1 Introduction
Fuzzy Sets theory, developed by Zadeh [1], is a useful and appropriate approach in
order to deal with imprecise and uncertain information in vague situations. After the
introduction of fuzzy sets, they have been very popular in almost all branches of
science [2]. Many researchers [2–14] have introduced many extensions of ordinary
fuzzy sets in the literature. These extensions have been utilized by numerous
researchers in recent years in the solution of multi-attribute decision-making problems
[2]. One of the latest extensions is Picture fuzzy sets (PFS). Picture Fuzzy Sets
(PFS) were developed by Cuong [14] and it is a direct extension of intuitionistic fuzzy
sets (IFS) that can model uncertainty using membership degree, non-membership
degree, and hesitant degree independently.
Wang and Li [15] introduced the theory of picture hesitant fuzzy set based on the
picture fuzzy sets and the hesitant fuzzy set. Sarwar Sindhu et al. [16] proposed a linear
programming model in order to find exact weights and construct a modified distance
based on similarity measure under picture fuzzy environment. Liang et al. [17] pre-
sented a MCDM method which is a combination of TODIM method with the
102 F. K. Gundogdu
ELECTRE method in a picture fuzzy environment. Thao [18] developed the entropy
measure for PFS and proposed the similarity measures for MCDM problems in order to
select suppliers. Tian et al. [19] proposed a picture fuzzy MCDM method and intro-
duced weighted picture fuzzy power Choquet ordered geometric operator and a
weighted picture fuzzy power Shapley Choquet ordered geometric operator.
The linear assignment method (LAM) was proposed by Bernardo and Blin [20],
inspiring from assignment problem in linear programming for multi-attribute decision-
making [21]. The basic idea of the LAM is that the combination of the criteria-wise
rankings into an overall preference ranking that produces an optimal compromise
among the several component rankings. Developing an extended linear assignment
method to solve multi-criteria decision-making (MCDM) problems under Pythagorean
fuzzy environment was the aim of [22]. In addition, Liang et al. [23] developed the
linear assignment method for interval-valued Pythagorean fuzzy sets. By extending the
traditional linear assignment method, Chen [24] developed an efficient method for
solving MCDM problems in the interval-valued intuitionistic fuzzy environment.
To the best of our knowledge, there is no research about extension and application
of linear assignment method in picture fuzzy environment. Therefore, the aim of this
paper is to develop a novel multi-attribute decision-making method based on linear
assignment approach with picture fuzzy sets and also show the useful application to site
selection of pest house. The proposed algorithm has the following contributions. First,
judgment values are given as picture linguistic terms, which can consider the hesitancy
degree of decision makers’ comments about alternatives and criteria. Second, linear
assignment method has been employed to rank alternatives to avoid the effect of
subjectivity.
In Sect. 2, the definitions of Picture fuzzy sets are presented. In Sect. 3, the Pic-
ture fuzzy linear assignment method are detailed step by step. In Sect. 4, an application
is given and in Sect. 5, the conclusion is given.
2 Picture Fuzzy Sets: Preliminaries
There are some definitions about PFS is given as follows with related equations.
~ p of the universe of discourse U is given by;
Definition 2.1: A PFS on a A
nD o
~p ¼
A u; ðlA~ p ðuÞ; IA~ p ðuÞ; vA~ p ðuÞÞu 2 U ð1Þ
where
lA~ p ðuÞ : U ! ½0; 1; IA~ p ðuÞ : U ! ½0; 1; mA~ p ðuÞ : U ! ½0; 1
and
0 lA~ p ðuÞ þ IA~ p ðuÞ þ vA~ p ðuÞ 1 8u 2 U ð2Þ

Picture Fuzzy Linear Assignment Method 103
Then, for each u, the numbers lA~ S ðuÞ; mA~ S ðuÞ and IA~ S ðuÞ are the degree of mem-
bership, non-membership and indeterminacy of u to A ~ S , respectively. v ¼ 1

lA~ p ðuÞ þ mA~ p ðuÞ þ IA~ p ðuÞ is called as a refusal degree [25].
Definition 2.2: Basic operators for Single-valued picture fuzzy sets;

n o
~p B
A ~ p ¼ l ~ þ lB~ l ~ lB~ ; I ~ IB~ ; v ~ vB~ ð3Þ
Ap p Ap p Ap p Ap p
n o
~p B
A ~ p ¼ l ~ lB~ ; I ~ þ IB~ I ~ IB~ ; v ~ þ vB~ v ~ vB~ ð4Þ
Ap p Ap p Ap p Ap p Ap p
k
~p ¼
kA 1 1 lA~ p ; IA~kp ; vkA~ p for k [ 0 ð5Þ
k k
~k ¼
A lkA~ p ; 1 1 IA~ p ; 1 1 vA~ p for k [ 0 ð6Þ
p
Definition 2.3: Single-valued Picture Fuzzy Weighted Averaging operator (PFWA)

P
n
with respect to, w ¼ ðw1 ; w2 ; . . .. . .; wn Þ; wi 2 ½0; 1; wi ¼ 1, is defined as;
i¼1
~ 1 ; . . .. . .:; A
PFWAw ðA ~ ~ ~ ~n
n Þ ¼nw1 A1 þ w2 A2 þn . . .. . . þ wn A
Q wi Q w i Q w i
n ð7Þ
¼ 1 ð1 lA~ ii Þ ; IA~ ; vA~
ii ii
i¼1 i¼1 i¼1
Definition 2.4: Score functions and Accuracy functions of sorting picture fuzzy
numbers are defined by;

1 .
~p ¼
Score A 1 þ 2lA~ p vA~ p IA~ p 2 ð8Þ
2

Accuracy A~ p ¼ l~ þ v~ þ I~ ð9Þ
Ap Ap Ap
~ p \B
Note that: A ~ p if and only if
~ p Þ\ScoreðB
(i) ScoreðA ~ p Þ or
~ ~ p Þ\AccuracyðB
~ p Þ and AccuracyðA
(ii) ScoreðAp Þ ¼ ScoreðB ~pÞ
104 F. K. Gundogdu
3 Picture Fuzzy Linear Assignment Method
The classical linear assignment method is extended to picture fuzzy linear assignment
model. The proposed PF-LAM is composed of several steps as given in follows.
Table 1 presents the linguistic terms and their corresponding picture fuzzy numbers.
Decision matrix whose elements show the judgments values of all alternatives with
respect to each criterion under picture fuzzy environment. Consider a group of k de-
cision makers, D ¼ fD1 ; D2 ; . . .; Dk g participated in a group decision making problem,
let X ¼ fx1 ; x2 ; . . .. . .xm gðm 2Þ be a discrete set of m feasible alternatives and C ¼
fC1 ; C2 ; . . .. . .Cn g be a finite set of n criteria and wj ¼ fw1 ; w2 ; . . .. . .wn g be the weight
P
n
vector of n criteria which proves 0 wj 1 and wj ¼ 1.
j¼1
Table 1. Picture fuzzy linguistic terms [25]

Linguistic terms ðl; p; vÞ
Very High Importance (VHI) (0.9, 0.0, 0.05)
High Importance (HI) (0.75, 0.05, 0.1)
Slightly More Importance (SMI) (0.6, 0.0, 0.3)
Equally Importance (EI) (0.5, 0.1, 0.4)
Slightly Low Importance (SLI) (0.3, 0.0, 0.6)
Low Importance (LI) (0.25, 0.05, 0.6)
Very Low Importance (VLI) (0.1, 0.0, 0.85)
Step 1: Collect the decision-makers’ evaluations for the alternatives and criteria based
on Table 1.
Step 2. Aggregate the individual decision matrices based on PFWA operator as given
in Eq. (7).
Step 3. Compute the elements of scored decision matrix by utilizing the picture fuzzy
score function (Eq. 8).
Step 4. Establish the rank frequency non-negative matrix bjk with elements that rep-
resent the frequency that Am is ranked as the mth criterion-wise ranking.
Step 5. Calculate and establish the weighted rank frequency matrix k, where the kik
measures the contribution of Am to the overall ranking. Note that each entry kik of the
weighted rank frequency matrix P is a measure of the concordance among all criteria
in ranking the mth alternative kth.
kik ¼ wi1 wi2 . . . wibmm ð10Þ

Step 6. Define the permutation matrix P as a square ðm mÞ matrix and set up the
following linear assignment model according to the Pik value. The linear assignment
model can be written in the following linear programming format:
Xm Xm
Max i¼1 k¼1
kik :Pik
Xm
s:t: k¼1
Pik ¼ 1; 8i ¼ 1; 2; . . .; m;
Xm
i¼1
Pik ¼ 1; 8k ¼ 1; 2; . . .; m;
Pik ¼ 0 or 1 for all i and k
Step 7. Solve the linear assignment model, and obtain the optimal permutation matrix
2 3
X1
6 X2 7
P
for all i and k. Calculate the multiplication of matrix P
:X ¼ P
:6 7
4 ... 5 and obtain
Xm
the optimal order of alternatives.
4 An Application to Pest House Location Selection
In this section, a numerical example is presented to illustrate feasibility and practical

advantages of new proposed method. Nowadays, Coronavirus Disease (COVID-19)
has emerged as a global problem since spread of the disease to March 2020, more than
one million people have been infected by COVID-19 virus. The disease is quickly
spreading between people during close contact. In the future, existing hospitals will not
be enough for the patients who are suffering from Coronavirus or similar dangerous
viruses. Each country has to establish pet houses. The aim of this problem is to select
the best site location in order to establish pest house in Turkey. The mostly preferred
five locations (X1: Ankara, X2: Izmir, X3: Istanbul-Atatürk Airport, X4: Istanbul-
Sancaktepe, and X5: Bursa) are evaluated as alternatives. Four criteria have been
determined in order to evaluate these alternatives. Criteria are logistic support oppor-
tunities to settlements (C1), economical situations (C2), population density (C3), and
proximity to settlements (C4). Three decision makers who have different significance
levels such as 0.3, 0.5, 0.2, are going to evaluate the above five possible alternatives
according to four criteria based on picture linguistic terms as presented in Table 2.
Aggregate the decision matrices using Eq. (7) into a single aggregated decision
matrix as given in Table 3. The weight of each criterion is aggregated based on Eq. (7)
and so that wj ¼ ½0:324; 0:091; 0:359; 0:226.
106 F. K. Gundogdu
Table 2. Assessments of decision-makers

DM1 C1 C2 C3 C4 DM2 C1 C2 C3 C4 DM3 C1 C2 C3 C4
X1 HI VHI VLI VHI X1 SMI HI EI HI X1 SMI SLI VHI SLI
X2 SLI VHI HI SLI X2 SLI SLI LI HI X2 HI LI VLI LI
X3 VLI HI EI SMI X3 HI HI SMI SMI X3 HI SMI HI HI
X4 SMI HI LI HI X4 HI EI LI EI X4 SMI VHI VLI VHI
X5 HI SLI HI SMI X5 HI SLI EI EI X5 SMI HI LI VLI
Table 3. Aggregated decision matrix

C1 C2 C3 C4
X1 (0.65, 0.00, 0.22) (0.77, 0.00, 0.12) (0.57, 0.00, 0.33) (0.77, 0.00, 0.12)
X2 (0.43, 0.00, 0.42) (0.60, 0.00, 0.28) (0.44, 0.00, 0.38) (0.58, 0.00, 0.24)
X3 (0.63, 0.00, 0.19) (0.73, 0.00, 0.12) (0.61, 0.00, 0.26) (0.64, 0.00, 0.24)
X4 (0.68, 0.00, 0.17) (0.71, 0.00, 0.17) (0.22, 0.00, 0.64) (0.71, 0.00, 0.17)
X5 (0.73, 0.00, 0.12) (0.43, 0.00, 0.42) (0.56, 0.07, 0.29) (0.47, 0.00, 0.43)
Calculate the score value of each alternative based on each criterion using Eq. (8).
The results are shown in Table 4.
Table 4. The score value of each alternative

C1 C2 C3 C4
X1 1.045 1.209 0.902 1.209
X2 0.721 0.962 0.753 0.953
X3 1.038 1.163 0.979 1.015
X4 1.097 1.119 0.401 1.119
X5 1.163 0.721 0.899 0.761
bij are computed and established the weighted rank frequency matrix kik , as shown
in Table 5. For example, consider k12 in the following: k12 ¼ wC2 þ wC4 ¼
0:091 þ 0:226 ¼ 0:317.
Table 5. Weighted rank frequency matrix kik

1st 2nd 3rd 4th 5th
X1 0,3170 0,0908 0,3244 0,0000 0,0000
X2 0,0000 0,0000 0,0000 0,6756 0,3244
X3 0,3586 0,0908 0,2262 0,3244 0,0000
X4 0,0000 0,5506 0,0908 0,0000 0,3586
X5 0,3244 0,0000 0,3586 0,0000 0,3170
The linear assignment model is constructed as follows. The objective function of

this binary mathematical model tries to maximize the sum of the weights of alternatives
by choosing the optimal order of them.
Max Z ¼ 0:3170P11 þ 0:0908P12 þ 0:3244P13 þ 0:6756P24 þ 0:3244P25 þ 0:3586P31

þ 0:0908P32 þ 0:2262P33 þ 0:3244P34 þ 0:5506P42 þ 0:0908P43 þ 0:3586P45
þ 0:3244P51 þ 0:3586P53 þ 0:3170P55
s.t.
P11 þ P12 þ P13 þ P14 þ P15 ¼ 1
P21 þ P22 þ P23 þ P24 þ P25 ¼ 1
P31 þ P32 þ P33 þ P34 þ P35 ¼ 1
P41 þ P42 þ P43 þ P44 þ P45 ¼ 1
P51 þ P52 þ P53 þ P54 þ P55 ¼ 1
P11 þ P21 þ P31 þ P41 þ P51 ¼ 1
P12 þ P22 þ P32 þ P42 þ P52 ¼ 1
P13 þ P23 þ P33 þ P43 þ P53 ¼ 1
P14 þ P24 þ P34 þ P44 þ P54 ¼ 1
P15 þ P25 þ P35 þ P45 þ P55 ¼ 1
Pik ¼ 0 or 1 for i ¼ 1; 2; 3; 4; 5; k ¼ 1; 2; 3; 4; 5.
The proposed mathematical model is solved by using GAMS 24.1.3 software and
the results are obtained. After solving the model, the results are P13 ¼ 1, P24 ¼ 1,
P31 ¼ 1, P42 ¼ 1 and P55 ¼ 1. The value of objective function is z ¼ 2:226. The
optimal ranking order of the five alternatives is X3 [ X4 [ X1 [ X2 [ X5 . Best loca-
tion for pest house is Istanbul-Ataturk Airport.
5 Conclusion
In the recent years, picture fuzzy sets have been very widespread in almost all branches.
Picture fuzzy sets are another extension of the ordinary fuzzy sets. PFS should satisfy
the condition that the sum of membership degree and non-membership degree and
hesitancy degree should be equal to or less than one. In this study, the classical linear
assignment model is extended to picture fuzzy linear assignment model and the novel
method is applied to site selection problem for pest house. It has been successfully
solved by picture fuzzy linear assignment model. The proposed PF-LAM method is
performed to get the optimal preference ranking of the alternatives according to a set of
criteria-wise rankings within the context of PFS.
For future studies, the proposed method can be applied to several decision support
systems and the illustrative example can be extended by real data.
108 F. K. Gundogdu
References
2. Kutlu Gündoğdu, F., Kahraman, C.: Spherical fuzzy sets and spherical fuzzy TOPSIS
3. Zadeh, L.A.: The concept of a linguistic variable and its application to approximate
reasoning. Inf. Sci. 8, 199–249 (1975)
5. Atanassov, K.T.: Geometrical interpretation of the elements of the intuitionistic fuzzy objects
Preprint IM-MFAIS (1989) 1-89. Sofia. Reprinted. Int. J. Bioautomation 20(S1), 27–42
(2016)
6. Garibaldi, J.M., Ozen, T.: Uncertain fuzzy reasoning: a case study in modelling expert
7. Grattan-Guinness, I.: Fuzzy membership mapped onto interval and many-valued quantities.
Zeitschrift fur mathematische Logik und Grundladen der Mathematik 22(1), 149–160 (1976)
8. Jahn, K.U.: Intervall-wertige Mengen. Mathematische Nachrichten 68(1), 115–132 (1975)
9. Sambuc, R.: Function U-Flous. University of Marseille, Application a l’aide au Diagnostic
en Pathologie Thyroidienne (1975)
10. Smarandache, F.: Neutrosophy: Neutrosophic probability, set, and logic: Analytic synthesis
& synthetic analysis (1998)
12. Yager, RR.: Pythagorean fuzzy subsets. In: Joint IFSA World Congress and NAFIPS Annual
Meeting, Edmonton, pp. 57–61 (2013)
13. Yager, R.: On the theory of bags. Int. J. Gen. Syst. 13(1), 23–37 (1986)
14. Cường, B.C.: Picture fuzzy sets. J. Comput. Sci. Cybern. 30(4), 409 (2014)
15. Wang, R., Li, Y.: Picture hesitant fuzzy set and its application to multiple criteria decision-
making. Symmetry 10(7), 295 (2018)
16. Sarwar Sindhu, M., Rashid, T., Kashif, A.: Modeling of linear programming and extended
TOPSIS in decision making problem under the framework of picture fuzzy sets. PLoS ONE
14(8), 14 (2019)
17. Liang, W., Dai, B., Zhao, G., Wu, H.: Performance evaluation of green mine using a
combined multi-criteria decision-making method with picture fuzzy information. IEEE
Access 7, 174139–174154 (2019)
18. Thao, N.X.: Similarity measures of picture fuzzy sets based on entropy and their application in
MCDM. Pattern Anal. Appl. 11, 1–11 (2019). https://doi.org/10.1007/s10044-019-00861-9
19. Tian, C., Peng, J., Zhang, S., Zhang, W., Wang, J.: Weighted picture fuzzy aggregation
operators and their applications to multi-criteria decision-making problems. Comput. Industr.
Eng. 137, 106037 (2019)
20. Bernardo, J.J., Blin, J.M.: A programming model of consumer choice among multi-attributed
brands. J. Consum. Res. 4(2), 111 (1977)
21. Razavi Hajiagha, S.H., Shahbazi, M., Amoozad Mahdiraji, H., Panahian, H.: A bi-objective
score variance based linear assignment method for group decision making with hesitant
fuzzy linguistic term sets. Technol. Econ. Dev. Econ. 24(3), 1125–1148 (2018)
22. Liang, D., Darko, A.P., Xu, Z., Zhang, Y.: Partitioned fuzzy measure-based linear
assignment method for Pythagorean fuzzy multi-criteria decision-making with a new
likelihood. J. Oper. Res. Soc. 15, 1–15 (2019)
23. Liang, D., Darko, A.P., Xu, Z., Quan, W.: The linear assignment method for multicriteria
group decision making based on interval-valued Pythagorean fuzzy Bonferroni mean. Int.
J. Intell. Syst. 33(11), 2101–2138 (2018)
24. Chen, T.Y.: The extended linear assignment method for multiple criteria decision analysis
based on interval-valued intuitionistic fuzzy sets. Appl. Math. Model. 38(7–8), 2101–2117
(2014)
25. Meksavang, P., Shi, H., Lin, S.M., Li, H.C.: An extended picture fuzzy VIKOR approach for
sustainable supplier management and its application in the beef industry. Symmetry 11(4),
468 (2019)
Simple Additive Weighting and Weighted
Product Methods Using Picture Fuzzy Sets
Fatma Kutlu Gundoğdu1(&) and Eda Bolturk2

1
Industrial Engineering Department, Turkish Air Force Academy,
National Defence University, 34149 Istanbul, Turkey
[email protected]
2
Department of Industrial Engineering, Istanbul Technical University,
34367 Macka, Istanbul, Turkey
Abstract. Multi criteria decision-making (MCDM) problem solutions have

been obtained with fuzzy sets for many years. In this manner, fuzzy sets and the
other extensions of fuzzy sets have been used in MCDM problems. Picture fuzzy
sets (PFS) is an effective extension of fuzzy sets in order to obtain results for
MCDM problems in uncertain environments. In this study, Simple Additive
Weighting (SAW) and Weighted Product Method (WPM) methods are extended
with Picture fuzzy sets (PFS). These methods are applied to an application and
future suggestions are shared in conclusion.
Keywords: Picture fuzzy sets SAW WPM MCDM
1 Introduction
PFS is proposed by Cuong [1] which is an extension of fuzzy sets and intuitionistic
fuzzy sets. PFS involve three membership function: positive membership, neutral
membership and negative membership. PFS is a tool like other fuzzy extensions in
order to solve problems under vagueness environment. To the best of our knowledge,
the picture fuzzy SAW and picture fuzzy WPM method have not been developed. In
order to find this gap, the literature review of PFS in MCDM is given in Table 1.
Table 1 shows that there is not any paper about Picture fuzzy SAW and Picture fuzzy
WPM methods. In this paper, the main aim is to develop the Picture fuzzy SAW and
WPM method. The rest of the chapter is organized as follows: In Sect. 2, the definitions
of PFS are presented. In Sect. 3, the Picture fuzzy SAW and Picture fuzzy WPM
methods are detailed step by step. In Sect. 4, an application is given and in Sect. 5, the
conclusion is given.
Simple Additive Weighting and Weighted Product Methods 111
Table 1. Picture fuzzy sets in MCDM.

Reference Main area
Zhang et al. [2] Picture 2-tuple linguistic information
Wang and Li [3] Picture hesitant fuzzy set
Sarwar Sindhu Linear programming model based on similarity measure with PFS
et al. [4]
Liang et al. [5] TODIM and ELECTRE with PFS
Thao [6] Entropy measure for PFS and related similarity measures
Tian et al. [7] Picture fuzzy MCDM-introduced weighted picture fuzzy power
Choquet ordered geometric operator and a weighted picture fuzzy power
shapley choquet ordered geometric operator
Lin et al. [8] A novel picture fuzzy MCDM
2 Definitions of Picture Fuzzy Sets
There are some definitions about PFS is given as follows with related equations [1].
~ p of the universe of discourse U is given by;
Definition 2.1: A PFS on a A
nD o
~p ¼
A u; ðlA~ p ðuÞ; mA~ p ðuÞ; pA~ p ðuÞÞu 2 U ð1Þ
where
lA~ p ðuÞ : U ! ½0; 1; mA~ p ðuÞ : U ! ½0; 1; pA~ p ðuÞ : U ! ½0; 1
and
0 lA~ p ðuÞ þ mA~ p ðuÞ þ pA~ p ðuÞ 1 8u 2 U ð2Þ
Then, for each u, the numbers lA~ S ðuÞ; mA~ S ðuÞ and pA~ S ðuÞ are the degree of mem-

bership, non-membership and hesitancy of u to A ~ S , respectively. q ¼ 1 l ~ ðuÞ þ
Ap
mA~ p ðuÞ þ pA~ p ðuÞÞ is called as a refusal degree [9].
Definition 2.2: Basic operators of Single-valued picture fuzzy sets;

n o
~p B
A ~ p ¼ l ~ þ lB~ l ~ lB~ ; p ~ pB~ ; v ~ vB~ ð3Þ
Ap p Ap p Ap p Ap p
n o
~p B
A ~ p ¼ l ~ lB~ ; p ~ þ pB~ p ~ pB~ ; v ~ þ vB~ v ~ vB~ ð4Þ
Ap p Ap p Ap p Ap p Ap p
k
~p ¼
kA 1 1 lA~ p ; pkA~ p ; vkA~ p for k [ 0 ð5Þ
112 F. K. Gundoğdu and E. Bolturk
k k
~k ¼
A lkA~ p ; 1 1 vA~ p ; 1 1 pA~ p for k [ 0 ð6Þ
p
Definition 2.3: Single-valued Picture Fuzzy Weighted Averaging operator (PFWA)

P
n
with respect to, w ¼ ðw1 ; w2 . . .. . .:; wn Þ; wi 2 ½0; 1; wi ¼ 1, is defined as;
i¼1
~ 1 ; . . .. . .:; A
PFWAw ðA ~ ~ ~ ~n
n Þ ¼n w1 A1 þ w2 A2 þ . . .. . . þ wn A
Q Q
n Q wi
n ð7Þ
¼ 1 ð1 lA~ ii Þwi ; vAw~ i ; pA~
ii ii
i¼1 i¼1 i¼1
Definition 2.4: Score functions and Accuracy functions of sorting picture fuzzy
numbers are defined by;

1 .
~p ¼
Score A 1 þ 2lA~ p vA~ p pA~ p 2 ð8Þ
2

~ p ¼ l~ þ v~ þ p~
Accuracy A ð9Þ
Ap Ap Ap
~p \ B
Note that: A ~ p if and only if
~ p Þ \ ScoreðB
(i) ScoreðA ~ p Þ or
~
(ii) ScoreðAp Þ ¼ ScoreðB ~ p Þ \ AccuracyðB
~ p Þ and AccuracyðA ~pÞ
3 The Proposed Methods
Scoring methods are the most frequently employed multi attribute decision-making
methods due to their easiness and effectiveness. In this section, SAW and WPM are
extended to their fuzzy versions by using single-valued picture fuzzy sets.
3.1 Simple Additive Weighting Method Using Picture Fuzzy Sets

A MCDM problem can be expressed as a decision matrix whose elements show the
evaluation values of all alternatives with respect to each criterion under picture fuzzy
environment. Let X ¼ fx1 ; x2 ; . . .. . .:xm g ðm 2Þ be a discrete set of m feasible
alternatives and C ¼ fC1 ; C2 ; . . .. . .Cn g be a finite set of criteria and wj ¼
fw1 ; w2 ; . . .. . .wn g be the weight vector of all criteria which satisfies 0 wj 1 and
Pn
wj ¼ 1.
j¼1
Step 1: Let Decision Makers (DMs) fill in the decision matrix based on the linguistic
terms given in Table 2.
Table 2. Picture fuzzy linguistic terms [9].

Linguistic terms ðl; p; vÞ
Equally Importance (EI) (0.5, 0.1, 0.4)
Step 2: Aggregate the judgments of each decision maker (DM) using Picture Fuzzy
Weighted Averaging operator (PFWA) as given in Eq. (7). Construct aggregated
picture fuzzy decision matrix based on the opinions of decision makers as given in
Eq. (10).
0 1
ðl11 ; p11 ; v11 Þ ðl12 ; p12 ; v12 Þ . . . ðl1n ; p1n ; v1n Þ
B ðl21 ; p21 ; v21 Þ ðl22 ; p22 ; v22 Þ . . . ðl2n ; p2n ; v2n Þ C

B C
~xij ¼ Cj ð~xi Þ mxn ¼ B
B: : : C
C
@: : : A
ðlm1 ; pm1 ; vm1 Þ ðlm2 ; pm2 ; vm2 Þ . . . ðlmn ; pmn ; vmn Þ
ð10Þ
Step 3: Decision makers also present the judgments about decision criteria as given in
Step 1. Decision makers evaluate the alternatives with respect to the criteria as if they
were benefit criteria such that they assign a lower linguistic term if it is a cost criterion.
Aggregate the judgments of each decision maker (DM) based on Eq. (7).
Step 4: Defuzzify the aggregated criteria weights based on the score function given in
Eq. (8) and normalize the aggregated criteria weights by using Eq. (11).

~ pj
Score w
p
j ¼ n
w ð11Þ
P
Score w ~ pj
J¼1
Step 5: Compute the fuzzy results of Picture Fuzzy Simple Additive Weighting method
(PFSAW) as given in Eq. (12) by utilizing Eqs. (3) and (5).
X
n X
n
PFSAWi ¼ ~xijw ¼ pj
~xij w 8i ð12Þ
j¼1 j¼1
Step 6: Calculate the score of each alternative by using Eq. (8) and put the alternatives
into order with respect to the decreasing values of score values.
3.2 Simple Weighted Product Method Using Picture Fuzzy Sets

The steps 1–4 applied in Sect. 3.1 are also used in this method in the same way and so,
first four steps are omitted here.
Step 5: Calculate the results of Picture Fuzzy Weighted Product Model (PFWPM) as
presented in Eq. (13) by utilizing Eqs. (4) and (6).
Y
n
p
w
PFWPMi ¼ ~xij j 8i ð13Þ
j¼1
Step 6: Compute the score of each alternative by using Eq. (8) and put the alternatives
into order with respect to the decreasing values of score values.
4 An Application
Selection among several internet server options is a very popular problem for most of
the people. Suppose that five internet provider options are presented to three people
(DM1, DM2, and DM3). The weights of these decision makers who have different
significant levels are 0.3, 0.5 and 0.2, respectively. These people request a help from
you for the selection among the following alternatives: Alternative 1 (X1), Alternative 2
(X2), Alternative 3 (X3), Alternative 4 (X4), Alternative 5 (X5). The selection criteria
can be speed of the internet (C1), quota of the preferred internet package (C2), cost per
month (C3), and access area of the server (C4). First of all, the assessments for the
criteria are collected from decision makers with respect to the goal, based on the
linguistic terms given in Table 1. All assessments are given in Table 3.
Table 3. Assessments of decision-makers

DM1 C1 C2 C3 C4 DM2 C1 C2 C3 C4 DM3 C1 C2 C3 C4
X1 VHI SMI SLI EI X1 SMI HI VLI HI X1 SMI SLI VLI SLI
X2 SLI LI LI LI X2 HI VLI LI SMI X2 HI LI VLI LI
X3 VHI SMI VHI VHI X3 VHI VHI HI VHI X3 HI HI HI HI
X4 SMI HI LI HI X4 HI EI LI EI X4 SMI HI SMI HI
X5 HI SLI HI SMI X5 HI SLI EI EI X5 HI EI HI SLI
These judgments are aggregated using single-valued PFWA operator by consid-

ering the importance levels of decision makers is presented in Table 4.
Table 4. Aggregated decision matrix

Alternatives C1 C2 C3 C4
X1 (0.74, 0.00, 0.18) (0.65, 0.00, 0.20) (0.17, 0.00, 0.77) (0.62, 0.00, 0.22)
X2 (0.66, 0.00, 0.17) (0.18, 0.00, 0.71) (0.22, 0.00, 0.64) (0.45, 0.00, 0.42)
X3 (0.88, 0.00, 0.06) (0.82, 0.00, 0.10) (0.81, 0.00, 0.08) (0.88, 0.00, 0.06)
X4 (0.68, 0.00, 0.17) (0.65, 0.07, 0.20) (0.34, 0.00, 0.52) (0.65, 0.07, 0.20)
X5 (0.75, 0.05, 0.10) (0.35, 0.00, 0.55) (0.65, 0.07, 0.20) (0.50, 0.00, 0.40)
The linguistic importance weights of the criteria assigned by DMs are shown in
Table 5. The weight of each criterion also obtained by using single-valued PFWA
operator. After the weights of the criteria have been determined, the defuzzified and
normalized criteria weights are calculated by using Eqs. (8) and (11) as given in
Table 5.
Table 5. Importance weights of the criteria

Criteria DM1 DM2 DM3 Weighted of each criterion Normalized of each criterion
C1 VHI SMI SMI (0.74, 0.00, 0.18) 0.324
C2 LI VLI LI (0.18, 0.00, 0.71) 0.091
C3 VHI VHI HI (0.81, 0.00, 0.08) 0.359
C4 SMI EI SLI (0.50, 0.00, 0.40) 0.226
By utilizing Eqs. (8), and (12) and SðPFSAWi Þ is calculated as in Table 6. The final
ranking is X3 [ X5 [ X4 [ X1 [ X2 .
Table 6. Single-valued picture fuzzy simple additive weighting model and ranking
Alternatives PFSAWi SðPFSAWi Þ Ranking
X1 (0.56, 0.00, 0.32) 0.898 4
X2 (0.45, 0.00, 0.38) 0.755 5
X3 (0.85, 0.00, 0.07) 1.319 1
X4 (0.57, 0.00, 0.27) 0.938 3
X5 (0.64, 0.00, 0.20) 1.036 2
Alternative 3 is the best alternative for decision makers with respect to single-
valued picture fuzzy simple additive weighting model. Based on Eqs. (8) and (13),
PFWPMi is calculated as in Table 7. The final ranking is the same with PFSAW
method.
Table 7. Single-valued picture fuzzy simple weighted product model and ranking
Alternatives PFWPMi SðPFWPMi Þ Ranking
X1 (0.41, 0.00, 0.48) 0.669 4
X2 (0.36, 0.00, 0.49) 0.620 5
X3 (0.85, 0.00, 0.07) 1.314 1
X4 (0.52, 0.02, 0.33) 0.852 3
X5 (0.60, 0.04, 0.26) 0.964 2
In order to test the reliability and robustness, we compared the proposed single-
valued picture fuzzy SAW and WPM models with interval-valued spherical fuzzy
TOPSIS model [10, 11]. We obtained the same ranking with the proposed methods in
this paper.
5 Conclusions
PFS, one of the recent extensions of ordinary fuzzy sets, let experts express their
judgments for a considered problem with a larger preference domain about the
membership, non-membership, and hesitancy, individually. In this paper, SAW and
WPM methods have been extended to single-valued picture fuzzy versions as a first
time. Both picture fuzzy SAW and picture fuzzy WPM methods have been applied to
an internet server selection problem. The obtained results are compared with spherical
fuzzy TOPSIS method and have obtained robust results. In future researches, the
proposed methods can be compared with other extensions of fuzzy sets with SAW and
WPM and sensitivity analysis can be reported.
References
2. Zhang, X., Wang, J., Hu, J.: On novel operational laws and aggregation operators of
Picture 2-tuple linguistic information for MCDM problems. Int. J. Fuzzy Syst. 20(3), 958–
969 (2018)
3. Wang, R., Li, Y.: Picture hesitant fuzzy set and its application to multiple criteria decision-
making. Symmetry 10(7), 295 (2018)
4. Sindhu, M.S., Rashid, T., Kashif, A.: Modeling of linear programming and extended
TOPSIS in decision making problem under the framework of picture fuzzy sets. PLoS ONE
14(8), 14 (2019)
5. Liang, W., Dai, B., Zhao, G., Wu, H.: Performance evaluation of green mine using a
combined multi-criteria decision-making method with picture fuzzy information. IEEE
Access 7, 174139–174154 (2019)
6. Thao, N.X.: Similarity measures of picture fuzzy sets based on entropy and their application in
MCDM. Pattern Anal. Appl. 11, 1–11 (2019). https://doi.org/10.1007/s10044-019-00861-9
7. Tian, C., Peng, J., Zhang, S., Zhang, W., Wang, J.: Weighted picture fuzzy aggregation
operators and their applications to multi-criteria decision-making problems. Comput. Ind.
Eng. 137, 106037 (2019). https://doi.org/10.1016/j.cie.2019.106037
8. Lin, M., Huang, C., Xu, X.: MULTIMOORA based MCDM model for site selection of car
sharing station under picture fuzzy environment. Sustain. Cities Soc. 53, 101873 (2020)
9. Meksavang, P., Shi, H., Lin, S.M., Li, H.C.: An extended picture fuzzy VIKOR approach for
sustainable supplier management and its application in the beef industry. Symmetry 11(4),
468 (2019)
10. Kutlu Gündoğdu, F., Kahraman, C.: A novel fuzzy TOPSIS method using emerging interval-
Evaluating Sustainable Vehicle Technologies
for Freight Transportation Using Spherical
Fuzzy AHP and TOPSIS
Miguel Jaller1(&) and Irem Otay2

1
Department of Civil and Environmental Engineering, Sustainable Freight
Research Center, Institute of Transportation Studies, University of California,
One Shields Avenue, Ghausi Hall, Room 3143, Davis, CA 95616, USA
[email protected]
2
Department of Industrial Engineering, Faculty of Engineering and Natural
Sciences, Eski Silahtarağa Elektrik Santrali, Istanbul Bilgi University,
Kazim Karabekir Cad. No: 2/13, Eyüpsultan, 34060 Istanbul, Turkey
[email protected]
Abstract. Freight transportation is vital for the economy and everyday life. It
brings the goods and services needed for industrial and manufacturing pro-
cesses, as well as those to be consumed by the population. However, the
vehicles (mostly diesel trucks) used are responsible for a disproportionate
amount of environmental externalities. Therefore, it is imperative to manage
transport demand, and foster the use of cleaner vehicles, fuels and technologies.
The most common alternatives include compressed (renewable) natural gas
(CNG/RNG), hybrid electric (HE), battery electric (BE) and fuel-cell hydrogen
(H2) vehicles. However, the technical and operational characteristics, market
readiness, and other factors related to these technologies can be very different.
Therefore, the most appropriate option for different uses (e.g., last mile, long-
haul distribution) and users’ preferences is not necessarily clear. Consequently,
this paper proposes Analytic Hierarchy Process (AHP) and The Technique for
Order of Preference by Similarity to Ideal Solution (TOPSIS) based on Spher-
ical fuzzy sets to evaluate the sustainable vehicle technology alternatives over
multiple criteria for freight transportation. Spherical fuzzy sets have been
receiving increasing attention because of their ability to better consider uncer-
tainty by defining membership functions on a Spherical surface and covering a
larger domain. Specifically, the authors evaluate the alternatives using five
criteria: Financial; Business & market-related; Environmental & legal; Main-
tenance & repair availability; and Safety & vehicle performance factors, and 21
sub-criteria. Moreover, the authors also performed sensitivity analysis.
Keywords: Freight transportation Vehicle and fuel pathways Spherical

Fuzzy AHP Spherical fuzzy TOPSIS Spherical fuzzy sets Multi-criteria
decision making
Evaluating Sustainable Vehicle Technologies 119
1 Introduction
The United States and other countries around the world have expressed their interest,
and some have already enacted regulations or plans to improve the environmental
efficiency of their freight transportation systems focusing on the use of zero and near-
zero emission vehicles (ZEVs) [1]. However, the technical and operational character-
istics, market readiness, and other factors related to these technologies can vary; and
the most appropriate option for different uses and users’ preferences is not necessarily
clear. Previous analyses have shown that besides the intensity of use of the vehicles
(e.g., yearly mileage, and duty cycles), purchase price, maintenance and operational
costs, different types of incentives could be determinant factors to advance their
adoption and use [2, 3]. Moreover, a better understanding of the behaviors and attitudes
of freight stakeholders will help define adequate programs and regulations that can
promote their deployment. Consequently, this study aims to use Multiple Criteria
Decision Making (MCDM) techniques to analyze the factors that contribute to the
prioritization and adoption of ZEVs in last mile distribution operations.
The literature review shows that MCDM techniques have been extensively used in
many applications and the different techniques provide a compromise in solutions with
respect to the process uncertainties, the expertise of the decision makers, and other
instrumental issues [4–9]. To overcome the limitations, Zadeh [7] introduced the fuzzy
set theory to deal with uncertainty, lack of information in human judgments, and
vagueness. In this study, the authors propose an MCDM technique integrating fuzzy
AHP and fuzzy TOPSIS methods using Spherical fuzzy sets considering the advan-
tages of handling uncertainty in a better way and covering a larger domain. To the best
of the authors’ knowledge, this is the first study analyzing ZEVs adoption using an
integrated multi-expert Spherical fuzzy MCDM method considering a wide range of
criteria and sub-criteria. Specifically, this study considered compressed (renewable)
natural gas (CNG/RNG), hybrid electric (HE), battery electric (BE) and fuel-cell
hydrogen (H2) vehicles, besides diesel vehicles. The model evaluates the alternatives
using five criteria: economic; business, incentives & market-related; environmental &
regulatory; infrastructure; and safety & vehicle performance factors. And include
twenty-one sub-criteria, e.g., total cost of ownership, payback period, and public/
private fueling/charging infrastructure availability.
The remaining of the study is organized as follows. Section 2 provides a literature
review regarding to the evaluation of zero emission alternatives. Section 3 describes
the overall methodology while Sect. 4 discusses the case implementation and empirical
results. The paper ends with a conclusion and discussion section.
2 Literature Review
There are some studies in the literature evaluating the adoption of alternative tech-
nologies in medium and heavy duty vehicles using quantitative and qualitative meth-
ods. [10] conducted a market penetration analysis of electric vehicles and discusses the
conditions that make them cost effective compared to conventional diesel vehicles. [11]
provided insights from focus groups and interviews in the heavy duty vehicle sector
120 M. Jaller and I. Otay
and the energy efficiency paradox of adopting more sustainable technologies due to the
high upfront costs of these technologies compared to the conventional ones. [12] used
freight data from California and developed a truck purchase decision choice model
considering vehicle/technology performance, vehicle capital and operating costs,
mileage and performance requirements, and other important purchase decision factors
for different types of trucks and fleets. Similarly, [13] estimated truck fleet choices
using structured surveys from fleets in China; they considered the factors influencing
their willingness to purchase alternative fuel vehicles, their timeframe, and purchase
prices. [2], using data from the Fleet DNA Project from the National Renewable
Energy Laboratory (NREL) conducted an extensive study of the total cost of ownership
for different vehicle classes in last mile delivery vocations.
Additionally, in the area of study of this paper, MCDM methods have been used to
evaluate different vehicle technologies in public and freight transport. For example,
[14] evaluated vehicle selection choices for public transport, and [15] developed a
Hierarchical Hesitant Fuzzy Linguistic model to evaluate alternative fuel heavy duty
vehicles. Moreover, [16] used PROMETHEE II and fuzzy TOPSIS to evaluate electric
vans for city logistics applications. They considered 4 criteria (i.e., performance, bat-
tery, engine, and price), which included 9 sub-criteria, and discussed different barriers
for these vehicles.
The following section provides details on Spherical Fuzzy Sets (SFS) and the
integrated Spherical Fuzzy AHP-TOPSIS methodology adopted in this study.
3 Methodology
3.1 Preliminaries on Spherical Fuzzy Sets (SFSs)
In Spherical fuzzy sets comprised of membership ðlðxÞÞ, non-membership ðvðxÞÞ, and
hesitancy ðpðxÞÞ parameters, the squared sum of these parameters cannot exceed “1”
while each of these parameters ranges from “0” and “1” [17].
~ S of the universe of discourse set U be a Spherical fuzzy set.
Definition 1. Let A
~ S ¼ fx; ðl1 ðxÞ; l1 ðxÞ; l1 ðxÞÞjx 2 U g where 0 l2 ðxÞ þ v2 ðxÞ þ p2 ðxÞ 1

A 1 1 1
8x 2 U ð1Þ
and l1 ðxÞ : U ! ½0; 1; v1 ðxÞ:U ! ½0; 1; and p1 ðxÞ:U ! ½0; 1
Definition 2. Addition ðA ~1 B ~1 B
~ 2 Þ, multiplication ðA ~ 2 Þ, multiplication by a scalar
~ ~ k
ðk A1 Þ and power functions ðA1 Þ are the basic arithmetic operations for
nD o
~1 A
A ~1 ¼ x; l ~ ðxÞ; v ~ ðxÞ; p ~ ðxÞ jx 2 U1 and B~2 B ~ 2 ¼ x; l ~ ðxÞ; v ~ ðxÞ;
A1 A1 A1 B2 B2
pB~ 2 ðxÞÞjx 2 U2 :gÞ of the universe of discourse sets U1 and U2 as discussed by [17]).
Definition 3. The formulation for Spherical Weighted Arithmetic Mean (SWAM) is
presented in Eq. (2) [17].
8" #1=2
< Y
n
~ S1 ; . . .; A
SWAMw ðA ~ Sn Þ ¼ 1 ð1 l2A~ Si Þwi ;
: i¼1
" #1=2 9 ð2Þ
Y
n Y
n Y
n =
mAw~ i ; ð1 l2A~ Si Þwi ð1 l2A~ Si p2A~ Si Þwi
i¼1
Si
i¼1 i¼1
;
Pn
where w ¼ ðw1 ; w2 . . .; wn Þ; wi 2 ½0; 1; i¼1 wi ¼ 1.
Definition 4. The value of score function can be calculated for a Spherical fuzzy

number A~s A~ s ¼ lA ; vAs ; pAs as follows [18]:
s
lA~ s þ 1 2vA~ s þ 1 pA~ s

~sÞ ¼
ScðA ð3Þ
3
3.2 Integrated Spherical Fuzzy AHP-TOPSIS Methodology

The steps of the proposed integrated Spherical fuzzy AHP-TOPSIS method are sum-
marized as follows [4, 5]:
Step 1: Generate an MCDM problem by representing a goal, a finite set of criteria
Cj ðXi Þ ¼ ðlij ; mij ; pij Þ; j ¼ 1; 2; . . .; n & sub-criteria, and an alternative set
X ¼ fx1 ; x2 ; . . .xm g.
Step 2: Collect pairwise comparison matrices based on the linguistic terms (in
Table 1) with their corresponding spherical fuzzy numbers, from decision maker(s).
Table 1. Linguistic scale with corresponding Spherical fuzzy numbers.

Linguistic terms ðl; m; pÞ Linguistic terms ðl; v; pÞ
Absolutely High (AH) (0.9, 0.1, 0) Medium Low (ML) (0.4, 0.6, 0.3)
Very High (VH) (0.8, 0.2, 0.1) Low (L) (0.3, 0.7, 0.2)
High (H) (0.7, 0.3, 0.2) Very Low (VL) (0.2, 0.8, 0.1)
Medium High (MH) (0.6, 0.4, 0.3) Absolutely Low (AL) (0.1, 0.9, 0)
Approximately Equal (AE) (0.5, 0.4, 0.4) E. Equal (EE) (0.5, 0.5, 0.5)
Step 3: Check the consistency of the pairwise comparison matrices by applying the
classical consistency analysis methods based on the score indices [5].
Step 4: Calculate the weights of the criteria using Spherical fuzzy AHP method.
Step 4.1: Use the SWAM operator to obtain the criteria’s Spherical fuzzy weights.
Step 4.2: Defuzzify the weights using Eq. (3) and normalize them by dividing each
defuzzified weight by the sum of the defuzzified criteria weights.
Step 5: Calculate the weights of sub-criteria through Step 4.1 and Step 4.2.
Step 6: Evaluate the alternatives by Spherical Fuzzy TOPSIS method.
Step 6.1: Ask decision maker(s) to fill out a Spherical fuzzy decision matrix (or
matrices) D ¼ ðCj ðXi ÞÞmxn using the linguistic scale.
Step 6.2: Aggregate the judgments of the decision makers using the SWAM
operator, and obtain the aggregated Spherical fuzzy decision matrix
Dagg ¼ ðCj ðXi ÞÞmxn
Step 6.3: Estimate the aggregated weighted spherical fuzzy decision matrix ðDwagg Þ.
0 1
ðlw11 ; vw11 ; pw11 Þ ðlw1n ; vw1n ; pw1n Þ
B .. .. .. C
Dwagg ¼ ðCj ðXiw ÞÞmxn ¼@ . . . A ð4Þ
ðlwm1 ; vwm1 ; pwm1 Þ ðlwmn ; vwmn ; pwmn Þ
Step 6.4: Defuzzify the values in the Dwagg matrix using Eq. (5).

pwij 2 pwij 2
ScoreðCj ðXiw ÞÞ ¼ 2lwij vwij ð5Þ
2 2
Step 6.5: Estimate Spherical Fuzzy Positive ðX PIS Þ and Negative Ideal Solutions
ðX NIS Þ by means of Eq. (6).

X PIS ¼ Cj ; max \Score Cj Xiw [ jj ¼ 1; 2. . .n ; ð6Þ
i

X NIS ¼ Cj ; min \ScoreðCj ðXiw ÞÞ [ jj ¼ 1; 2. . .n
i
Step 6.6: Calculate the distances between each alternative and ideal solutions.
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
1 Xn 2 2 2
dðXi ; X PIS Þ ¼ ðl Xi l X PIS Þ þ ðv Xi v X PIS Þ þ ðp Xi p X PIS Þ ð7Þ
2n i¼1
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 Xn 2 2 2

dðXi ; X Þ ¼
NIS
ðl X i
l X NIS Þ þ ðvXi vX NIS Þ þ ðpXi pX NIS Þ ð8Þ
2n i¼1
Step 7: Finally, derive the closeness coefficient ratio ðCCRðXi ÞÞ for each alterna-
tive, and rank the alternatives based on the descending values.
dðXi ; X NIS Þ
CCRðXi Þ ¼ ð9Þ
dðXi ; X PIS Þ þ dðXi ; X NIS Þ
4 Implementation and Empirical Results
Based on the findings from the literature review, previous work, and informal com-
munications with experts, fleets, vehicle manufacturers, and other organizations, the
authors developed the hierarchical structure of the problem. In the model, the authors
evaluated five criteria (C1–C5) and 21 sub-criteria (see Table 2) to evaluate Diesel
(A1), Hybrid Electric (A2), CNG/RNG (A3), Battery Electric (A4), and Fuel Cell (H2)
(A5) vehicles. The authors designed a survey instrument and administered to three
experts (E1, E2, and E3). After receiving the data, the authors checked for consistency
using the score indices proposed by [17]. For illustration purposes, Table 3 shows the
pairwise evaluations of criteria from the experts. After the linguistic values were
converted to their corresponding Spherical fuzzy numbers, the evaluations were
aggregated using SWAM operator in Eq. (2).
Table 2. The list of criteria and sub-criteria.

C1: Financial factors C23: Purchase voucher incentives
C2: Business & Market-related factors C24: Non-financial incentives
C3: Env. & legal factors C25: Secondary market development
C4: Maintenance & repair availability C31: Env. & health considerations
C5: Safety & performance factors C32: Regulatory environment
C11: Total cost of ownership C41: Public fueling/charging infrastructure availab.
C12: Initial purchase cost C42: Depot fueling/charging infrastructure availab.
C13: Maintenance & repair costs C43: Maintenance and repair shop availability
C14: Resale value C51: Driving range
C15: Fuel costs C52: Fueling/charging time
C16: Payback period C53: Vehicle payload
C21: Brand image C54: Vehicle safety
C22: Market penetration C55: Vehicle reliability
By following the steps of the methodology, the authors estimated the Spherical
fuzzy weights of criteria and sub-criteria (omitted due to space limitations). The
weights were defuzzified using Eq. (3) and normalized (see the weights in Table 4). In
the analyses, the weights of all of the experts were taken as 1/3. The results indicate that
economic factors are the most important, followed by safety and performance indi-
cators, with maintenance, repair and refueling in the third place.
Table 3. Pairwise comparison matrix of criteria.

Criteria C1 C2 C3 C4 C5
E1 E2 E3 E1 E2 E3 E1 E2 E3 E1 E2 E3 E1 E2 E3
C1 EE EE EE VH E VH VH MH VH H ML H H L H
C2 EE EE EE E H MH ML ML ML E L ML
C3 EE EE EE ML L ML ML VL ML
C4 EE EE EE H ML ML
C5 EE EE EE
Table 4. The weights of criteria and sub-criteria.

Criteria Weights Sub-criteria Weights Sub-criteria Weights Sub-criteria Weights
C1 0.27 C11 0.24 C22 0.25 C42 0.37
C2 0.17 C12 0.20 C23 0.23 C43 0.37
C3 0.13 C13 0.14 C24 0.17 C51 0.23
C4 0.21 C14 0.09 C25 0.15 C52 0.15
C5 0.22 C15 0.18 C31 0.43 C53 0.17
C16 0.15 C32 0.57 C54 0.17
C21 0.19 C41 0.26 C55 0.27
Then, the authors estimate the decision matrices (omitted due to space limitations)
collected from the experts. The judgments of the experts were aggregated, the positive and
negative ideal solutions were determined, and distances were estimated. The distances
were dðXi ; X PIS Þ ¼ f0:041; 0:044; 0:051; 0:076; 0:084g and dðXi ; X NIS Þ ¼ f0:088;
0:070; 0:070; 0:047; 0:045g. Finally, CCR(xi) values were calculated and the alterna-
tives were ranked as presented in Table 5.
Table 5. Relative closeness to ideal solutions (CCR) and ranking.

A1 A2 A3 A4 A5
CCRðxi Þ 0.6832 0.6166 0.5792 0.3847 0.3488
Rank 1 2 3 4 5
The results show that battery electric and fuel cell are in a separate preference
group, with their CCR values to be almost half of the preferred alternative (diesel
trucks). These results are consistent with previous findings which indicate that the
preferred alternative continues to be diesel vehicles; however, the results show promise
to hybrid electric vehicles. The authors also performed sensitivity analyses to evaluate
the potential changes in the solutions for the changes in the criteria weights by con-
sidering 50 scenarios. The results indicate low sensitivity with respect to the changes in
the weight of economic factors. The CRR become more prominent at the extreme
alternatives (diesel and fuel cell). Similar results are observed for changes in the
business incentives and market-related factors criteria. The results also show more
variability for the changes to the environmental and regulatory factors.
5 Conclusion and Future Research
The study proposes an integrated Spherical fuzzy AHP and TOPSIS method for ana-
lyzing sustainable vehicle technology considering last mile deliveries. The proposed
fuzzy model evaluates compressed (renewable) natural gas (CNG/RNG), hybrid
electric, battery electric and fuel-cell hydrogen vehicles. The ultimate ranking estimated
by the model resulting in diesel vehicles to be the preferred alternative, and battery
electric and fuel cell vehicles lagging. However, the results show that there is an
acknowledgment that under a stricter environmental regulatory framework, the pre-
ferred choices are battery electric and fuel cell vehicles. In the study, the sensitivity
analysis is also applied and its results are discussed. For future studies, vehicle tech-
nologies can be evaluated for a long haul option, and it is also recommended to use
Pythagorean, Hesitant, or Neutrosophic fuzzy sets, and compare the results.
Acknowledgments. The authors would like to thank the U.S. Department of Transportation,
and the National Center for Sustainable Transportation and the Institute of Transportation Studies
at the University of California Davis for the funding support for this research. The authors also
appreciate the experts surveyed in this research.
References
1. California Air Resource Board. Advanced Clean Truck (ACT) Program. https://ww2.arb.ca.
gov/our-work/programs/advanced-clean-truck/about. Accessed 10 Feb 2020
2. Jaller, M., Pineda, L., Ambrose, H.: Evaluating the use of zero-emission vehicles in last mile
deliveries. UC Davis Institute of Transport Studies, University of California, Davis, CA
(2018)
3. Davis, B.A., Figliozzi, M.A.: A methodology to evaluate the competitiveness of electric
delivery trucks. Transp. Res. Part E Logist. Transp. Rev. 49(1), 8–23 (2013)
4. Kahraman, C., Kutlu Gundogdu, F., Onar, S.C., Oztaysi, B.: Hospital location selection
using spherical fuzzy TOPSIS. In: 2019 Conference of the International Fuzzy Systems
Association and the European Society for Fuzzy Logic and Technology (EUSFLAT 2019).
Atlantis Press (2019)
5. Kutlu Gündoğdu, F.K., Kahraman, C.: A novel spherical fuzzy analytic hierarchy process
and its renewable energy application. Soft Comput. 24(6), 4607–4621 (2020)
6. Zadeh, L.A.: Fuzzy sets and systems. Presented at the Symposium on System Theory,
Polytechnic Institute of Brooklyn (1965)
8. Otay, İ., Jaller, M.: Multi-expert disaster risk management & response capabilities
assessment using interval-valued intuitionistic fuzzy sets. J. Intell. Fuzzy Syst. 38(1),
835–852 (2020)
9. Otay, I., Jaller, M.: Multi-criteria and multi-expert wind power farm location selection using
a Pythagorean fuzzy analytic hierarchy process. In: Proceedings of International Conference
on Intelligent and Fuzzy Systems, pp. 905–914. Springer (2019)
10. Pelletier, S., Jabali, O., Laporte, G.: Battery electric vehicles for goods distribution: a survey
of vehicle technology, market penetration, incentives and practices. https://www.cirrelt.ca/
DocumentsTravail/CIRRELT-2014-43.pdf. Accessed 19 May 2016
11. Klemick, H., Kopits, E., Wolverton, A., Sargent, K.: Heavy-duty trucking and the energy
efficiency paradox: evidence from focus groups and interviews. Transp. Res. Part A Policy
Pract. 77, 154–166 (2015)
12. Miller, M., Wang, Q., Fulton L.: Truck Choice Modeling: Understanding California’s
Transition to Zero-Emission Vehicle Trucks Taking into Account Truck Technologies,
Costs, and Fleet Decision Behavior (2017)
13. Zhang, Y., Jiang, Y., Rui, W., Thompson, R.G.: Analyzing truck fleets’ acceptance of
alternative fuel freight vehicles in China. Renew. Energy 134, 1148–1155 (2019)
14. Aydin, S., Kahraman, C.: Vehicle selection for public transportation using an integrated
multi criteria decision making approach: a case of Ankara. J. Intell. Fuzzy Syst. 26(5), 2467–
2481 (2014)
15. Yavuz, M., Oztaysi, B., Onar, S.C., Kahraman, C.: Multi-criteria evaluation of alternative-
fuel vehicles via a hierarchical hesitant fuzzy linguistic model. Expert Syst. Appl. 42(5),
2835–2848 (2015)
16. Wątróbski, J., Małecki, K., Kijewska, K., Iwan, S., Karczmarczyk, A., Thompson, R.G.:
Multi-criteria analysis of electric vans for city logistics. Sustainability 9(8), 1453 (2017)
18. Otay, I., Kahraman, C., Oztaysi, B., Onar, S.C.: Score and accuracy functions for different
types of spherical fuzzy sets. In: Proceedings of FLINS/ISKE 2020: The 14th Interna-
tional FLINS Conference on Robotics and Artificial Intelligence and the 15th International
Conference on Intelligent Systems and Knowledge Engineering, Cologne, Germany (2020)
Evaluating Strategic Entry Decisions
Using Spherical Fuzzy Sets
Sezi Cevik Onar(&), Cengiz Kahraman, and Basar Oztaysi
Industrial Engineering Department, Istanbul Technical University,

Macka, Istanbul, Turkey
[email protected]
Abstract. In today’s economy, companies have to constantly develop new

strategies and have to choose whether to expand into new markets, stabilize in or
exit from the existing markets. The entry in to the new markets is a complex
strategic decision; it does not only involve the entry mode decision but also in
some occasions involve partner selection. When the entry modes such as
mergers, acquisitions, joint ventures or strategic alliances are selected, partner
selection become the major determinant of the success. Attributes such as
partners’ resources, technological diversity, common benefits or the size of the
partners have a significant impact on the success of the entry decisions. Yet,
under different environmental conditions partners may show different perfor-
mances. It is not easy to evaluate partners under these uncertainties. Spherical
fuzzy sets enable dealing with hesitancy and uncertainty. In this study, a min-
imax regret based spherical fuzzy decision making approach is developed for
evaluating the partners in the market entry decisions.
Keywords: Strategic decisions Market entry Spherical fuzzy sets Minimax

regret
1 Introduction
Rationality is one of the most important characteristics of a strategic decision making

process. The rationality in this process refers to the necessity of taking logical steps as
we proceed along the path. When examining this rationality, it is also necessary to pay
attention to the importance of strategic decisions made by managers. According to
many experts, decision makers classify strategic decisions when revaluating them. One
of the characteristics of a strategic decision is the importance of the decision: Not all
strategic decisions have the same importance. The second characteristics is the hesi-
tation in the decision. The core of strategic making process involves uncertainty [1].
Many studies have shown that hesitation in decisions has an impact on the strategic
decision making process. The motivation in the decision is another characteristic.
When decision makers classify strategic decisions, they see them as opportunities in a
crisis environment. Especially while selecting a partner for a new market entry they
should consider the importance, hesitancy and motivation. Yet, this is not an easy
process since an opportunity can become a threat under different conditions.
128 S. C. Onar et al.
In the partner selection process several attributes should be considered. These

attributes may have different impacts on different environmental conditions. For
instance, partner size is an important aspect while making strategic decisions [2].
Although it seems like a positive attribute, partner size can be good or bad depending
on the environmental conditions. Under a declining economy size can be burden since
it will diminish the agility but for a growing economy it can create many benefits. It is
very important to determine the company resources in order to see the internal potential
of the companies [2]. There is a positive relation between the success of joint ventures
and the company resources especially under crisis. Yet, if the resources have low
liquidity this may result with a false decision. Tobin’s q value is a frequently used value
in determining the intangible resources of firms. Equity, short and long term debts,
liquidity, including inventory, are calculated by taking into account the carrying value
of total assets. Partner’s knowledge-based resources is an important aspect for the
partner selection [3]. Partners’ technological portfolios are very important for the
companies when in the entry decisions. Usually, the partners’ technological portfolio is
calculated according to the patent classes of the patents they hold, taking into account
the years. As technological diversity increases, innovative performance of companies
increases. If the technological diversity of the companies in cooperation is at a medium
level, it contributes much more to the company’s innovation. As a result of the part-
nership, the partners share some of their talents among them and try to benefit them
more by applying to their operations in the field of partnership. These common benefits
can be shared on issues such as resource sharing, technological diversity, and partners
can also contribute to each other financially [2]. Like the other attributes, common
benefits can be good or bad depending on the environmental conditions.
The nature of strategic decisions always involves some level of uncertainty [4].
Whether the economy going to be decline, stay stable or grow often depends on many
complex parameters that cannot be fully known before the decision making process.
Moreover, evaluating the companies by considering many attributes that have inter
correlations is very hard. Usually decision makers have hesitancies in the evaluations
[5–7]. Spherical fuzzy sets are the extensions of fuzzy sets where the fuzziness is
defined with the degree of membership, the degree of non-membership and the degree
of hesitancy. The total sum of the squares of these parameters should be less than one.
The hesitancies in evaluating partner companies can be represented better with the
spherical fuzzy sets. Minimax regret approach tries to minimize the maximum regret in
a decision making process. In the strategic decisions, such as partner selection in the
market entry minimax regret approach can be useful since it considers the opportunity
costs. In this study, we utilized spherical fuzzy sets for evaluating partners in the
market entry decision and we used minimax regret approach for selecting the partners.
The rest of the paper is organized as follows, Sect. 2 gives the preliminaries of the
spherical fuzzy sets and the utilized minimax regret approach. The proposed approach
is applied to a strategic entry decision in Sect. 3. Section 4 concludes and gives further
studies.
Evaluating Strategic Entry Decisions Using Spherical Fuzzy Sets 129
2 Spherical Fuzzy Sets and Spherical Fuzzy Maximin Regret

Approach
f Þ are the extensions of fuzzy sets that can be represented

A Spherical fuzzy sets ð SF
with lSF ðr Þ the degree of membership, #SF ðr Þ the degree of non-membership and
pSF ðrÞ the degree of hesitancy [8].
f ¼ fhr; ðlSF ðr Þ; #SF ðr Þ; pSF ðr ÞÞijr 2 U g

SF ð1Þ
where the parameters have the relation in Eq. (2).
0 l2SF ðr Þ þ #2SF ðr Þ þ p2SF ðr Þ 1 ð2Þ
f can be defined as follows [8, 9]:

The basic operations on SF
0:5
SF f 2 ¼ l2 þ l2 l2 l2
f 1 SF ; #SF1 #SF2 ;
SF1 SF2 SF1 SF2
0:5 ð3Þ
1 l2SF2 p2SF1 þ 1 l2SF1 p2SF2 p2SF1 p2SF2
0:5
f 1 SF
SF f 2 ¼ lSF lSF ; #2 þ #2 #2 #2 ;
SF1 SF2 SF1 SF2
ð4Þ
1 2
0:5
1 #2SF2 p2SF1 þ 1 #2SF1 p2SF2 p2SF1 p2SF2
k 0:5
f
k SF 1 ¼ 1 1 lSF1
2
; #kSF1 ;
k k 0:5 ð5Þ
1 lSF1 1 lSF1 pSF1
2 2 2
where k [ 0.
f the scoring function in Eq. (6) can be utilized.
In order to defuzzify SF

f ¼ lSF pSF1 2 ð#SF1 pSF1 Þ2
Score SF ð6Þ
1
f Eq. (7) can be utilized.

In order to obtain the distances between two SF
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 Xn 2 ffi
f 1 ; SF
D SF f2 ¼ lSF1 lSF2 þ ðvSF1 vSF2 Þ þ ðpSF1 pSF2 Þ 2 2
ð7Þ
2n i¼1
The minimax regret approach considers the opportunity cost in the decision making
process [10, 11]. The best decision is considered as the one that minimizes regret or the
opportunity cost. It is beneficial for a risk neutral decision maker who wants to be
rational.
In the proposed approach the objective is to minimize the maximum regret. The
steps of the proposed model can be given as follows.
Step 1. Evaluate every alternative under different conditions by using spherical
fuzzy sets. The decision maker can either use the scale given in Table 1 or assign their
own membership, non-membership and hesitancy values and obtain SF f ijk , where SF
f ijk
th th th
denotes the k decision maker’s evaluation of i alternative under j condition.
Table 1. Linguistic evaluation scale with Spherical fuzzy numbers

Linguistic scale ðl; v; pÞ
Absolutely favorable (AF) (0.9, 0.1, 0.1)
Very favorable (VF) (0.8, 0.2, 0.2)
Favorable (F) (0.7, 0.3, 0.3)
Slightly favorable (SF) (0.6, 0.4, 0.4)
Neutral (N) (0.5, 0.5, 0.5)
Slightly unfavorable (SUF) (0.4, 0.6, 0.4)
Unfavorable (UF) (0.3, 0.7, 0.3)
Very unfavorable (VUF) (0.2, 0.8, 0.2)
Absolutely unfavorable (AUF) (0.1, 0.9, 0.1)
Step 2. Aggregate the evaluations of the decision makers by using Spherical

Weighted Arithmetic Mean (SWAM) operator.
f ij ¼ SWAMð SF
SF f ij1 ; . . .; SF
f ijP Þ ¼
h Qp wk i1=2 Q
1 k¼1 1 l2SFijk ; pk¼1 vSFijk ;
ð8Þ
hQ wk Q wk i1=2
p p
k¼1 1 lSFijk k¼1 1 lSFijk pSFijk
2 2 2
Pp
where wk define the weight of the decision maker k and k¼1 wk ¼ 1; wk 2 ½0; 1;
P
n
w ¼ ðw1 ; w2 . . .. . .:; wn Þ; wi 2 ½0; 1; wi ¼ 1.
i¼1
Step 3. Define the best alternative under every condition j, ðj ¼ 1; ::; mÞ by using
the score function defined in Eq. (6).
n o
f ¼ SF
SF f ij i ¼ 1; 2; ::; l
f ; max Score SF ð9Þ
j j
where l is the number of alternatives.

f and SF
Step 4. Calculate the distances between SF f ij for every i; j using the
j
function defined in Eq. (7) and obtain the regret Rij for ith alternative under jth
condition.
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

1 Xn 2 2 2
Rij ¼ lSFij lSFj þ vSFij vSFj þ pSFj pSFj ð10Þ
2n i¼1
Step 5. Calculate the maximum regret Max Ri for each alternative.

Max Ri ¼ max Rij j for j ¼ 1; 2; . . .; l ð11Þ
Step 6. Select the alternative that provides minimum Max Ri .
3 Evaluation of a Foreign Direct Investment with a Spherical

Fuzzy Maximin Regret Approach
The top management of companies must consider interactions with partner companies
when they are making strategic decisions such as joint ventures, mergers, acquisitions,
and strategic alliances. When the companies decide to create a partnership, they have
an idea to complete some of their requirements from the partner company. They want
to use their needs such as labor requirement, market requirement, capital requirement,
and strategic asset needs in the best way by sharing resources among themselves.
A company wants to evaluate an international foreign entry decision. The mangers
of the company want to evaluate these companies but the expected performances of the
companies depend on the environmental conditions. In order to see the applicability of
the proposed approach one of the managers evaluate five partners for the foreign direct
investment decision. The decision makers are expecting to have either one of the
environmental conditions, a declining economy, a stable economy or a growing
economy. The evaluations are done based on using the scale given in Table 1. The
associated Spherical evaluations are given in Table 2.
Table 2. Spherical fuzzy evaluation of alternatives

Criteria Declining economy Stable economy Growing economy
Partner company 1 SF(0.7,0.3,0.3) SF(0.5,0.5,0.5) SF(0.3,0.7,0.3)
The decision maker wants to minimize the opportunity cost. Therefore, the regret
values and the maximum regret values are calculated by using the proposed approach.
Table 3 summarizes the regret scores and the maximum regret score.
Table 3. Regret scores and the maximum regret score

Criteria Declining economy Stable economy Growing economy Maximum Regret
Partner company 1 0.125 0.52 0.64 0.64
Partner company 2 0 0.24 0.375 0.375
Partner company 3 0.525 0 0.24 0.525
Partner company 4 0.46 0.24 0.115 0.46
Partner company 5 0.525 0.52 0 0.525
Candidate partner 5 shows a very favorable performance under a growing econ-

omy. Yet, when the economy is declining or stable it may not provide favorable results.
On the other hand, candidate partner 2 will create a better performance under a
declining economy. With the proposed methodology, the amount of regret is calculated
and the candidate with the minimum regret is selected. For this application, partner
company 2 is the alternative that will minimize the maximum regret and should be
selected.
4 Conclusion
Firms often face multiple different strategic options. The most important issue to be
considered these options is to evaluate whether the decision will best meet the com-
pany’s future expectations. Spherical fuzzy sets enable us better deal with uncertainty
and hesitancy. In this study, Spherical fuzzy sets are utilized for evaluating strategic
decisions. A foreign market entry decision with uncertainties have been evaluated by
using spherical fuzzy sets. In order to deal with uncertainty, the minimax regret
approach is used. The proposed approach enables us dealing with both hesitancy and
uncertainty. In order to show the applicability of the approach, it has been applied to a
foreign entry decision. The alternatives are evaluated with the proposed approach.
For the future studies, the model can be converted to a multi-criteria decision
making problem. It will be beneficial to compare this approach with other methods. In
the modification of the study the model can be enhanced by assigning weights to the
different environmental conditions and sensitivity analysis can be applied to see the
impact of weights in the decision making process. The application can be improved by
using more alternatives and conditions.
References
1. Warren, K.: Strategic Management Dynamics. Wiley, New York (2008)
2. Kumar, M.V.S.: Are joint ventures positive sum games? The relative effects of cooperative
and non-cooperative behavior. Strateg. Manag. J. 32(1), 32–54 (2011)
3. Chen, D., Park, S.H., Newburry, W.: Parent contribution and organizational control in
international joint ventures. Strateg. Manag. J. 30(11), 1133–1156 (2009)
4. Cevik Onar, S., Oztaysi, B., Kahraman, C.: Strategic decision selection using hesitant fuzzy
TOPSIS and interval type-2 fuzzy AHP: a case study. Int. J. Comput. Intell. Syst. 7(5),
1002–1021 (2014)
5. Kahraman, C., Çevik Onar, S., Öztayşi, B.: Engineering economic analyses using
intuitionistic and hesitant fuzzy sets. J. Intell. Fuzzy Syst. 29(3), 1151–1168 (2015)
6. Estrella, F.J., Onar, S.C., Rodríguez, R.M., Oztaysi, B., Martínez, L., Kahraman, C.:
Selecting firms in university technoparks: a hesitant linguistic fuzzy TOPSIS model for
heterogeneous contexts. J. Intell. Fuzzy Syst. 33(2), 1155–1172 (2017)
7. Cevik Onar, S., Oztaysi, B., Kahraman, C.: Multicriteria evaluation of cloud service
providers using Pythagorean fuzzy TOPSIS. J. Multiple-Valued Logic Soft Comput.
30(2–3), 263–283 (2018)
8. Gündoğdu, F.K.: Principals of spherical fuzzy sets. Adv. Intell. Syst. Comput. 1029, 15–23
(2020)
method. J. Intell. Fuzzy Syst. 36, 337–352 (2019)
10. Savage, L.J.: The theory of statistical decision. J. Am. Stat. Assoc. 46(253), 55–67 (1951)
11. Renou, L., Schlag, K.H.: Minimax regret and strategic uncertainty. J. Econ. Theory 145(1),
264–286 (2010)
Spherical Fuzzy Cost/Benefit Analysis of Wind
Energy Investments
Sezi Cevik Onar(&), Basar Oztaysi, and Cengiz Kahraman

{cevikse,infus}@itu.edu.tr
Abstract. Energy investments are highly risky due to the uncertainties of the
energy market. Over the past two decades, the usage of wind energy has been
increased dramatically. Appropriate wind energy investment valuations are vital
for both the public and private sectors. The fluctuations in the energy prices, the
uncertainties in the production levels and the changes in the investment costs
harden the wind energy investment decision. Using the traditional investment
valuations overlooks the necessary managerial flexibility of investments,
uncertainty and risk concepts. The spherical fuzzy sets enable us dealing with
this uncertain structure with the hesitancy. In this study, a new Spherical fuzzy
cost benefit analysis is developed and modified to the wind energy investment
evaluation. In order to show the applicability of the proposed method a wind
energy investment is analyzed with it. This new approach by taking into account
the hesitancy may enable investments to be evaluated with more flexibility and
accuracy.
Keywords: Spherical fuzzy sets Cost benefit analysis Engineering

economics Wind energy investment analysis
1 Introduction
Wind energy is considered as one of the most important energy sources for the tran-
sition from fossil based traditional energy sources. In Europe, the total wind power
investment is €52 billion in 2019 [1]. Yet, the amount of investments for the new
projects is €19 billion which is less than in 2018. The improvements in the wind energy
technologies along with the economies of scale, increased the capacity factor while
decreasing the installation costs. In 2018, the capacity factor of wind energy invest-
ments raised to 35% from 26% in 2010 [2].
Although wind energy investments are reliable investments for the medium to long
term, the crisis such as the crisis due to COVID-19 pandemic slow down the reliable
energy investments. Modeling the uncertainty and hesitancy in wind energy invest-
ments become one of the biggest challenges for the decision makers. Figure 1 shows
the change in the total installed costs, levelized cost of energy and the average capacity
factors are given in Fig. 1 [2].
In literature, many studies focus on the modelling of uncertainty in energy
investments. Dawis and Owens [3] utilized the real options analysis in estimating the
Spherical Fuzzy Cost/Benefit Analysis of Wind Energy Investments 135
40
35
30
25
20
15
10
5
0
Weighted average total installed costs (2018 USD/100kW)

Weighted average capacity factor (%)
Weighted average levelized cost of energy (2018 US Cent/kWh)
Fig. 1. The wind energy investments statistics [2].
value of renewable energy and electricity generating technologies exposed to variable

fossil prices. The main findings of this study are that the real value of renewable energy
investments cannot be obtained by using traditional valuation techniques, and high-
level valuation methods should be used. Similarly, Siddiqui et al. [4] utilized the real
options analysis for the benefit of the renewable energy technology developments. The
researchers emphasized that the deterministic analyzes ignore uncertainties in renew-
able energy costs, technical risks underlying R&D processes, and the possibilities of
regulation that should be made in proportion to the changing state of the world. Fuzzy
sets have been widely used for modeling the uncertainty in renewable energy invest-
ments. Cevik Onar and Kilavuz [5] utilized real options for modelling wind energy
investments in Turkey.
The fuzzy sets are also widely utilized for modelling the uncertainty and hesitancy
in renewable energy investments [6–9]. Spherical fuzzy sets are excellent tools that can
deal with hesitancy and uncertainty by enabling defining membership, non-
membership and hesitancy values independently. Cost benefit analysis is a useful
engineering economics tool but it needs to be modified for dealing with uncertainty and
hesitancy [10]. In this study, spherical fuzzy cost benefit analysis is utilized for eval-
uating wind energy investments.
The rest of the paper is organized as follows: The second section first gives the
basic definitions of Spherical fuzzy sets. Section 3 gives the details of Spherical fuzzy
Cost Benefit analysis. Section 4 applies the proposed methodology to a wind energy
investment. Section 5 concludes and gives further suggestions.
2 Spherical Fuzzy Sets
Spherical fuzzy developed by Gundogdu and Kahraman [11] can be defined as follows
[11–14]:
Let f Sp be a Spherical fuzzy set on the universe U, it can be defined as in Eq. (2)
and (3).

f
Sp ¼ s; lsp ðsÞ; #sp ðsÞ; psp ðsÞ js 2 U ð1Þ
where lsp ðsÞ : U ! ½0; 1; lsp ðsÞ : U ! ½0; 1; lsp ðsÞ : U ! ½0; 1, lsp ðsÞ denotes the
degree of membership, #sp ðsÞ denotes the degree of membership and psp ðsÞ is the
degree of non-membership and hesitancy of s to f Sp.
0 l2sp ðsÞ þ #2sp ðsÞ þ p2sp ðsÞ 1 ð2Þ
The basic operations on the Spherical fuzzy sets are given in Eq. (3–7) [3–7].
0:5
~Sp1 f
Sp 2 ¼ l2sp1 ðsÞ þ l2sp2 ðsÞ l2sp1 ðsÞl2sp2 ðsÞ ; #sp1 ðsÞ#sp2 ðsÞ;
0:5 ð3Þ
1 l2sp2 ðsÞ p2sp1 ðsÞ þ 1 l2sp1 ðsÞ p2sp2 ðsÞ p2sp1 ðsÞp2sp2 ðsÞ
0:5
~Sp1 ~Sp2 ¼ ls ðsÞls ðsÞ; #2 ðsÞ þ #2 ðsÞ #2 ðsÞ#2 ðsÞ ;
p1 p2 sp1 sp2 sp1 sp2
0:5 ð4Þ
1 #2sp2 ðsÞ p2sp1 ðsÞ þ 1 #2sp1 ðsÞ p2sp2 ðsÞ p2sp1 ðsÞp2sp2 ðsÞ
k 0:5
k k 0:5
k~
Sp ¼ 1 1 l2sp ðsÞ ; #sp ðsÞk ; 1 l2sp ðsÞ 1 l2sp ðsÞ p2sp ðsÞ
ð5Þ
where k [ 0.
Spherical fuzzy sets can be defuzzified by using the scoring function defined in
Eq. (6).
2 2
Score f
Sp ¼ lsp ðsÞ psp ðsÞ #sp ðsÞ psp ðsÞ ð6Þ
In order to aggregate and defuzzify the Spherical fuzzy sets Spherical Weighted
Arithmetic Mean (SWAM) aggregation operator can be utilized.

SWAM f Sp n ¼ w1 ~Sp1 þ . . . þ wn f
Sp 1 ; . . .; f Sp n ¼
2 Q wi 0:5 Q 3
* +
1 ni¼1 1 l2spi ðsÞ ; ni¼1 #spi ðsÞwi ; ð7Þ
6 7
4 wi 0:5 5
Qn wi Q
n
i¼1 1 l 2
spi ð s Þ i¼1 1 l 2
spi ð s Þ p 2
spi ð s Þ
Pn
where w ¼ ðw1 ; w2 . . .. . .:; wn Þ; wi 2 ½0; 1; i¼1 wi ¼ 1.
3 Spherical Fuzzy Cost Benefit Analysis
The engineering economic analysis can be enhanced by using fuzzy sets [10, 15]. In the
Spherical fuzzy cost benefit analysis, Spherical numbers are utilized for defining the
parameters. These parameters are initial funding ( f IF Sp ), the uniform annual costs
g Sp ), uniform annual income ( UAI
( UAC g Sp ), the life of the investment ( m e Sp ), the interest
f
rate (~rSp ), and the scrap value ( SV Sp ). In order to represent the hesitancy of the decision
makers, spherical fuzzy values are assigned. These parameters can be defined as
follows:
nD E D E D Eo
f
IF Sp ¼ if1 ; f
Sp 1 ; if2 ; f
Sp 2 ; . . .; ifn ; f
Sp n ð8Þ
nD E D E D Eo
g Sp ¼
UAC uac1 ; e
Sp1 ; uac2 ; f
Sp 2 ; . . .; uacn ; f
Sp n ð9Þ
nD E D E D Eo
g Sp ¼
UAI uai1 ; f
Sp 1 ; uain ; f
Sp 2 ; . . .; uain ; f
Sp n ð10Þ
nD E D E D Eo
f Sp ¼
SV sv1 ; f
Sp 1 ; sv2 ; f
Sp 2 ; . . .; svn ; f
Sp n ð11Þ
nD E D E D Eo
e Sp ¼
m m1 ; f
Sp 1 ; m2 ; f
Sp 2 ; . . .; mn ; f
Sp n ð12Þ
n D E o
~rSp ¼ r1 ; ~Sp1 ; r2 ; f
Sp 2 ; rn ; ~
Spn ð13Þ
~
BSp
By using these parameters benefit/cost ~ Sp
C
analysis of as single alternative can be
defined as follows:

~ Sp
B g Sp P ; ~rSp ; m
UAI e Sp
¼ A
ð14Þ
~ Sp f
C g Sp P ; ~rSp ; m
IF Sp þ UAC f Sp P ; ~rSp ; m
e Sp SV e Sp
A F
or
" #
em
g Sp ð 1 þ ~rSp Þ Sp 1
UAI
~ Sp em
~rSp ð1 þ ~rSp Þ Sp
B
¼ " # ð15Þ
~ Sp
C em Sp m Sp
f g Sp ð1 þ ~rSp Þ 1 SV
IF Sp þ UAC f 1 þ ~rSp e
em e
r
~rSp ð1 þ ~rSp Þ Sp
Sp
For an incremental Spherical cost benefit analysis can be calculated as follows:

gSp
DB g Sp P ; ~rSp ; m
DUAI e Sp
¼
A
ð16Þ
f Sp DIF
DC g Sp þ DUAC
g Sp P ; ~rSp ; m g Sp P ; ~rSp ; m
e Sp DSV e Sp
A F
or
" #
em
g ð1 þ ~rSp Þ Sp 1
DUAI Sp
f em
DB ~rSp ð1 þ ~rSp Þ Sp
¼ " # ð17Þ
f
DC em Sp m Sp
g Sp ð1 þ ~rSp Þ 1 D SV
g Sp þ D UAC
DIF f ~rSp 1 þ ~rSp e
~rSp ð1 þ ~rSp Þ
em Sp
Where nD E D E D Eo
S
IF Sp ¼ nj¼1 SWAM if1 ; f
f Sp 1 ; if2 ; e
Sp2 ; . . .; ifn ; f Sp n
S nD E D Eo
g Sp ¼ n SWAM uac1 ; f
UAC Sp 1 ; uac2 ; ~Sp2 ; . . .; uacn ; e Sn
j¼1
S n D E o
g Sp ¼ n SWAM uai1 ; ~Sp1 ; uai2 ; f
UAI Sp 2 ; . . .; uain ; ~ Sn
j¼1
S n D E D Eo
f S ¼ n SWAM sv1 ; ~Sp1 ; sv2 ; f
SV Sp 2 ; . . .; sv n ; f
Sp n
j¼1
Sn nD E D E D Eo
f f
e Sp ¼ j¼1 SWAM m1 ; Sp 1 ; m2 ; Sp 2 ; . . .; mn ; Sp n
m f
S n D E o
~rS ¼ nj¼1 SWAM r1 ; ~Sp1 ; r2 ; f
Sp 2 ; . . .; rn ; ~Spn
4 Application
A wind energy investment in the Marmara Region of Turkey are evaluated by using the
proposed methodology. Three experts evaluate a wind tribune alternative, the weights
of the decision makers are 0.4, 0.4 and 0.2. The possible values of the alternatives and
the evaluations done by the experts are given in Table 1.
The aggregated values are defuzzified and summed by the proposed methodology.
Table 2 gives the defuzzified investment parameters.
The benefit B ~ Sp ¼ $10,052,820 and the cost C ~ Sp ¼ $6,291,507.84 of the wind
energy investment are calculated by using the proposed approach. By using the these
~ Sp
B
parameters benefit/cost ~ Sp
C
¼ 1.598 is calculated. This wind energy investment is a
favorable investment.
Table 1. Spherical fuzzy wind investment parameters.

Parameters Values Experts’ evaluations
E1 E2 E3
0.4 0.4 0.2
f
IF Sp $2,000,000 (0.8, 0.2, 0.2) (0.7, 0.3, 0.3) (0.6, 0.4, 0.4)
$2,500,000 (0.6, 0.4, 0.4) (0.8, 0.2, 0.2) (0.7, 0.3, 0.3)
$3,000,000 (0.4, 0.6, 0.4) (0.5, 0.5, 0.5) (0.6, 0.4, 0.4)
g Sp
UAC $330,000 (0.5, 0.5, 0.5) (0.6, 0.4, 0.4) (0.7, 0.3, 0.3)
$340,000 (0.4, 0.6, 0.4) (0.8, 0.2, 0.2) (0.4, 0.6, 0.4)
$350,000 (0.7, 0.3, 0.3) (0.7, 0.3, 0.3) (0.9, 0.1, 0.1)
g Sp
UAI $850,000 (0.4, 0.6, 0.4) (0.2, 0.8, 0.2) (0.8, 0.2, 0.2)
$900,000 (0.6, 0.4, 0.4) (0.5, 0.5, 0.5) (0.4, 0.6, 0.4)
$950,000 (0.3, 0.7, 0.3) (0.7, 0.3, 0.3) (0.6, 0.4, 0.4)
fS
SV $4,000 (0.5, 0.5, 0.5) (0.7, 0.3, 0.3) (0.5, 0.5, 0.5)
$5,000 (0.4, 0.6, 0.4) (0.8, 0.2, 0.2) (0.9, 0.1, 0.1)
$6,000 (0.7, 0.3, 0.3) (0.5, 0.5, 0.5) (0.7, 0.3, 0.3)
er S 5% (0.4, 0.6, 0.4) (0.4, 0.6, 0.4) (0.3, 0.7, 0.3)
6% (0.2, 0.8, 0.2) (0.4, 0.6, 0.4) (0.5, 0.5, 0.5)
7% (0.7, 0.3, 0.3) (0.7, 0.3, 0.3) (0.4, 0.6, 0.4)
e Sp
m 18 (0.5, 0.5, 0.5) (0.5, 0.5, 0.5) (0.9, 0.1, 0.1)
19 (0.8, 0.2, 0.2) (0.7, 0.3, 0.3) (0.6, 0.4, 0.4)
20 (0.2, 0.8, 0.2) (0.3, 0.7, 0.3) (0.5, 0.5, 0.5)
Table 2. Defuzzified values of the investment parameters.

Parameters Defuzzified values
f
IF Sp $2,508,536
g Sp
UAC $340,080
g Sp
UAI $900,154
fS
SV $4,986
er S 5.98%
e Sp
m 18.98
5 Conclusion and Further Suggestions
Wind energy investments are one of the fastest-increasing investment areas in

renewable energy. The installed wind capacity has been increasing not only due to the
increasing awareness on the sustainability and the consequent governmental supports
but also due to the decrease in the costs. Wind energy production is available and has
great potential in many parts of the world. Despite wind energy has huge potential,
wind energy investments are hard to evaluate. Fuzzy sets are excellent tools for dealing
with uncertainty [16, 17]. Spherical fuzzy sets, one of the newest extensions of fuzzy
sets enable us to deal with uncertainty and hesitancy. In this study, we propose a novel
spherical fuzzy cost/benefit analysis for evaluating uncertain wind energy investments.
The applicability of the model is shown by applying it to a wind energy investment
decision.
For further studies, several wind energy investments can be compared by using the
proposed cost/benefit analysis. The results of the study can be compared with the other
engineering economic analysis. A sensitivity analysis can be applied to see the effect of
decision maker weights in the evaluation results.
References
1. Wind Europe Report. https://windeurope.org/about-wind/reports/financing-and-investment-
trends-2019/#findings. Accessed 20 Apr 2020
2. Irene Stats. https://www.irena.org/publications/2019/May/Renewable-power-generation-
costs-in-2018#RestrictedModal. Accessed 20 Apr 2020
3. Davis, G., Owens, B.: Optimizing the level of renewable electric R&D expenditures using
real options analysis. Energy Policy 31(15), 1589–1608 (2003)
4. Siddiqui, A.S., Marnay, C., Wiser, R.H.: Real options valuation of US federal renewable
energy research, development, demonstration, and deployment. Energy Policy 35(1), 265–
279 (2007)
5. Cevik Onar, S., Kılavuz, T.N.: Risk analysis of wind energy investments in Turkey. Hum.
Ecol. Risk Assess. 21(5), 1230–1245 (2015)
6. Çoban, V., Onar, S.Ç.: Modelling solar energy usage with fuzzy cognitive maps. Intell. Syst.
Ref. Libr. 113, 159–187 (2017)
7. Çoban, V., Cevik Onar, S.: Energy management maturity model based on fuzzy logic. Adv.
Intell. Syst. Comput. 1029, 1034–1041 (2020)
8. Onar, S.C., Oztaysi, B., Otay, I., Kahraman, C.: Multi-expert wind energy technology
selection using interval-valued intuitionistic fuzzy sets. Energy 90, 274–285 (2015)
9. Çoban, V., Çevik Onar, S.: Pythagorean fuzzy engineering economic analysis of solar power
plants. Soft. Comput. 22, 5007–5020 (2018)
10. Kahraman, C., Onar, S.C., Oztaysi, B.: A comparison of wind energy investment alternatives
using interval-valued intuitionistic fuzzy benefit/cost analysis. Sustainability 8(2), 118
(2016)
12. Gündoğdu, F.K.: Principals of spherical fuzzy sets. Adv. Intell. Syst. Comput. 1029, 15–23
(2020)
13. Gündoğdu, F.K., Kahraman, C.: Extension of codas with spherical fuzzy sets. J. Multiple-
14. Gündoǧdu, F.K., Kahraman, C.: A novel spherical fuzzy analytic hierarchy process and its
renewable energy application. Soft. Comput. 24, 4607–4621 (2020)
15. Kahraman, C., Çevik Onar, S., Öztayşi, B.: Engineering economic analyses using
16. Estrella, F.J., Onar, S.C., Rodríguez, R.M., Oztaysi, B., Martínez, L., Kahraman, C.:
Selecting firms in University technoparks: a hesitant linguistic fuzzy TOPSIS model for
heterogeneous contexts. J. Intell. Fuzzy Syst. 33(2), 1155–1172 (2017)
17. Cevik Onar, S., Oztaysi, B., Kahraman, C.: Multicriteria evaluation of cloud service
providers using Pythagorean fuzzy TOPSIS. J. Multiple-Valued Logic Soft Comput. 30(2–
3), 263–283 (2018)
A Fuzzy Pricing Model for Mobile
Advertisements by Using Spherical
Fuzzy AHP Scoring
Basar Oztaysi(&), Sezi Cevik Onar, and Cengiz Kahraman
Industrial Engineering Department, İstanbul Technical University,

34367 İstanbul, Turkey
[email protected]
Abstract. Beacon technology enables applications that provide location-based

services such as notifications, tracking, and navigation via smart mobile devices.
One of the most commonly adopted business models based on beacon tech-
nology is location-based advertisements. In this model, target customers are
defined, and the most appropriate advertisement is sent to the customer based on
his/her location. While the conversion rates of this model are high, there is a
need for a dynamic pricing model that calculates the value of an advertisement
to an advertiser in a given location. In this study, we propose a price of
advertisement pricing model by using the Spherical Fuzzy AHP Scoring
method.
Keywords: Location-based systems AHP Scoring Spherical fuzzy sets
1 Introduction
As a result of the development of mobile technology, tools, and the role of marketing
has evolved. Understanding the needs and preferences of the customers has become
vital to provide efficient communication [1]. What is even more, personalized mar-
keting messages and recommendations have become very popular, and the customers
even want to have these messages within stores [2]. The companies started to integrate
online and offline channels to provide unified communication messages.
Location detection can be maintained by various different technologies, such as
GPS, RFID, WiFi, Beacon, ZigBee [3]. Today, advertising screenings on mobile media
can be displayed via SMS or on mobile applications as banners. What is more, cus-
tomized advertisements related to the location of users can be made. From the per-
spective of the advertiser, they can target individuals based on various segmentations,
time periods, and locations. As a result, this business model needs dynamic pricing
because the value of a single advertisement may be different for different advertisers.
This study aims to introduce a dynamic pricing model using fuzzy MCDM. Various
fuzzy MCDM methods have been applied to different decision models such as fields
such as health systems [4], engineering problems [5], quality management [6], public
transportation [7] firms selection [8], disaster response [9], urban transformation [10].
One of the most commonly used MCDM methods is the Analytic Hierarchy Process
A Fuzzy Pricing Model for Mobile Advertisements 143
(AHP) proposed by Saaty [11] for modeling complex decision models. AHP method is
based on the pairwise comparison; however, as the number of alternatives increases,
the decision matrices become huge, and the evaluations become inconsistent. AHP
Scoring method is an extension of AHP that can be used for scoring the alternatives
instead of making pairwise comparisons. The spherical fuzzy set is a new fuzzy
extension that is defined by three parameters; membership, non-membership and
hesitancy parameters. Each of the parameters can be defined between 0 and 1, but their
squared sum is at most equal to 1 [12]. With this structure, spherical fuzzy sets provide
a better and more flexible definition for linguistic terms. In this paper, we propose a
Spherical Fuzzy AHP Scoring method and apply it to the mobile advertisement pricing
model. By doing this, the traditional AHP method is extended to handle the vagueness
and uncertainty in linguistic definitions in the pricing problem.
The remainder of this paper is structured as follows: in the second section literature
review is given. The third section basically explains the methodology, and the fourth
section provides a real-world example with sample calculations. And finally, the
conclusions are given in the final section.
2 Literature Review
Dynamic pricing is a pricing approach in which businesses do not have fixed prices for
products and goods. Instead, the prices are determined based on current market
demands. Businesses generally use advanced algorithms to dynamically change prices
depending on competitor pricing, supply and demand, and other external factors in the
market. In hospitality, tourism, electricity industries, dynamic pricing is highly adop-
ted. Bitran and Caldentey [13] analyze the concept from the income management
concept and give an overview of the constraints and objectives. In income manage-
ment, the main goal is to obtain the most appropriate price policy by considering the
demand variability in a specific market, price elasticity, and competitor company
products. In the field of e-commerce, dynamic pricing refers to the pricing mechanisms
in which the price of a product is generated by an auction method or by the seller
evaluating the special offers from the customer [14]. The problem we focus on is
different from the two approaches mentioned above. The company wants to set a
pricing methodology where the price is determined based on a predefined factor
according to the decision-makers’ preferences. Thus we propose the Spherical fuzzy
AHP Scoring method for this problem.
Analytic Hierarchy Process is one of the most popular Multicriteria Decision
Making (MCDM) methods. AHP is introduced by Saaty [11] as a structured approach
for modeling and solving complex decision problems. This method allows constructing
the decision making criteria as a hierarchy, calculating the weights of the criteria and
alternatives, and it also stresses the consistency of the comparison of alternatives.
The traditional AHP method uses crisp numbers to represent decision-makers’
evaluations. The method is later extended to involve fuzzy logic for providing a
mathematical strength to capture the uncertainties accompanying the human cognitive
process [15]. The initial fuzzy AHP algorithm [16] uses triangular fuzzy membership
functions. Later Buckley [17] proposed AHP with trapezoidal fuzzy numbers and used
144 B. Oztaysi et al.
the geometric mean method to derive fuzzy weights. Chang [18] proposes using the
extent analysis method for the synthetic extent values of the pairwise comparisons by
utilizing triangular fuzzy numbers. In one of the recent studies In recent years, other
extensions of fuzzy sets are used to propose novel AHP methods. The novel fuzzy AHP
methods involve; fuzzy AHP with type 2 fuzzy sets [19, 20], intuitionistic fuzzy AHP
with intuitionistic fuzzy sets [21], fuzzy AHP with hesitant fuzzy sets [22, 23], fuzzy
AHP with interval-valued intuitionistic sets [24], neutrosophic AHP method [25, 26],
Pythagorean fuzzy AHP method [27], and Spherical Fuzzy AHP [28].
3 Spherical Fuzzy AHP Scoring
The spherical fuzzy set is proposed by Gundogdu and Kahraman [12] and involves
three parameters; membership, non-membership, and hesitancy parameters. Each of the
parameters can be defined between 0 and 1, but their squared sum is at most equal to 1.
~ S of the universe of discourse U1 be as
Let U1 be a universe. Let a spherical fuzzy set A
follows:
n o
~ S ¼ x; l ~ ðxÞ; m ~ ðxÞ; p ~ ðxÞ x 2 U1
A ð1Þ
AS AS AS
where
lA~ S ðxÞ : U1 ! ½0; 1; mA~ S ðxÞ : U1 ! ½0; 1; pA~ S ðxÞ : U1 ! ½0; 1
and
0 l2A~ S ðxÞ þ m2A~ S ðxÞ þ p2A~ S ðxÞ 18x 2 U1 ð2Þ
For each x, the lA~ S ðxÞ; mA~ S ðxÞ and pA~ S ðxÞ are the degrees of membership, non-
membership, and hesitancy of x to A ~ S , respectively.
In this study, a new spherical fuzzy AHP scoring method is proposed based on
Spherical Fuzzy AHP proposed by [28]. The steps of the proposed methodology are as
follows:
Step 1. The hierarchical structure is constructed. In this step, a hierarchical structure
consisting of at least three levels is developed.
Step 2. Pairwise comparisons are constituted using spherical fuzzy judgment
matrices based on the linguistic terms given in Table 1.
Step 3. The consistency of the pairwise comparison matrix is checked. In order to
do this, the linguistic terms in the pairwise comparison matrix are transformed into
their corresponding score indices. Then, the classical consistency check is applied.
Table 1. Linguistic measures of importance used for pairwise comparisons.

ðl; v; pÞ Score Index (SI)
Absolutely More Importance (AMI) (0.9, 0.1, 0.0) 9
Very High Importance (VHI) (0.8, 0.2, 0.1) 7
High Importance (HI) (0.7, 0.3, 0.2) 5
Slightly More Importance (SMI) (0.6, 0.4, 0.3) 3
Equally Importance (EI) (0.5, 0.4, 0.4) 1
Slightly Low Importance (SLI) (0.4, 0.6, 0.3) 1/3
Low Importance (LI) (0.3, 0.7, 0.2) 1/5
Very Low Importance (VLI) (0.2, 0.8, 0.1) 1/7
Absolutely Low Importance (ALI) (0.1, 0.9, 0.0) 1/9
Step 4. The spherical fuzzy local weights of criteria are calculated. The weight of
each criterion is obtained by using SWAM operator given in Eq. (3). The weighted
arithmetic mean is used to compute the spherical fuzzy weights.
D Q
SWAMw ðAS1 ; . . .. . .. . .; ASn Þ ¼ w1 AS1 þ w2 AS2 þ . . .. . . þ wn ASn ¼ ½1 ni¼1 ð1 l2ASi Þwi 1=2 ;
Q n wi Q n Qn E
2 wi 2 wi 1=2
i¼1 vASi ; ½ i¼1 ð1 lASi Þ i¼1 ð1 lASi pASi Þ
2
ð3Þ
Where w ¼ 1=n
Step 6. In AHP scoring, instead of comparing the pairwise comparisons, the pos-
sible interval values that the alternative may take is defined for each criterion.
Step 7. Step 4 is applied comparing the interval values with respect to each crite-
rion. To this end, a pairwise comparison matrix is formed for each criterion
involving the categories. After getting the evaluations from the experts, the weights
of the categories are defined.
Step 8. The hierarchical layer sequencing is established to obtain global weights.
The spherical fuzzy weights at each level are aggregated from the bottom level
(alternatives) to top-level (goal).
The difference from the classical AHP method, AHP Scoring method, provides a
formula composed of global weights of alternative categories in each criterion. In this
case, spherical fuzzy global preference weights are computed by using Eq. (4).
Q
n
~ Sij ¼ A
A ~ Si1 A
~ Si2 . . . A
~ Sin 8i
j¼1
12 ð4Þ
* +
lA~ S lA~ S ; v2A~ þ v2A~ v2A~ v2A~ ;
~ S11 A
i:e: A ~ S12 ¼ 11 12 S11 S12 S11 S12
1=2
1 v2A~ p2A~ þ 1 v2A~ p2A~ p2A~ p2A~
S12 S11 S11 S12 S11 S12
Step 9. The final global weights are defuzzified and normalize the values to obtain
the final equation for dynamic pricing.
4 Application
An Beacon technology provider in Turkey wants to develop a dynamic pricing system

for increasing the effectiveness of advertisements used in the shopping malls. This
technology provider is actively working with four big shopping malls located in
Istanbul. The shopping malls are dealing with more than 10 million customers every
year. The company seeks to determine the optimal price for each advertisement. To this
end, four criteria are determined. The criteria are as follows:
• Content Quality: Based on the quality of the message and visuals used in the
advertisement, a level of quality is determined by the area experts.
• Daily Potential: The potential of the advertisement represents the potential target
market of the selected advertisement. If the target market specified by the advertiser
is very specific, the daily potential will be low.
• The frequency of Communication: The criterion shows when the last time the
targeted person was last contacted with the selected campaign. If it is too soon, the
frequency will be high.
• User Interest: The probability of the target person to be interested in the content.
This part of the study comes from a machine learning model.
The intervals for each criterion is defined as Very Low, Low, Medium, High, and
Very High.
At the first step of the application, the four criteria are pairwise compared to
calculate their relative importance. The pairwise comparison matrix is given in Table 2;
the associated Spherical fuzzy sets are given in Table 3.
Table 2. Pairwise comparison of the criterion with respect to the goal.

C1 C2 C3 C4
C1 EI SLI SLI SLI
C2 SMI EI HI SLI
C3 SMI LI EI SMI
C4 VHI SMI SLI EI
The next step is to calculate the aggregated values by using SWAM method. Later
the defuzzified values are calculated, and finally, they are normalized, and the weights
of the criteria are obtained (Table 4).
Then the intervals defined for each criterion are pairwise compared. Due to page
limitations, only one of them is presented here. For Content Quantity five intervals (L1,
L2, L3, L4, L5) are defined. The pairwise comparison matrix for Content Quantity are
given in Table 5, and the associated spherical fuzzy sets are given in Table 6.
Table 3. Spherical fuzzy representations of the expert evaluations.

C1 C2 C3 C4
C1 (0.5, 0.5, 0.4) (0.4, 0.6, 0.3) (0.4, 0.6, 0.3) (0.4, 0.6, 0.3)
C2 (0.6, 0.4, 0.3) (0.5, 0.5, 0.4) (0.7, 0.3, 0.2) (0.4, 0.6, 0.3)
C3 (0.6, 0.4, 0.3) (0.3, 0.7, 0.2) (0.5, 0.5, 0.4) (0.6, 0.4, 0.3)
C4 (0.8, 0.2, 0.1) (0.6, 0.4, 0.3) (0.4, 0.6, 0.3) (0.5, 0.5, 0.4)
Table 4. Aggregated values, defuzzified values and normalized weights.

SWAM Defuzzified values Normalized weights
(0.428, 0, 0.332) 10.686 0.188
(0.571, 0, 0.303) 15.332 0.270
(0.522, 0, 0.315) 13.737 0.242
(0.619, 0, 0.277) 16.947 0.299
Table 5. Pairwise comparison of the interval value with respect to C1.

L1 L2 L3 L4 L5
L1 EI HI HI VHI AMI
L2 LI EI SMI HI AMI
L3 LI SLI EI SMI HI
L4 VLI LI SLI EI HI
L5 ALI ALI LI LI EI
Table 6. Spherical Fuzzy Representations of the expert evaluations.

L1 L2 L3 L4 L5
L1 (0.5, 0.5, 0.4) (0.7, 0.3, 0.2) (0.7, 0.3, 0.2) (0.8, 0.2, 0.1) (0.9, 0.1, 0)
L2 (0.3, 0.7, 0.2) (0.5, 0.5, 0.4) (0.6, 0.4, 0.3) (0.7, 0.3, 0.2) (0.9, 0.1, 0)
L3 (0.3, 0.7, 0.2) (0.4, 0.6, 0.3) (0.5, 0.5, 0.4) (0.6, 0.4, 0.3) (0.7, 0.3, 0.2)
L4 (0.2, 0.8, 0.1) (0.3, 0.7, 0.2) (0.4, 0.6, 0.3) (0.5, 0.5, 0.4) (0.7, 0.3, 0.2)
L5 (0.1, 0.9, 0) (0.1, 0.9, 0) (0.3, 0.7, 0.2) (0.3, 0.7, 0.2) (0.5, 0.5, 0.4)
Then the aggregated values are obtained by using SWAM method, and the results
are defuzzified and normalized (Table 7).
The steps are applied for the criteria, and the interval values are obtained as given in
Table 8.
Table 7. Aggregated values, defuzzified values and normalized weights of the intervals.
SWAM Defuzzified values Normalized weights
(0.76, 0.246, 0.187) 21.87 1
(0.685, 0.335, 0.229) 19.386 0.886
(0.534, 0.479, 0.291) 14.544 0.665
(0.474, 0.55, 0.269) 12.865 0.588
(0.307, 0.724, 0.24) 7.928 0.362
Table 8. Criteria weights and the priorities of the intervals.

C1 C2 C3 C4
Weights 0.188 0.270 0.242 0.299
Very high 1 1 1 1
High 0.886 0.915 0.734 0.898
Medium 0.665 0.687 0.653 0.667
Low 0.588 0.618 0.545 0.598
Very low 0.362 0.374 0.401 0.366
The results in Table 8 are used to find the dynamic pricing. For example, if an
advertisement has the values L1 = Medium, L2 = High, L3 = Very High, L4 = High
the price parameter is calculated by summing up the weighted scores as given in the
following.
0:118 0:665 þ 0:270 0:915 þ 0:242 1 þ 0:299 0:898 ¼ 0:836
The final price can be found by multiplying this parameter with the predefined
maximum price.
5 Conclusions
In this study, we focus on mobile advertisement dynamic pricing by using Spherical

Fuzzy AHP Scoring. The model is based on four main criteria namely; the content
quality, daily potential, frequency of communication and user interest. As a future
study, this study can be extended by other criteria. Besides, the same problem can be
solved by using other multicriteria decision-making techniques, and the results can be
compared.
References
1. Celikkan, U., Somun, G., Kutuk, U., Gamzeli, I., Cinar, E.D., Atici, I.: Capturing
supermarket shopper behavior using smartbasket. In: Snasel, V., Platos, J., El-Qawasmeh, E.
(eds.) Digital Information Processing and Communications, pp. 44–53. Springer, Berlin
(2011)
2. Merad, D., Aziz, K.E., Iguernaissi, R., Fertil, B., Drap, P.: Tracking multiple persons under
partial and global occlusions: application to customers’ behavior analysis. Pattern Recogn.
Lett. 81, 11–20 (2016)
3. Dogan, O., Öztaysi, B.: In-store behavioral analytics technology selection using fuzzy
decision making. J. Ent. Inf. Manage. 31(4), 612–630 (2018)
4. Otay, İ., Oztaysi, B., Onar, S.C., Kahraman, C.: Multi-expert performance evaluation of
healthcare institutions using an integrated intuitionistic fuzzy AHP & DEA methodology.
Knowl. -Based Syst. 133, 90–106 (2017)
5. Kahraman, C., Gülbay, M., Kabak, Ö.: Applications of fuzzy sets in industrial engineering: a
topical classification. In: Kahraman, C. (ed.) Fuzzy Applications in Industrial Engineering,
1–55. Springer, Heidelberg (2006)
6. Kaya, İ., Kahraman, C.: Development of fuzzy process accuracy index for decision making
problems. Inf. Sci. 180(6), 861–872 (2010)
7. Kaya, İ., Oztaysi, B., Kahraman, C.: A two-phased fuzzy multicriteria selection among
public transportation investments for policy-making and risk governance. Int. J. Uncertainty
Fuzziness Knowl. Based Syst. 20, 31–48 (2012)
8. Estrella, F.J., Cevik Onar, S., Rodríguez, R.M., Oztaysi, B., Martínez, L.: Selecting firms in
university technoparks: a hesitant linguistic fuzzy TOPSIS model for heterogeneous
contexts. J. Intell. Fuzzy Syst. 33(2), 1155–1172 (2017)
9. Oztaysi, B., Behret, H., Kabak, Ö., Sarı, I.U., Kahraman, C.: Fuzzy inference systems for
disaster response. In: Vitoriano, B., Montero, J., Ruan, D. (eds.) Decision Aid Models for
Disaster Man and Emergencies, pp. 75–94. Atlantis Press, Paris (2013)
10. Oztaysi, B., Cevik Onar, S., Kahraman, C.: Fuzzy multicriteria prioritization of Urban
transformation projects for Istanbul. J. of Int. Fuzzy Syst. 30(4), 2459–2474 (2016)
11. Saaty, T.L.: The Analytic Hierarchy Process. McGraw-Hill, New York (1980)
13. Bitran, G., Caldentey, R.: An overview of pricing models for revenue management. Manuf.
Serv. Oper. Manage. 5, 203–229 (2003)
14. Turban, E., King, D., Lee, J.K., Liang, T.P., Turban, D.C.: Electronic Commerce: A
Managerial and Social Networks Perspective, 8th edn. Springer, Switzerland (2015)
15. Kahraman, C., Kaya, İ.: A fuzzy multicriteria methodology for selection among energy
alternatives. Expert Syst. Appl. 37(9), 6270–6281 (2010)
16. Van Laarhoven, P.J., Pedrycz, W.: A fuzzy extension of Saaty’s priority theory. Fuzzy Sets
Syst. 11(1–3), 229–241 (1983)
17. Buckley, J.J.: Fuzzy hierarchical analysis. Fuzzy Sets Syst. 17(3), 233–247 (1985)
18. Chang, D.Y.: Applications of the extent analysis method on fuzzy AHP. Eur. J. Oper. Res.
95(3), 649–655 (1996)
19. Kahraman, C., Öztayşi, B., Sarı, İ.U., Turanoğlu, E.: Fuzzy analytic hierarchy process with
interval type-2 fuzzy sets. Knowl.-Based Syst. 59, 48–57 (2014)
20. Oztaysi, B., Onar, S.C., Kahraman, C.: Prioritization of business analytics projects using
interval type-2 fuzzy AHP. In: Kacprzyk, J., Szmidt, E., Zadrożny, S., Atanassov, K.,
Krawczak, M. (eds.) Advances in Fuzzy Logic and Technology, pp. 106–117. Springer,
Cham (2017)
21. Wu, J., Huang, H.B., Cao, Q.W.: Research on AHP with interval-valued intuitionistic fuzzy
sets and its application in multicriteria decision making problems. Appl. Math. Model. 37
(24), 9898–9906 (2013)
22. Oztaysi, B., Onar, S.Ç., Boltürk, E., Kahraman, C.: Hesitant fuzzy analytic hierarchy
process. In: 2015 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), pp. 1–7.
IEEE, Istanbul (2015)
23. Bolturk, E., Çevik Onar, S., Öztayşi, B., Kahraman, C., Goztepe, K.: Multi-attribute
warehouse location selection in humanitarian logistics using hesitant fuzzy AHP. Int.
J. Anal. Hierarchy Process 8, 271–298 (2016)
24. Tooranloo, H.S., Iranpour, A.: Supplier selection and evaluation using interval-valued
intuitionistic fuzzy AHP method. Int. J. Procurement Manage. 10(5), 539–548 (2017)
25. Abdel-Basset, M., Mohamed, M., Sangaiah, A.K.: Neutrosophic AHP-Delphi Group
decision making model based on trapezoidal neutrosophic numbers. J. Ambient Intell.
Humaniz. Comput. 9(5), 1427–1443 (2017)
26. Bolturk, E., Kahraman, C.: A novel interval-valued neutrosophic AHP with cosine similarity
measure. Soft. Comput. 22(15), 4941–4958 (2018)
27. Ilbahar, E., Karaşan, A., Cebi, S., Kahraman, C.: A novel approach to risk assessment for
occupational health and safety using Pythagorean fuzzy AHP & fuzzy inference system. Saf.
Sci. 103, 124–136 (2018)
28. Kutlu Gündoğdu, F., Kahraman, C.: A novel spherical fuzzy analytic hierarchy process and
its renewable energy application. Soft. Comput. 24(6), 4607–4621 (2019)
A Framework for Selection of the Best Food
Waste Management Alternative by a Spherical
Fuzzy AHP Based Approach
Aysu Melis Buyuk(&) and Gul Tekin Temur
Faculty of Engineering and Natural Sciences, Department of Management

Engineering, Bahcesehir University, 34353 Besiktas, Turkey
[email protected],
[email protected]
Abstract. In today’s sustainable business environment, managers and

researchers have been attached high importance on food waste management.
Especially for developing countries, the readiness level is comparatively low
belongs to policies and low consciousness. That forces all stakeholders in the
food waste supply chain benefit from some methods that can be beneficial for
improving their decision-making skills. Because this decision needs to handle
multi-faceted dimensions simultaneously under linguistic and uncertain envi-
ronment, a fuzzy based multi criteria decision making (MCDM) approach-based
methodology is proposed in this study. As a novel fuzzy approach, spherical
fuzzy sets (SFS) are utilized to select the best alternative. In the proposed model,
SFSs are combined with a traditional MCDM approach, analytic hierarchy
process (AHP) with a name of Spherical Fuzzy AHP. A numerical example is
conducted in order to indicate how the model can be used for food waste
management. By the help of the proposed model, it is aimed to create a sup-
portive tool for decision makers in the related institutions such as municipalities.
Keywords: Food waste Fuzzy sets Multi criteria decision making

Spherical fuzzy sets Spherical Fuzzy Analytic Hierarchy Process
1 Introduction
Food waste management (FWM) has become an important issue today, which is a sub-
part of Solid Waste Management. Food waste can be defined as a reduction in the
amount or quality of food as a result of decisions and acts taken by retailers, food
service providers and buyers. When global consumption is considered, 1.3 tons food
wasted, which equals to one third of all comestible food [1]. Food waste management
includes minimization of food wastes, transporting, recycling and disposal operations.
The main hierarchy of FWM starts with prevention and decreasing the amount of food
wasted as the primary method and followed by reaching food to people and animals
[2]. If these are unavailable to apply, main treatment options such as Composting,
Landfilling, Anaerobic Digestion and Incineration can be chosen. As a result of the
decreasing level of natural resources, waste management has become more visible and
technological developments in this area lead people to provide more efficient methods
152 A. M. Buyuk and G. T. Temur
to use in food waste [3]. Ineffective ways of managing food waste process can cause
negative consequences such as increased cost and impact on environment and health. In
this research, the main aim is to provide a framework that is helpful for decision makers
to select the best food waste management treatment with a consideration of a set of
criteria. This selection needs to handle multi-faceted conflicting criteria simultaneously,
therefore a multi criteria decision making (MCDM) approach is beneficial to use. In the
proposed framework, a novel fuzzy method named Spherical Fuzzy Analytic Hierarchy
Process (SFAHP) is recommended in order to deal with the uncertainty caused from
linguistic evaluations of the criteria set. This study contributes to the relevant literature
by (1) offering comprehensive criteria set for food waste treatment selection, and
(2) using SFAHP as an evolved fuzzy approach as a first time for food waste
management.
In the last decade, with the improvement of new methods and updated old methods,
the role of MCDM in different fields is increased [4]. Over the last decade, researchers
have contributed to the food waste management process by proposing several decision
approaches. One of them is the assessment of food waste management treatments. In
most of the researches, criteria are defined considering life cycle assessment (LCA) to
examine the environmental effects of methods [5]. LCA is also considered to compare
two main methods named composting or anaerobic digestion and scenarios created
from these alternatives [6]. In another study, food waste management treatment options
are chosen by using the AHP method considering environmental socio-cultural, tech-
nical, and economic criteria [7]. Preference Ranking Organization Method for
Enrichment Evaluations (PROMETHEE) is another MCDM method used with eco-
nomic, environmental, social and technological criteria specifically for in-sink food
wastes [8]. In the literature, some researches are found that food waste handled as a
category of solid waste. In one of these researches, solid waste disposal scenarios are
created and ranked by using Technique for Order Preference by Similarity to Ideal
Solution (TOPSIS) [9]. Also in some cases, MCDM is used for location selection of
waste disposal facility [10]. It is observed that although there is a high interest on food
waste management, there is a lack of study that deals with vagueness originated from
subjective assessments. Therefore, in this study, a framework including Spherical
Fuzzy AHP approach integrated MCDM model is proposed.
The rest of this study is organized as follows. Section 2 explains the treatment
alternatives. Section 3 explains the criteria set. Section 4 and 5 develops the SFAHP
based model for the selection. Section 6 concludes the paper and offer some future
directions.
2 Food Waste Management Treatments

2.1 Composting
Composting is one of the major options that is used as a food waste management treat-
ment. In the biological environment, composting can be explained as the emergence of
compost material that is soil-like. Compost occurs when organic materials break down by
microorganisms, snails, insects and earthworms in the existence of oxygen [11].
A Framework for Selection of the Best Food Waste Management Alternative 153
2.2 Anaerobic Digestion

Anaerobic digestion is a complex microbial process that contains a series of metabolic
reactions. In these reactions, with the breakdown of organic matter, food wastes turn
into biogas and organic fertilizer. Hence, the end of the process methane occurs, it can
be considered as one of the potent renewable energy sources [12].
2.3 Landfilling
Landfills are areas where food wastes are stored and disposed of in a traditional way.
The outcomes of landfilling are gas and leachate generation resulting from microbial
decomposition, climatic conditions, characteristics of refuse and operations [13].
2.4 Incineration
Incineration is an alternative option that recover energy from food waste, which used
for producing electricity, steam and heating in energy plants. It can be defined as the
controlled burning of solid, liquid, or gaseous combustible wastes [14].
3 Criteria Set of Decision Model
For selection of the best food waste treatment option, 13 sub-criteria are determined
which is classified under 5 main criteria as follows (the hierarchical structure of criteria
set is given in Fig. 1):
Infrastructural Criteria (C1):

• Capacity (C11) represents the amount of waste that can be disposed in the facility
[15].
• Adaptability to System (C12) explains how compatible the alternative is [7].
• Land Usage Efficiency (C13) explains comparison the amount of land use and
amount of food waste be treated [5].
• Work Safety (C14) explains work environment of employees and technical safety of
disposal alternative.
• Technological Maturity (C15) explains the current technological status of country
and ability to combine with disposal alternative [5].
Governmental Criteria (C2):
• Law & Regulations (C21) explains what are the regulations and limitation that can
affect the result of selection of treatment.
• Political Support (C22) refers if a treatment method got support from government
or political institutions.
Fig. 1. The hierarchical structure of criteria set.
Economic Criteria (C3):

• Initial Investment Cost (C31) includes set up cost of disposal technologies and
expenses made in the opening of disposal area [15].
• Collection & Transportation Cost (C32) includes collection of wastes and trans-
portation cost to center for selected treatment [15].
• Operational Cost (C33) includes regular expenses made for process of disposal and
maintenance of tools and technologies [15].
Environmental Criteria (C4):
• Air & Water Pollution (C41) explains effects of treatments to air and water pollution
with a consideration of gasses and leachate they generated [7].
• Ecological Impact (C42) explains what the consequences of treatments are to living
organisms like humans, plants and animals and their balance with environment [5].
• Energy Consumption (C43) explains the amount of energy consumed by applica-
tions of each treatments.
4 Research Methodology
In this study, Spherical Fuzzy AHP (SFAHP), which is firstly introduced by Kutlu
Gundogdu and Kahraman [16], is utilized. The reason for selecting SFAHP is, in
spherical fuzzy sets (SFS), the decision makers are allowed to characterize all the
parameters in a larger space. That increases the strength of the methodology against the
uncertainties.
Following phases are tracked within the given methodology:

Phase 1: Pairwise comparisons of criteria set are done by conducting surveys.
Surveys are evaluated by experts. Specifically designed linguistic variables given by
Kutlu Gundogdu and Kahraman [16] is utilized for measurement.
Phase 2: The consistencies of evaluations are checked Firstly, by using score
indices (SI) given in Table 1, the matrices are defuzzied. Then, the consistency ratio
(CR) is calculated by using Eqs. (1), and (2) (CI: consistency index; kmax : the largest
eigenvector, n: the number of criteria, RI: random index);
kmax n
CI ¼ ð1Þ
n1
CI
CR ¼ ð2Þ
RI
If CR is higher or equal than 0.1, the evaluations are considered inconsistent. The
surveys have to be conducted again until it becomes lower than 0.1.
Phase 3: Spherical Weighted Arithmetic Mean (SWAM) is used to aggregate the
assessments of experts. The operator is also utilized to obtain fuzzy local weights of
main and sub criteria. Calculation of SWAM operator is given in Eq. (3) with respect
to weights (lA~ ; vA~ ; and pA~ address membership, non-membership, and hesitancy
~
degrees, respectively, to A).
~ ~ ~ ~ ~
Pn SWAMW ðAS1 ; . . .; ASn Þ ¼ w1 AS1 þ w2 AS2 þ . . . þ wn ASn , (w1 2 ½0; 1, and
i¼1 w1 ¼ 1)
8" #1 " #1 9
< Yn wi 2 Y n Yn wi Yn wi 2 =
1 1 l2A~ si ; vAw~ i ; 1 l2A~ si 1 l2A~ si p2A~ si
: i¼1 i¼1
Si
i¼1 i¼1
;
ð3Þ
Phase 4: For defuzzification of fuzzy weights, the score function (S) has been used
(shown in Eq. (4)) and the obtained values are normalized by dividing each value by
total value (shown in Eq. (5)).
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi

pA~ S 2 vA~ S 2
~ sj ¼
S w 100 3l ~ pA~ S ð4Þ
AS
2 2

S w~ sj
s

wj ¼ wj ¼ P ð5Þ
n
j¼1 S ~ sj
w
Phase 5: The local weights are multiplied with the weights of corresponding parent
criteria. The obtained weights are global, and it is ready to use for alternative
evaluation.
The same process is repeated for evaluation of alternatives, and the results of
alternatives are multiplied with global weights to find the prioritization.
5 Selection Procedure
In order to show how the proposed methodology can be used in practice, a numerical
example is generated. Assume that three experts from food waste industry have been
selected and the survey comparing all levels of hierarchy (main criteria, sub criteria and
alternatives) is conducted. After the experts have a consensus on pairwise comparisons,
the consistency is checked (as mentioned at Phase 2). New iterations are applied if the
matrices are inconsistent. Assume that the consensually agreed pairwise comparisons
~ s (spherical fuzzy values), S w
are as given in Table 1 and Table 2. The values of w ~ sj
s (crisp values) are also given in the table. A compu-
(value of score function) and w
tational example that uses Eqs. (3–5) is shown for main criteria C1 as follows:
Table 1. Pairwise comparisons of main criteria.

C1 C2 C3 C4
C1 0.5 0.4 0.4 0.7 0.3 0.2 0.5 0.4 0.4 0.6 0.4 0.3
C2 0.3 0.7 0.2 0.5 0.4 0.4 0.3 0.7 0.2 0.4 0.6 0.3
C3 0.5 0.4 0.4 0.7 0.3 0.2 0.5 0.4 0.4 0.6 0.4 0.3
C4 0.4 0.6 0.3 0.6 0.4 0.3 0.4 0.6 0.3 0.5 0.4 0.4
CR = 0.016
Table 2. Pairwise comparisons of main criteria (cont.).

~s
w Sw~ SJ s
w
C1 0.59 0.37 0.33 15.93 0.29
C2 0.39 0.59 0.30 10.13 0.18
C3 0.59 0.37 0.33 15.93 0.29
C4 0.49 0.49 0.33 12.93 0.24
CR = 0.016
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

1 1 1 1
lC1 ¼ 1 ð1 0:52 Þ4 ð1 0:72 Þ4 ð1 0:52 Þ4 ð1 0:62 Þ4 ¼ 0:59
1 1 1 1
vC1 ¼ 0:44 0:34 0:44 0:44 ¼ 0:37
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u 1 1
ð1 0:52 Þ4 ð1 0:72 Þ4 ð1 0:52 Þ4 ð1 0:62 Þ4
1 1
pC1 ¼t 1 1 1 1 ¼ 0:33
ð1 0:52 0:42 Þ4 ð1 0:72 0:22 Þ4 ð1 0:52 0:42 Þ4 ð1 0:62 0:32 Þ4
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
u u
" 2 2 #
0:33 0:37
~ sC1 ¼ t100
S w ð3 0:59Þ 0:33 ¼ 15:93
2 2
15:93
~ SC1 ¼
w ¼ 0:9
ð15:93 þ 10:13 þ 15:93 þ 12:93Þ
Similar steps are followed for evaluation of sub criteria. As an example, pairwise
comparisons of economic sub criteria are shared in Table 3. The local and global
weights are also given in Table 4.
As it is expected, economic and infrastructural considerations are found as the most
important criteria. It is good to see that weight of environmental criteria is very close to
these criteria. It is also surprising that although governmental criteria is the least
important one, its subcriterion titled “law & regulations” has the highest importance
Table 3. Pairwise comparisons of economic sub criteria.

C31 C32 C33 ~s
w Sw~ SJ s
w
C31 0.5 0.4 0.4 0.7 0.3 0.2 0.6 0.4 0.3 0.61 0.36 0.30 16.80 0.41
C32 0.3 0.7 0.2 0.5 0.4 0.4 0.4 0.6 0.3 0.41 0.55 0.29 10.91 0.26
C33 0.4 0.6 0.3 0.6 0.4 0.3 0.5 0.4 0.4 0.51 0.46 0.34 13.60 0.33
CR = 0.034
among all sub criteria. In order to complete the selection, same procedure is completed
for food waste treatment alternatives (composting (A1), landfilling (A2), anaerobic
digestion (A3) and incineration (A4)). Each alternative is evaluated pairwisely in terms
of sub criteria. Based on the calculations, anaerobic digestion (A3) is selected as the
Table 4. The weights of main and sub criteria.

Main criteria Local weights of Sub Local weights of Global weights of
main criteria criteria sub criteria sub criteria
Infrastructural 0.29 C11 0,25 0,073
(C1) C12 0,13 0,037
C13 0,17 0,048
C14 0,21 0,061
C15 0,24 0,071
Governmental 0.18 C21 0,66 0,122
(C2) C22 0,34 0,063
Economic (C3) 0.29 C31 0,41 0,118
C32 0,26 0,077
C33 0,33 0,095
Environmental 0.24 C41 0,44 0,103
(C4) C42 0,32 0,075
C43 0,24 0,057
best treatment option with the weight of 0.28. The total rank of alternatives are as
follows: A3 > A4 > A2 > A1.
6 Conclusion
In this study, a framework using SFAHP based model is proposed for selection of the
best food waste treatment option. To our knowledge, it is the first study that deals with
the vagueness in food waste treatment option selection problem area with SFS. Within
the model, four main criteria including infrastructural, governmental, economic and
environmental criteria; and their thirteen sub criteria are chosen belongs to relevant
literature. A numerical example is studied in order to indicate how the model can be
used in practice. Economic and infrastructural criteria are found as the most important
criteria. Among options, composting is selected as the best. In further studies, a real
application on food waste treatment option selection can be done and a sensitivity
analysis can be conducted in order to reveal how the output of the model changes
belongs to changes on evaluations.
References
1. FAO: Global food losses and food waste – Extent, causes and prevention (2011)
2. U.S EPA: Food Recovery Hierarchy. Retrieved from United States Environmental
Protection Agency (2013).https://www.epa.gov/sustainable-management-food/food-
recovery-hierarchy
3. Balaban, Y., Birdoğan, B.: Analitik Ağ Süreci Yaklaşımıyla En Uygun Katı Atık Bertaraf
Sisteminin Belirlenmesi: Trabzon İli Örneği. Atatürk Üniversitesi, pp. 183–194 (2010)
4. Velasquez, M., Hester, P.: An analysis of multi-criteria decision making methods. Int.
J. Oper. Res. 10(2), 56–66 (2013)
5. Hung, M.L., Ma, H.W., Yang, W.F.: A novel sustainable decision making model for
municipal solid waste management. Waste Manag. 27(2), 209–219 (2007)
6. Bernard, A., Jansen, J.: A life cycle approach to the management of household food waste –
a Swedish full-scale case study. Waste Manag. 31(8), 1879–1896 (2011)
7. Bababola, M.: A multi-criteria decision analysis of waste treatment options for food and
biodegradable waste management in Japan. Environments 2(4), 471–488 (2015)
8. Erceg, O., Margeta, J.: Selection of Food Waste Management Optıon by Promethee Method
(2019)
9. Coban, A., Ertis, İ., Cavdaroglu, N.: Municipal solid waste management via multi-criteria
decision making. J. Clean. Prod. 180, 159–167 (2018)
10. Mourmouris, J.C.: MCDM for location and evaluation of waste disposal facilities in
metropolitan areas: the case of East Africa. Int. J. Environ. Technol. Manage. 6(1–2), 180–
192 (2006)
11. Ellen Macarthur Foundation.: Towards the circular economy, vol. 2: opportunities for the
consumer goods sector (2013)
12. Kumar, A., Samadder, S.: Performance evaluation of anaerobic digestion technology for
energy recovery from organic fraction of municipal solid waste: a review. Energy 197,
117253 (2020)
13. El-Fadel, M., Findikakis, A., Leckie, J.: Environmental impacts of solid waste landfilling.
J. Environ. Manage. 50(1), 1–25 (1997)
14. Leal, F., Kovaleva, M.: Methods of food waste reduction. In: Food Waste and Sustainable
Food Waste Management in the Baltic Sea Region, pp. 51–80. Springer (2015)
15. Arıkan, E., Simsit, Z., Vayvay, O.: Solid waste disposal methodology selection using
MCDM methods and an application in Turkey. J. Clean. Prod. 142, 403–412 (2017)
16. Kutlu Gundogdu, F., Kahraman, C.: A novel spherical fuzzy analytic hierarchy process and
its renewable energy application. Soft. Comput. 24, 4607–4621 (2020)
Using Spherical Fuzzy AHP Based Approach
for Prioritization of Criteria Affecting
Sustainable Supplier Selection
Yagmur Unal(&) and Gul T. Temur
Faculty of Engineering and Natural Sciences, Department of Engineering

Management, Bahcesehir University, Besiktas, Istanbul, Turkey
[email protected],
[email protected]
Abstract. In today’s highly compeller business environment, the issue of

sustainable supplier selection has received considerable attention due to
changing face of business and rising awareness on both social and environ-
mental considerations. The companies can have the potential to execute their
core competencies efficiently and in sustainable manner, but it does not guar-
antee the success of the good or service that they offer; it also depends on the
strength of the other actors in the chain. As a field of study, selection of sus-
tainable suppliers has remained popular, because it needs to develop a sys-
tematic procedure dealing with the conflicting quantitative and qualitative
criteria simultaneously. Several tools and methodologies can enhance firms’
efficiency for solution of this problem. Nevertheless, there is a lack of sus-
tainable supplier selection studies that use an evolved fuzzy approach to our
knowledge. In this study, to help the advancement of some limited tools, one of
the evolved fuzzy approach named Spherical Fuzzy Sets is utilized for priori-
tizing the criteria affecting sustainable supplier selection. The proposed
methodology integrates Spherical Fuzzy Sets and analytical hierarchy process
(AHP) and take into consideration four main criteria: traditional, economic,
social and environmental criteria. Interviews are conducted with three experts
who have valuable experiences on this study field. The figured out ranking of
the criteria can be used as a guide by the researchers and industrial experts who
are responsible for sustainable supplier selection.
Keywords: Analytic Hierarchy Process (AHP) Spherical Fuzzy Sets

Sustainable supplier selection
1 Introduction
In today’s global environment, management of supply chain is highly critical for

organizations. An organization’s supply chain has a big influence on its day-to-day
operations, as it is the integrated distribution network between the actors. Due to its
integrated structure, the effectiveness of supply chain and satisfaction of customers are
highly dependent on suppliers [1]. With the competitive and globalizing structure of
production, supply chain network seems to be strengthened by the selection of eligible
Using Spherical Fuzzy AHP Based Approach for Prioritization 161
supplier in order to deal with issues relevant to the supply chain, from simple to
complicated [2, 3].
As stated by Kannan [4], supply chain success and the performance of suppliers are
strongly correlated in practice; the improved performance of the supplier tends to
increase the productivity of the whole supply chain [4]. Correspondingly, the man-
agement of appropriate supplier selection criteria for an organization is vital due to its
profound influence on the public image in an industry [5, 6].
As companies begin to seek more improvements within their organization, they
start to understand the need to consider emerging trends on global scale, such as
sustainability. Sustainable business strategies have begun to be applied both internally
and externally through the efforts on finding appropriate business partners [5, 7] in the
face of governmental regulations and increasing awareness of society [6, 8]. Supplier
contribution for the implementation of the sustainability framework in supply chain
activities is essential as suppliers play a significant role in accomplishments of the
entire supply chain [2, 9]. Sustainable supplier selection thus becomes more of an issue
to enable organizations to deal with environmental, economic and social challenges.
Supplier selection problem is a multi-criteria decision making (MCDM) problem
intrinsically. MCDM approaches are useful methods for dealing with this problem in
practice [9], because it enables to overcome the conflicting criteria affecting the
selection procedure. However, the linguistic nature of MCDM approaches is a chal-
lenge, as it is difficult to quantify [10]. According to a systematic review of literature
undertaken by Govindan et al. [11], fuzzy analyses prevail among various different
approaches of MCDM in the selection of sustainable suppliers in order to deal with the
linguistic terms. The potential of fuzzy logic to deal with qualitative and vague con-
cepts and its ability to produce concrete outputs (crisps) with the aid of its mathematical
tools further illustrates the widespread use of fuzzy approaches in existing literature [1].
In addition, eligibility of suppliers regarding sustainability issues is hard to analyze due
to vague and complex nature of sustainability framework.
As a consequence of these stated features of sustainable supplier selection process,
fuzzy set approaches seem to be more satisfactory for researchers. Therefore, in this
study, a model for prioritizing the criteria affecting supplier selection process is pro-
posed. In the proposed model, a novel approach titled spherical fuzzy analytic hier-
archy process (SFAHP) is utilized. To our knowledge, there is no study on sustainable
supplier selection process using SFAHP. In order to rank the criteria, surveys were
conducted with the experts which have valuable expertise. The next part of the paper
will also provide further information on why fuzzy approaches match the sustainable
supplier selection challenge quite well by showing the existing contributions to the
literature.
The reminder of this article is structured as follows. Section 2 illustrates the rele-
vant literature. Section 3 explains the determination of the criteria. Section 4 gives the
details about the proposed methodology. Section 5 computes the weights of the criteria
and reveals their prioritization. Section 6 summarizes the study and gives recom-
mendations for further studies.
162 Y. Unal and G. T. Temur
2 Literature Review
The selection of a convenient supplier is essential but has high degree of complexity
due to inclusion of numerous new factors [5]. Because of the inclusion of the vague
sustainability factor, this decision has become even more compelling to resolve [1] and
therefore, fuzzy based approaches are efficient to deal with subjectivity of such kind of
problems. Therefore, a brief overview of the intersection between Fuzzy Set applica-
tions and Sustainable Supplier Selection problem is given in this section with the
revision of existing literature. Since the problem of supplier selection seemed to be
gradually addressed from a sustainable viewpoint over last decade, we have revisited
the work of the last 10 years, between 2009–2020.
Based on the literature review findings provided by [1, 4, 11, 12]. Fuzzy set theory
appeared to be extended with the integration of other tools such as ANP, AHP,
DEMATEL, TOPSIS, AHP-VIKOR and Fuzzy Axiomatic Design (FAD). In the light
of these literature review sections from various papers, it is safe to state that fuzzy set
approaches with different tools seemed to be one of the most common ones utilized in
sustainable/green supplier selection. Due to the limited space, reviewed papers related
with supplier selection that are using fuzzy sets are summarized in Table 1.
Table 1. Literature review on fuzzy MCDM approaches used for sustainable supplier selection.
Author(s) Proposed fuzzy approach
Lee et al. [13] Fuzzy extended AHP (FEAHP)
Büyüközkan and Çifçi [3] Fuzzy extension of ANP method
Govindan et al. [9] Fuzzy TOPSIS
Shen et al. [14] Fuzzy TOPSIS
Kannan et al. [15] Fuzzy TOPSIS
Kannan et al. [16] Fuzzy Axiomatic Design (FAD)
Gupta and Barua [17] Fuzzy TOPSIS
Ghadimi et al. [1] Fuzzy Inference System
Khan et al. [7] Fuzzy Shannon Entropy
Fuzzy Inference System
Awashi et al. [18] Fuzzy AHP
Fuzzy VIKOR
Tirkolaee et al. [12] Fuzzy ANP
Fuzzy DEMATEL
Fuzzy TOPSIS
As it is seen from the literature review, ordinary fuzzy sets have been extended and
have been used in various research papers. In this paper, we adopt the Spherical Fuzzy
approach, which is one of new and lately extended versions of fuzzy logic.
3 Determination of the Criteria
In the proposed model, belongs to literature review, four main criteria are selected for
sustainable supplier selection: economic (C1), quality (C2), social (C3) and environ-
mental (C4).
3.1 Economic Criteria (C1)

The goal is to assess the capacity of the supplier to regulate cost and pricing practices
and to understand the competitive position of the supplier in the industry in order to
measure the economic efficiency of the supplier. It includes traditional cost (C1) and
financial reliability in sector (C12).
3.2 Quality Criteria (C2)

The goal is to assess the capacity of supplier to handle the product, service and delivery
viewpoints of quality in order to evaluate the supplier’s potential to meet with the
quality and performance requirements. It includes quality of product (C21), quality of
service (C22), quality of delivery (C23), quality of organization (C24).
3.3 Social Criteria (C3)

The goal is to assess the capacity of supplier to take actions in accordance with society
from the perspective of stakeholder, employee and community in order to understand
social development level of the supplier. It includes stakeholder’s perspective (C31),
employee’s perspective (C32) and society’s perspective (C33).
3.4 Environmental Criteria (C4)

The goal is to assess the capacity of the supplier to take action on environmental
standards and regulations in order to determine the degree of environmental devel-
opment of the supplier. It includes environmental management (C41), current envi-
ronmental impact (C42) and green image (C43).
4 Spherical Fuzzy Analytic Hierarchy Process (SFAHP)
Fuzzy logic appears to be useful in the context of sustainable supplier selection due to
its ability to cope with vague, difficult to quantify details based on expert opinions and
finally to give concrete answers (crisps) [19]. To prioritize the criteria and sub-criteria
within the model, spherical fuzzy AHP (SFAHP) method that is adapted from [20] is
utilized. The methodology can be summarized as follows:
Step 1. The questionnaires for pairwise comparisons of criteria are shared with
three experts who have valuable sectoral expertise in this study field. Linguistic vari-
ables given in [20] are used for evaluation. They are shown in Table 2.
Step 2: The consistency check is done by using the defuzzified matrices. In order to
defuzzify the evaluations, score indices (SI) given in Table 2 are utilized. The rules
developed by Saaty [21] are considered for consistency. If the value is unsatisfactory,
in other words if it is higher than 0.1, the survey is conducted again until the results are
consistent.
Step 3: There are two main aggregation operators: Spherical Weighted Geometric
Mean (SWGM) and Spherical Weighted Arithmetic Mean (SWAM). In our study,
SWAM is preferred to aggregate the evaluations of experts and to have fuzzy local
weights of main and sub-criteria. In Eq. (1) SWAM P operator is given with respect to
weights, w ¼ ðw1 ; w2 ; . . .; wn Þ (w1 2 ½0; 1, and ni¼1 w1 ¼ 1).

SWAMW A ~ S1 ; . . .; A
~ Sn ¼ w1 A ~ S1 þ w2 A ~ S2 þ . . . þ wn A
~ Sn
h wi i12
Qn wi i12 Q hQ wi Q ð1Þ
1 i¼1 1 l2A~ ; ni¼1 vAw~ i ; n
i¼1 1 l 2
~
A
n
i¼1 1 l2
A~ p 2
~
A
si Si si si si
Table 2. Linguistic measures for pairwise comparisons.

(µ, v, p) Score Index (SI)
Absolutely Higher Importance (AHI) 0.9 0.1 0 9.00
0.85 0.15 0.04 8.00
Very High Importance (VHI) 0.8 0.2 0.1 7.00
0.75 0.25 0.14 6.00
High Importance (HI) 0.7 0.3 0.2 5.00
0.65 0.35 0.23 4.00
Slightly More Importance (SMI) 0.6 0.4 0.3 3.00
0.55 0.45 0.3 2.00
Equally Important (EI) 0.5 0.4 0.4 1.00
0.45 0.55 0.3 0.50
Slightly Lower Importance (SLI) 0.4 0.6 0.3 0.33
0.35 0.65 0.23 0.25
Lower Importance (LI) 0.3 0.7 0.2 0.20
0.25 0.75 0.14 0.17
Very Low Importance (VLI) 0.2 0.8 0.1 0.14
0.15 0.85 0.04 0.13
Absolutely Lower Importance (ALI) 0.1 0.9 0 0.11
Step 4: In order to defuzzify the spherical fuzzy criteria weights, the score function
(S) has been used (given in Eq. (2)) and the calculated values were normalized by using
Eq. (3).
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi

pA~ S 2 vA~ S 2

~ sj ¼
S w 100 3l ~ pA~ S ð2Þ
AS
2 2

S wsj
sj ¼ P
wj ¼ w ð3Þ
n
j¼1 S wsj
Step 5: To reveal the global weights of the sub-criteria, the local weights are
multiplied with the weights of corresponding parent criteria. The obtained global
weights are ranked and the priority of each sub-criterion can easily be noticed from this
rank.
5 Prioritization of the Criteria
As a first step, the questionnaires are sent to three experts who have deep experiences in
business world. They are informed in detail about the aim of the study, meanings of the
criteria and the structure of the questionnaire. At the first iteration, the consistency
ratios of all experts are found satisfactory. Step 3 thru 5 are followed to find the global
weights that refer to the importance of each sub-criterion.
Spherical fuzzy (~ ws ) and crisp weights (ws ) of aggregated matrices are given in
Tables (3, 4, 5, 6 and 7). As an example for utilization of SWAM operator and
computation of score function that are given Eq. (1) and Eq. (2) respectively, com-
putations for C1 from aggregated matrix of main criteria are given as follows.
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

1 1 1 1
lC1 ¼ 1 ð1 0:52 Þ4 ð1 0:412 Þ4 ð1 0:782 Þ4 ð1 0:642 Þ4 ¼ 0:62
1 1 1 1
vC1 ¼ 0:44 0:64 0:224 0:374 ¼ 0:37
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u ð1 0:52 Þ14 ð1 0:412 Þ14 ð1 0:782 Þ14 ð1 0:642 Þ14 ð1 0:52 0:42 Þ14 ð1 0:412 0:272 Þ14
pC1 ¼t 1 1 ¼ 0:25
ð1 0:782 0:132 Þ4 ð1 0:642 0:182 Þ4
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
u u
" 2 2 #
0:25 0:37
~ sC1 ¼ t100
S w ð3 0:62Þ 0:25 ¼ 17:37
2 2
By following the similar steps, score functions of other main criteria are all sub-
criteria are computed. Then the results are normalized to find the weights as mentioned
in Step 5. The weights of main criteria and global weights of all sub criteria can be seen
at Table 8.
Table 3. Aggregated pairwise comparisons of main criteria.

C1 C2 C3 C4 ~s
w Sw~ SJ s
w
Cl 0.50 0.40 0.40 0.41 0.60 0.21 0.78 0.22 0.13 0.64 0.37 0.18 0.62 0.37 0.25 17.37 0.283
C2 0.61 0.39 0.27 0.50 0.40 0.40 0.86 0.14 0.05 0.85 0.16 0.07 0.75 0.24 0.20 21.53 0.351
C3 0.25 0.76 0.15 0.16 0.85 0.06 0.50 0.40 0.40 0.46 0.54 0.28 0.38 0.61 0.28 9.94 0.162
C4 0.49 0.54 0.19 0.18 0.83 0.09 0.56 0.44 0.28 0.50 0.40 0.40 0.47 0.53 0.28 12.57 0.205
Table 4. Aggregated pairwise comparisons of economic subcriteria.

C11 C12 ~s
w Sw~ SJ s
w
C11 0.50 0.40 0.40 0.66 0.34 0.25 0.59 0.37 0.33 16.05 0.586
C12 0.36 0.65 0.26 0.50 0.40 0.40 0.44 0.51 0.35 11.26 0.414
Table 5. Aggregated pairwise comparisons of quality subcriteria.

C21 C22 C23 C24 ~s
w Sw~ SJ s
w
C21 0.50 0.40 0.40 0.59 0.42 0.27 0.66 0.34 0.23 0.82 0.18 0.08 0.67 0.32 0.25 18.84 0.317
C22 0.45 0.56 0.28 0.50 0.40 0.40 0.65 0.36 0.23 0.77 0.24 0.17 0.62 0.37 0.27 17.30 0.291
C23 0.36 0.65 0.24 0.44 0.58 0.27 0.50 0.40 0.40 0.64 0.26 0.24 0.50 0.48 0.30 13.63 0.229
C24 0.18 0.82 0.09 0.32 0.70 0.22 0.37 0.63 0.25 0.50 0.40 0.40 0.37 0.62 0.28 9.65 0.162
Table 6. Aggregated pairwise comparisons of social subcriteria.

C31 C32 C33 ~s
w Sw~ SJ s
w
C31 0.50 0.40 0.40 0.47 0.53 0.30 0.63 0.40 0.16 0.54 0.44 0.30 14.78 0.322
C32 0.54 0.46 0.30 0.50 0.40 0.40 0.69 0.33 0.16 0.59 0.39 0.29 16.15 0.352
C33 0.62 0.43 0.12 0.49 0.55 0.19 0.50 0.40 0.40 0.54 0.45 0.26 14.93 0.326
Table 7. Aggregated pairwise comparisons of environmental subcriteria.

C41 C42 C43 ~s
w Sw~ SJ s
w
C41 0.50 0.40 0.40 0.48 0.52 0.30 0.78 0.22 0.11 0.63 0.36 0.27 17.46 0.388
C42 0.55 0.46 0.30 0.50 0.40 0.40 0.81 0.19 0.12 0.66 0.33 0.27 18.43 0.409
C43 0.22 0.78 0.12 0.24 0.78 0.15 0.50 0.40 0.40 0.35 0.62 0.28 9.16 0.203
Based on the calculations, “traditional cost” is found to be the most important sub
criterion. That means it has the highest influence on supplier selection. This finding
shows that economic considerations have still been attached high priority although the
selection is made within a sustainable manner. On the other hand, “green image” is the
least important sub criterion, but it does not mean that environmental considerations are
unsatisfactory. Because other two environmental criteria (environmental management
(C41) and current environmental impact (C42)) have valuable importances. Rather than
marketing principles for increasing the image, real operations to protect the environ-
ment are found as more important.
Table 8. The weights of main and sub criteria.

Main criteria Local weights of Sub Local weights of Global weights of
main criteria criteria sub criteria sub criteria
Economic (C1) 0.28 C11 0.59 0.166
C12 0.41 0.117
Quality (C2) 0.35 C21 0.32 0.111
C22 0.29 0.102
C23 0.23 0.080
C24 0.16 0.057
Social (C3) 0.16 C31 0.32 0.052
C32 0.35 0.057
C33 0.33 0.053
Environmental 0.20 C41 0.39 0.079
(C4) C42 0.41 0.084
C43 0.20 0.042
6 Conclusion
In this study, a generalizable model is proposed for prioritizing the criteria affecting
sustainable supplier selection. An evolved fuzzy approach titled Spherical Fuzzy
Analytic Hierarchy Process (SFAHP) is utilized in the model. Within the model, four
main criteria including economic, quality, social and environmental criteria; and their
twelve sub criteria are taken into account. It is not surprising that “traditional cost” is
found as the most important sub criterion affecting this selection procedure. However,
it is good to reveal that two environmental considerations (C41, C42) have similar and
more importance than two quality considerations. (C23, C24). It is suggested that the
reason is originated from increasing consciousness on environmental protection. In the
future, a real case on supplier selection can be conducted and a sensitivity analysis can
be made in order to see how the results are changed belongs to changes on evaluations.
References
1. Ghadimi, P., Dargi, A., Heavey, C.: Sustainable supplier performance scoring using audition
check list based fuzzy inference system: a case application in automotive spare part industry.
Comput. Ind. Eng. 105, 12–27 (2017)
2. Kuo, R.J., Wang, Y.C., Tien, F.C.: Integration of artificial neural network and MADA
methods for green supplier selection. J. Clean. Prod. 18, 1161–1170 (2010)
3. Büyüközkan, G., Çifçi, G.: Evaluation of the green supply chain management practices: a
fuzzy ANP approach. Prod. Plann. Control 23(6), 405–418 (2012)
4. Kannan, D.: Role of multiple stakeholders and the critical success factor theory for the
sustainable supplier selection process. Int. J. Prod. Econ. 195, 391–418 (2018)
5. Bai, C., Sarkis, J.: Integrating sustainability into supplier selection with grey system and
rough set methodologies. Int. J. Prod. Econ. 124, 252–264 (2010)
6. Luthra, S., Govindan, K., Kannan, D., Mangla, S.K., Garg, C.P.: An integrated framework
for sustainable supplier selection and evaluation in supply chains. J. Clean. Prod. 140, 1686–
1698 (2017)
7. Khan, S.K., Kusi-Sarpong, S., Arhin, F.K., Kusi-Sarpong, H.: Supplier sustainability
performance evaluation and selection: a framework and methodology. J. Clean. Prod. 205,
964–979 (2018)
8. Wu, C., Barnes, D.: An integrated model for green partner selection and supply chain
construction. J. Clean. Prod. 112, 2114–2132 (2016)
9. Govindan, K., Khodaverdi, R., Jafarian, A.: A fuzzy multi criteria approach for measuring
sustainability performance of a supplier based on triple bottom line approach. J. Clean. Prod.
47, 345–354 (2013)
10. Jain, N., Singh, A.R.: Sustainable supplier selection under must-be criteria through fuzzy
inference system. J. Clean. Prod. 248, 119–275 (2020)
11. Govindan, K., Rajendran, S., Sarkis, J., Murugesan, P.: Multi criteria decision making
approaches for green supplier evaluation and selection: a literature review. J. Clean. Prod.
98, 66–83 (2015)
12. Tirkolaee, E.B., Mardani, A., Dashtian, Z., Soltani, M., Weber, G.W.: A novel hybrid
method using fuzzy decision making and multiobjective programming for sustainable-
reliable supplier selection in two-echelon supply chain design. J. Clean. Prod. 250, 119517
(2020)
13. Lee, A., Kang, H.Y., Hsu, C.F., Hung, H.C.: A green supplier selection model for high-tech
industry. Expert Syst. Appl. 36, 7917–7927 (2009)
14. Shen, L., Olfat, L., Govindan, K., Khodaverdi, R., Diabat, A.: A fuzzy multi criteria
approach for evaluating green supplier’s performance in green supply chain with linguistic
preferences. Resour. Conserv. Recycl. 74, 170–179 (2013)
15. Kannan, D., Jabbour, A.B., Jabbour, C.J.C.: Selecting green suppliers based on GSCM
practices: Using fuzzy TOPSIS applied to a Brazilian electronics company. Eur. J. Oper.
Res. 233, 432–447 (2014)
16. Kannan, D., Govindan, K., Rajendran, S.: Fuzzy axiomatic design approach based green
supplier selection: a case study from Singapore. J. Clean. Prod. 96, 194–208 (2015)
17. Gupta, H., Barua, M.K.: Supplier selection among SMEs on the basis of their green
innovation ability using BWM and fuzzy TOPSIS. J. Clean. Prod. 152, 242–258 (2017)
18. Awasthi, A., Govindan, K., Gold, S.: Multi-tier sustainable global supplier selection using a
fuzzy AHP-VIKOR based approach. Int. J. Prod. Econ. 195, 106–117 (2018)
19. Phillis, Y.A., Andriantiatsaholiniaina, L.A.: Sustainability: an ill-defined concept and its
assessment using fuzzy logic. Ecol. Econ. 37, 435–456 (2001)
its renewable energy application. Soft. Comput. 24(6), 4607–4621 (2019)
21. Saaty, T.L.: Axiomatic foundation of the analytic hierarchy process. Manage. Sci. 32, 841–
855 (1986)
Estimation and Prediction
Estimating Shopping Center Visitor Numbers
Based on Various Environmental Indicators
Cagatay Ozdemir1(&), Sezi Cevik Onar2(&),

and Selami Bagriyanik 3(&)
1
Digital Business Solutions, Turkcell, İstanbul, Turkey
[email protected],
[email protected]
2
Industrial Engineering Department, ITU, İstanbul, Turkey
[email protected]
3
Turkcell Technology, Turkcell, İstanbul, Turkey
[email protected]
Abstract. The value of data is gaining importance in recent years both in

Turkey and globally through increase data production. As data began to gain
importance, data mining began to change and evolve. With help of data mining,
companies started to determine their customer management strategies based on
intelligent data analysis. Literature reviews in this area show that many data
analysis models studied in the field of customer management. When a more
detailed literature review is made, it is observed that the number of sources
where demand estimation and location analysis applied together with the
machine learning algorithms is very low in all sectors and almost none in the
shopping mall. Within the scope of this study, a new model has been developed
by combining location analysis and demand forecasting models that will esti-
mate the number of customers for shopping malls in order to overcome this
deficiency in the literature. This model was strengthened with estimation
algorithms and tested to generalize this model to all shopping malls. In this
study conducted through a large-scale technology and communications services
provider company, it was found that environmental factors had a significant
effect on the number of customers going to shopping centers.
Keywords: Shopping mall Customer strategy Predictive analysis Machine

learning Location analysis Demand model Regression model Intelligent
data analysis
1 Introduction
One of the important economic indicator for countries is the number of shopping cen-
ters. With the increase in the number of shopping centers, competition also increases and
because of this, shopping centers start to differentiate their services and campaigns in
order to stand out from the competition [1]. The increasing number of shopping malls
causes many shopping malls to be unable to reach the desired number of customers and
this situation constitutes a major obstacle for the shopping malls to reach the expected
revenues [2]. Since the beginning of the investment, the main targets of the shopping
172 C. Ozdemir et al.
centers are; establishing the shopping mall in the right location, attracting customers to
this location and reaching expected revenue [3, 4]. The two most important obstacles to
reaching estimated revenue are not estimating number of customers correctly and could
not attracting them to the shopping mall [5]. One of the most important strategy to stand
out from the competition between shopping malls and ability to reach the expected
revenue is the customer management strategy [6]. In order to provide right customer
management, companies have developed many strategies. Identifying the target audi-
ence, determining the needs of these audiences and reaching out these audiences through
the right communication channels are the most critical issues for shopping centers [7, 8].
Customer management strategies have started to be data-driven due to increasing data
diversity and volume in behalf of digital transformation.
In recent years, with the increasing importance of data in all sectors, the value of the
data also started to increase for the shopping mall sector [6–9]. Shopping centers have
started to use this data, which is produced and stored, in order to estimate the number of
customers and expected revenue for following days [10]. When the literature study was
carried out specifically for shopping centers, it was found that there were few publi-
cations on demand forecast of customer numbers and there were almost no publications
where demand forecast of customer numbers was combined with location analysis. In
this study, it is used by combining the estimation of the number of customers and
location analysis, which are critical for the shopping mall sector and are not available in
the literature. The location analysis was carried out together with the large-scale
technology and communications services provider company. The number of customers
on a daily basis for the selected shopping centers was reached through the signaling
data at the locations [11, 12]. This study has two main objectives. The first objective is
originality, that is estimating the number of customers for the selected location by
combining location analysis, customer demand forecasting analysis and machine
learning algorithm that have not been studied thoroughly in the literature. The second
objective is contributing to the strategic plans and marketing actions of the shopping
centers with the estimation the number of customers and analysis of customer behavior
according to the environmental factors.
In the next part of the study, the details of the mentioned literature review and
analytical details of case study will be discussed. In the last part, the results and dis-
cussion are included.
2 Background
Previous studies for shopping centers have generally been customer-based. Therefore,
there is a need for a study that includes the estimation of the number of customers on a
daily basis and examines the environmental factors affecting the number of customers.
In order to perform this study, it is necessary to evaluate customer demand forecast
analysis and location analysis combined. As of 2010, the total number of studies
including location analysis and customer demand forecast analysis for all sectors in the
literature is 169. If this analysis is narrowed further, the number of studies for all
sectors decreases to 13 by combining location analysis, customer demand forecast
Estimating Shopping Center Visitor Numbers 173
analysis and machine learning algorithm. When we review these 13 studies on a

sectoral basis, the number of resources for retail sectors decreases to 6 and the number
of resources in shopping centers decreases to 1. Table 1 and 2 shows the literature
reviews details below. When we examine this research further, customer faces are
observed through camera images and their behavior will be estimated [13]. With this
estimation, marketing actions are planned according to the customer’s next expected
behavior. Since the number of resources in this area is limited and it cannot be
determined what affects the estimation of the number of customers per day for shop-
ping centers, and implementation studies have been started in this area. In this study,
machine learning algorithms, which can be used in location analysis and customer
demand forecasting model, were tested. However, the data consists of daily customer
numbers and environmental information for the selected shopping center and location-
based data sets are available instead of person-based data. In addition to these, all data
types are numerical. Considering all these limitations, since the classification structure
cannot be established, artificial neural networks, support vector machines, naïve bayes
and decision-tree learning techniques cannot be applied in this study. In the light of this
information and the literature review, it was found that the most appropriate machine
learning technique for the application was regression analysis.
Table 1. Customer management strategy literature review (Method breakdown)

Customer strategy Total 2010–2019 2010–2019
models year
Total Total Machine learning Total Machine learning ML +
(ML) (ML) Regression
Location analysis 37,200 18,900 1,720 17,400 1,380 509
Demand analysis 90,400 22,500 1,860 19,000 1,620 878
Location + Demand 313 243 19 169 13 8
analysis
Table 2. Customer management strategy literature review (Sectors breakdown)

2010–2019
Customer strategy models Retail Banking Telco Mall Production Energy Health
with ML
Location analysis 203 190 104 28 442 503 571
Demand analysis 254 225 92 21 693 779 556
Location + Demand analysis 6 1 1 1 6 7 5
3 Case Study
The main purpose of the shopping centers is estimating the number of customers and
attracting new customers. With estimating the number of customer in following days,
the customer-specific strategic plans for increasing the revenue can be developed. In line
with the main scope of the application, the number of individual customers on a daily
basis will be determined based on the signaling at the shopping center location. Together
with the obtained environmental data, the number of individual customers in the past
period will be analyzed and the number of customers going to the selected shopping
center will be estimated when environmental conditions change in the following days. In
this way, it will be possible to observe which variables affect the number of customers
going to the shopping center, which environmental factors are meaningful and how the
behavior of the customers changes.
3.1 Data Sources

One of the biggest problems in the big data analysis is the fact that most of the data is
unstructured and the available data is not clean, processable and analyzable. Although
it was challenging to get the data within the scope of this study, the data was collected
and made ready for analysis with data cleaning methods. Within the scope of this
application, the number of individual customers on a daily basis for selected shopping
malls was achieved by leveraging the analysis ability of the large-scale technology and
communications services provider company’s signaling data. In addition, the following
data types have been accessed, cleaned and corrected for the analysis. Table 3 shows
the variable categories and data types in detail.
Table 3. Variable list and details

Category Category of variables Variables Metric Data type
1 Total number of customer 1 Numerical Output
2 Daily temperature 2 Numerical Input
3 Avg. daily temperature 1 Numerical Input
4 Daily amount of rainfall 1 Numerical Input
5 Financial market data 5 Numerical Input
6 Hourly average traffic density 8 Numerical Input
7 Daily average traffic density 1 Numerical Input
8 Internet trend 4 Numerical Input
3.2 Data Models for Predictive Analysis

In addition to the number of customers of the selected shopping center data in the past
period, the environmental and financial data collected and the data set was enriched. By
using machine learning algorithm on this data set, it was requested to estimate the
number of customers and any output variable to be selected for the next day. Linear
regression analysis was used as the machine learning algorithm since it was the most
suitable model for our data set. The following formulas explain the multiple regression
analysis and cost function (1)–(2).
yi ¼ b0 þ b1 xi1 þ b2 xi2 þ þ bp xip þ 2 ð1Þ
1 Xm i ðiÞ
2
J ð h0 ; h1 Þ ¼ i¼1
h x y ð2Þ
2m
This study is a predictive analysis since the next day of the number of customers are
estimated. For this machine learning algorithm, 80% of the data set has been used for
training the model and 20% of the data has been used for testing the performance of the
model. This model does not include only one regression analysis. Multiple regression
models are generated to find the model that best predicts the number of customers or
the selected output variable. Since each regression model contains different combina-
tions of variables, it is aimed to find out which variable has the most significant effect
by analysis. Therefore, the variables were chosen to be appropriate for the regression
analysis. Studies involving this machine learning algorithm were performed on Python
application. In this context, seven regression analyzes were carried out using random
values and the success rates were tested after the model setup. Adj R2, Success Rate,
VIF, Durbin- Watson (Auto Correlation) are used to evaluate the efficiency of pre-
diction model whose definition are described as following formula (3)–(5).
P
ð1 R2 Þðn 1Þ SSRES ðyi by i Þ2
R2adj ¼ 1 R2 ¼ 1 ¼ 1 Pi ð3Þ
nk1 SSTot y i Þ2
i ð yi b

1 jPrediction Actualj
Success Rate ¼ Avg: ð4Þ
Actual
PT 2

1 t¼2 ð e t e t1 Þ
VIF ¼ DW ¼ PT 2 ð5Þ
1 R2i t¼1 et
Table 4 shows the model details below.

Table 4. KPI comparison of regression models

Regression Selected variables Adj R2 Success Multi Auto
models rate collinearity correlation
name
Model 1 Shopping Center Internet Trend, 95.70% 98.10% No Positive
Monthly CPI Change, Daily Night
Temperature and Daily Average Traffic
Density
Model 2 Shopping Center Internet Trend, 95.70% 97.90% No Positive
Monthly CPI Change, Daily Night
Temperature, Daily Average Traffic Density
and E-Commerce Internet Trend
Model 3 Shopping Center Internet Trend, Daily Daily 95.40% 90.50% Yes Positive
Gram Gold/TL Parity, Daily Sunrise
Temperature, Daily Night Temperature and
Daily BIST Parity
Model 4 Shopping Center Internet Trend, Daily 95.60% 77.50% No Positive
Sunrise Temperature, Daily Amount of
Rainfall and Daily Average Traffic Density
Model 5 Shopping Center Internet Trend, Daily 95.60% 96.50% Yes Positive
Sunrise Temperature, Daily BIST Parity and
Monthly CPI Change
Model 6 Shopping Center Internet Trend, Daily 95.40% 97.30% Yes Positive
Sunrise Temperature, Daily BIST Parity and
Daily Gram Gold/TL Parity
Model 7 Daily Average Temperature, Daily BIST 95% 91.50% Yes Positive
Parity and Daily Average Traffic Density
3.3 Data Models Evaluation

The seven linear regression analysis created within the scope of the prediction analysis
mentioned above shows us; The VIF rate is very high when at least two of the financial
data and at least two of the temperature data are included in the same analysis. This
reveals that the linearity between the variables is high. When the results of the
regression analysis are considered, the ideal combination is to choose one of the
financial data and one of the temperature data. Considering the R-adjusted, F test,
statistical estimation and success rate, the first and second regression analysis turns out
to be the most effective analyzes. When variables were normalized, it was observed that
the error rate in the first and second regression analyzes decreased to a minimum and
acceptable level. According with the above mentioned scopes it reveals (model 1)
Shopping Center Internet Trend, Daily Night Temperature, Monthly CPI (Customer
Price Index) Change and Daily Average Traffic Intensity should be used in order to best
estimate the customer numbers of ongoing days.
Strengthening Data Models. In this study the main purpose was the testing the
extensiveness and applicability of the first and second regression analysis based on
machine learning algorithm for established shopping centers. To measure the validity
of these tests, mentioned models above also applied to new two shopping centers.
In order to make the mentioned measurement, the internet trends data of the shopping
centers and the number of customers from the previous period were collected. Other
outsourcing data were taken from the data set at the first shopping center. When the
results of the regression analysis tests for new shopping malls examined, it was
revealed that the model was reasonable with the F-test and high R-adjusted results.
Moreover, if we examine in more detail, these results also revealed the suitability of
Durbin-Watson and VIF values in the table below. Finally, comparing the models
between different shopping malls it is observed that the success rate of 90% in both
malls. This rate represents a high success of the analysis. Although the success rate of
both regression models was higher for the first shopping mall, the error rate was higher
for other shopping malls where the analysis was conducted. The reason for the fact that
the success rate is higher in the first shopping center is that the model was established
according to the first shopping center. Nevertheless, the application of the model to the
other two shopping centers and the acceptable error rate show us the success, validity
and dissemination of the regression analysis.
When both regression models are examined, it is seen that the first regression
analysis model is more successful in estimating. Table 5 shows the strengthening data
model details below. MSE and F-Test definition are described as following formula (6).
ðSSTotal SSRes Þ=ðk 1Þ 1 Xn

F¼ Fk1;nk MSE ¼ ðy by i Þ2
i¼1 i
ð6Þ
ðSSRes Þ=ðn k Þ n
Table 5. Strengthening details of first and second regression model

KPI Model 1 Model 2
First Second Third First Second Third
shopping mall shopping mall shopping mall shopping mall shopping mall shopping mall
Success 98.10% 95.20% 93.70% 97.90% 94.90% 93.60%
rate
Adj R2 95.70% 96.60% 97.10% 95.70% 96.70% 97.10%
MSE 0.82% 1.46% 1.63% 0.83% 1.45% 1.64%
F test Significant Significant Significant Significant Significant Significant
Multi No No No No No No
collinearity
Auto Positive Positive Positive Positive Positive Positive
correlation
4 Discussion
This study will benefit in the accurate estimation of the number of customers for next
days, which is one of the most important problems in shopping centers. In this way,
shopping centers will have the opportunity to reach their expected revenue by planning
their operations and strategies correctly. The use of location analysis in the estimation
of the customer numbers for shopping malls, which is lacking in the literature was
applied together with this study and contributed to the literature. In addition to these,
with this study, how the number of customer numbers of the shopping center changes
according to the environmental climate, financial and internet trends and the estimation
of the next day based on these variables are examined. According to the results of the
analysis, it was found that the variables mentioned in the case study section had a
significant effect on the number of customers. Furthermore, the model was found to be
successful when applied to other shopping centers. This indicates the potential for the
generalizability of the model.
In this study, the model and variables that should be used in estimating the number of
customers for the next day are examined with the help of the combined demand
forecasting model and location model of customer management strategies which are
strengthened with a machine learning algorithm. Regression analysis was found to be
the most appropriate machine learning technique for the model. Within the scope of the
analysis studies, the first and second regression models mentioned above are the best
models to estimate the number of customers. Based on these results, the best envi-
ronmental variables predicting the number of customers for the next day are Shopping
Center Internet Trend, Daily Night Temperature, Monthly CPI Change, Daily Average
Traffic Density. Based on these results, shopping centers can estimate the number of
customers for the next day with a high success rate by looking at environmental
variables. In order to put these environmental variables into their models, they need to
follow the weather information of Istanbul, the traffic density information,
daily/weekly/monthly change information of the finance market, daily internet trend
change of the shopping mall and e-commerce. Shopping centers will be able to provide
both cost optimization and marketing strategies by planning the next day with these
models. Shopping centers will be able to plan the number of security guards by esti-
mating the customer numbers arriving. They will be able to draw up the plans of the
number of stores and advisors in the shopping center and increase the number of
employees when reaching the high customer number. In addition, shopping centers will
be able to plan and optimize energy costs and other costs according to these customer
numbers. From the point of view of marketing strategies, it will be possible to create a
good experience for new customers and provide existing customer to stay loyal with
help of the campaigns. Shopping centers will also have the chance to change their
strategic plans according to market situation variability or traffic density variability.
In the future studies in order to strengthen this study, the number of customers in all
shopping centers in the selected province should be examined, the scope of the analysis
should be deepened and the data used should be enriched. At the same time, other
potential machine learning algorithms should be applied and the success rate should be
increased. In addition to these, this analysis was performed on a 217-day data set and the
results were based on these data. However, the validity of this analysis may be reduced if
the climate changes substantially, financial variables change in an unusual situation and
if the internet trend is collected in a different format. For that reason, new variables, data
sets and new regression algorithms (SVM regression and ANN regression etc.) may be
needed to strengthen these estimation models in the subsequent analyzes.
References
1. Chebat, J., Sirgy, M., Grzeskowiak, S.: How can shopping mall management best capture
mall image. J. Bus. Res. 63(7), 735–740 (2010)
2. Şekerkaya, A., Cengiz, E.: Kadın tüketicilerin alışveriş merkezi tercihlerinin belirlenmesi ve
bir pilot araştırması. Öneri 9(34), 41–55 (2010)
3. Hedhli, K.E., Chebat, J.-C., Sirgy, M.J.: Shopping well-being at the mall: construct,
antecedents, and consequences. J. Bus. Res. 66(7), 856–863 (2013)
4. Cheng, E.W.L., Li, H., Yu, L.: The analytic network process (ANP) approach to location
selection: a shopping mall illustration. Constr. Innov. 5(2), 83–97 (2005)
5. Zolfani, S.H., Aghdaie, M.H., Derakhti, A., Zavadskas, E.K., Varzandeh, M.H.M.: Decision
making on business issues with foresight perspective; an application of new hybrid MCDM
model in shopping mall locating. Expert Syst. Appl. 40(17), 7111–7121 (2013)
6. Wang, L., Fan, H., Wang, Y.: Sustainability analysis and market demand estimation in the
retail industry through a convolutional neural network. MDPI Sustain. 10(6), 1–19 (2018)
7. Gauvreau, C.A., Kairy, D., Mazer, B., Guindon, A., Dorze, G.L.: Rehabilitation strategies
enhancing participation in shopping malls for persons living with a disability. Disabil.
Rehabil. 40(8), 917–925 (2018)
8. El-Adly, M.: Shopping malls attractiveness: a segmentation approach. Int. J. Retail Distrib.
Manage. 35(11), 936–950 (2007)
9. Fayyad, U., Piatetsky-Shapiro, G., Smyth, P.: From data mining to knowledge discovery in
databases. AI Mag. 17(3), 37–54 (1996)
10. Wesley, S., LeHew, M., Woodside, A.G.: Consumer decision-making styles and mall
shopping behavior: building theory using exploratory data analysis and the comparative
method. J. Bus. Res. 59(5), 535–548 (2006)
11. Eiselt, H., Marianov, V.: Applications of Location Analysis. Springer, London (2015)
12. Eiselt, H., Marianov, V.: Foundations of Location Analysis. Springer, London (2011)
13. Huotari, V.: Depth camera based customer behaviour analysis for retail: Master’s thesis.
Oulu: University of Oulu (2015)
Predicting Customers’ Next-to-Be Purchased
Products
Buse Mert , Defne İdil Eskiocak(&) , and Işıl Öztürk
Migros Ticaret A.Ş, Ataşehir, 34758 İstanbul, Turkey

{buseme,defnee,isilo}@migros.com.tr
Abstract. Maximizing customer satisfaction and increasing sales and profit is

the first and the most crucial goal of the retail sector in an increasingly com-
petitive environment and rapidly evolving technology. We aim to improve
customer experience and to customize their journey by creating personalized
offers and baskets, allowing them to use their time more effectively and making
them feel as we know them. Using machine learning, we create a customer-
specific product (SKU) list by estimating which of the previously purchased
products would be re-purchased within a specific timeframe by considering the
customers’ previous purchase patterns. Our machine learning model uses cus-
tomer information, product information, and customer-product information
derived from the loyalty card program, sales data and promotion data. Currently,
we use the outcome of the model to create promotional offers, both at the
product level and the category level. Recent results show that the utilization
rates of individual promotions at the category level increased by more than 5%.
To the best of our knowledge, previous studies show that the models are mostly
trained infrequently with sampled data. In this study, we use our entire data to
train our model weekly to capture any changes to the customers’ behavior.
Keywords: Basket prediction Reorder prediction Promotion customization
1 Introduction
Especially in recent years, developing technology is creating more and more compe-
tition in the retail sector. Customer-centric companies are one step ahead of the other
companies in this race. The largest companies around the world are strengthening their
analytical capabilities by making considerable investments to achieve success in being
customer-oriented. These companies store all the data that they can obtain in addition
to enhancing their analytical efficiency. They produce great success in customer sat-
isfaction and loyalty with their analytical projects using the data that they archived. To
give some examples to these projects from tech giants, Netflix recommends the content
which would be best suited for each customer. The streaming company achieves this by
building models using customer’s in-app activities as well as their likes and dislikes
[1]. Netflix also uses personalization as a tool to make their customers spend more time
in the application. They personalize the order of their recommendations, the con-
struction of the customer’s home page and notifications. Personalization is difficult
because even though a company may have large amounts of data, the data points for
Predicting Customers’ Next-to-Be Purchased Products 181
each customer may not be as extensive [2]. While challenging, companies are working
to do more in the field of personalization to increase customer satisfaction and loyalty
[3]. For example, beyond the usual content recommendations, Netflix personalizes the
artwork of its content to catch their customers’ attention [4].
In the retail sector, creating shopping lists for customers is an example of a cus-
tomizable application. Researchers can approach this problem on a segment-level or an
individual-level. According to previous studies, there is no definite champion approach
[5]. In this paper, we conducted our analysis on an individual-level using machine
learning algorithms. As far as we know, previous research shows that the models are
mostly trained infrequently with sampled data. In this study, we use our entire data to
train our model weekly to capture any changes to the customers’ behavior.
Section 2 introduces our methodology of approaching the problem, our data pre-
processing steps, explanation of our predictive features and the results. Section 3
presents the discussion of the study and concludes the paper with suggestions for future
research.
2 Methodology
This study aims to predict customers’ next-to-be purchased products by analyzing their
past shopping behavior using machine learning algorithms. We use the last 12 months
of data to create each customer-product pairing’s features and the last 15 days of data to
label the bought pairs to be able to create a binary classification problem.
To perform this analysis, we use transactional data of loyalty card-holding cus-
tomers. Rather than concluding a segment-level analysis, we approach this problem on
an individual-level. By doing so, we create features to explain the behavior of cus-
tomers and customer-product pairings. Therefore, we aim to find the best-suited
products for each customer.
We approach this work as a classification problem. The intended outcome is to be
able to generate binary labels to each customer-product pairing so that we can deter-
mine which products will likely to be bought by customers. We use PySpark on a
Hadoop cluster to create our model since traditional methods cannot process our data.
We discuss this problem in the following section about data preprocessing.
We use a Random Forest Classifier to create our model. The result of the model
gives us a score between 0 and 1 for each customer-product pairing. We observe that
the range of the scores is different for each customer due to their shopping habits. The
scores of the customers frequently using their loyalty cards are seen to be higher than
the scores of the customers with less frequency. We use three metrics to measure
frequency; the number of purchased products, the number of visits, the number of
purchased products. As shown in Fig. 1, customer A who has a higher frequency has
higher scores than customer B. Table 1 displays the metrics for measuring frequency.
The number of unique products which customer A has purchased is 2.67 times
more than customer B. Customer A has visited the retailer 3.15 more times than
customer B. The number of purchased products of customer A is 5.16 times more than
that of customer B. These metrics inform us of the nature of the relationship between a
182 B. Mert et al.
Fig. 1. Range of the scores for two customers with different frequency metrics
Table 1. Frequency metrics for two customers

Customer Number of purchased Number of Quantity of purchased
products visits products
A 604 183 2410
B 226 58 467
customer and the retailer. According to these metrics, customer A has a stronger
relationship with the retailer. The range of their scores reflects this association.
As seen in Fig. 1, the maximum score of customer B is approximately in the middle
of customer A’s score range. This case indicates that determining a static threshold to
label scores as 0 and 1 would not be ideal for every customer-product pairing. To avoid
mislabeling customer-product pairings, we decide to calculate dynamic thresholds for
every product category. This dynamic calculation allows us to differentiate thresholds
for fast-moving and slow-moving product categories. After the calculation, we remove
the customer-product pairings with a lower score than the product category’s threshold.
To be able to create a balanced product list according to the customers’ needs, for each
customer, we choose the product with the highest score in a class and the products with
the five highest scores in a category. Figure 2 presents an example of a product tree.
This method enables us to list only the highest scored product from a class, for
example, the Cavendish Banana in the class of bananas.
Fig. 2. Example of a product tree
2.1 Data Preprocessing

As mentioned in the section above, we prepare the data on the customer-product level.
We transform our transactional data into customer-product based data by removing the
time dimension. The transactional data spans 12 months of loyalty card-holding cus-
tomers’ activity. If we assume that the time is t and we train our model at this moment,
the training period is between t-365 days and t-15 days (Fig. 3). The 15-day difference
to the present moment allows us to label the bought customer-product pairings
accordingly.
Fig. 3. Visualization of the training period and label assignment
After the preparation, we obtain the training and scoring data in the sizes shown in
Table 2.
Table 2. Training and scoring data sizes

Number of observations Training data Scoring data
(customer-product pairings) 615812448 797409480
Customers 4198330 7842324
Products 24344 25297
184 B. Mert et al.
2.2 Predictive Features

During the data preparation, we created 119 features divided into three groups. The first
group consists of product features. These include customer-independent information
such as life cycle, life span, stockpiling status of a product and provides input to the
model on a product basis. Table 3 contains several examples of product features and
their descriptions.
Table 3. Several examples of product features

Feature Description
prod_orders The number of transactions that the product is a part of them
prod_first_orders Number of customers who purchase the product at least one time
prod_second_orders Number of customers who purchase the product at least two times
prod_reorder_ratio Product re-purchase rate
The second feature group is the customer features. We created these features to
provide input to the model on a customer basis independently of products. Some of
these inputs give information about the customer’s preference of online/offline chan-
nels and spending level among the others. Table 4 contains several examples of cus-
tomer features and their descriptions.
Table 4. Several examples of customer features

Feature Description
cust_online_spend Customer’s online spending level
cust_mean_period Average number of days between the customer’s transactions
cust_avg_order_item Average number of products on the customer’s transactions
Third and final feature group is the customer-product features. We created these to
provide input to the model about the behavior and dynamic of customer-product
pairings. For instance, the consumption cycle of a product can be 5 days; meanwhile, a
customer’s consumption speed of said product could be 3 days. Customer-product
features are aimed to inform the model of such dynamics. Table 5 contains several
examples of customer-product features and their descriptions.
Table 5. Several examples of customer-product features

Feature Description
cp_avg_cons_day Customer’s average consumption time of the product
cp_per_order Average number of transactions between the customer’s
purchase of the product
cp_orders_since_last_order Number of transactions the customer has made since the last
purchase of the product
Presently, we use the results of the model to generate customized promotional offers
for loyalty card-holding customers. We evaluate the success of the model by comparing
the utilization rates with the previous customized offers created before the current
model. As shown in Table 6, we compare two different customized promotions with
their previous versions. We can see an increase in the utilization rates by at least 7%.
Table 6. Comparison of utilization rates

Promotion type Previous application’s utilization rate Current model’s utilization rate
Promotion #1 22% 32%
Promotion #2 23% 30%
3 Conclusion
The study shows higher accuracy in specific product categories as well as specific
customer segments. The predictions containing fast-moving products are more accurate
than the prediction of products which belong to slow-moving categories. We also see a
similar case for specific customer segments. The predictions for customers who use
their loyalty cards frequently are more accurate than the predictions for less frequent
customers.
After observing the success of the model by examining the utilization rates of
customized offers, the work of integrating the model results on online channels as a
customized basket for each customer has started. In further studies, we aim to display
these generated customized baskets not only on online channels but also on offline
channels by way of push notifications at the moment that a customer enters the
retailer’s store.
The objective of this work was to predict the customers’ next-to-be purchased
products. We executed the study with loyalty card-holding customers’ transactional
data stretched over 12 months. We approached this work as a classification problem
and used a Random Forest Classifier to create our model. After the model deployment,
the increased utilization rates of customized promotions demonstrate the success of the
model. In further studies, we aim to use the online channel feedback such as click-
stream data, likes/dislikes of recommendations to create new features. Also, we intend
to try new methods, particularly Neural Networks, to improve our results.
References
1. Amatriain, X., Basilico,J.: Netflix Recommendations: Beyond the 5 stars (Part 1). https://
netflixtechblog.com/netflix-recommendations-beyond-the-5-stars-part-1-55838468f429.
Accessed 18 Mar 2020
2. Basilico, J.: Recent Trends in Personalization: A Netflix Perspective. https://www.slideshare.
net/justinbasilico/recent-trends-in-personalization-a-netflix-perspective. Accessed 18 mar 2020
3. Riecken, D.: Personalized views of personalization. Commun. ACM 43(8), 26–28 (2000)
186 B. Mert et al.
4. Chandrashekar, A., Amat, F., Basilico, J., Jebara, T. : Artwork Personalization at Netflix.
https://netflixtechblog.com/artwork-personalization-c589f074ad76. Accessed 18 Mar 2020
5. Peker, S., Kocyigit, A., Eren, P.E.: An empirical comparison of customer behavior modeling
approaches for shopping list prediction. In: 41st International Convention on Information and
Communication Technology, Electronics and Microelectronics (MIPRO), pp. 1220–1225
(2018)
Estimating Breathing Levels of Asthma
Patients with Artificial Intelligence Methods
Asım Sinan Yüksel1(&) and Fatma Gülşah Tan2

1
Computer Engineering Department, Süleyman Demirel University,
Isparta, Turkey
[email protected]
2
Vocational School Computer Programming Department,
Manisa Celal Bayar University, Manisa, Turkey
Abstract. Pollen contains highly allergic proteins. One of the major causes of
allergic diseases is the pollen in the air we breathe. Asthma patients are known
to show allergic reaction to pollens. Therefore, they need to be more careful and
avoid the factors that trigger asthma. In this study, the first step was taken to
develop an artificial-intelligence-based decision support system to improve the
quality of life of asthma patients. Finding the ratio of pollen in nature is a long
process, and it is measured by using different instruments and hours of calcu-
lations in laboratories. In this study, Adaptive Network-Based Fuzzy Extraction
System (ANFIS) and normalization methods were applied to meteorological
data to estimate the breathing levels of asthma patients according to the amount
of pollen in the air. The data were tested via the Artificial Neural Networks
method, and it was found that the model produced better results that are very
close to real values when compared to the results in ANFIS.
Keywords: ANFIS Artificial neural networks Asthma Prediction

Normalization
1 Introduction
One of the most important causes of asthma and respiratory diseases is pollen. The
distribution of pollen varies depending on its type. The pollen of the Oleaceae family,
which is a type of tree pollen, is very important, especially in places where trees are
abundant [1]. Olea pollen is one of the most important factors that causes seasonal
respiratory allergies, especially in Mediterranean countries.
Weather condition is the most important factor that determines the amount of pollen
released into nature. For example, the increase in heat and pressure is a feature that
increases the distribution of pollen in hot, dry and windy weather, while the amount of
pollen dispersed in rainy weather decreases as it rises above sea level [10]. Today,
pollen retention devices such as “Lanzoni and Burkard” are used in accordance with
the volumetric method in the studies for the detection of pollen in the air [11]. In this
study, we aimed to estimate the breathability of asthma patients by using meteoro-
logical values and Oleaceae pollen numbers from the Antalya province in 2006–2007.
Meteorological factors that affect the distribution of pollen were determined and
monthly breathability estimations for patients with asthma profiles were predicted using
188 A. S. Yüksel and F. G. Tan
MATLAP ANFIS simulation and Artificial Neural Networks (ANN). As a result, the
first steps of an ANN based system aiming to improve the quality of life of asthma
patients in their region have been taken.
The remainder of this paper is organized as follows: Sect. 2 describes our method,
data collection process, model development and feature selection processes in detail;
Sect. 3 highlights the theoretical and practical implications of our study; and Sect. 4
provides our concluding remarks and makes suggestions for future research.
2 Method
The data for estimating the breathing levels should be standardized as the first step. In
the second step, a training model is produced by using the relevant data. Figure 1
shows the flow of the system for prediction.
Fig. 1. General estimation model.
2.1 Adaptive Neural Network Based Fuzzy Inference System (ANFIS)

Fuzzy logic and neural networks are the methods applied in the development of
intelligent systems [4]. The purpose of the ANFIS environment is to convert input
variables to output variables by using an optimization algorithm [7]. In the ANFIS
model, the aim is to optimize these parameters by using the specified parameters.
Figure 2 shows an overview of ANFIS architecture.
Fig. 2. ANFIS architecture [6].

Estimating Breathing Levels of Asthma Patients 189
Figure 3 shows the ANFIS model for this study.
Fig. 3. ANFIS model of the amount of pollen.
The results shown in Fig. 3 (which is the amount of pollen in the air) were
interpreted using the table suggested by the National Allergy Office of the American
Academy of Asthma Allergy & Immunology. It is stated that people can breathe under
specific circumstances. Table 1 shows the pollen ratings according to the amount of air
to be used for estimation. If the amount of pollen in the air is very high, the ability of
asthma patients to breathe decreases.
Table 1. Rating by amount of pollen (m3) in the air [12].

Tree-shrub pollen Herbaceous pollen Meadow pollen
0 No 0 No 0 No
1–14 Low 1–9 Low 1–4 Low
15–89 Middle 10–49 Middle 5–19 Middle
90–1499 High 50–499 High 20–199 High
>1500 Very high >500 Very high >200 Very high
2.2 Min-Max Normalization

Normalization is the process of finding the minimum and maximum values in the data
set and converting the other data to these values by a linearity approach. In the
minimum-maximum normalization, all other values are normalized in the range 0–1,
with the maximum value being 1 and the smallest value being 0 [13]. The limitation of
the value range reduces the standard deviation value. The purpose of normalization is
to make the training of the data set be used in artificial neural networks efficiently.
xi xmin
x0 ¼ ð1Þ
xmax xmin
According to the above equation;

x0 = Data obtained as a result of normalization,
xi = Input value in the data set,
xmin = The smallest number in the input,
xmax = It represents the largest number in the input [14].
2.3 Dataset Preparation

Data from the Antalya province were divided into two sets (training set and test set) by
normalizing certain 2006–2007 meteorological data and pollen data. Some of the data
and the parameters used are given in Table 2.
Table 2. Pollen number estimation data [1].

Temperature Total Wind speed Relative Amount of olea
precipitation (m/sn) humidity (%) pollen
(mm)
13.3 78.2 2.6 71.4 0.0
17.2 87.3 2.6 63.7 4.4
– – – – –
– – – – –
– – – – –
12.9 247.2 2.6 59.5 0.0
11.4 113.8 2.7 57.4 0.0
12.0 152.7 2.8 67.6 0.0
In Table 3, the parameters, minimum-maximum values and the type of variables

are given.
Table 3. Parameters and properties used.

Parameter Unit Maximum value Minimum value Variable type
Temperature C 29.7 9 Input
Total precipitation mm 494.7 0 Input
Wind speed m/sn 3.5 1.9 Input
Relative humidity % 71.4 50.9 Input
Olea pollen m3 9.3 0 Output
As shown in Fig. 3 and Table 2, the model consists of four inputs and one output.
2.4 ANFIS Prediction Model

After preparing the data set, the Gauss Function (gauss2mf), which has a low error rate,
is selected as the membership function. Variables to be used for temperature, total
precipitation, wind speed and relative humidity were determined as “high” and “low”.
The textual expression of the output variable was determined as “very high”, “high”,
“low”, and “very low”. After the training process, a 0.36 error rate was obtained as a
result of the trial conducted on the test data.
Table 4. Sample rules created for the model.

1 Temperature high AND precipitation low AND wind high AND humidity high while pollen
high
2 Temperature low AND precipitation low AND wind high AND humidity low IF pollen
high
3 Temperature is high AND precipitation is high AND wind is low AND humidity is high
while pollen is low
4 Temperature is high AND precipitation is high AND wind is high AND humidity is high
while pollen is very low
While creating the model, the number of textual expressions for each variable was
two. Accordingly, 16 rules were created for the model that had four inputs and one
output. Four sample rules obtained from the model are shown in Table 4.
An estimate is produced by comparing the results of the model tested on meteo-
rological data with the pollen amount (m3) given in Table 1. For example, if the
amount of pollen in the air is 142 m3, this is considered high and asthma patients will
have low breathing ability.
2.5 Artificial Neural Networks (ANN)

The ANN method is the basis of studies in the field of artificial intelligence. Unlike
other techniques, ANN is a successful mathematical method consisting of intercon-
nected neurons that can make decisions based on data and tolerate errors even in cases
of missing data [5]. Therefore, ANN method has the ability to manage complex and
large systems because of this feature [9]. The model of ANN consists of an input layer
that has one or more hidden layers and an output layer. Figure 4 shows the model of a
typical ANN.
According to ANN, each link has a weight value and these weight values change
when new samples arrive. Its purpose is to find the weights that will give the correct
outputs for each sample shown in the model [2].
The data used in the training of ANFIS were also used during the training of ANN.
The ANN application was executed with a feed-forward and reverse-spread neural
network [3]. In this study, a three-layer ANN is designed with an input layer consisting
of four neurons, a hidden layer consisting of 16 neurons and an output layer consisting
of one neuron. 70% of the data set was used for training and 30% was used for testing
purposes.
Fig. 4. ANN model [8].
3 Results and Discussion
According to the data obtained from the results, the R-square value was close to 1. The
fact that this value is close to 1 indicates that the test results are acceptable. Figure 5
shows the R values of the training and test data.
Fig. 5. R value of ANN model.
When the network model was tested using Mean Square Error (MSE) approach,
which is one of the popular error evaluation methods, 0.3689 and 0.035063 values were
obtained for ANFIS and ANN, respectively.
The preliminary information that asthma patients will have will provide conve-
nience for the precautions to be taken by these patients. In order to strengthen the
decision-making mechanism of the model, additional criteria such as the person’s
asthma history and the factors that cause allergies can be applied. In addition, users can
receive real time notifications about daily pollen density by moving the developed
decision support system to the mobile platform.
4 Conclusion
In this study, it was observed that Olea pollen had the highest concentrations in late
spring and early summer. In the Mediterranean and Aegean coasts, in the southern
Marmara regions, especially in May and June, olive trees produce a lot of pollen in the
air. These periods are critical for individuals with pollen allergies.
At the same time, the increase in heat and pressure in those months, along with the
fact that the air is dry or windy, increase the pollens distribution, while height and rainy
weather reduce the amount of pollen.
In this study, breathing levels were estimated by using monthly meteorological data
of Antalya province. In addition, the model can be trained to use weekly or daily
meteorological data, and it can show whether people can breathe or not depending on
the pollen status in the air before going out. Knowing the amount of pollen density that
asthma patients should pay attention to will improve their quality of life.
According to our results, ANFIS and ANN produced similar results. Considering
the limited data, the accuracy values obtained with both models were found to be quite
high; therefore, the breathing levels of asthma patients can be predicted with high
accuracy by using our developed model. On the other hand, one weakness of our study
is that it only considers one pollen type. As a future work, pollen types and patient
diversity can be increased and more accurate decision-making mechanisms can be
established.
References
1. Bıçakcı, A., Altunoğlu, M., Tosunoğlu, A., Çelenk, S., Canıtez, Y., Malyer, H., Sapan, N.:
Türkiye’de Oleaceae Familyasına Ait Allerjenik Olea (Zeytin Ağacı) ve Fraxinus (Dişbudak
Ağacı) Polenlerinin Havadaki Dağılımları. Asthma Allergy Immunol (2009)
2. Chopra, S., Yadav, D., Chopra, A.: Artificial neural networks based indian stock market
price prediction: before and after demonetization. J. Swarm Intel. Evol. Comput. 8, 2 (2019)
3. Gonzalez-Fernandez, I., Iglesias-Otero, M., Esteki, M., Moldes, O., Mejuto, J., Simal-
Gandara, J.: A critical review on the use of artificial neural networks in olive oil production,
characterization and authentication. Crit. Rev. Food Sci. Nutr. 59, 1913–1926 (2019)
4. Hocaoğlu F., Kurban M.: Adaptif Ağ Tabanlı Bulanık Mantık Çıkarım Sistemi ile Eskişehir
Bölgesi İçin Güneşlenme Süreleri Tahmini. Elektrik Elektronik Mühendisliği Bölümü
Mühendislik Mimarlık Fakültesi Anadolu Üniversitesi İki Eylül Kampüsü, Eskişehir (2013)
5. İleri, R., Latifoğlu, F., İçer, S.: Artificial Neural Network Based Diagnostic System For
Melanoma Skin Cancer. In: 2019 Medical Technologies National Conference, İzmir (2019)
6. Anfis, J.S.R.: Adaptive-network based fuzzy inference systems. IEEE Trans. Syst. Man
Cybern. 23(03), 665–685 (1993)
7. Kaya, E., Karaboga, D.: Adaptive network based fuzzy inference system (ANFIS) training
approaches: a comprehensive survey. Artif. Intell. Rev. 52, 2263–2293 (2018)
8. Nasıroğlu, B., Nasıroğlu, H., Şahin, M., Yıldız, B., Peştemalci, V.: Yapay Sinir Ağı Ve
Noaa/Avhrr Uydu Verilerini Kullanarak Hava Sıcaklığının Tahmini. Uzaktan Algılama-Cbs
Sempozyumu (2016)
9. Pinheiro, A.C.T., Mahoney, R., Privette, J.L., Tucker, C.J.: Development of a daily long
term record of Noaa-14 Avhrr land surface temperature over Africa. Remote Sens. Environ.
103, 153–164 (2006)
10. Polen Allerjisi. https://www.toraks.org.tr/halk/News.aspx?detail=2483. Accessed 2020
11. Ege Üniversitesi Polen Alerji Bülteni, https://polenalerji.ege.edu.tr/tr-8684/polenler.html.
Accessed 2020
12. The Spanish Aerobiology Network- Polen Sayımları. https://polenalerji.ege.edu.tr/tr-8473/
skala.html. Accessed 2020
13. Velchamy, I., Subramanian, R., Vasudevan, V.: A five step procedure for outlier analysis in
data mining. Eur. J. Sci. Res. 75, 327–339 (2012)
14. Yavuz, S., Deveci, M.: İstatiksel Normalizasyon Tekniklerinin Yapay Sinir Ağın Perfor-
mansına Etkisi. Erciyes Üniversitesi İktisadi ve İdari Bilimler Fakültesi Dergisi, pp. 167–187
(2012)
The Problem of Selection with the Fuzzy
Axiomatic Design of MEMS Based Sensors
in Industry 4.0 Predictive Maintenance Process
Arif Sercan Erez1, Mehmet Cakmakci2(&), and Rasit Imrenci3

1
Graduate School of Natural and Applied Sciences, Dokuz Eylül University,
Konak, Turkey
[email protected]
2
Industrial Engineering Department, Dokuz Eylül University, Konak, Turkey
[email protected]
3
Dedem Mekatronik Inc., Izmir, Turkey
[email protected]
Abstract. The aim of this study is to investigate the Fuzzy Axiomatic Design
Method of MEMS (Micro Electro-Mechanical System) based sensors in the field of
nanoscience, which is specific to the different parameters (sound, vibration, etc.) of
industrial machinery etc. in the scope of Industry 4.0 predictive maintenance.
Within the scope of the study, which is the continuation of the ongoing project, it is
aimed to advance the literature search in a special R & D center, to carry out
university-industry cooperation studies, and to conduct the tests of the appropriate
MEMS based sensor in the predictive maintenance processes of the companies. In
order to determine the most suitable method (lithography, etc.) used in the pro-
duction of sensor based sensors and shaping their design/modeling, experimental
studies will be carried out. After the modeling/patterning process, the dimensional
and similar parameters of the sensor base will be optimized and the most suitable
design will be selected by the axiomatic design method. Finally, it will be evaluated
whether the selections made for MEMS based sensors are suitable for NEMS
(Nano Electromechanical System) based sensors at nanoscale.
Keywords: MEMS (Micro Electro-Mechanical Systems) Preventive

maintenance Axiomatic design and fuzzy system relationship Industry 4.0
1 Introduction
Today, engineers face multi-criteria decision-making problems in many industries.

When designing a product or process, there are many alternatives to choose from by
comparing design criteria. In this context, Fuzzy Axiomatic Design (FAD) has become
the most commonly used decision making method to obtain a more scientific system,
process and product design form in the design process. In this study, this method will
be used to select the optimum design model for micro and nano electrical mechanical
system (MEMS and NEMS) based sensors.
196 A. S. Erez et al.
In this study, design verification and analysis studies will take place in the design
optimization process. The MEMS/NEMS-based sensors selected as the optimum
design will be a tool for predictive maintenance and “talking machines”, which is the
sub-branch of industry 4.0 processes, which are becoming increasingly popular today.
In this study, the concepts of industry 4.0 and predictive maintenance will be men-
tioned in general terms for the first time in the literature. As will be seen in the sections
that follow, this study is developed using appropriate methodology. Then, decision
trees will be drawn to define the objectives, cover design alternatives and the main
criteria and sub-criteria included in these alternatives. Design and system ranges will be
created for each criterion. Linguistic variables are used for criteria that cannot be
expressed in concrete expressions of design and system ranges, and this linguistic
variable will be converted into triangular fuzzy numbers. Then, the intersection area of
the design alternative and the system range of each criterion of each alternative are
calculated and information content will be obtained from this area calculation. All
calculations will be made using the MATLAB R2019A version.
2 Material and Method
The new process, called Industry 4.0, contains a structure that will completely change
the relations of production and consumption. On the one hand, it defines the production
systems that instantly adapt to the changing needs of the consumer, and on the other
hand, it defines the automation systems that are in constant communication and
coordination with each other [1] and encourages close cooperation between various
disciplines in product development [2]. Internet of Things (IoT), which is one of the
sub-topics of Industry 4.0 concept, is a collection of physical works including
embedded systems of electrical, mechanical, computer and communication mecha-
nisms that provide internet-based communication and data exchange. It started with the
idea of labeling and tracking “objects” with low-cost sensor technologies such as radio
frequency identification (RFID) devices. When the labeling and tracking of objects are
moved to the industrial area, the processes of monitoring the machines, taking data
from them during the working of the machines, and converting these data into
meaningful inferences emerge. The next step of this process is to collect and interpret
machine operating data and use them in predictive maintenance applications.
2.1 Predictive Maintenance

The maintenance and repairs of the machines in the industry and the possible downtime
costs of the facility are quite costly. Due to an unexpected malfunction, the machines
may be inoperable for a very long time. In this case, it causes production losses and an
increase in costs in companies that carry out mass production.
In order not to stop production, turning to predictive maintenance methods, which
can be determined using various test and analysis methods, has reached a very
important point in cost calculations. It produces vibration in response to the force
applied against the machine due to each machine component produced. These vibra-
tions are converted into numerical data within the time axis using Fast Fourier
The Problem of Selection with the Fuzzy Axiomatic Design 197
Transform techniques. The error manifests itself with symptoms such as differences in
vibration values and exceeding certain threshold values when any malfunction occurs.
With predictive maintenance, generally undamaged examinations such as oil analysis,
infrared thermography measurement, ultrasonic testing, and vibration measurement and
analysis are performed.
In the predictive maintenance applications within the scope of Industry 4.0 pro-
cesses, the industrial vibration sensor, flow etc. components are used for the mea-
surement of the relevant parameters. Devices/sensors based on NEMS (Nano Electric
Mechanical Systems)/MEMS (Nano Electric Mechanical Systems) based on NEMS
(Nano Electric Mechanic Systems)/MEMS (Nano Electric Mechanic Systems), which
have been revealed as a result of the studies carried out at nano and microscale, have
the advantages of macroscale devices/sensors/replaces sensors [3].
3 MEMS – Micro Electro Mechanical Systems
MEMS is a technology that allows the development of small products. Sensors are
devices that collect information from the environment by measuring mechanical,
thermal, biological, chemical, optical and magnetic data. Then, the data obtained from
the sensors are directed to the access devices by the electronics and we can easily
perform the environmental controls we aim for, such as motion, positioning, adjust-
ment, pumping, filtering. Because, the production of MEMS devices can be made
relatively economically from small silicon chips with a high level of reliability and with
complex method of mass production [4–6], just like in integrated circuits. The
requirements of the photolithographic method are given in Fig. 1.
Fig. 1. Representation of the photolithographic method

The process begins with the choice of a base material and geometry. Typically a
single crystalline silicon layer with diameters between “4 and 8” is used (see Fig. 1-a).
The backing is then covered with a photosensitive polymer called photoresist (see
Fig. 1-b). The mask with precisely transparent areas is used to create a shadow on the
photoresist when light is passed [7]. These non-transparent areas under ultraviolet light
bombardment are chemically destroyed (see Fig. 1-c). After the bombardment, the
photoresist is immersed in a solution (developer). This solution chemically removes
bombarded areas or non-bombarded areas from photoresist (see Fig. 1-d). Once the
layer has dried, the photoresist is used as a mask for the next storage (see Fig. 1-e) or
scraping (see Fig. 1-f). Finally, the photoresist is removed and the micro processed
backing is obtained as a result (see Fig. 1-g and 1-h). Optimization processes must be
carried out at the design stage before producing MEMS solver devices. In this context,
how to carry out optimization processes with the fuzzy axiomatic design method,
which is one of the multi-criteria decision making methods in the fields of industrial
engineering, will be explained.
4 Axiomatic Design
Axiomatic design is a method developed to make the system, process and product
design more scientific and systematic [8]. The purpose of the proposed AD Method is
to base the designs on a scientific basis. As a result of the literature research, there are
many case studies ranging from the selection of the sealing material to the selection of
the supplier, and the evaluation of the personnel alternatives in the business procure-
ment process. In another study, the problem of choosing a supplier for the white goods
manufacturer was discussed and the Fuzzy AD method was used to solve the problem
[9, 10]. In some cases, such as this study, the system and design ranges of functional
requirements cannot always be expressed in a certain range. It can be stated to be above
or below a certain value, and these values can be represented by triangular or trape-
zoidal numbers. First of all, the definition of the problem was made within the scope of
MEMS/NEMS design selection problem. Design and system ranges will be created for
each criterion. Linguistic variables are used for criteria that cannot be expressed in
concrete expressions of design and system ranges, and this linguistic variable will be
converted into triangular fuzzy numbers. Then, the intersection area of the design
alternative and the system range of each criterion of each alternative are calculated and
information content will be obtained from this area calculation. All calculations will be
made using the MATLAB R2019A version.
4.1 Introduction to Axiomatic Design Methodology

The axiomatic design method puts product and process designs into a systematic.
Axioms are proposals that cannot be proved but are accepted as correct and these are
limited. In this method, two axioms are used. These are independence and information
axioms. These axioms provide a proportional basis for creating product designs and
choosing the best of the solutions offered [9].
Step 1. The information axiom is expressed as probability values ðpi Þ. Therefore,

the most likely option is the best design. In the MEMS/NEMS design type selection
problem study, alternatives and their relevant criteria are determined first and the value
of these alternatives is called the information value ðI i Þ to the probability to meet
functional requirements. In other words, this value can be defined by the following
equation:

1
I i ¼ log2 ð1Þ
pi
The greater the probability of this equation, the smaller the value of the information The
total information content is the sum of n-FR probabilities. In other words, if the common
area coincides with the system range, the calculation function value is 0. If this value is
infinite, this option is undesirable because it cannot meet the functional requirement at
all. Figure 2 shows the probability density function for the FR system range [8].
Fig. 2. Probability density function of a FR expressed as discrete variable
Step 2. As a second step, alternatives and their respective criteria are evaluated and
the system range (SR) and design ranges (DR) are determined. The junction area
between DR and SR is shown as the common area (CR), and this area is where only
functional needs are met. As a result, by realizing the designated purpose of the design
by dividing the area under the system area into the area under the common area:
common interval
pi ¼ ð2Þ
system interval
Step 3. The information content ðI i Þ from Eq. 3 is calculated as follows:

system interval
I i ¼ log2 ð3Þ
common interval
4.2 Axiomatic Design Method with Fuzzy Numbers

In some cases, in multi-criteria decision making problems, design and system ranges
cannot be given in a numerical range. For example: for an alternative cost criterion, the
design range must be between $5 and $10, which can be expressed in a numerical
range. However, weak, good and similar verbal expressions are used instead of
numerical expressions for the customer satisfaction criterion. These expressions (lin-
guistic variables) can be represented by triangular or trapezoidal numbers. Thus, lin-
guistic variables are placed in a numerical range value. The system includes five
conversion scales such as ‘Bad-Fair - Good - Very Good – Excellent’ for the intangible
criteria in Fig. 3.
Fig. 3. Representation of triangular numbers
With the help of the graphic shown above, the information content of the alter-
natives is calculated with the following equation:

Triangular Fuzzy Area of the System Range
I i ¼ log2 ð4Þ
Common Area
The graphical representation of the common area representing the intersection of

the system spectrum and the design spectrum is as follows (see Fig. 4) [11]. The
general steps of the fuzzy axiomatic design method to be used in the MEMS/NEMS
design selection problem are as follows:
– Determination of design alternatives and criteria of MEMS/NEMS based sensors
– Evaluation of design alternatives and criteria of MEMS/NEMS based sensors
– Numerical transformation of linguistic variable evaluations for design criteria of
– MEMS/NEMS based sensors
– Collection of information content scores from the criteria for each alternative
– Choosing the best MEMS/NEMS based sensor design alternative
5 Application of Fuzzy Axiomatic Design on the MEMS
The urge of having Micro Electro Mechanical Systems (MEMS) based solutions are
tremendously increasing in modern times as it ensures several promising features like
miniaturization product cost advantages, bulk processing, low power consumption and
reliability etc. The perception that the future of MEMS is flavored with full of com-
mercial prospects; commissioning of billions of MEMS based sensors as ears and eyes
of Internet of things (IoT) have driven the industries for employing it in creating
competitive edges. Pressure sensors have been known in the form of strain gauges over
several decades. The piezo resistivity in silicon and germanium paved a way in
attempting the miniaturization of pressure sensors and other mechanical sensors [12].
The MEMS capacitive pressure sensors are working on the basis of the principle of
parallel plate capacitor, wherein a thin diaphragm of the order of few microns acts as
the top plate and lower plate is provided by a thin layer of suitable material patterned
lithographically on substrate. In most of the micromachined capacitive pressure sensor,
silicon substrate is used as sensor structure and polysilicon or polymer materials like
polymide, kapton polymide, su-8, liquid crystal polymers are used as diaphragm
membrane as adopted in. Such a system when encounters pressure, top plate gets
deformed; distance between the plates varies and it leads to the change of capacitance
that could be sensed through suitable electronic circuits. The square shaped membranes
of three materials such as silicon nitride, silicon dioxide and polysilicon have been
modelled with the parameters.
Fig. 4. Intersection area of system range and design range
According to the information axiom calculation formulas given in the previous sec-
tions, 6 different alternative mems sensor membranes were evaluated according to
different criteria. As a result of this evaluation, the information values of each criterion
were found and the total information values of each alternative were obtained. As a
result of this evaluation, it was determined that the size criterion had no effect on the
process of choosing the best alternative since it took the same value for all alternatives.
As can be seen in Table 1, the alternative with the lowest information content value is
the Silicon Nitride-1 alternative. In other words, it is this alternative that we know the
most about the design process and includes the least random process. So the best
alternative is the MEMS sensor membrane design Silicon Nitride-1 alternative. Within
the scope of this study, different mems sensor membrane alternatives according to
different criteria have been ranked according to the fuzzy axiomatic design method and
the best alternative has been selected. For the first time in the literature, many important
decision-making methods for industrial engineering have been applied to micro-sized
MEMS systems. Thanks to the axiomatic design method used in this study, the
manufacturer design range for the most alternative MEMS device is matched to the
customer system range. In subsequent studies, thanks to the information obtained from
this study, nano-sized NEMS sensor systems or completely different types of NEMS
systems can be evaluated with more detailed design criteria. With this study, NEMS
and MEMS sensors with optimized design can be obtained and used successfully in
preventive maintenance processes in machine systems. Thus, uninterrupted commu-
nication of industrial objects can be provided. As a result, this study provides micro and
nano-sized multi-disciplinary contribution to the concept of industry 4.0, which is one
of the most important issues of today’s industry.
Table 1. Calculation of information content for valve alternatives

Alternatives
Ii Silicon Silicon Poly Silicon Silicon Poly
Nitride-1 Dioxide-1 Silicon-1 Nitride-2 Dioxide-2 Silicon-2
ISize 0,00 0,00 0,00 0,00 0,00 0,00
IThickness 1,32 1,32 1,32 0,00 0,00 0,00
IYoung 0,00 1,74 0,00 0,00 1,74 0,00
Modulus
IMass Density 0,00 0,00 0,86 0,00 0,00 0,86
IPoisson Ratio 0,00 1,32 0,29 0,00 1,32 0,29
IMax. 0,78 0,00 0,29 2,18 0,78 2,18
Displacement
IStress at 0,00 0,00 0,00 0,78 0,78 0,78
P
Edges
I 2,10 4,38 2,76 2,96 4,61 4,12
References
1. Cakmakci, M., Kucukyasar, M., Aydin, E.S., Aktas, B., Sarikaya, M. B., Turanoglu Bekar,
E.: KANBAN optimization in relationship between industry 4.0 and project management
approach. In: Bolat, H., Temur, G. (eds.) Agile Approaches for Successfully Managing and
Executing Projects in the Fourth Industrial Revolution, pp. 210–227. IGI Global, Hershey
(2019). https://doi.org/10.4018/978-1-5225-7865-9.ch011
2. Herter, J., Ovtcharova, J.: A model based visualization framework for cross discipline
collaboration in industry 4.0 scenarios. Procedia CIRP 57, 398–403 (2016)
3. Bılgıner, S.: Endüstri 4.0 ile Kestirimci Bakım. https://www.endustri40.com/endustri-4–0-

ile-kestirimci-bakim/. Accessed 26 December 2019
4. Cornelius, T. L.: Multimems Microsystems Manufacturing Cluster Multimems Process
Design Handbook Version 3.0, PartIIIa. 1st Edition, SensoNor (2003). Springer
Science + Business Media, Inc., USA
5. Yıldız, H.A., Önel, A., Trabzon, L., Karakaş, H.C., Uçar, N.: On the design and analysis of a
simple smart shirt for monitoring sudden infant death in babies. In: International Scientific
Conference Intelligent Ambience and Well-Being, Tampere, Finland (2015)
6. Judy, J.W.: Microelectromechanical systems (MEMS): fabrication, design and applications.
Smart Mater. Struct. 10, 1115–1134 (2001)
7. Trabzon, L., Lukat, K., Jankowski, I., Duerr, P., Schenk, H.: Measurement of charging under
DUV laser by a test chip for MOEMS and the mechanism of charging. In: Proceedings of the
6th Euspen International Conference, Wien, Austria, (2006)
8. Suh, N.P.: Axiomatic Design Advances and Applications, 7th edn. Oxford University Press,
New York (2001)
9. Ozel, B., Ozyörük, B.: Supplier selection with fuzzy axiomatic design. Gazi Univ. J. Fac.
Eng. 22(3), 415–423 (2006)
10. Güngör, F.: Selection of seal material by axiomatic design method. El Cezerî J. Sci. Eng. 4
(1), 1–10 (2016)
11. Kulak, O., Kahraman, C.: Fuzzy multi-attribute selection among transportation companies
using axiomatic design and analytic hierarchy process. Inf. Sci. 170(7), 191–210 (2005)
12. Saravanan, V.K.: Finite element analysis of membranes for MEMS capacitive pressure
sensors. Int. J. Eng. Res. Technol. (IJERT) 5(10), 330–335 (2016)
Fuzzy Based Noise Removal, Age Group
and Gender Prediction with CNN
Ali Tunç1(&), Sakir Taşdemir2, and Murat Köklü2

1
Kuveyt Türk Participation Bank Konya R&D Center, Konya, Turkey
[email protected]
2
Computer Engineering of Technology Faculty, Selcuk University,
Konya, Turkey
Abstract. Facial recognition systems, which are a type of biometric systems,

use the person’s facial features to identify people, and determine the information
in the position between the eyes, nose, cheekbones, jawline, chin and the like
and reveal a personal numerical model. The main purpose of this study is to
determine the gender and age group of a selected image. By making necessary
studies on the image, it is aimed to separate the gender of the person and to
extract child or adult information. This study is proposed as research that per-
forms noise reduction using the fuzzy logic fire filter algorithm and classifies the
result by gender using a convolutional neural network (CNN), matrix comple-
tion and deep learning techniques. CNN algorithm, which is one of the deep
facial recognition algorithms that can be used in the recognition of facial images,
is the algorithm used in the study. In this study, the data obtained after the
cleaning of the noises in the photographs were classified. The fuzzy logic fire
filter algorithm is used to remove the noise in the image, and the image is
classified according to gender and age range with the help of CNN and Ten-
sorFlow algorithms. The system has been trained using certain learning and test
data sets and these algorithms. The image entered into the system is converted
into a matrix expression and compared with the matrix expressions on the
trained system. According to the comparison result, the image is classified as
male, female, boy or girl, which matrix is closer to the value group. The data we
obtained as a result of our work and the success rates show the applicability of
the study. The main purpose of the study is to make gender and age group
estimations according to the image with CNN classifications.
Keywords: Fuzzy logic Deep learning Image processing Gender predict

Age classification CNN
1 Introduction
With the technological developments, the use of computer-aided systems has increased
for the reporting, processing, monitoring, analysis, and finalization of these data. The
analysis of computer-aided images and the transformation of these images into infor-
mation have been carried to higher levels with the development of technological tools
with artificial intelligence and image processing algorithms. Image processing algo-
rithms that provide level feature extraction have determined the dominant and
Fuzzy Based Noise Removal, Age Group and Gender Prediction with CNN 205
meaningful features of the existing images and videos, and have caused a systematic
classification or inference processes by using these features with various machine
learning algorithms. Artificial neural networks, which has been a favorite one of the
field in recent years, has transitioned from networks to deep networks in many areas
with the rapid development of technology and development in GPUs. With this
approach, it has been successfully used in a wide range of images from image pro-
cessing, data definition, natural language processing to applications to be used in
healthcare. Especially in the analysis of biometric data, many areas such as face
recognition, facial expression analysis, gender, and race classification have been con-
ducted. With the widespread use of the internet and mobile technologies, meaningful
expressions of facial expressions, emotion analysis, age, gender and race estimation
studies can yield very high results. The main purpose of this study is to make CNN
classification and gender and age group estimates of the pictures cleaned with Fuzzy
Logic. In this context, firstly, information was given on fuzzy logic, CNN and Ten-
sorFlow, necessary literature studies were presented about the subject, and then details
about the study and its results were explained. The innovative aspect of the study is the
comparison of noise removal with fuzzy logic in grayscale images and compared to the
result of the CNN study in RGB images.
2 Face Recognition
Facial recognition processes are an extremely important field of study because people
are identified and provide information about their characters. Security is used effec-
tively in a wide range of fields, especially in areas such as education, health, law, and
entertainment. Facial recognition systems have been examined in many respects and
solutions to many problems have been introduced by using various techniques, algo-
rithms, and methods on the subject [1]. Face detection processes create an active
workspace in areas such as personal recognition, human-machine interaction, emotion
analysis, facial animation and robotics [2]. The face recognition process can be divided
into basic 3 steps: detecting a particular face with a complex background (face
detection), localization of these faces, removing features from facial areas, and finally
recognition and verification [3]. People register to the system by measuring one or
more physical/behavioral features in one or more periods. Facial features are filtered on
the data saved in the system and noisy data is cleared. Attribute vectors are removed
from normalized faces, simplified and robust, classifiers are trained and validity tests
are completed. Subsequent facial expressions are classified using this training data, and
the resulting output is produced. Facial recognition algorithms generally determine
about 80 points on a human face. In this context, points are used to measure the
variables of basic character traits in the person. The system works by capturing data
according to the points determined in the digital image of a person’s face and recording
the obtained data as the front layer. The face recognition system can operate in two
modes, face recognition and face verification [4]. The most important point to achieve a
successful result in face recognition systems is to work on low noise data that will
reveal the contours of the face. There are many studies in the field of face recognition in
the literature. Vitor Albiero and et al. present a comprehensive analysis of how and
206 A. Tunç et al.
why face recognition accuracy differs between men and women [5]. Mehdi Hashemi
and Margeret Hall have worked to extract personality traits using data on facial images
[6]. Shubham Mittal and Shiva Mittal studied deep learning-based solutions for
automatic detection of gender from face images of a well-balanced dataset [7].
Khurram Zeeshan Haider and et al. studied face detection, noise removal, facial fea-
tures determination, and gender classification [8]. Nisha Srinivas and et al. worked on
the performance of a child who faces recognition [9].
3 Method
3.1 Fuzzy Logic

Fuzzy logic is a logic structure formed in 1961 with an article prepared by Lütfü
Aliasker Zade. In classification studies according to the classical approach, an entity is
either a member of a cluster or not. When this concept is expressed mathematically, “1”
takes the value when the entity is an element of the set and “0” when the element is not
an element of the set. Fuzzy logic is a more extended representation of the classical set
representation. In the fuzzy asset set, each asset has a membership rating. The mem-
bership degree of the assets can be any value in the range (0, 1). Unlike conventional
clusters, the degree of membership of elements in inflected clusters can vary infinitely
in the range [0, 1]. Fuzzy logic - rather than full logical approaches, the underlying
logic base is to find applications that change from the process medical diagnostic
check. In conclusion, in fuzzy logic, everything including truth is a become of degree
[10]. The fuzzy logic approach is used in many areas such as control systems, pro-
duction lines, image processing areas, motion detection, simulation, medical studies,
predictions, classification studies. There are many studies in the field of fuzzy logic in
the literature. L. Szilagyi and at ai. present a new algorithm for fuzzy segmentation of
MR brain images for the medical image processing area [11]. Jaspreet Kaur and Preeti
Gupta worked on real-time image processing processes with a fuzzy logic-based noise
filter [12]. F. Farbiz and his friends have worked with the fuzzy logic-based filter, in the
area that cleaned the image noises by protecting the edge and image details efficiently
[13]. F. Russo and G. Ramponi have worked to remove noise in the pictures with fuzzy
logic [14]. Orhan Yaman and his friends studied a model-based approach is proposed
for the recognition of the pantograph type used in electric trains. They studied contact
point problems such as arcing and excessive contact force can be detected with the
fuzzy model [15].
3.2 Convolutional Neural Network (CNN)

Convolution Neural Network (CNN) is a type of multilayer perceptron (Multi-Layer
Perceptron MLP). It is based on filtering that has been put forward as an approach
based on animal vision system [16]. It makes the features of the image clear with the
determined filtering structure and by using them at different values, it provides the
features that have less poverty [17]. CNN is one of the most widely used deep learning
algorithm models in image analysis processes, image processing applications and
computer vision studies in recent years. CNN starts with the application of the con-
volution process to the matrix obtained from the pixel values of the image in its first
stage. Convolution process enables the extraction of different levels of attributes to the
filter type to which it is applied. The model, which learns the attributes in the first
layers, transfers the most dominant pixels and attributes learned from the first layer to
the next layers. This step is given the names of feature learning or transfer learning.
The CNN algorithm consists of input, convolution, rectified linear unit (RELU),
pooling (flattening), fully connected layers (FCL) layers and activation function stages.
There are many studies in the field of CNN in the literature. Bartłomiej Hebda and
Tomasz Kryjak made studies with the deep convolutional neural network (DCNN)
architecture, on age and gender estimation of the facial images on the image [18]. Gil
Levi and Tal Hassner conducted studies that showed a significant performance increase
in age and gender classification using the deep convolutional neural network (DCNN)
algorithm [19]. Rajeev Ranjan and et al. studied a multi-purpose algorithm that can
make face recognition, face alignment, gender recognition and age estimation using the
convolutional neural network (CNN) algorithm [20]. Steve Lawrence and his friends
made a study comparing other methods their performance on image sampling by
combining SOM and CNN algorithms with a hybrid algorithm [21].
3.3 Tensor Flow - Keras

TensorFlow is an open-source and one of the most used libraries in the field of deep
learning. Thanks to its flexible structure, it allows you to deploy calculations, using one
or more CPUs, GPUs, regardless of platform, with a single API. Based on Python, this
framework supports many languages such as C++, Java, C#, Javascript and R as well
as Python today. With open-source TensorFlow developed by Google, artificial
intelligence applications with deep learning support can be developed. A network
model is created with Keras and this model is trained with training data. After the
training process is completed, the expected value is estimated with the test data and the
performance of the model is queried. The basic structure in Keras is called the model.
The simplest model type is the Sequential model, which consists of a stack of layers.
Although this is the commonly used structure; there are also model types that different
layers can be created using.
4 Study and Results and Discussion
Within the scope of the study, the success and performance metrics in gender esti-
mation and classification processes are presented over the image data. It was tried to
recognize and analyze the data in the image by using facial recognition functions on the
incoming image data. Technically, code was developed in the python environment to
process, train and test data. The APIs needed on python, which is the ideal program-
ming language especially for algorithms running on computers with high configuration
features such as Deep Learning, have been installed and the work to be done has been
developed step by step. The quality of the images was done as the first step, and the
analysis of the data on the images was carried out as the second step. In this context,
208 A. Tunç et al.
two different algorithms have been used. Firstly, the fuzzy logic algorithm was used to
correct the corrupt data in the images and to clear the noise. With the Fuzzy Logic
algorithm, the Fire Filtering method was added for the lost data on the images and the
unnecessary noisy data was removed. The reason for choosing the Fuzzy Logic Fire
algorithm in the study is to test the contribution of the operations to be performed on
the images before the CNN classification study.
The example of the image whose noise is cleared with the help of the Fuzzy Logic
fire filter algorithm is shown in Fig. 1. With this algorithm, reduction noise of image
data has been performed and data pre-processing techniques have been used to obtain
more successful results. One of the most important items in image processing studies is
to prepare the data correctly and to include them in the studies by pre-processing
techniques. The fuzzy Logic fire filter algorithm has been chosen to remove the noise in
the image, so it is aimed to obtain an optimum result.
Fig. 1. Sample image of the Fuzzy Logic algorithm, cleaned by using fire filter methods.
The reason for choosing CNN algorithm in the study is that it is one of the highest
algorithms for feature extraction and classification study. Another stage is the classifi-
cation study by analyzing the image with the deep learning algorithm. In the study, the
images prepared to create training data were recorded to the system in certain dimen-
sions, and the system was trained on these images by making 120 X 160 dimensions for
each image. A python language was used for the classification study and a three-layer
model was created following the CNN deep learning algorithm. A four-class classifi-
cation study was carried out by applying the Convolutional layer, Pooling layer, Flatten
layers. Role and Sigmoid algorithms are compiled on the CNN layer. The CNN structure
we created in our study is as shown in Fig. 2.
Fig. 2. CNN architectural structure in the study.
For the images to be recognized from different angles, the images in the training set
are translated and introduced to the system with the rescale function. For each clas-
sification category, 60 images, including male, female, girl and boy, were prepared, and
240 images were used in the training phase of the study. Models created for Training
and Test data are introduced to the system as these four classes. 180 of the images were
used for educational purposes and 60 of them were used to create the accuracy value.
Relevant python objects were created for the determined training and test data, and the
learning and test classes and learning model systems were created. Then, the learning
model created in the system was developed in a way that can be used as a model in all
applications independent of the project. A feedback feature selection layer has been
created with the created epochs value. With the three convolutional layers installed, the
images have been extracted with a certain feature attribute. When an image with a clear
background suitable for the given measurements is given with the study we obtained, it
had achieved as a result of classification with 93% success.
Table 1. Time and accuracy comparisons of algorithms.

Function Accuracy Total time Avg epoc time
CNN Classification (Gray Scale) 90% 224,262 s 4,485 s
Fuzzy Logic + CNN Classification (Gray Scale) 91% 219,413 s 4,388 s
CNN Classification (RGB) 93% 166,011 s 3,458 s
210 A. Tunç et al.
Fig. 3. Last 10 epoch accuracy graphic of algorithms.
With the Fuzzy Logic algorithm, doing the work with CNN classification by
removing noise has increased the success of the classification study. The results
obtained according to RGB and Gray Scale images with the help of CNN Classification
and Fuzzy Logic + CNN Classification algorithms are shown in Table 1 and Fig. 3.
In the study conducted with CNN in RGB format, the success rate was higher in a
shorter time. In the Gray Scale format, the Fuzzy Logic + CNN algorithm was tried
according to the CNN algorithm, increasing the success rate and reducing the time.
When the CNN classification is used alone in grayscale images, the duration of success
is low. When the CNN classification is done after cleaning the noise with Fuzzy Logic,
the success rate has increased and the duration has decreased. Because CNN classifi-
cation was successful to feature extraction, the working time decreased and the success
rate increased in RGB images. The main reason for this is the difficulty of CNN in
classification attribute extraction in grayscale images. Since CNN provides the feature
extraction through its layers, a structure to be constructed correctly is very important
for producing successful results. Besides, correct training data and test data should be
created for classification. The success of the study is directly proportional to the
accuracy and quality of the data in the training and test set. In Fuzzy Logic with black
and white images or high noise images, by classifying the noise with the CNN algo-
rithm after cleaning the noise, the success rate in the study increased. In high color
quality images, classification can be made directly with the CNN algorithm.
5 Conclusion and Future Works
The increasing use of technology has made the need to analyze data with computer
systems inevitable. In particular, studies on authentication and analysis of personal data
have become an increasingly important field of study. With these high-performance
algorithms that produce fast results and high performance, facial information can be
analyzed very quickly and access to desired results can be achieved. With this study,
we have been able to estimate the gender of facial information and determine whether a
child or adult can be determined with successful results. Especially, the estimation of
gender and the formation of the age group are important in terms of conducting
customer segmentation in many sectors. The next planned work is to estimate the age
of the person in the image statement with a high success rate. The age estimation will
also contribute to the grouping studies that will be created. Another study planned to be
carried out is to estimate the gender and age of each person individually in facial
expressions of more than one person in an image. In this way, by interpreting the
pictorial expressions involving more than one person, gender and age ratios can be
analyzed.
References
1. Zhao, W., Chellappa, R., Phillips, P.J., Rosenfeld, A.: Face recognition: a literature survey.
ACM Comput. Surv. 35, 399–459 (2003)
2. Hariri, W., Tabia, H., Farah, N., Benouareth, A., Declercq, D.: 3D facial expression
recognition using kernel methods on Riemannian manifold. Eng. Appl. Artif. Intell. 64, 25–
32 (2017)
3. Cevikalp, H., Neamtu, M., Wilkes, M., Barkana, A.: Discriminative common vectors for
face recognition. IEEE Trans. Pattern Anal. Mach. Intell. 27(1), 4–13 (2005)
4. Li, S.Z., Jain, A.K.: Handbook of Face Recognition. Springer (2005)
5. Albiero, V., Krishnapriya, K.S., Kushal, V., Kai, Z, Michael C.K., Kevin W.B.: Analysis of
gender inequality in face recognition accuracy. In: The 2nd Workshop on Demographic
Variation in the Performance of Biometric Systems at WACV (2020)
6. Mahdi, H., Margeret, H.: Criminal tendency detection from facial images and the gender bias
effect. J. Big Data 7, 2 (2020)
7. Shubham, M., Shiva, M.: Gender Recognition from facial images using convolutional neural
network. In: 2019 Fifth International Conference on Image Information Processing (ICIIP).
IEEE Xplore (2020)
8. Khurram, Z.H., Kaleem, R.M., Shehzad, K., Tabassam, N., Sohail, J.: Deepgender: real-time
gender classification using deep learning for smartphones. J. Real-Time Image Proc. 16, 15–
29 (2019)
9. Nisha, S., Matthew, H., Kevin, G., Harleen, A., Michael, K., Karl, R.: Exploring automatic
face recognition on match performance and gender bias for children. In: 2019 IEEE Winter
Applications of Computer Vision Workshops (WACVW). IEEE Xplore (2019)
10. Zadeh, L.A.: Fuzzy logic. Comput. J. 21(4), 83–93 (1988)
11. Szilagyi, L., Benyo, Z., Szilagyi, S.M., Adam, H.S.: MR brain image segmentation using an
enhanced fuzzy C-means algorithm. In: Proceedings of the 25th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society (IEEE Cat.
No.03CH37439). IEEE Xplore (2004)
12. Kaur, J., Kaur, P.: Fuzzy logic based adaptive noise filter for real time image processing
applications. Int. J. Comput. Appl. 58(9), 1–5 (2012)
13. Farbiz, F., Menhaj, M.B., Motamedi, S.A., Hagan, M.T.: A new fuzzy logic filter for image
enhancement. IEEE Trans. Syst. Man Cybern. Part B (Cybern.) 30(1), 110–119 (2000)
14. Russo, F., Ramponi, G.: A fuzzy filter for images corrupted by impulse noise. IEEE Signal
Process. Lett. 3(6), 168–170 (1996)
15. Yaman, O., Karaköse, M., Aydın, I., Akın, E.: Detection of pantograph geometric model
based on fuzzy logic and image processing. In: 22nd Signal Processing and Communications
Applications Conference (SIU). IEEE Xplore (2014)
16. Hubel, D.H., Wiesel, T.N.: Receptive fields and functional architecture of monkey striate
cortex. J. Physiol. 195(1), 215–243 (1968)
17. Fukushima, K., Miyake, S.: Neocognitron: a self-organizing neural network model for a
mechanism of visual pattern recognition. In: Amari, S., Arbib, M.A. (eds.) Competition and
cooperation in neural nets, pp. 267–285. Springer, Heidelberg (1982)
212 A. Tunç et al.
18. Hebda, B., Kryjak, T.: A compact deep convolutional neural network architecture for video
based age and gender estimation. In: Federated Conference on Computer Science and
Information Systems (FedCSIS). IEEE Xplore (2016)
19. Levi, G., Hassner, T.: Age and gender classification using convolutional neural networks. In:
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pp. 34–
42 (2015)
20. Ranjan, R., Sankaranarayanan, S., Castillo, C., Chellappa, R.: An all-in-one convolutional
neural network for face analysis. In: 12th IEEE International Conference on Automatic Face
& Gesture Recognition (FG 2017). IEEE Xplore (2017)
21. Lawrence, S., Giles, C., Tsoi, C., Back, A.: Face recognition: a convolutional neural-network
approach. IEEE Trans. Neural Netw. 8(1), 98–113 (1997)
Click Prediction in Digital Advertisements:
A Fuzzy Approach to Model Selection
Ahmet Tezcan Tekin(&), Tolga Kaya, and Ferhan Çebi

Istanbul, Turkey
{tekina,kayatolga,cebife}@itu.edu.tr
Abstract. The fuzzy logic theorem is inherently used effectively in expressing

current life problems. So, using fuzzy logic in machine learning is getting
popular. In machine learning problems, especially using digital advertisement
data, products/objects are being trained and predicted together, but this can
cause worse prediction performance. A significant commitment of our research
is, we propose a new approach for ensembling prediction with fuzzy clustering
in this study. This approach aims to solve this problem. It also enables flexible
clustering for the objects which have more than one cluster’s characteristics. On
the other hand, our approach allows us ensembling boosting algorithms which
are different types of ensembling and very popular in machine learning because
of their successful performance in the literature. For testing our approach, we
used an online travel agency’s digital advertisements data for predicting each
hotel’s next day click amount, which is crucial for predicting marketing cost.
The results show that ensembling the algorithms with a fuzzy approach has
better performance result than applying algorithms individually.
Keywords: Click prediction Fuzzy model selection Fuzzy regression

Ensemble learning OTA
1 Introduction
Although the history of machine learning algorithms is analyzed, their entry into the
literature is quite old. Still, the frequency of use has increased, especially with the
development of technology and the number of operations that computers can perform
per second. Both academic and private-sector problems are trying to be solved with
different machine learning methods and the popularity of the subject is increasing day
by day. Notably, the size and content of the data generated in the digital environment
are overgrowing and if these data can be processed, the solutions of many problems can
be shed light on their results.
Machine learning problems, which usually try to be solved in the private sector, are
mostly demand, sales, cost, customer movement, etc. prediction is focused. As a result,
whatever the forecast is for, the aim is to support the decision-making process and to
predict many uncertain events that may occur in the future. This estimation process;
businesses have an essential place in shaping their information and strategic operations
214 A. T. Tekin et al.
for future situations. Regardless of the type of problem in the estimation process, the
aim is to have a minimum error rate.
To achieve this goal, many methods have been tried in the literature and one of the
most applied methods today is ensemble learning. The purpose of this method is; It is
to combine multiple algorithms to close their weaknesses and increase the success rate
in forecasting. For this purpose, regardless of whether the problem is classification or
regression, several methods have been proven in the literature and proved to be
successful.
Clustering, which is a sub-branch of machine learning, can be used in conjunction
with a classification or regression problem, as well as clustering the values in the
observation set on its own. The aim here is; Again, to apply machine learning algo-
rithms on observation sets that are similar to each other and to increase the success rate.
The concept of fuzzy logic is also a frequently used method for machine learning
problems. One of the main reasons for this is that the answers to real-life problems are
not just 0 and 1 singular values, as in machine learning approaches that address real-life
problems. In the current life, we try to explain the answer to many problems with the
values between 0 and 1.
In this study, fuzzy logic was used for bagging and stacking ensemble methods.
The purpose of this application is clustering the observation set which has similar
behavioral characteristics. But the real issue is the items are subjected to fuzzy clus-
tering and involve belonging to more than one cluster at the same time; it is to identify
this and at the same time make the group members and estimate the success rates of the
algorithms with the most optimal validation rate for these cluster variations. In this
way, the prediction process will be performed for the items that are similar to each
other regardless of the overall success of the model, and the overall success of the
model is tried to be increased.
In the paper, Sect. 2 deals with ensemble learning and fuzzy approach in machine
learning literature studies. Section 3 describes our proposed methodology and mod-
elling details. Finally, conclusions of the study are briefly described and future works
are presented in the last section.
2 Literature Review
Ensemble learning is the method of training more than one algorithm in the same
problem and trying to solve the same problem. Learning takes place by trying to
combine to create and use a series of hypotheses, unlike the approaches of individual
machine learning algorithms [1]. The main purpose of ensemble learning is to achieve
better predictive performance than individual machine learning algorithms.
Ensemble learning algorithms generally consist of multiple algorithms called basic
algorithms. The generalization ability of ensemble learning algorithms is more robust
than basic algorithms. Ensemble learning was first applied to classification problems
and was later adapted to regression problems [2]. Unfortunately, most of the time,
successful techniques in classification problems are not valid for regression problems
[5]. In other words, current research in ensemble learning methods used for classifi-
cation is not suitable to provide an overview of current regression problem approaches.
Click Prediction in Digital Advertisements 215
As the working principle; More than one model is created and the samples to be
estimated are given as input to these models. Outputs are passed through a voting
mechanism and the final prediction process is made. But one of the most important
issue is ensemble learning steps differ in each ensembling method [12].
Many studies in the literature have emphasized that they have higher success rates
compared to a single basic algorithm [3, 4]. Also, since ensemble learning approaches
on regression algorithms are different than classification problems, Medes-Moreira
et al. [5] examined current approaches to the learning ensemble for regression.
There are 3 main approaches suggested in the literature in the ensemble learning step
of regression problems. These approaches are; Stochastic Gradient Boosting [6, 10],
Standard Bagging [7] and the combined form of these two methods are the Bagging and
Stochastic Gradient Boosting methods. Bagging and Stochastic Gradient Boosting is
also known as another type of MultiBoosting approach [8]. Adaboost [9, 11] is a kind of
ensemble algorithm which is based on random sampling like Stochastic Gradient
Boosting.
Apart from these approaches, other ensemble learning methods recommended in
the literature and used in machine learning problems are voting ensemble[13] and
stacking ensemble [14].
Machine learning algorithms mainly aim to extract information from data and for
this purpose, they are used in traditional clustering, classification and correlation
methods [15]. Since fuzzy set theory aims to produce more flexible results, it is
widespread to use with machine learning methods. Since the fuzzy set theory can
model missing and inaccurate data as a feature, it is used in different steps from
machine learning steps such as data processing, feature engineering and modelling.
Fuzzy Logic [16–19] is a method that allows defining intermediate values between
traditional evaluations such as true/false or yes/no. Fuzzy logic, which is also a fre-
quently used method for classification problems, appears in the literature as fuzzy
classifiers. Classes can often be specified with numerical expressions. In this case, a
fuzzy classification system can be described by means of a rule base. To the left of each
rule is a combination of values of linguistic variables that define a particular class [20].
The integer variable that represents the same class is on the right [21].
The fuzzy logic theorem is also frequently encountered in the literature with cluster
problems from unattended learning problems. Clustering algorithms can generally be
dealt with in two classes, precise and flexible clustering [22]. In the exact clustering
method, each observation in the test data set belongs to only one cluster. In compar-
ison, in the flexible clustering method, one observation may belong to more than one
cluster [15]. Flexible clustering method is also included in the literature as fuzzy
clustering. In the fuzzy clustering method, the membership level is calculated for each
observation value. This level is between 0 and 1.
There are different approaches proposed in the literature for fuzzy clustering. These
methods are Fuzzy C Means Clustering [23, 24], Possibilistic C Means Clustering [25],
Fuzzy Possibilistic C Means Clustering [26, 27] and Possibilistic Fuzzy C Means
Clustering [29]. But the most popular one is Fuzzy C Means Clustering and it is used in
this study also. This approach, first proposed by Dunn [23] in 1973 and later organized
by Bezdek [24] in 1981. This approach consists of two main steps.
1) Calculation of cluster centers.

2) Calculating the distance of each observation set to these centers by Euclidian
distance calculation method and assigning them to the centers.
This algorithm assigns a membership value between 0 and 1 for each set of our
observations. The turbidity parameter also calculates the degree of turbidity within the
cluster. Therefore, thanks to this method, if an observation value involves belonging to
more than one cluster at the same time, this situation is determined and its degree is
calculated.
3 Proposed Methodology and Modelling
The dataset which is used in this study is related to an online travel agency’s meta-search
platform performance report which indicates digital advertisements’ performance metrics
daily. In this study, we aim to predict each hotel’s next day click count which will be got
from metasearch engine. This prediction indicates the marketing cost which is crucial for
companies. Some of the base features which are used in this study are impression, click,
beat, meet, lose, unavailability ratio, rating and click-through rate information of the hotel
etc. The dataset which is used in this study consists of 18 base features and 1214377 rows.
Initially, data cleaning techniques were applied to the files. For this purpose, the features
which are hotel URL, hotel name, etc. which could not be used for machine learning
algorithms have been wiped out of the data. Also, redundant rows which should not be on
the data have been wiped out, like more than one row, which is about a hotel that day. After
that, measures were added to the data collection to enrich the data. In the data enrichment
step; certain functions with shifted and rolling average were added to the dataset. Simi-
larly, the hotels were sorted as city or summer hotels as indicated by their location.
In the feature engineering step, the average values of some of the features are shift-
ed and rolled average as 3, 7, 15, 30 and 45-day values were added to the dataset
because of the problem is a kind of time series problem. After the data cleaning
techniques and feature engineering step, we acquired 134 features and 1208151 rows as
a dataset which is used in the modelling section.
This study’s principal contribution is in the modelling section. We proposed a new
ensembling method for the modelling section. The steps of the new model were
explained briefly below.
1) The dataset is formed according to objects which have different characteristic
features. This object can be a product, a hotel etc.
2) Fuzzy Clustering techniques are applied to the data for finding related groups for
each object. Fuzzy clustering techniques supply that if an object is related to more
than one group, this object is processed according to its group information. In this
step, a threshold value is set for membership value.
3) In this step, candidate models with selected parameters are applied to the whole
dataset and predictions are stored in data frames.
4) Most successful models (with a selected threshold success criteria) are determined
for each fuzzy cluster.
5) The weighted averaging method is applied for the predictions of each fuzzy cluster.
The main reason for this approach is each object in observation dataset has different
characteristics. These characteristics features are price, sales frequency, seasonality etc.
So, using the same model for predictions of each object causes a certain amount of
prediction error. There are several studies in the literature for preventing this issue.
They generally apply clustering techniques which are especially K-Means [28, 29]
clustering. But K-Means clustering which iş a member of hard clustering techniques
assign each object to one cluster. But this approach ignores if the object has more than
one cluster’s characteristic features. So, Fuzzy C-Means clustering was applied to the
dataset and FPC [30] scores were evaluated for different k initial seed values. Five was
selected for k number for fuzzy clustering with nearly 0.9 FPC score. Cluster details are
shown in Table 1.
Also, for the regression modelling, three of the most successful boosting algorithms
which are XGBoost, Light GBM and CatBoost algorithms with different hyperpa-
rameters were applied to the dataset separately. Prediction results were saved in sep-
arate data frames for each observation. Then performance results were evaluated for
each cluster with each algorithm. Performance results of the algorithms for each cluster
were briefly described in Table 2. For the prediction performance, Root Mean Squared
Error (RMSE) metric, which is a prevalent metric in regression problems, was used.
The results indicate that each model with different parameters has different performance
results in different clusters. So, ensembling top three models with parameters option of
these models according to their reciprocal proportion of RMSE seems to be a more
valid solution for better prediction.
In the next step, the ensembled prediction was generated and it was compared with
other model and parameters group. The details of the comparison are shown in Table 3.
The results show us the ensembled approach have better performance on the pre-
diction stage than using a model – parameter tuples individually.
Table 1. Fuzzy cluster details.

Fuzzy member groups Unique hotel count Fuzzy member groups Unique hotel count
1 2762 3,4 96
1,2 329 4 641
2 3880 4,5 53
2,3 411 5 201
3 847
Table 2. Model and parameters prediction results on groups (RMSE) (max_depth:md,

learning_rate:lr, depth:d).
Model Parameters Grp Grp Grp Grp Grp Grp Grp Grp Grp
1 1,2 2 2,3 3 3,4 4 4,5 5
Xgboost md:6, lr:0,01 19,97 21,89 20,02 19,71 15,04 26,39 14,96 19,69 20,36
Xgboost md:6, lr:0,1 5,72 5,56 5,51 6,45 5,76 6,73 21,02 24,56 20,98
Xgboost md:8, lr:0,01 23,58 18,09 21,35 16,52 7,18 6,87 5,54 27,33 25,09
Xgboost md_depth:8, 25,18 31,75 18,14 8,36 26,98 8,41 7,41 25,48 29,62
lr:0,1
Catboost d:4, lr:0,01 7,98 7,66 2,45 1,87 27,89 22,82 5,61 21,49 27,69
Catboost d:4, lr:0,03 21,67 21,18 21,67 6,72 4,98 22,55 21,05 5,58 4,94
Catboost d:8, lr:0,01 21,95 21,96 21,96 14,75 5,58 6,19 5,8 6,41 26,73
Catboost d:8, lr:0,03 4,28 5,15 6,82 6,83 6,45 4,78 6,13 4,26 4,98
LightGBM md:6, lr:0,01 5,6 4,99 5,75 4,77 4,78 22,1 16,22 5,22 5,78
LightGBM md:6, lr:0,1 29,98 29,53 24,61 2,01 1,8 5,74 17,9 25,23 2
LightGBM md:8, lr:0,01 2,02 2,02 1,69 6,93 7,86 32,21 32,46 24,39 24,94
LightGBM md:8, lr:0,1 14,81 6,13 6,72 6,41 6,46 4,92 14,86 5,51 6,85
Table 3. Overall prediction results with each model and parameters (RMSE) (max_depth:md,
learning_rate:lr, depth:d).
Model Parameters Overall RMSE Model Parameters Overall RMSE
Xgboost md:6, lr:0,01 19,84 Catboost d:8, lr:0,03 5,93
Xgboost md:6, lr:0,1 12,03 LightGBM md:6, lr:0,01 8,41
Catboost d:4, lr:0,01 13,98 LightGBM md:8, lr:0,1 8,09
Catboost d:4, lr:0,03 14,61 Ensembled Model 4,31
Catboost d:8, lr:0,01 15,07
In this study, we intended to predict click amount of digital advertisements for an

online travel agency. For this purpose, we proposed a new ensemble method which is
based on fuzzy clustering and model selection for prediction. In general approach, after
the data preprocessing and feature engineering step, different models with different
parameter are applied to the dataset and chosen the best performance model for pre-
diction. But this approach can fail for some of the objects which don’t have similar
characteristics in the dataset. Our method is usable for products/hotels which are
objects that have different characteristics like price, seasonality etc.
In our approach, in the first step, we clustered all objects in the dataset with a fuzzy
approach. We chose fuzzy clustering instead of hard clustering methods because
objects in the dataset can be related to more than one group. Then the whole possible
algorithms with parameters were applied to each cluster separately. Predictions and
validation performance results were stored to separate data frames. In the final step, the
most successful three algorithms were ensembled with the weighted average of pre-
dictions according to the performance result.
The results show us ensembling with fuzzy approach has better performance than
applying algorithms individually. For future work, our new approach will be tested
with the data, which is indifferent domains for testing its generalizability. Also, our
approach shows us this method can be used for missing value prediction in the dataset.
Because of filling missing value is essential in machine learning, this can also increase
the prediction performance.
References
1. Zhou, Z.H.: Ensemble learning. In: Li, S.Z., Jain, A. (eds.) Encyclopedia of Biometrics.
Springer, Boston (2009)
2. Ridgeway, G., Madigan, D., Richardson, T.: Boosting methodology for regression problems.
In: The Seventh International Workshop on Artificial Intelligence and Statistics, pp. 152–
161. Morgan Kaufmann (1999)
3. Che, D., Liu, Q., Rasheed, K., Tao, X.: Decision tree and ensemble learning algorithms with
their applications in bioinformatics. In: Arabnia, H.R., Tran, Q.N. (eds.) Software Tools and
Algorithms for Biological Systems, vol. 696, pp. 191–199. Springer, New York (2011)
4. Yu, H., Ni, J.: An improved ensemble learning method for classifying high-dimensional and
imbalanced biomedicine data. IEEE/ACM Trans. Comput. Biol. Bioinf. 11(4), 657–666
(2014)
5. Mendes-Moreira, J., Soares, C., Jorge, A.M., de Sousa, J.F.: Ensemble approaches for
regression: a survey. ACM Comput. Surv. 45(1), 1–40 (2012)
6. Friedman, J.H.: Stochastic gradient boosting. Comput. Stat. Data Anal. 38, 367–378 (1999)
7. Breiman, L.: Bagging predictors. Mach. Learn. 24(2), 123–140 (1996)
8. Webb, G.I.: Multiboosting: a technique for combining boosting and wagging. Mach. Learn.
40(2), 159–196 (2000)
9. Freund, Y., Schapire, R.: Experiments with a new boosting algorithm. In: Machine Learning:
Proceedings of the 13th International Conference, pp. 148–156 (1996)
10. Moisen, G.G., Freeman, E.A., Blackard, J.A., Frescino, T.S., Zimmermann, N.E., Edwards,
T.C.: Predicting tree species presence and basal area in Utah: a comparison of stochastic
gradient boosting, generalized additive models, and tree-based methods. Ecol. Model. 199
(2), 176–187 (2006)
11. Bauer, E., Kohavi, R.: An empirical comparison of voting classification algorithms: bagging,
boosting, and variants. Mach. Learn. 36, 105–139 (1999)
12. Yildirim, P., Birant, K.U., Radevski, V., Kut, A., Birant, D.: Comparative analysis of
ensemble learning methods for signal classification. In: 2018 26th Signal Processing and
Communications Applications Conference (SIU) (2018)
13. Džeroski, S., Ženko, B.: Is combining classifiers with stacking better than selecting the best
one? Mach. Learn. 54, 255–273 (2004)
14. Bulut, F.: Örnek tabanlı sınıflandırıcı topluluklarıyla yeni bir klinik karar destek sistemi.
J. Faculty Eng. Architect. Gazi Univ. 32(1), 65–76 (2017)
15. Sujamol, S., Ashok, S., Kumar, U.K.: Fuzzy based machine learning: a promising approach.
CSI Commun. Knowl. Digest for IT Community 41(8), 21–25 (2017)
16. Zadeh, L.A.: Information and control. Fuzzy sets 8, 338–353 (1965)
17. Zadeh, L.A.: Outline of a new approach to the analysis of complex systems and decision
processes. IEEE Trans. Syst. Man Cybern. 3(1), 28–44 (1973)
18. Zadeh, L.A.: Fuzzy algorithms. Info. Ctl. 12, 94–102 (1968)
19. Kumar, M., Misra, L., Shekhar, G.A: Survey in fuzzy logic: an introduction. Int. J. Sci. Res.
Dev. 3(6), 822–824 (2015)
20. Holeček, P., Talasová, J., Stoklasa J.: Fuzzy classification systems and their applications. In:
Proceedings of the 29th International Conference on Mathematical Methods in Economics,
pp. 266–271. Janská Dolina, Slovakia (2011)
21. Kuncheva, L.I.: Fuzzy Classifier Design. Springer, Heidelberg (2000)
22. Nasibov, E., Ordin, B.: An incremental fuzzy algorithm for data clustering problems.
J. Balıkesir Üniversitesi Fen Bilimleri Enstitüsü Dergisi 21, 169–183 (2019)
23. Dunn, J.C.A.: Fuzzy relative of the ISODATA process and its use in detecting compact well-
separated clusters. J. Cybern. 3(3), 32–57 (1973)
24. Bezdek, J.C.: Pattern Recognition with Fuzzy Objective Function Algorithms. Kluwer
Academic Publishers, Dordrecht (1981)
25. Krishnapuram, R., Keller, J.M.: The possibilistic C-means algorithm: insights and
recommendations. IEEE Trans. Fuzzy Syst. 4(3), 385–393 (1996)
26. Pal, N.R., Pal, K., Bezdek, J.C.: A mixed C-means clustering model. In: IEEE International
Conference Fuzzy Systems, pp. 11–21 (1997)
27. Pal, N.R., Pal, K., Keller, J.M., Bezdek, J.C.: A possibilistic fuzzy C-means clustering
algorithm. IEEE Trans. Fuzzy Syst. 13(4), 517–530 (2005)
28. Pollard, K.S., Van Der Laan, M.J.: A Method to Identify Significant Clusters in Gene
Expression Data. U.C. Berkeley Division of Biostatistics Working Paper Series 107 (2002)
29. Kaufman, L., Rousseeuw, P.: Finding Groups in Data: An Introduction to Cluster Analysis.
Wiley, New York (1990)
30. Xie, X.L., Beni, G.: A validity measure for fuzzy clustering. IEEE Trans. Pattern Anal.
Mach. Intell. 13(8), 841–847 (1991)
Neural Fuzzy Inference Hybrid System
with Support Vector Machine for Identification
of False Singling in Stock Market Prediction
for Profit Estimation
Bhupinder Singh(&) and Santosh Kumar Henge
School of Computer Science and Engineering,

Lovely Professional University, Phagwara, Punjab, India
[email protected]
Abstract. The Stock market prediction is one of the biggest challenges in the
global market. The volatility in the movement of stock prices deteriorates the
interest of the investor and trader. The main reason of weakness of direction
prediction accuracy of trader is due to buy or selling stock based on false signals
that always result in loss of capital. The false singling based perdition is the
biggest problems in stock market prediction. Identification of false signals in the
stock market prediction will remove some sort of noise with the implication of
intelligent system based algorithms, which are used to build for solving the
specific problem in specific domain and it not extendable for solving some
specific uncertainties in same specific problem. The machine learning algo-
rithmic based neural networks, support vector machines and decision trees
techniques will more helpful for detecting future stock values based on historical
data and concurrent data. The blended technology of neural fuzzy inference
hybrid system is deriving more flexible solutions for predicting the stock market
values. This research has identified in-depth gaps in techniques that had not been
explored earlier by previous researchers and it is proposing the blended tech-
nologies of neural fuzzy inference hybrid system along with support vector
machines to reduce complexities in stock market prediction. This research is
more helpful for the stock traders whose depends on intelligent trading system
that help them to take efficient buy or sell decisions based on specific conditions.
Keywords: Stock Predictability (SP) Neural fuzzy inference hybrid system

(NFIHS) Support Vector Machine (SVM) Cross Validation (CV) Profit
Estimation (PE) Time Frames (TF) Machine Learning (ML) Tuning
Parameters (TP)
1 Introduction
Biggest question is our mind whether investing or trading. The ugly truth is that
investing products like mutual funds or stocks does not give us real positive results.
The random movement of stock of company leads to unpredictable results. The noise in
stock market can be referred to as volatility in the market. Trading is a considered to be
better choice for making profits with high risks. Fundamental Analysis and Technical
222 B. Singh and S. K. Henge
analysis [5] play an essential role in picking up the best stock for Stock value of
Company. Stock market movement is referred as random walk by some of researchers
in back 1990s. The conventional advantage that is used to access the confidence of the
prediction is random walk method. Non-Linear system [6] movements have become a
prominent tool in improving the prediction accuracy of time varying system such as the
intraday scenario. Neuro-Fuzzy Models are an breathtaking tool for no-linear system.
The system tools for assembling Neuro fuzzy models are based on the models of
combination of algorithms from the area of neural network, pattern recognition and
regression analysis. This research has implicated with the performance analysis of two
companies which is measured based the random months for computational purpose.
Motive behind case study is to find the maximum loss and maximum profit, so as to
finalize the stock company for buying or selling of shares using the heuristic intelligent
system (Table 1).
Parameters: 1 h time frame, 1 h high price, 1 h low price
Table 1. Describes the Performance analysis of two Blue chip Companies in respective months
Company name Jan-19 Feb-19 July-19 Nov-19 Dec-19 Jan-20 Feb-20
TCS −57 140 29 65 58 122 291
MARUTI 794 279 276 194 −13 463 333
2 Machine Learning Methods
Recent Studies which has explored through Machine Learning and Soft Computing
Methodologies for stock prediction on the basis of Financial Ratio’s. Best Supervised
Methods for prediction lies between Financial Prediction are Feed Forward Neural
Network, Random Forest, Adaptive Neural Fuzzy Inference System and Support
Vector Machine. These types of methods use samples of data that is needed for creating
a hope for the underling function that had produced all of the other data [10]. Taking
out a deduction from different samples which are given to the model is the main aim for
this. Feed-Forward Neural Network [8, 9] is the versatile form of neural Network
Architecture. Random Forest is very flexible Algorithm that can be utilized for both
classification and regression Problems. It Construct multiple decision trees during the
data fitting process. Random Forest takes the mean value of output values of all
decision trees for regression task [7]. Support Vector Machine take linear models to put
non-linear boundaries through relationship of mapping of Input vectors into high
dimensional space. An optimal separating Hyper-plane is constructed in the new space.
The maximum margin hyper-planes gives prominent separation between the decision
classes [12]. Adaptive Neural Fuzzy Inference System [1] is a method that is used to
tune an existing rule base with a supervised learning algorithm based on collection of
training data from consecutive dataset [4]. Fuzzy rules may be extracted from expert
knowledge and it is recommended to create an initial model. Too many rules may lead
to a complex model with superfluous fuzzy rules that compromises the integrity of the
model [3] (Fig. 1).
Neural Fuzzy Inference Hybrid System with Support Vector Machine 223
Fig. 1. Structure for a two-input TSK model with four rules [11]
The Model selection involves the following two choices:

1. Selection of Input Variables.
2. Number and type of membership functions, number of rules.
3 Methodology
The initial step of running the neuro-fuzzy program sequences are used for searching
the model complexity. The MAT-LAB software tool has for implicating the neuro-
fuzzy identification and data analysis execution sequences. To train the model, the train
and target inputs from the data set used to search for the model complexity is used to
train the model. The trained model is then saved to be evaluated. The training is similar
to the MLP training in that the gradient descent method has used for finding the
minimum error.
The testing set of the data is then used to evaluate the predictive power of the
model. The error is also evaluated using various error measures. Basically, the main
objective of this paper is to collect the stock information for some previous years and
then accordingly predict the results for the predicting what would happen next. So, for
its going to use of two well-known techniques neural network and data mining for
stock market prediction. Therefore, combining both these techniques could make the
prediction more suitable and much more reliable [10].
Level 1: Stock prediction based on Blue Chip Company based on Regression
Analysis and fuzzy Rules.
Level 2: Stock prediction based on trend of Nifty 50 gainers and Losers in Intraday
Scenario.
Level 3: Stock Prediction based on Blue Chip Company and Nifty 50 gainers and
losers with conditional Stop-loss (Fig. 2).
Data Collection
Data Selection using stepwise regression

Analysis RF
SVM
Extraction of Fuzzy Rules
ANN
ANFIS
Machine Learning Algorithm
Evaluation
Fig. 2. The Stock recommendation system based execution scenarios and stages.
4 Experiments
Building a dataset involves collecting data from various sources and putting them
together. Sample Stocks used for experiments were chosen from Nifty 50 stock. This
research is considering IT Stock Wipro for the experimental purpose. First Step is to
import the data from the dataset. The sample structure of data set is given below (Fig. 3
and Table 2):
Table 2. Describes sample dataset structure

Date Open High Low Close Volume Adj Close
11/21/2012 361.6 368 360.9 366.8 663800 366.8
11/20/2012 361 363.8 356.2 361.75 666300 361.75
11/19/2012 358.35 362.4 355.15 358.15 625200 358.15
11/16/2012 361 365.8 358 359.1 576200 359.1
11/15/2012 368.2 369.35 360.25 361.75 564900 361.75
According to Graham’s Study on missing values can problem in data handling and
thus create invalid conclusion. The raw fundamental data fetched online for stocks have
a considerable fraction of data entries missing. Previous applications of machine
learning for stock prediction use different metrics for performance evaluation. The
metrics are selected based on how the models are utilized for predicting stock per-
formance: For a regression model, the absolute or relative return of a stock at some time
in the future is estimated. Metrics such as root mean square error (RMSE), mean square
error (MSE) and mean error (MAE) are usually used for evaluating the accuracy of the
Fig. 3. Experimental scenarios with training data
Fig. 4. Implementation and experimental scenarios of ANFIS on Wipro dataset

regression model. The model is trained for 200 epochs for optimal validation results.
The model is then tested on the test data. 0 stands for back-propagation and 1 for
hybrid. Once, Training to the data is complete, and then ANFIS is applied to the data.
Last Step is to calculate the Error associated with ANFIS [2, 4] (Fig. 4).
5 Conclusion
The future stock prediction can be considered as success only if stock prediction
accounts to be Profitable. Global factors are also associated with be the volatile
movements of stock market. So, Investing is not advisable and Trading can be con-
sidered to be safe in every type of market condition. Success and failure go side by
side. Our Methodology emphasized on improving the performance of the Model so that
it may reduce the risks of failure. This research also represents the comparative study of
different machine learning algorithms associated with stock market prediction. This
research carefully split our data into training, validation and testing sets and made sure
that it did not accidentally snooped the test data or over-fit the models. Finally, research
has employed the feature selection technique in order to remove unreliable features and
reduce model complicity. Future work can be extended to Stock market prediction with
trading with trailing stop-loss. More data could potentially improve model performance
as well as conclusiveness of our results.
References
1. Chakraborty, A., Mukherjee, D., Dutta, A., Chakraborty, A., Kole, D.K.: An effective stock
price prediction technique using hybrid adaptive neuro fuzzy inference system based on grid
partitioning. In: Berretti, S., Thampi, S., Dasgupta, S. (eds.) Intelligent Systems Technolo-
gies and Applications, pp. 275–284. Springer, Cham (2016)
2. Do, Q., Trang, T.V.: Forecasting Vietnamese stock index a comparison of hierarchical
ANFIS and LSTM. Decis. Sci. Lett. 9(2), 193–206 (2020)
3. Nai, B.B., Minuvarthini, M., Sujithra, B, Mohandas, V.P.: Stock market prediction using a
hybrid neuro-fuzzy system. In: International Conference on Advances in Recent Technolo-
gies in Communication and Computing, pp. 243–247. IEEE (2010)
4. Nayak, S.C., Misra, B.B.: A chemical-reaction-optimization-based neuro-fuzzy hybrid
network for stock closing price prediction. Financ. Innov. 5, 38 (2019)
5. Pahwa, K., Agarwal, N.: Stock market analysis using supervised machine learning. In:
International Conference on Machine Learning, Big Data, Cloud and Parallel Computing
(Com-IT-Con), India, 14–16 February 2019. IEEE (2019). 978-1-7281-0211-5/19. https://
doi.org/10.1109/comitcon.2019.8862225
6. Kalra, S., Shankar Prasad, J.: Efficacy of news sentiment for stock market prediction. In:
2019 International Conference on Machine Learning, Big Data, Cloud and Parallel
Computing (Com-IT-Con), India, 14th–16th February 2019 (2019). https://doi.org/10.1109/
comitcon.2019.8862265
7. Ahmed, M.K., Wajiga, G.M., Blamah, N.V., Modi, B.: Stock market forecasting using ant
colony optimization based algorithm. Am. J. Math. Comput. Model. 4(3), 52–57 (2019)
8. Rahman, M.O., Hossain, M.S., Junaid, T.S., Forhad, M.S.A., Hossen, M.K.: Predicting
prices of stock market using gated recurrent units (GRUs) neural networks. Int. J. Comput.
Sci. Netw. Secur. 19(1), 213–222 (2019)
9. Lv, D., Yuan, S., Li, M., Xiang, Y.: An empirical study of machine learning algorithms for
stock daily trading strategy. Math. Prob. Eng. 1–30 (2019)
10. Prashanth, G., Raval, S., Siddharth, K.S., Raj, P., Dixit, S.: Over view on stock predictor
using machine learning and deep learning (2019)
11. Huang, M., Ma, Y., Wan, J., Wang, Y., Chen, Y., Yoo, C.: Improving nitrogen removal
using a fuzzy neural network-based control system in the anoxic/oxic process. Environ. Sci.
Pollut. Res. 21(20), 12074–12084 (2014)
12. Olson, D.L., Wu, D.: Predictive Data Mining Models. Computational Risk Management
Series. Springer (2017)
Estimation of Potential Locations of Trade
Objects on the Basis of Fuzzy Set Theory
Alekperov Ramiz Balashirin Oglu1(&)

and Salahli Vuqar Mamadali Oglu2(&)
1
Information Technology and Programming,
Azerimed LLC, Baku AZ1147, Azerbaijan
[email protected]
2
Department of Computer Engineering,
Odlar Yurdu University, Baku AZ1072, Azerbaijan
Abstract. In the present article, we have considered the issue of selection of

potential locations of trade objects as a multi-factor decision-making in the
conditions of uncertainty by applying the theory of fuzzy sets. Examples were
given to assess the location options of commercial enterprises and the potential
location options of the facilities on these samples were assessed on the basis of a
fuzzy extraction method in the notation MATLAB\Fuzzy Inferences Systems.
Keywords: Commercial establishments Multi-criteria decision-making

Fuzzy sets Fuzzy logic derivation Multi-factor decision-making
1 Introduction
The main purpose of commercial enterprises is to provide goods or services to cus-

tomers in order to meet consumer needs. Thus [1–4]:
• Due to the large number of sales, end customer needs are met at the place of
residence or workplace. At the same time, large volumes of products from the
supplier are broken down into smaller parts that meet the wishes and needs of the
last customer;
• Through trade, the producer receives information about the demand for goods,
which is the most sensitive indicator in the regulation of production;
• New markets are being developed and new products are being promoted through
trade;
• The commercial activity performs the advertising functions of the producers’
products.
The most common and major mistake when choosing a method of selecting and
evaluating the location of new branches and outlets in commercial enterprises is to
evaluate only one location. The essence of this parable is to identify and select the best
object with maximum potential from several different proposals.
Most retailers in our country build their successful trading network on an intuitive
approach. This method is justified as long as it allows the owners of the firm to participate
Estimation of Potential Locations of Trade Objects on the Basis of Fuzzy Set Theory 229
directly in the selection of the scope of work and facilities. However, the economic
situation requires that the existing methods be used to reduce the risks when deciding to
open new facilities.
Choosing which approach or method to use is not a simple matter for every retailer.
Research and experience on such projects show that the response depends on a good
understanding of management requirements, the reasons for the successes and failures
of past experiences, the data obtained from business analysis and the methods used to
assess site-facility efficiency.
Modern methods of solving this problem [3, 4]:
1. Financial analysis;
2. Checklists
3. Analogical comparison approach
4. Use of the Huff gravitation model
5. Regression analysis methods.
In the research work, these methods were analyzed in detail and their main features
were investigated:
1. The method of financial analysis - consists of the evaluation of the proposed object
on the basis of current rental prices and the selection of an expert on the basis of
these indicators. In one way or another, the rental price is taken into account when
deciding for each location.
2. Checklists - Most are used to pre-select potential areas for retailing. Based on the set
of parameters, there are a number of procedures adopted to assess the location of
any object: the area of an object, the convenience of access roads, the proximity of
competitors, the presence of shop windows and billboards, etc. In fact, these are
mandatory place requirements that can be seen in retail outlets or in advertisements
looking for “rental housing”. Sometimes, in order to decide on the selection of the
most successful locations, the company’s experts measure and evaluate each
parameter and then rank the objects according to the accumulated points.
3. Analog comparison method - is to calculate the possible sales volume for a new
object (by comparison) according to the object of the same distribution network
with a similar feature.
4. The method of using the Huff gravitational model. Based on the assumption that the
buyer chooses a store to shop in two ways: the area of the store (the larger the
better) and the distance to the exit (the smaller the better). Moreover, the depen-
dence on these parameters repeats Newton’s law of gravitation. By knowing the
speed of consumption of products, the number of residents and the availability of
the facility, the geography of the population in the area, the location and coordinates
of the main competitors, you can calculate the potential of the site in terms of future
traffic. This is a classic geomarketing tool.
5. Regression analysis method. This allows you to estimate the location of any
selected parameter, for example, in relation to other parameters of trade turnover.
This method is more statistical and accurate than checklists and analog comparison
methods. To implement this method, it is necessary to identify, measure and analyze
the factors that affect the efficiency of the store and determine the degree of impact
230 A. R. B. Oglu and S. V. M. Oglu
on the final or intermediate indicator. The simplest example of such a dependency is

the assessment of the conversion rate of visitors into buyers. Everyone understands
that there is a connection here, but after research, only a few retailers can turn this
dependence into a formula. It often turns out that this connection is not simple, and
without learning it, a big mistake can be made in calculating the sales potential of
the future store.
As a result of the analysis of the above methods and research, socio-economic
factors influencing the selection of new branches and outlets in commercial enterprises
[4] have been identified and grouped (see Table 1).
Table 1. Criteria for evaluating the location options of commercial enterprises

Buyers (potential, Obstacles Competition Expenses
real)
• Quantity • Appearance • Level of competition • Construction
• Income level • Flood of people • Types of competing costs
• Cost structure • Obstacles (d/y, major firms • Rent
• Population highways) • Saturation index • Salary level
density • Location type • Proximity of • Taxes
• Life style • Transport exchange competitors • Transportation
• Private vehicles • Condition of roads (market leaders, Large costs
• Parking firms)
• Public transport
The analysis revealed that:

1. The success of a business is determined by its location in the most crowded places.
These are, first of all, the city center, urban centers, main transport routes, subway
zones, central avenues and streets. However, the purchase price of a lease or facility
is also a very important factor for an enterprise.
2. Surveys conducted among the population revealed that the following factors
influenced the choice of the place of shopping [6]: price - 38%, quality - 38%, range
- 42%, staff - 27%, ease of accommodation - 35%, service - 10%, services - 27%,
purchase incentives - 5%, advertising - 15%, atmosphere - 18%, reputation - 20%.
Basically, several possible options for the placement of commercial facilities are
mentioned, and one of the alternatives is preferred by making an objective comparison
between them.
The analysis of the above-mentioned methods for solving the problem of choosing
the potential location of the trade object and the factors influencing this selection
process allow us to argue that this issue is a matter of choosing alternatives in a multi-
criteria environment and implemented under various socio-economic factors.
The analysis of the available methods once again shows that the criteria that affect
the implementation of the task are the most subjective, uncertain, difficult to formalize.
Recent problems are solved on the basis of the application of the theory of non-
volatile sets [5], which allows to take into account the knowledge and experience of
specialists and to take into account the uncertainty and difficult formalized factors.
In this regard, we have proposed to consider the issue of selection of potential
locations of trade objects as a matter of decision-making in the conditions of uncer-
tainty by applying the theory of fuzzy sets [6].
2 The Method of Solving the Problem of Multi-criteria

Decision-Making on the Location of Commercial Facilities
on the Basis of the Application of Fuzzy Set Theory
Trading associations usually help to develop a rating system that helps to choose the
geographical location of a particular object. In the example, Table 1 proposes a system
of 10 employee assessments that allows to assess any prospective geographical location
of the enterprise for the repair of household appliances [4].
As shown in Table 2, all alternatives scored the same number of points. From this
point of view, in such cases, it is necessary to give preference to the knowledge and
experience of managers and make a choice taking into account the weight of the
criteria. As noted, the best way to solve this type of problem is fuzzy logic methods [5,
6]. From this point of view, the task we performed for the case was considered as a
point-by-point evaluation of alternatives under conditions of uncertainty.
Thus, as in [4], suppose that the managers identified the possible placement options
of the objects as in Table 2 and evaluated them on different scores on 8 criteria.
As can be seen from Table 2, the alternative placement options for the object are
marked as A- a1, B - a2 and C- a3, respectively. The normal concept here is the N
column in Table 2 - the intervals defined by managers for each criterion.
It is proposed to conduct a numerical (point) evaluation of alternatives on the basis
of the following judgments, which do not constitute the following contradictions and
are formed by managers:
R1: If K1 = “High flow of people passing by the facility” and K2 = “Relative
purchasing power per inhabitant is low” and K3 = “If the number of customers in the
area of influence is small” and K4 = “The sales area of the facility is very large” and
K5 = “Sales If the area allows a full view” K6 = “If the number of stops around the
object is small” and K7 = “If home delivery is high” and K8 = “If public transport is
3 min away” Then Q = “It is enough to place the object in these coordinates”;
R2: If K1 = “High flow of people passing by the facility” and K2 = “Relative
purchasing power per inhabitant is high” and K3 = “If there is a large number of
customers in the area of influence” and K4 = “If the sales area of the facility is very
large” and K5 = “Sales If the area allows less full view” and If the area allows less full
view” and K6 = “If the number of stops around the object is high” and K7 = “If home
delivery is high” and K8 = “If public transport is 3 min away” Then Q = “It is very
expedient to place the object in these coordinates”;
Table 2. Examples for assessing the location options of commercial enterprises

No. Criteria (K) Evaluation (Q – satisfactory)
Norms Cretier’s Alternative options (V)
(N) weight (w) A(a1) B(a2) C(a3)
K2 Relative purchasing For a 2 100 66.67 75
power for a resident resident
200
K3 Number of customers 8000 3 66.67 79.17 95.83
in the impact zone
K1 The flow of passers-by 500 1 62.33 83.33 50
within 1 h
K4 The width of the sales Min 4 79.17 83.33 91.67
area of the object 800 kV.m
K5 Area of sales outlets Min 5 100 100 83.33
40 kV.m.
K6 Number of stops Min 10 5 75 91.67 100
places
K7 Possibility of delivery – 6 96 83.33 90
K8 Public transport 3 min 7 100 91.67 93.34
away
Sums for comparison 679.17 679.17 679.17
R3: If K1 = “High flow of people passing by the object” and K2 = “Relative

purchasing power per inhabitant is high” and K3 = “If the number of customers in the
area of influence is high” and K4 = “If the object has a large sales area” and K5 =
“Sales If the area allows a full view” and K6 = “If there are many stops around the
object” and K7 = “If home delivery is high” and K8 = “If public transport is 3 min
away” then Q = “It is more expedient to place the object in these coordinates”;
R4: If K1 = “High flow of people passing by the object” and K2 = “Relative
purchasing power per inhabitant is high” and K3 = “If there is a large number of
customers in the area of influence” and K4 = “If the object has a small sales area” and
K5 = “Sales If the area allows less full view” and K6 = “If the number of stops around
the object is small” and K7 = “If home delivery is low” and K8 = “If public transport
is less than 3 min” then Q = “It is expedient to place the object in these coordinates”;
R5: If K1 = “Low flow of people passing by the object” and K2 = “Relative
purchasing power per inhabitant is low” and K3 = “If the number of customers in the
area of influence is small”. Then Q = “It is not expedient to place the object in these
coordinates”;
R6: If K1 = “Low flow of people passing by the facility” and K2 = “Relative
purchasing power per inhabitant is average” and K3 = “If the number of customers in
the area of influence is average” and K4 = “If the sales area of the facility is average”
and K6 = “If the number of stops around is middle” and K8 = “If public transport is
average than 3 min” Then Q = “It is enough to place the object in these coordinates”.
Restoring the classification (application of fuzzy) of the terms in the

left part of the
rules considered on the basis of the criteria lðuÞ ¼ exp ðu 10Þ2 =r2k (k = 1 8)
K1, K2, K3, K4, K5, K6, K7, K8 on the reference vector of fuzzy sets (a1, a2, a3)
(k = 1¸ 8) with the help of the Gaussian membership function [5, 6] (see Fig. 1) (where
the values for rk are selected depending on the degree of significance of the criteria K1,
0.8
0.6
0.4
0.2
0
0 20 40 60 80 100
Fig. 1. Gauss membership function
K2, K3, K4, K5, K6, K7, K8, for our case r = 20 + w).
For each criterion, three term sets were used, taking into account their weight ratios
(see Table 2): Low, Middle, High. Figure 2 shows a graphical representation of the
term-sets corresponding to the criterion of human flow K1 - passing through the object.
Expediency of choosing alternatives - term-sets used to describe the linguistic variable
Y and their P = {0; 0.1; 0.2; …; 1} the following have been selected as membership functions
e PURPOSEFUL, M E
that characterize a discrete set: E- e - MORE PURPOSEFUL, S E e-
e
ENOUGH (PURPOSEFUL), V E - VERY PURPOSEFUL, U E - NOT PURPOSEFUL. e
Figure 3 shows a graphical representation of the term sets used to describe the linguistic
variable Y.
According to Table 1, the calculated values for the high-term set of the Gaussian
membership function for each alternative and criterion are given:
HUMAN FLOW THROUGH THE OBJECT FOR 1 h:
e 1 ¼ 0:0400 þ 0:5325 þ 0:0035 ;

K ð1Þ
a1 a2 a3
NUMBER OF CUSTOMERS IN THE EFFECT ZONE:
e 2 ¼ 1 þ 0:1007 þ 0:2749 ;
K ð2Þ
a1 a2 a3
Fig. 2. K1 - belonging functions of the crite- Fig. 3. Appropriateness of the choice alterna-
rion of human flow passing through the object tives Functions of the linguistic concept
RELATED BUYING POSSIBILITY

FOR A RESIDENT:
e 3 ¼ 0:1225 þ 0:4403 þ 0:9677 ;

K ð3Þ
a1 a2 a3
WIDE RANGE OF SALE OF THE OBJECT:
e 4 ¼ 0:4703 þ 0:6173 þ 0:8865 ;

K ð4Þ
a1 a2 a3
AREA OF SHOPPING SALES:
e 5 ¼ 1 þ 1 þ 0:6411 ;
K ð5Þ
a1 a2 a3
NUMBER OF STATIONS:
e 6 ¼ 0:3679 þ 0:8949 þ 1 ;
K ð6Þ
a1 a2 a3
DELIVERY OPPORTUNITY:
e 7 ¼ 0:9766 þ 0:6629 þ 0:8625 ;

K ð7Þ
a1 a2 a3
PUBLIC TRANSPORT:
e 8 ¼ 1 þ 0:9092 þ 0:9410 ;
K ð8Þ
a1 a2 a3
Given these formalisms, let us express the fuzzy rules outlined above as follows:
R1: IF X1 = Ke 1 is high and X2 = K e 2 is low and X3 = K e 3 is low and X4 = K e 4 is
e e e
high and X5 = K 5 is high and X6 = K 6 is low and X7 = K 7 is high and X8 = is high
e
then Y = S E;
R2: IF X1 = Ke 1 is high and X2 = K e 2 is high and X3 = K e 3 is high and X4 = K e4 s
e e e
high and X5 = K 5 is low and X6 = K 6 is high and X7 = K 7 is high and X8 = K 8 is e
high then Y = V E;e
e
R3: IF X1 = K 1 is high and X2 = K e 2 is high and X3 = K e 3 is high and X4 = Ke 4 is
high and X5 = K e 5 is high and X6 = K e 6 is high and X7 = K e 7 is high and X8 = K
e 8 is
high then Y = M E; e
R4: IF X1 = Ke 1 is high and X2 = K e 2 is high and X3 = K e 3 is high and X4 = Ke 4 is
e e e
low and X5 = K 5 is low and X6 = K 6 is low and X7 = K 7 is low and X8 = K 8 is e
low then Y = E;e
R5: IF X1 = Ke 1 is low and X2 = K e 3 is low and X3 = K e 4 is low then Y = U E;e
e e e
R6: IF X1 = K 1 is low and X2 = K 2 is middle and X3 = K 3 is middle and X4 = K e4
is middle X6 = K e 6 is middle and X8 = K e 8 is middle then Y = S E; e
The formula for calculating the membership functions le ðuÞ for each alternate
Ri
(i = 1 6) for the left-hand sides of each rule is given below:
n o
Ri : lLi ðaÞ ¼ min lKj ðaÞ ; ði ¼ 1 6; j ¼ 1 8Þ ð9Þ
Finally, the rules can be summarized as follows:

R1: ıf X = e
L1 then Y = e
S E;
R2: ıf X = e
L2 then Y = e
V E;
R3: ıf X = e
L3 then Y = e
M E;
R4: ıf X = e
L4 then Y = e
E;
R5: ıf X = e
L5 then Y = e
U E;
R6: ıf X = e
L6 then Y = e
S E;
Based on this rule, the evaluation of the alternatives given in Table 2 was carried
out in the notation MATLAB\Fuzzy Inferences Systems [7–11] (see Fig. 4). The
created logical derivation module used minimum, maximum and centroid, respectively,
as methods of implication, aggregation and defazification. Figure 5 graphically shows
the flow of people passing by the K1 facility for 1 h and the effect of the K6 parking
rules on the final numerical evaluation of the alternatives in Fig. 6 and Fig. 7. As can
be seen from the figures, the K1 and K6 criteria strongly influence the final decision.
Fig. 4. Rules determining the dependence of Fig. 5. Graphical interpretation of the effect of
the location of the trade object on the criteria. K1 and K2 criteria on the evaluation of alternatives
Fig. 6. Dependence of the location of the Fig. 7. Dependence of the location of the
object on the criterion of the flow of passers-by. facility on the criterion of the number of stops
3 Conclusion
The evaluation of the alternatives given in Table 2 in the notation MATLAB\Fuzzy

Inferences Systems described above allowed the following results to be obtained: for
the alternative a1 - Y1 = 0.8460; for the alternative a2 - Y2 = 0.8730; for the alternative
a3 – Y3 = 0.8360. In descending order: a3 ! 0.8360 < a1 ! 0.8460 < a2 ! 0.8730.
As you can see, the best object among these alternatives is the second object. Thus,
as can be seen from Table 2, the majority of respondents noted that this object was
superior to the others in terms of both K1 and K6 criteria. Although the total score
scores calculated in Table 2 are the same in all three alternatives. On the other hand,
other criteria also have an advantage in the rules given by managers, and the alternative
a2 is in some cases weaker than the alternatives a1 and a3 on these criteria.
The results obtained are more rhetorical, showing that in conditions of uncertainty,
it is effective to evaluate the location of new branches and outlets in commercial
enterprises on the basis of fuzzy set theory.
On the other hand, it is useful to use fuzzy set theory when selecting the location of
objects, instead of a simple scoring method, based on the knowledge and experience of
managers, taking into account other simple but uncertain and difficult-to-formalize
criteria and making judgments based on them is reasonable and economically viable.
References
1. Brizhasheva, O.V.: Marketing torgovli [Mapкeтинг тopгoвли], UlGTU, Ulyanovsk (2007)
2. Kiselev, V.M.: Metodicheskiy apparat marketingovykh kommunikatsiy v mestakh prodazh:
obonyatel’nyy merchandayzing. Marketing v roznichnoy torgovle [Meтoдичecкий aппapaт
мapкeтингoвыx кoммyникaций в мecтax пpoдaж: oбoнятeльный мepчaндaйзинг.
Mapкeтинг в poзничнoй тopгoвлe], FBK-PRESS, Moskva (2004)
3. Hiduke, G., Ryan, J.D.: Small Business: An Entrepreneur’s Business Plan 9th Edition.
South-Western College Pub. (2013)
4. Naumov, V.H.: Marketing sbyta [Mapкeтинг cбытa]. https://www.marketing.spb.ru/read/
m11/index.htm. Accessed 26 Mar 2020
5. Zadeh, L.A.: Fuzzy logic, neural networks, and soft computing. Commun. ACM 37, 77–84
(1994)
6. Borisov, A.N.: Modeli prinyatiya resheniy na osnove lingvisticheskoy peremennoy [Moдeли
пpинятия peшeний нa ocнoвe лингвиcтичec-кoй пepeмeннoй], Zinatne, Riqa (1982)
7. Beale, M.H., Hagan, M.T., Demuth, H.B.: Neural Network Toolbox. User’s Guide, Natick:
Math Works, Inc. (2014)
8. Dyakonov, V.: Matematicheskiye pakety rasshireniya MATLAB [Кpyглoв B. Maтeмa-
тичecкиe пaкeты pacшиpeния MATLAB. Cпeциaльный cпpaвoчник], SPb, Piter (2001)
9. Zimmermann, H.-J.: Fuzzy Set Theory and its Applications, 4th edn. Kluwer Academic
Publishers, Dordrecht (2001)
10. Piegat, A.: Nechetkoye modelirovaniye i upravleniye [Heчeткoe мoдeлиpoвaниe и
yпpaвлeниe], Binom, Labarotoyiya znaniy, Moscow (2013)
11. Alekperov, R.B., Ibrahimova, K.A.: Neural network modeling and estimation of the
effectiveness of the financing policy impact on the socio-economic development of the
socio-educational system. In: 13th International Conference on Theory and Application of
Fuzzy Systems and Soft Computing, ICAFS-2018, Advances in Intelligent Systems and
Computing, vol. 896, pp. 754–759. Springer, Cham (2018)
Surface Roughness Prediction Using ANFIS
and Validation with Advanced
Regression Algorithms
M. A. Vinod Kumar(&)
Larsen & Toubro Limited, Bengaluru, India

[email protected]
Abstract. For mass production, mainly automation is used, in which cutting

parameters are set to obtain required Surface Roughness (SR). The parts like IC
Engine piston, cylinders require very smooth surface finish. The same is the case
of sleeves, collets etc., of machine parts. These are made by automatic
machining operations. To get approximate value of required SR, the cutting
parameters that are to be set with help of Adaptive Neuro Fuzzy Inference
System (ANFIS) that is designed by using Fuzzy Logic Toolbox. The Fuzzy
Logic Toolbox is a collection of functions built on the MATLAB numeric
computing environment. It provides tools to create and edit fuzzy inference
systems (FIS) within the framework of MATLAB. ANFIS constructs a relation
between given parameters (input data and output data), when it is trained with
experimentally predetermined values. It consists of different functions, of which
bell and triangular membership functions are used for our purpose. The com-
parison of accuracy of predicted values for both membership functions are
performed using testing data. The training and testing data were obtained per-
forming operation on CNC lathe for 50 work pieces of which 40 were used for
training ANFIS and the remaining 10 were used for comparing the accuracy of
both Bell and Triangular membership functions. The detailed analysis and
procedure are presented. Results predicted using ANFIS are later verified and
validated using supervised learning algorithms. Here, we are adapting Advanced
regression algorithm using Python language to validate the results from ANFIS.
Variance in results from ANFIS and regression algorithm are analyzed and
evaluated.
Keywords: Regression Supervised learning ANFIS CNC lathe Surface

roughness MATLAB Python
1 Introduction
Surface Roughness (SR) contributes an important role in improving the fatigue

strength, corrosion resistance and creep life of a component. SR also affects several
functional attributes of parts, such as contact causing surface friction, wearing, light
reflection, heat transmission, ability to distributing and holding a lubricant, and coating.
When it comes to parts with good SR, the right input parameters for manufacturing the
Surface Roughness Prediction using ANFIS and Validation 239
component plays an important role in achieving the desired SR. It is important to arrive
at an empirical relation between SR and Input parameters.
We have got several methods in market to develop empirical relation between SR
and cutting parameters like speed, feed and depth of cut. Idea in this article is to
establish a relation and predict SR using Adaptive Neuro Fuzzy Inference System
(ANFIS), a tool kit in MATLAB software. Based on the predictive results from ANFIS,
we also tried to validate the results using regression algorithms from Python.
Here, stock 50 aluminum parts are turned using CNC turning machine and SR of
the parts are measured experimentally. An effort was made to obtain Adaptive Infer-
ence System in predicting the SR based on the training set of data provided by
experimental results. We also used the multiple regression analysis in Python to val-
idate the predictive results from ANFIS and understand the linearity behavior. Devi-
ation in the predictive results from the actual SR values gives an understanding of the
effectiveness of the algorithms used for the prediction. Based on these techniques or
algorithms we can get to know the input parameters for achieving the desired outputs
required during the mass production well before the production starts.
The subsequent sections directly focus on the application phase and the theoretical
background is implied wherever needed.
2 Application Phase
2.1 Training Data

Based on the cutting parameter combinations, 50 turning experiments were performed.
The SR value was measured by SR tester with a sampling length of 8 mm. The three
parameters which define the SR of the component in turning operation include speed,
feed and depth of cut. Varying these three independent parameters, the SR of the
component varies. Other parameters like coolant, equipment vibration etc., are con-
sidered outside the purview of our analysis.
FANUC CNC lathe is programmed; a total of 40 sets of data were selected from the
total of 50 sets obtained in the turning experiments for the purpose of training ANFIS
in Table 1. This training adjusts the membership function parameters. In some cases,
data is collected using noisy measurements, and the training data cannot be repre-
sentative of all the features of the data that will be presented to the model. Also, the
same training set data is used in establishing the regression model for the validation of
the predictive results obtained using ANFIS.
2.2 Adaptive Neuro Fuzzy Inference System (ANFIS)

The intelligent systems combine knowledge, techniques and methodologies from
various sources. These intelligent systems possess humanlike expertise within a
specific domain, adapt themselves and learn to do better in changing environments. In
neuro fuzzy inference systems, neural network recognizes patterns and helps to adapt
for the changing environments. Fuzzy inference systems incorporate human knowledge
240 M. A. V. Kumar
Table 1. Training data obtained by turning operation

S N in D in F in SR S N in D in F in SR
No rpm mm mm/min No rpm mm mm/min
1 1000 1 45 1.63 21 1156 1 45 1.36
2 1000 1 48 1.44 22 1156 1 48 1.41
3 1000 1 51 1.28 23 1156 1 51 1.44
4 1000 1 54 1.33 24 1156 1 54 1.5
5 1031 0.6 45 2.04 25 1187 0.6 45 2.59
6 1031 0.6 47 1.21 26 1187 0.6 47 1.15
7 1031 0.6 52 1.01 27 1187 0.6 52 1.37
8 1031 0.6 55 1.37 28 1187 0.6 55 1.34
9 1062 0.7 46 1.58 29 1219 0.7 46 2.07
10 1062 0.8 49 1.32 30 1219 0.8 49 1.16
11 1062 0.8 53 1.54 31 1219 0.8 53 1.15
12 1062 0.6 55 2.45 32 1219 0.6 55 1.38
13 1094 0.9 45 1.98 33 1250 0.9 45 2.17
14 1094 0.9 48 1.69 34 1250 0.9 48 1.23
15 1094 0.9 50 1.75 35 1250 0.9 50 1.25
16 1094 0.9 54 1.49 36 1250 0.9 54 1.24
17 1125 0.5 46 1.27 37 1300 0.5 46 1.21
18 1125 0.5 47 0.07 38 1300 0.5 47 1.32
19 1125 0.5 50 1.25 39 1300 0.5 50 1.17
20 1125 0.5 53 1.26 40 1300 0.5 53 0.64
and perform interfacing and decision-making. ANFIS is tuned with hybrid learning
algorithm based on the collection of input – output data.
A training database with regarding to cutting parameters and SR is essential to build
an ANFIS for predicting SR of turning operation performed on CNC lathe. The
experiment is carried out using a CNC lathe with a tungsten carbide tool and working
on Aluminum bars as mentioned above. Cutting speed (N), feed rate (f) and depth of
cut (d) were selected as cutting parameters to analyze their effect on SR (Ra). The
ranges of cutting parameters were selected as shown in Table 2.
Table 2. Range of cutting parameters

Cutting Parameter Range
Cutting speed (rpm) 1000–1300
Depth of cut (mm) 0.5–1.0
Feed (mm/rev) 45–55
The fuzzy inference process we are using is Takagi-Sugeno-Kang, method of fuzzy

inference, the most common methodology. Here, we are using trimf and gbell mem-
bership functions for training.
A Triangular MF is specified by three parameters {a, b, c}:
8 9
>
> 0; x a >
>
< xa =
a x b
f ðx; a; b; cÞ ¼ ba ð1Þ
>
>
cx
bxc > >
: cb ;
0 cx
and a generalized bell MF (or bell MF) is specified by three parameters {a, b, c}:
1
f ðx; a; b; cÞ ¼ xc2b ð2Þ
1þ
a
2.3 Multiple Regression Analysis Using Python

Regression is a statistical approach to modelling the relationship between a dependent
variable and a given set of independent variables. Here, in this scenario, we have 3
independent variables in the form of speed, feed and depth of cut, which drives the SR
parameter (dependent variable).
Here, we are not considering the noise parameters as independent variable which
can be a reason in getting the deviation of predictive results from the actual values of
SR. Also, the deviation in the results may get reduced as we start using a bigger sample
size for training the regression model.
Figure 1 below shows a sample regression model code from Python. Here, we
made use of Anaconda distribution of Python language to support the functionalities
required for creating a regression model. We used 40 samples mentioned in Table 1 for
establishing the regression model. Once, the model is obtained using the training data
set, we made predictions for the testing data using the regression model.
Model validation is needed to cross validate the fuzzy inference system using testing
data set. The testing data set is useful in checking the generalization capability of the
resulting fuzzy inference system. Hence, the other 10 sets were used for testing after
training was completed to verify the accuracy of predicted values of SR. Figure 2,
shows the flow chart for predicting the SR using ANFIS. Also, the validation is done
using the regression model in Anaconda Python distribution for scientific computing.
In this study, Speed, depth of cut and feed are the inputs and Ra are the output of
the system. Triangular shape and bell shape are used for the MF distribution of the
input variables. First order TSK fuzzy inference system is used in this work. The
number of fuzzy rules in a fuzzy system is related to the number of fuzzy sets for each
input variables. The three inputs (N,d,f) of the fuzzy inference system are classified into
242 M. A. V. Kumar
Fig. 1. Sample regression model code from python
3 fuzzy sets. Therefore, maximum number of rules for this system can be 27. Thus, a
typical rule will look as follows in Eq. (3).
SR ¼ pðiÞ NðiÞ þ qðiÞ dðiÞ þ rðiÞ f ðiÞ þ cðiÞ ð3Þ
where p(i), q(i), r(i) and c(i) are the design parameters referred as consequent param-
eters; N(i) is cutting speed, d(i) is depth of cut and f(i) is feed rate.
During training of ANFIS, 40 sets of experimental data were used. Once the FIS is
generated with the training set of data, the fuzzy system adapts and obtains the con-
sequent parameters based on the experimental data provided, thus becomes the base for
predicting the SR for the testing cases.
A regression model is also established using the 40-training data set from Ana-
conda Python distribution. Based on this regression model, predictive results were
obtained for the testing data set for further validation of the results.
Fig. 2. Flow chart for predicting the SR using ANFIS.
Simulation result shows that the average error of prediction of SR in triangular

shaped MF as 6.931%, average error of prediction of SR in bell-shaped MF as 4.466%
and average error of prediction of SR in regression analysis as 14.082%.
Below Fig. 3 shows a typical diagram representing the surface view for the input
parameters (Speed and Depth of cut) against the Output parameter (SR).
The average error values obtained for the triangular MF, bell-shaped MF and
regression analysis infers that the error of prediction of SR in regression analysis,
triangular shaped MF is more than that of bell-shaped MF. This means that the
accuracy rate obtained by using bell-shaped MF (95.534%) is higher than the accuracy
rate obtained by using triangular shaped MF (93.069%), regression analysis (85.918%).
This is because SR exhibits non-linearity with respect to cutting parameters.
244 M. A. V. Kumar
Table 3. Comparison of SR predicted from ANFIS, regression analysis and experimental values
of the test cases
Cutting Depth Feed rate Surface ANFIS Predicted Value Multiple
Speed of cut (mm/min) Roughne ss Traingular MF Bell MF Regression Pre
(rpm) (mm) (SR) dicte d Value
SR Abs % SR Abs % SR Abs %
Error Error Error
1000 1 55 1.46 1.69 15.753% 1.32 9.589% 1.53403 5.07%
1080 0.6 45 1.5 1.57 4.667% 1.42 5.333% 1.28667 14.22%
1100 1 45 2.22 2.15 3.153% 2.29 3.153% 1.54477 30.42%
1100 0.5 55 1.42 1.46 2.817% 1.4 1.408% 1.2904 9.13%
1115 1 48 1.8 1.7 5.556% 1.87 3.889% 1.56998 12.78%
1150 0.9 47 1.54 1.42 7.792% 1.6 3.896% 1.52549 0.94%
1175 1 55 1.49 1.54 3.356% 1.46 2.013% 1.64461 10.38%
1200 0.7 47 1.24 1.32 6.452% 1.2 3.226% 1.43435 15.67%
1225 0.8 47 1.32 1.44 9.091% 1.26 4.545% 1.51151 14.51%
1278 1 48 1.31 1.45 10.687% 1.21 7.634% 1.67298 27.71%
Fig. 3. Surface view for speed and depth of cut against SR
4 Conclusion
An ANFIS based first-order Takagi, Sugeno and Kang fuzzy inference system is used
to predict SR in turning operation performed on CNC lathe. By employing the hybrid
learning algorithm, ANFIS can obtain the optimal triangular and bell- shaped mem-
bership functions of the fuzzy system. Also, for further validation a regression algo-
rithm is used to establish a model. A total of 40 sets of experimental data are used for
training in ANFIS and regression modeling. After the training is completed, another 10
sets of data are used as testing data. SR values predicted by ANFIS; regression model is
compared with the actual measurement values derived from the 10 data sets in order to
determine the error. The error of SR values predicted by ANFIS with triangular
membership function is 6.931% (accuracy 93.069%) and error of SR values predicted

by regression analysis is 14.082% (accuracy 85.918%). In contrast, the error by ANFIS
with bell shaped membership function is 4.466% (accuracy 95.534%). The comparison
indicates that the adaptation of bell- shaped membership functions in ANFIS achieved
very satisfactory accuracy and hence can be implemented in the mass production
process where highly precise SR components are to be produced. Also, it is found that
SR exhibits a non-linearity with respect to the cutting parameters (Table 3).
Predicting Likelihood to Purchase of Users
for E-commerce
Çağlar İçer1(&), Deniz Parmaksız1, and Altan Çakır2

1
Insider (useinsider), 34415 Istanbul, Turkey
[email protected]
2
Istanbul Technical University, 34467 Istanbul, Turkey
Abstract. The gap between the marketer and the customer is increasing in the
recent years. The marketers are not able to accurately segment customers. Pre-
dictive modelling and auto-optimization technologies will be disrupting the
digital customer experience delivery space. Being able to predict the future
behavior of online and mobile visitors, the gap between marketers and customers
will decrease. In this study, with the clickstream data that have been collected
from the users on the websites, machine learning models will be created to predict
each and every users’ likelihood to purchase, so that the marketers can target only
those users, in order to have higher ROI’s in advertising world.
Keywords: Customer behavior Purchase prediction Behavior analytics

Machine learning E-commerce Digital retail
1 Introduction
1.1 Predictive Ad-Audiences with Likelihood to Purchase

In today’s world, every day brings new challenges for companies that urgently need to
battle in an increasingly competitive and changing environment. As your business
grows, traditional approaches, especially in the ecommerce context, simply doesn’t
work due to lack of customer trends and shopping behaviors. In this manner, knowl-
edge is a major source of competitive advantage and value creation to the development
of dynamic capabilities, and this is reflected in organizational success.
All e-commerce businesses produce and rely on large volumes of information –
click counts, shopping records, interactions with customers and other products inter-
ests, product details, stock requirements and so on. From this perspective, it’s too much
information to keep track of - let alone use effectively - without the right systems.
Business responsibilities and tasks can be delegated as your business grows, however
without solid management information systems the company cannot manage effec-
tively this operation. The larger your business grows, the harder it is to ensure that
information is shared and different knowledge, based on the data driven, functions
work together effectively. Putting the right infrastructure in place is an essential part of
helping your business to grow.
Insider, an innovative growing Turkish mid-size enterprise company, studies to
develop a Predictive Ad-Audiences (PAA) under Insider’s Growth Management
Predicting Likelihood to Purchase of Users for E-commerce 247
Platform (GMP), which helps marketers go beyond conversion optimization and help
them increase their ROI in acquisition to Activation, Retention, and Revenue channels,
while driving loyalty through relevant and personalized experiences and customer
expectations. The study is offering Predictive Ad Audiences (PAA) product based on
Custom Conversion Prediction (CCP) features and Likelihood to Purchase (LTP) al-
gorithm, based on the machine learning, that leads digital marketers to spend less time
and budget to find right visitors. The CCP features accelerate the process of acquiring
right visitors. That is the challenge for the digital marketers in order to catch up right
customer while spending less time, and budget while achieving high ROI and ROAS.
Predictive Ad Audience (PAA) will be faster, personalized and accessible compared to
what is readily available today. This study intends to discuss a predictive framework
for customer purchase behavior in the business-to-business (B2B) context. This
framework, developed in Insider, known as predicting likelihood to purchase of Users
(LTP), can be understood as a xx-stage process.
In the following sections, we will discuss challenges in the predictive approach in
detecting user behaviors in ecommerce challenges and the results we achieved with the
combined experimental approach in cloud-based set-up at Insider. The paper is orga-
nized as follows. The literature review for predicting likelihood to purchase models and
our approach are described in Sect. 1.2. Section 2 begins with a brief summary of the
system architecture of our set-up, discussing the LTP model for PAA approach with AI
based methodology. The methodologies for obtaining business-to-business model are
discussed between possible outcomes. An API based product is employed on these
approaches. The systematic experiments and evaluation metrics are summarized in
Sect. 2.3. Finally, the results, interpretation, and summary of the study are presented in
Sect. 3.
1.2 Challenges in the Ecommerce and Solution

Nowadays the world of the analysis is changing quickly, the reason behind is the
development of new intelligence systems able to respond to different questions thanks
the improvements of so-called machine learning and the numbers of data available by
Smart Devices and IoT. This mix of computing power and data to be analyzed created
the opportunity to design new algorithms. In this forecast is born the concept of
Predictive analysis. In digital marketing, acquiring new customers is getting more and
more expensive whereas lifetime values are getting lower. Using predictive likelihood
to purchase approach - Predictive Ad Audiences (PAA) – it is helping companies and
agencies to use digital marketing budgets more effectively by providing them with a
very valuable segments (LTP and CCP) which will bring more conversions (converting
to sales), hence improving their Return on Ad Spend (ROAS). The idea is increasing
the efficiency of technology, based with the data collection and subsequently different
way of optimizations, creating more personalized and relevant messages for increas-
ingly selective buyers and optimizing customized marketing campaigns.
248 Ç. İçer et al.
2 Related Work and Approach

2.1 Predictive Ad-Audiences
Predictive Ad Audiences (PAA) provides marketers go beyond conversion optimiza-
tion and help them increase their ROI in acquisition channels, while driving loyalty
through relevant and personalized behaviour. Predictive Ad Audiences (PAA) released
on machine learning algorithms and Artificial Intelligence AI to create ready- to-use
segments based on the future behaviors of visitors, which are continuously improved
and based on new research findings, latest industry trends and customer expectations.
One of the segments of PAA is Likelihood to Purchase (LTP) algorithm in which
identifies the likelihood of future purchase outcomes based on historical data. The goal
is to go beyond knowing what has happened to provide a best assessment of what will
happen in the future. Likelihood to Purchase algorithm is one of the machine learning
algorithms designed to be used at all stages of the purchase pattern thanks to that
Predictive Ad Audience offer. By tracking purchase patterns, our algorithm calculates
the expected purchase date range of visitors.
1. Monitor and analyze each visitor’s purchase history to predict their conversion date
range;
2. LTP algorithm learns from the buyer behavior and predicts the visitors that are more
likely to make a purchase within the next 7 days. The conversion means for LTP,
how many visitors are likely to buy something on a site (Fig. 1).
Fig. 1. The workflow of Predictive Add Audiences
Another segment of PAA is Custom Conversion Prediction (CCP) is the segment

that is a ML-based algorithm which predicts users who are likely to make any chosen
custom conversions (other than purchase tendency in which LTP segments determines)
within 7 days. As CCP learns from the behaviors of users who complete a custom
conversion which is defined by the end user, determines what kind of behaviors lead to
that conversion (e.g. which pages are visited, from which device they come from, on
which pages they land on initially etc.) and gives a coefficient to those behaviors
through its model.
2.2 Data Flow and System Architecture

Many technological services provided by Amazon Web Services are used in opera-
tional usage at Insider. Amazon Elastic MapReduce (EMR) is one of the analytic
services that basically hosts Apache Hadoop framework that is entirely built on
Amazon Elastic Compute Cloud (EC2) and Amazon Simple Storage Service (S3).
Thereby, there is no need to deal with the intricacy of installing Hadoop or opera-
tionally dealing for it. In general, EMR is used as a distributed computing environment
(cluster) that number of its master and core nodes can be scaled up or down on demand
without the need to establish complex structure of hardware. EMR cluster is able to
process several tasks of massive datasets in parallel within few minutes as well.
Moreover, Elastics search, MySQL, Redis and Cassandra solutions or more can be
added to EMR cluster and real time data processing. It plays an important role in seeing
data while it is distributed among nodes, monitoring the usage of memory and virtual
central processing unit (vCPU) cores, and detecting the running and failed applications
and their tasks.
Fig. 2. The system architecture Insider data pipeline on AWS

In this manner, each user is assigned a unique score regardless of whether the user
is logged in or anonymous. Whenever a user comes to the website, him or her unique
score changes thanks to our real time solution based on him/her behavior on the
website. Once him/her score passes the threshold that is determined automatically by
the AI (artificial intelligence) model of PAA, we say that the user is expected to make
the custom conversion defined within the next 7 days. With difference of LTP segments
of PAA in which learns from the buyer behavior and predicts the visitors that are more
likely to make a purchase, Custom Conversion Prediction (CCP) will learn rest of
activities of the customer behaviors on side. As a result of prediction schema and data
pipeline, digital marketers will obtain to segment users with the following tags:
– Likelihood to Click on a button
– Likelihood to Adding to Basket
– Likelihood to Sign Up
– Likelihood to Apply to a Job
– Likelihood to Apply for a Test Drive Likelihood to click on Ancillary Products
– Likelihood to «x» behavior that visitors (customers) have onside.
Currently, many online businesses are failing because- of low Life Time
Value/Customer Acquisition Cost. LTV refers to Life Time Value is an estimate of the
average revenue that a customer will generate throughout their lifespan as a customer.
Customer Acquisition Cost (CAC) is the spending associated in convincing a customer
to buy a product/service. Digital marketers can able to make right decisions toward
their customers. Our system architecture provides tackles the several problems affected
SaaS predictive tools by analyzing three points of view as point solution, inhouse
solution and marketing clouds.
1. With Point solution PAA can target directly right uses/visitors without extra effort;
2. No need extra IT and software expenses that European customer carry. They can
track their right audiences simple by using PAA algorithms. PAA provide inhouse
and manageable solution without extra IT, software expenditure;
3. Marketers do not need to spend expensive marketing clouds to keep all visitors to
analyze visitor’s behavior to figure out right audiences. PAA provides real time
tracking. PAA track all visitors and give outcome to the marketers who is the right
audiences. PAA is doing this process real time thanks to AI and ML models.
2.3 Approach and Results

Using the data flow pipeline in Fig. 2, LTP algorithm is the perfect tool to predict a user’s
desires in the advertising solutions. In this study we use a cognitive approach by detecting
the value of the machine learning algorithm and the segmented output to be valuable on
advertisement and marketing space earlier than other. Advertising researchers have a
long-standing interest in understanding both the degree and type of cognitive elaboration
that occurs when costumers are exposed to persuasive messages. In this approach is linked
to deep learning method that makes effective the approach. It is able to single out one user
in an online crowd, a person who may initially look like a user acting chaotically, but in
fact has the biggest potential to finalize a purchase. This is possible because PAA defines
every potential client who searches for a product slightly different than usual models. It
references history and sees that the person changed his or her behavior dynamically. It
then delivers extremely precise conversion probabilities, learning not only from one user,
but every user in the network. Consumers become as rational decision-makers and that
advertising provides consumers with information utility by reducing the need to search for
other information about a brand.
Our aim to optimize marketers’ efforts in digital platform to track right users, to
optimize digital marketing budgets and ad spending. Also, it provides compact marketing
technology that digital marketers use it without excessive IT spending. A cognitive
approach works as path to persuasion as follows: Cad Advertising cognition ! Attitude
toward advertising (Ad) ! Brand cognition (Cb) ! Brand Attitude (Ab) ! Purchase
Intention (PI). It is able to be used it on various channel (Fig. 3).
Fig. 3. Schema of advertising prediction likelihood purchase of users
Digital platforms give rise to a humongous amount of data on a daily basis, making
it vital for companies to have the necessary tools to track and sift through all the data to
find the valuable nuggets that will assist in their decision making. AI applications such
as deep learning, machine learning, adaptive learning, natural language processing, and
computer vision are utilized in marketing endeavors to enhance the efficacy of digital
marketing. These tools help companies filter out unnecessary data and zero in on the
valuable data with ease, allowing them to formulate more detailed and better thought
out competitive strategies that allow them to make the most of the tools at their disposal
shown as an example in Fig. 4.
Current stage of development PAA with LTP will validate a robust ecosystem of
prepackaged APIs, open-source software and cloud-based platforms in shown Fig. 2. It
is helping accelerate AI adoption, bringing new capabilities to speed up, scale and
personalize marketing campaigns in more economical ways.
Fig. 4. LTP Histogram- distribution of users with respect to LTP scores and cohort metrics- how
many users are segmented as LTP high every day, conversion rate day by day
3 Summary and Future Work
Development in the marketing, especially in the ecommerce, has made numerous

significant contributions to the prediction of customer purchase behavior in contrast to
the traditional business context. However, new methods and techniques must be
developed to perform the predictive task in the e-commerce context.
The main need of digital marketers is to bring more conversions of any kind and
increase ROAS (Return on Ad Spend) on ad channels. It means they need to know
which audience to target for each conversion type. As a conclusion of this study, with
predictive components of PAA (Predictive Ad Audience), our solution allows simul-
taneously run multiple conversion prediction algorithms on the cloud-based system.
The keys ensuring successful commercial exploitation are our stakeholder and strategic
partners, special focus will be on targeting B2B models. We analyzed target telecom,
airlines, automotive, publisher traffic in PAA solution to make high quality advertising
information accessible to everyone every user on the internet.
Hesitant Fuzzy Sets
Webpage Design Based on Generation
Differences Using Hesitant Fuzzy Sets
M. Çağrı Budak(&) and Sezi Çevik Onar(&)
Industrial Engineering Department, ITU, İstanbul, Turkey

{budakm17,cevikse}@itu.edu.tr
Abstract. Online shopping has a vital place in shopping behavior. Various

factors have an impact on individuals’ motivations for online shopping. In this
study, the online shopping motivation of individuals who have experienced that
earlier has been evaluated with regards to the Technology Acceptance Model,
Diffusion of Innovation Theory, and Extended Unified Theory of Acceptance
and Use of Technology. An integrated model is developed by using the inno-
vativeness and relative advantage factors from the Diffusion of Innovation
Theory; perceived usefulness, perceived ease of use, attitude towards use and
usage intention factors from Technology Acceptance Model and finally habit,
hedonic motivation and social influence from Extended Unified Theory of
Acceptance and Use of Technology. All those factors and relationships between
factors were tested using PLS-SEM. The model was employed to test the online
shopping behavior of Generation X, Y, and Z and revealed the differences
among those 3 groups. After finding out differences and significant relationships
among generations, the hesitant fuzzy set method is used to explain how
webpage design should be for each generation. The study, finally, has explained
which factors are given priority when building a webpage for target generation
based on factors of the proposed model.
Keywords: Generation cohort Hesitant fuzzy sets Web-page design

PLS-SEM
1 Introduction
Shopping via internet or online shopping has a crucial impact on our lives. Shoppers
want to go through more alternatives and reduce their shopping time. In 2019, the e-
commerce retailing market size in Turkey was approximately $31.5 billion dollars and
it has grown nearly 42% compared to 2018 [1]. Goods and services are sold by the
company to the consumer directly through the internet, this system called B2C e-
commerce or retailing e-commerce [2]. In the light of growth B2C e-commerce in
Turkey, this paper is about online shopper behavior and web-page design suggestions
by considering online shopping behavior. Online shopping behavior has been tested
with an integrated model which is proposed by using factors of diffusion of innovation
theory [3], technology acceptance model [4] as well as, finally, extended unified theory
of acceptance and use of technology [5]. In order to run model, after the data collection,
online shoppers are classified based on their generations. Generations are determined as
258 M. Çağrı Budak and S. Çevik Onar
Generation X (Gen X), Generation Y (Gen Y) and Generation Z (Gen Z) based on

generation cohort theory. The study aims finding out significant relationships and
differences among generations and explains how to web-page design should be. Main
contribution of the study to the literature is the proposed model.
The structure of the paper is as follows. Section 2 presents diffusion of innovation
theory (DOI), technology acceptance model (TAM) and unified theory of acceptance
and use of technology (UTAUT2) and factors which are used in the proposed model.
Gen X, Y and Z is also explained in this section. Section 3 explains the fundamental
concept of research model, sample data, partial least squares structural equation
modelling (PLS-SEM) and hesitant fuzzy cognitive mapping (HFCM). Scenarios are
generated and suggestions are given regards to scenarios towards web-page design and
results obtained from PLS-SEM are shown in Sect. 4. Concluding remarks are pre-
sented in Sect. 5 some future suggestions are given about web-page design and online
shoppers behavior.
2 Literature Review
DOI explains spreading of new idea or new technology on social system via which
channels. Rogers [6] claims that innovation itself, communication, time and social
system have an impact on the diffusion of innovation. An innovation goes through the
information, conviction, decision, implementation and verification phases [7]. In this
study two factors are used which are defined in DOI namely innovativeness and
relative advantage.
TAM is the model which explains the adoption of new technologies [4]. TAM is
derived from Theory of Reasoned Action [9]. Main skeleton of the model which is
proposed in this study is TAM. All factors of TAM are used.
Unified Theory of Acceptance and Use of Technology (UTAUT) is adapted from
TAM originally. UTAUT2 is broad version of UTAUT. UTAUT2 tries to measure
adoption of new technologies or ideas. In this study, hedonic motivation, habit and
social influence factors are used from UTAUT2.
2.1 Factors
In this section, general concept of factors which are used in this study are explained
briefly. Attitude towards use (ATU) is the degree to which shoppers like online
shopping idea [10]. Behavioral Intention (BI) is intent or determination towards online
shopping [11]. In this study, BI is defined as urge to continue online shopping. Hedone
is a word which comes from ancient Greek era and means pleasure [14] and hedonic
motivation (HM) is to take pleasure in doing something [5]. From online shopping
perspective HM is the degree of pleasure that customers take from online shopping.
Habit (HBT) is an act that a person performs automatically after many learnings [5]. In
this paper, operational definition of HBT is adapted to want to shop online because of
past experiences. Adopting a new idea before any other person in a social system [6] or
making a decision without affected from others [8] is called innovativeness (IN).
Perceived Ease of Use (PEoU) defined as “the degree to which a person believes that a
Webpage Design Based on Generation Differences Using Hesitant Fuzzy Sets 259
particular system would be free of effort.” [4]. From online shopping perspective,
PEoU is perception of customer how easy shopping in online [10]. Perceived Use-
fulness (PU) defined as “the degree to which a person believes that using a particular
system would enhance his or her job performance.” [4]. From online shopping view-
point, PU is a sense of fulfilment or benefit when shopping in online [12, 13]. Relative
advantage (RA) is defined as “the degree to which an innovation is perceived as being
better than the idea it supersedes.” [6]. In this study, RA is, in this case, adapted as the
degree of favoring the online shopping idea over any other online shopping methods.
Social Influence (SI) is an effect upon individual that comes from other people [5].
2.2 Generations
Generations can be classified as the individuals are born with same era, who are
affected from same historical and social phenomenon. Individuals within same group
have similar acts, opinion and attitude. In this study covers three generations: Gen X,
Gen Y and Gen Z. Members of Gen X are born between 1964–1979. Compared to the
western world, members of Gen X in Turkey do not meet with technology on their
early age. Gen X are characterized as individualistic and pragmatist [16, 17].
Members of Gen Y are born between 1980–1996. Most of the members Gen Y who
are born in Turkey grow with same technology level as peers that live in western world
[15–17].
Members of Gen Z are born in 1996 and later. Gen Z has born in high level
technology and digital era. Gen Z is very good at using technology and see the
technology as their limbs [16–18].
3 Research Model and Methodology
This study builds on the question that “what is the motivation of online shoppers to
continue online shopping?” and “what is the differences between generations?” After
determining the factors impact on online shopping study, the aims to make suggestion
upon web-page design using these factors.
3.1 Research Model and Sample Data

In order to test consumer behavior for online shopping, the model is developed. The
proposed model consists of 9 factors which are borrowed from DOI, TAM and
UTAUT2. All nine factors are adapted to context of this study. Figure 1 shows that
relationships between factors. The arrows show that relationships and points out the
relationship direction. For instance, increase on SI has an impact on HM degree. This
structure is valid for every relationship presented in the model. All relationships are
constituted as a result of the literature review and a new model has been revealed.
Fig. 1. Research model
After the model construction, data collection procedure started. A questionnaire is

prepared based on the measurement question of the literature. The questionnaire which
consists of 38 questions and 7 likert scale consists of 38 questions and 7 points Likert
scale is prepared based on the measurement question on the literature. 1 indicates full
disagreement and indicates full agreement. 822 responses are obtained, 111 of which
are invalid and 711 of which are valid. Generational breakdown of participants is as
follows: 86 for Gen X, 426 for Gen Y and 199 for Gen Z.
3.2 Methodology
PLS-SEM and HFCM methods are used to cover main purposes of the study. PLS-
SEM is used to reveal the significance of relationships between factors and HFCM is
used to make suggestions towards web-page building.
PLS-SEM
To analyze hypotheses which are represented by arrows in Fig. 1 and validity of
proposed integrated model, PLS-SEM method is used. The method consists of two
steps. In first step reliability and validity are checked. If all reliability and validity
values higher than minimum expected value, the hypotheses are decided to be exam-
ined in second step [19]. The model data should ensure internal consistency reliability,
convergent validity and discriminant validity.
Hesitant Fuzzy Cognitive Maps
HFCM is an approach that enables to make interpretation about uncertain conditions.
There are 4 steps; development of the model, collecting suggestion from experts,
development of fuzzy envelope for hesitant fuzzy linguistic sets and operation of
HFCM [21].
In the study, network model development and casual relationships between factors
are determined by using literature. Figure 1 in Sect. 3 represents the network model
and casual relationships. To compare relationships among factors, f2 values are
employed instead of expert suggestion. The f2 values, which can be seen at Table 1 in
Sect. 4, have been transformed to linguistic terms and these linguistic terms have been
used to generate trapezoidal membership function through fuzzy envelope operation.
4 Results
The hypotheses significance and strength of relationship namely f2 are shown at

Table 1. 8 out of 16 relationships for Gen X, 12 out of 16 relationships for Gen Y, and
12 out of 16 relationships for Gen Z are statistically significant. Relationships between
I ! BI, HBT ! BI and RA ! BI do not show statistical significance for all
generations.
f2 values represent the impact between casual relationships. Definition of f2 values
in a manner of linguistic terms as these: <0.02 represents there is almost no impact;
.02–0.015 represents there is weak impact; 0.15–0.35 represents there is average impact
and >0.35 represents high impact between relationships [20]. From this point of view
general structures of HFCM is created. The following part consists of the simulation of
divergent scenarios under HFCM model and suggestions towards the web-page design.
The scenarios have been run under hyperbolic tangent function and k value, which
Table 1. The significance of hypotheses and f2 values.

Generation X Generation Y Generation Z
Hypothesis f2 Significancy Hypothesis f2 Significancy Hypothesis f2 Significancy
PU ! ATU 0.183 Significant PU ! ATU 0.118 Significant PU ! ATU 0.154 Significant
PEoU ! 0.001 Non- PEoU ! 0.12 Significant PEoU ! 0.087 Significant
ATU significant ATU ATU
PEoU ! 0.101 Significant PEoU ! 0.064 Significant PEoU ! 0.221 Significant
PU PU PU
PEoU ! BI 0.002 Non- PEoU ! 0.015 Significant PEoU ! 0.026 Significant
significant BI BI
RA ! BI 0.004 Non- RA ! BI 0.002 Non- RA ! BI 0 Non-
significant significant significant
RA ! PU 0.718 Significant RA ! PU 0.374 Significant RA ! PU 0.352 Significant
IN ! BI 0.002 Non- IN ! BI 0.001 Non- IN ! BI 0.013 Non-
IN ! PEoU 0.084 Significant IN ! 0.106 Significant IN ! 0.013 Non-
PEoU PEoU significant
HM ! 0.087 Significant HM ! 0.072 Significant HM ! 0.087 Significant
ATU ATU ATU
HM ! 0.152 Significant HM ! 0.316 Significant HM ! 0.21 Significant
HBT HBT HBT
HBT ! BI 0.012 Non- HBT ! BI 0.003 Non- HBT ! BI 0.002 Non-
HBT ! 0.064 Non- HBT ! 0.008 Non- HBT ! 0.038 Significant
ATU significant ATU significant ATU
SI ! HM 0.173 Significant SI ! HM 0.16 Significant SI ! HM 0.113 Significant
SI ! ATU 0.041 Non- SI ! ATU 0.083 Significant SI ! ATU 0.055 Significant
significant
SI ! AF 0.055 Non- SI ! AF 0.054 Significant SI ! AF 0.108 Significant
significant
ATU ! BI 3.374 Significant ATU ! BI 1.847 Significant ATU ! BI 1.605 Significant
represents time dependent changes, is taken as 0.25. For each generation a scenario has
been generated. For Gen X “what if PEoU lacks for the web-page?” and for Gen Y
“what if HM lacks for the web-page?” and for Gen Z “what if SI lacks for the web-
page?” are the questions that have been asked.
4.1 Scenarios
This scenario is generated for Gen X and simulates the lack of PEoU at a web-page.
Figure 2 shows lack of PEoU condition. The lack of PEoU decreases ATU which
represents positive idea towards online shopping, in the beginning and BI which
represents the intention to shop decreases afterwards. With regards to this scenario, the
web-page has to seems wieldy and emphasise and convince Gen X web-page is easy
shopping.
0.5
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
-0.5
-1
ATU BI HBT HM IN PEOU PU RA SI
Fig. 2. HFCM simulation for Gen X.
0.5
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
-0.5
-1
ATT BI HBT HM IN PEOU PU RA SI
Fig. 3. HFCM simulation for Gen Y.

0.5
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
-0.5
-1
ATT BI HBT HM IN PEOU PU RA SI
Fig. 4. HFCM simulation for Gen Z.
The second scenario is about Gen Y and HM. Figure 3 shows lack of HM situation.
When HM does not exist for shopping Gen Y does not want to shop online. ATU and
BI decrease rapidly in short term. To attract Gen Y more, a web-page should contain
hedonic items. Gen Y has to be entertained when they are shopping online.
Third scenario is simulated for Gen Z. Figure 4 shows absence of SI case. The lack
of SI decreases ATU after first iteration and BI fallowing ATU at other iteration. To
increase attractability of web-page for Gen Z, Gen Z has to be convinced about most of
the people use this web-page. For that purpose influencer may be usefull to affect
Gen Z.
5 Conclusion
In this study, we focused on factors that affect online shopping and web-page design.
For these purposes, we used PLS-SEM and HFCM methods. First of all, we developed
9-factor-model which is gathered from 3 main models. We examined their relationships
and we found statistically significant relations for each generation separately. Solutions
obtained from PLS-SEM is used to generated scenario. Effect size between significant
relationships were transformed to linguistic terms for HFCM method. For each gen-
eration, a different scenario was generated. One factor is taken away from the model
which affects the ATU and BI. Simulation has been shown and the result of simulation
interpreted briefly for each generation separately.
Although the study analyzes various possible conditions, it has several limitations.
Increasing sample size may changes either the relationship between factors or simulation
of the absence of PEoU for Gen X. To construct HFCM, PLS-SEM results were used.
Future studies may also construct the HFSM construction with an expert opinion.
And also, future research may remove factors from the model that are non-significant
for all factors and additional factors may be added in the model.
References
1. Göl, H., İlhan, E., Ot, İ.: E-Ticaretin Gelişimi, Sınırların Aşılması ve Yeni Normlar.
TUBİSAD (2019)
2. Turban, E., King, D.R., Liang, T.P., Turban, D.C.: Electronic Commerce: A Managerial and
Social Networks Perspective. Springer, Cham (2015)
3. Rogers, E.: M: Diffusion of Innovation. Free Press of Glencoe, New York (1962)
4. Davis, F.D.: Perceived usefulness, perceived ease of use and user acceptance of information
technology. MIS Q. 13, 319–340 (1989)
5. Venkatesh, V., Thong, J.Y., Xu, X.: Consumer acceptance and use of information
technology: extending the unified theory of acceptance and use of technology. MIS Q. 36,
157–178 (2012)
6. Rogers, E.M.: Diffusion of Innovation. Free Press, New York (1983)
7. Demir, K.: Rogers`ın Yeniliğin Yayılması Teorisi ve İnternetten Ders Kaydı. Kuram ve
Uygulamada Eğitim Yönetimi 47, 367–391 (2006)
8. Lee, S.: An integrated adoption model for e-books in a mobile environment: evidence from
South Korea. Telematics Inform. 30, 165–176 (2013)
9. Lingyun, Q., Dong, L.: Applying TAM in B2C E-commerce research: an extended model.
Tsinghua Sci. Technol. 13, 265–272 (2008)
10. Vijayasarathy, L.R.: Predicting consumer intentions to use on-line shopping: the case for an
augmented technology acceptance model. Inf. Manag. 41, 747–762 (2004)
11. Crespo, A.H., Rodriguez, I.A.: Explaining B2C e-commerce acceptance: an integrative
model based on the framework by Gatignon and Robertson. Interact. Comput. 20, 212–224
(2008)
12. Lestari, D.: Measuring e-commerce adoption behaviour among gen-Z in Jakarta, Indonesia.
Econ. Anal. Policy 63, 103–115 (2019)
13. Liébana-Cabanillas, F., Alonso-Dos-Santos, M.: Factors that determine the adoption of
Facebook commerce: the moderating effect of age. J. Eng. Technol. Manag. 44, 1–18 (2017)
14. Tilley, J.J.: Hedonism. Encyclopedia of Applied Ethics, 2nd edn, pp. 566–573. Academic
Press, San Diego (2018)
15. Lissitsa, S., Kol, O.: Generation X vs. Generation Y–A decade of online shopping.
J. Retailing Consum. Serv. 31, 304–312 (2016)
16. Kurz, C.J., Li, G., Vine, D.J.: Are millennials different? In: Handbook of US Consumer
Economics, pp. 193–232. Academic Press, San Diego (2019)
17. İlhan, Ü.D.: Kuşaklar Arası Çalışma Değerleri Algısı ve Örgüte Duygusal Bağlılık: X ve Y
Kuşaklarının Farklılıkları Üzerine Bir Araştırma. Doctoral dissertation. Retrieved from YÖK
Ulusal Tez Merkezi. (Accession No: 560693) (2019)
18. Priporas, C.V., Stylos, N., Fotiadis, A.K.: Generation Z consumers’ expectations of
interactions in smart retailing: a future agenda. Comput. Hum. Behav. 77, 374–381 (2017)
19. Anderson, J.C., Gerbing, D.W.: Structural equation modeling in practice: a review and
recommended two-step approach. Psychol. Bull. 103, 411–423 (1988)
20. Nunnally, J.C., Bernstein, I.H.: Psychometric Theory. McGraw-Hill, New York (1994)
21. Çoban, V., Onar, S.Ç.: Analysing of solar energy generation capacity using fuzzy cognitive
maps. Int. J. Comput. Intell. Syst. 10, 1149–1167 (2017)
A Hesitant Fuzzy Linguistic Group Decision
Making Model for Energy Storage Unit
Selection
Ahmet Aktas(&) and Mehmet Kabak
Department of Industrial Engineering, Gazi University, 06570 Ankara, Turkey

[email protected]
Abstract. Hybrid energy systems are considered to be an appropriate solution

for sustainable and clean energy supply. Due to the variability of energy
amounts that can be generated by using renewable resources, energy intermit-
tencies and voltage fluctuations may occur in these systems. Placement of
energy storage units into hybrid energy systems are suggested to overcome the
mentioned problems. There are several types of energy storage units at different
technical conditions and prices, and consideration of several criteria simulta-
neously makes the selection of the most appropriate unit a multiple-criteria
decision making problem. Moreover, evaluation of decision elements is usually
not possible with precise quantitative values and decision makers sometimes feel
hesitant on the assessment of decision elements. The main aim of this study is to
propose a group decision making model based on Hesitant Fuzzy Linguistic
Terms Set to determine the most appropriate energy storage unit in hybrid
energy systems to overcome the aforementioned difficulties. Energy storage unit
selection criteria were determined by a literature review, then criteria evaluation
of experts were collected and the importance degree of each criterion was
determined by pairwise comparisons. The obtained results of the study would
support storage unit investment decisions in hybrid energy systems.
Keywords: Hesitant Fuzzy Linguistic Term Sets Energy storage unit

selection Hybrid energy systems
1 Introduction
Energy is a vital part of human life. The rapid growth of economy around the world and
increase of population lead a great accrual in energy demand [1]. Moreover, some
experiences like global warming and depletion of fossil fuels have emerged worries
about energy supply and people started to give more attention to the efficient use of fossil
fuels and renewable energy resources [2]. Renewable energy resources are generally
defined as the resources of energy which can be recycled and do not pollute the envi-
ronment [3]. These resources are classified as traditional and new renewables groups by
experts. Large hydropower plants and directly burnt biomass represents traditional
renewables group while new renewable energy includes small hydropower plants, solar
energy, wind energy, biomass energy, geothermal energy and ocean energy, etc. [4].
266 A. Aktas and M. Kabak
Increasing the exploitation of renewable resources gets more importance due to

generation costs, fossil fuels and environmental issues [5]. Renewable energy is con-
sidered among the most promising solutions for energy supply in today’s and future
world.
Since renewable resources promise great potential, there are some issues must be
solved in order to supply intermittent energy without voltage fluctuations before the
designation of an energy system containing these resources. Renewables are dependent
on climate conditions and energy generation amounts are variable [6]. To overcome the
voltage fluctuation and intermittency problem, the combination of different resources
and/or grid connection with/without storage units is suggested. This kind of energy
system is called as hybrid energy system. The effects of variability of generation
amounts caused by the climate conditions on energy supply is tried to be minimized by
hybrid energy system.
In this study, the selection of the most appropriate energy storage unit decision,
which is an important decision at the design stage of hybrid energy systems, is con-
sidered. This is an important strategic decision for energy network investments,
because there are various alternative storage units with different qualifications on
several attributes such as capacity levels, prices, etc. Looking at the decision at this
point of view makes it possible to model it as a multi-criteria decision-making
(MCDM) problem. Decision problems, which contain a finite number of criteria and a
finite number of alternatives, can be modeled as a multi-criteria decision-making
problem [7].
The main aim of this study is to develop a scientific approach for energy storage
unit selection in order to provide a decision support to decision makers. Decision
makers may feel hesitate about the assessment of decision elements, to expose this
situation, a hesitant fuzzy linguistic terms set (HFLTS) based group decision making
model was proposed. Energy storage unit selection criteria were determined by liter-
ature review and importance degree of selected criteria were calculated by the proposed
model.
The rest of the paper is organized as follows. In Second Part, a literature review on
multi-criteria decision-making applications of energy storage systems is presented with
the considered criteria in these studies. A short explanation of hesitant fuzzy linguistic
terms set and proposed decision making approach are given in the Third Part. Fourth
Part consists of an application of the proposed approach to calculate the importance
degree of the selected criteria. This study is concluded in the Fifth Part with further
research suggestions.
2 Literature Review
Evaluation of energy storage systems has taken attention of researchers in the last
decade. Different methods were applied by researchers to determine the best option
among several energy storage systems.
An assessment model for thermal energy storage alternatives based on Fuzzy
Technique of Order Preference by Similarity to Ideal Solution (F-TOPSIS) approach
was proposed [6]. Daim et al. [8] determined energy storage technology evaluation
A Hesitant Fuzzy Linguistic Group Decision Making Model 267
criteria by Fuzzy Delphi (F-Delphi) technique, then obtained the criteria weights by
Analytic Hierarchy Process (AHP) and evaluated alternative technologies by Fuzzy
Consistent Matrix method. Raza et al. [9] determined the best energy storage alter-
native by using a weighted sum based multi-criteria decision making method.
Ozkan et al. [10] utilized a hybrid method based on Type-2 Fuzzy (T2F) Sets, AHP
and TOPSIS methods on energy storage alternative selection. Large-scale storage
alternative assessment was made by using weighted sum method [11]. Alternative
energy technologies were prioritized by Ren [12] using a hybrid Interval Analytic
Hierarchy Process (I-AHP) and Intuitionistic Fuzzy Combinative Distance-based
Assessment Method (IF-CODAS) methodology. Zhao et al. [13] determined the best
battery energy storage system by using a hybrid method including F-Delphi for criteria
determination, Entropy and Best – Worst Method (BWM) for criteria and sub-criteria
weighting, and VIKOR method for alternative evaluation. Hesitant Fuzzy (HF) AHP –
TOPSIS model was used by Acar et al. [14] to analyze the sustainability of energy
storage systems. Zhang et al. [15] used Intuitionistic Fuzzy Full Multiplicative Form of
Multi-Objective Optimization by Ratio Analysis (IF-MULTIMOORA) technique for
the assessment of energy storage technologies. Zhao et al. [16] analyzed battery energy
storage systems by using F-Delphi, BWM and Fuzzy Cumulative Prospect Theory
hybrid methodology.
As we can understand from the literature summary, researchers have evaluated
energy storage systems by different methods so far. None of these studies utilizes
HFLTS methodology. The criteria taken into consideration in these studies were also
different. Indicators related to technical, economic, environmental and social aspects of
energy storage systems are confronted. Commonly considered criteria can be listed as
follows:
• Capital cost
• Operating and maintenance costs
• Energy cost
• Efficiency
• Energy density
• Life time
• Reliability
• Storage capacity
• Technology maturity
• Environmental impact
• Social acceptability
3 Proposed Methodology
Experts evaluating the elements in decision problems defined under uncertainty cannot
express their knowledge by using a single term. When experts hesitate about several
values for a linguistic variable, the fuzzy linguistic approach is very limited. This is
caused just because of the need to more complex linguistic terms to define their
thoughts than the terms in the defined linguistic term set.
Rodriguez et al. [17] proposed a decision model, which manages the linguistic
expressions that are represented by Hesitant Fuzzy Linguistic Term Sets. The aim of
the HFLTS is to improve the elicitation of linguistic information, mainly when experts
hesitate among several values to assess linguistic variables. The proposed evaluations
can semantically be represented by means of HFLTS and generated by a context-free
grammar.
Yavuz et al. [18] developed a multi-criteria group decision-making approach based
on HFLTS which can handle complex multi-criteria problems with hierarchical
structure and their approach use a fuzzy representation for comparative linguistic
expressions based on a fuzzy envelope for HFLTS.
Readers may refer to the related papers for detailed expressions of HFLTS and the
group decision making approach. Due to the page restrictions, we provide the appli-
cation steps in the application part of the study.
4 Application of the Proposed Methodology
Investment on energy systems are strategic and important decisions. So, they have to be
made after careful consideration. In this part, a storage unit selection model is presented
by using the group decision making methodology based on HFLTS, which is proposed
in [18]. This method is firstly being used for energy storage alternative evaluation.
Application steps of the model goes on as follows:
Step 1: Definition of goal, criteria and alternatives related to the decision making
problem.
The aim of this study is to determine the best energy storage unit. 4 main and 11
sub-criteria are determined after a literature review on MCDM applications in energy
storage systems. Hierarchical structure of the problem is given in Fig. 1.
Step 2: Definition of linguistic terms and context – free grammar.
According to Yavuz et al. [18] the production rules for context – free grammar can
be expressed as follows, where “|” symbol indicates alternative elements:
P = {I ::= <primary term> | <composite term>
<composite term> ::= <unary relation> <primary term> | <binary relation> <pri-
mary term> <conjunction> <primary term>
<primary term> ::= s0 | s1 | … | sg |
<unary relation> ::= lower than | greater than | at least | at most
<binary relation> ::= between
<conjunction> ::= and}.
Capital Cost (C11)

Economic (C1) Opera ng and Maintenance Cost (C12)
Energy Cost (C13)
Efficiency (C21)
Energy Density (C22)
Life me (C23)
Best energy storage unit Technical (C2)
Reliability (C24)
Technology Maturity (C25)
Storage Capacity (C26)
Social (C3) Social Acceptability (C31)
Environmental (C4) Environmental Impact (C41)
Fig. 1. Hierarchical structure of the problem
Primary terms used in context–free grammar in this study are presented in Table 1.
Table 1. Linguistic scale used in the study

Importance degree Linguistic term
0 No importance (n)
1 Very low importance (vl)
2 Low importance (l)
3 Medium importance (m)
4 High importance (h)
5 Very high importance (vh)
6 Absolute importance (a)
Step 3: Collecting preferences of experts for criteria and sub-criteria.

Two experts’ preferences are collected to assess main and sub-criteria. One of the
experts is the manager of a company in renewable energy sector and the other expert is
an academician working on renewable energy system planning. Collected preferences
of the experts for main criteria are given in Table 2 and Table 3, respectively. Due to
the paper limitations, only preferences and calculations for main criteria evaluation are
given in this study.
Table 2. Evaluation of main criteria by Expert 1.

C1 C2 C3 C4
C1 – between l and m at least h between m and vh
C2 between m and h – at least vh at least h
C3 at most l at most vl – is n
C4 between vl and m at most l is a –
Table 3. Evaluation of main criteria by Expert 2.

C1 C2 C3 C4
C1 – at most m at least m between m and h
C2 at least m – at least h between m and vh
C3 at most m at most l – at most vl
C4 between l and m between vl and m at least vh –
Step 4: Transformation of experts’ preferences into HFLTS.

In this step, collected preferences of experts are transformed into HFLTS by using
grammar rules.
Step 5: Obtaining envelopes for each HFLTS.
Each envelope shows the lower and upper term of HFLTS.
Step 6: Determination of optimistic and pessimistic collective preferences by using
a linguistic aggregation operator.
Optimistic and pessimistic collective preferences are calculated by using arithmetic
means and presented in Table 4, as follows:
Table 4. Optimistic and pessimistic collective preferences for main criteria

Optimistic Pessimistic
C1 C2 C3 C4 C1 C2 C3 C4
C1 – 2.000 4.000 3.000 – 0.667 2.333 2.000
C2 3.333 – 4.000 3.667 2.000 – 3.000 2.333
C3 1.667 1.000 – 0.333 0.000 0.000 – 0.000
C4 2.000 1.667 4.000 – 1.000 0.333 3.667 –
Step 7: Building the vector of intervals for collective preferences.

Vector of intervals are built by the arithmetic mean of rows in pessimistic and
optimistic aggregate preferences matrices.
Step 8: Obtaining priority values by normalization of interval values.

Interval midpoints are calculated and normalized. Normalized values show the
importance degree of main criteria. Calculations in Step 7 and Step 8 for the impor-
tance degree of main criteria are given in Table 5 as follows:
Table 5. Interval utilities and criteria weights

Interval Midpoints Weights
utilities
C1 1.667 3.000 2.333 0.292
C2 2.444 3.667 3.056 0.382
C3 0.000 1.000 0.500 0.063
C4 1.667 2.556 2.111 0.264
According to Table 6, the most important criterion among four main criteria is
Technical criteria. It is followed by Economic, Environmental and Social.
Same steps are followed for sub-criteria of economic and technical criteria. Since
social and environmental criteria consist one sub-criteria for each, no calculations are
needed for their sub-criteria. The importance degree of sub-criteria is presented in
Table 6, as follows:
Table 6. Weights of sub-criteria

Main Sub- Weight of main Local weight of sub- Global weight of sub-
criteria criteria criteria criteria criteria
C1 C11 0.2917 0.3472 0.1013
C12 0.3889 0.1134
C13 0.2639 0.0770
C2 C21 0.3819 0.1778 0.0679
C22 0.0583 0.0223
C23 0.1583 0.0605
C24 0.2000 0.0764
C25 0.1389 0.0530
C26 0.2667 0.1019
C3 C31 0.0625 1.0000 0.0625
C4 C41 0.2639 1.0000 0.2639
Table 6 shows the importance degree of each sub-criterion. These results show that
the most important factor for the selection of an energy storage unit is Environmental
Impact (C41). It is followed by Operating and Maintenance Costs (C12) and Tech-
nology Maturity (C26). Energy Density (C22) is seen as the least important factor.
5 Conclusion
Energy investments require a high amount of capital and their effects last for a long
time. For this reason, energy investments should be considered carefully. In this study,
a group decision making methodology based on hesitant fuzzy linguistic terms set is
proposed for the energy storage unit selection decisions in hybrid energy systems.
Weights (importance degree) of selection criteria are determined by using the proposed
approach. The obtained results will support the decisions of energy system investors. In
further studies, this study can be extended by evaluation of alternative energy storage
systems.
Acknowledgements. Ahmet Aktas was a visiting researcher at University of Coimbra during a

part of this study. The author acknowledges the support of the Scientific and Technological
Research Council of Turkey (TUBITAK-2214-A program).
References
1. Kousksou, T., Bruel, P., Jamil, A., El Rhafiki, T., Zeraouli, Y.: Energy storage: applications
and challenges. Sol. Energy Mater. Sol. Cells 120, 59–80 (2014)
2. Poullikkas, A.: Implementation of distributed generation technologies in isolated power
systems. Renew. Sustain. Energy Rev. 11, 30–56 (2007)
3. Tükenmez, M., Demireli, E.: Renewable energy policy in Turkey with the new legal
regulations. Renew. Energy 39, 1–9 (2012)
4. Peidong, Z., Yanli, Y., Jin, S., Yonghong, Z., Lisheng, W., Xinrong, L.: Opportunities and
challenges for renewable energy policy in China. Renew. Sustain. Energy Rev. 13, 439–449
(2009)
5. Rehman, S., Al-Hadhrami, L.M., Alam, M.M.: Pumped hydro energy storage system: a
technological review. Renew. Sustain. Energy Rev. 44, 586–598 (2015)
6. Cavallaro, F.: Fuzzy TOPSIS approach for assessing thermal-energy storage in concentrated
solar power (CSP) systems. Appl. Energy 87, 496–503 (2010)
7. Kabak, M., Sağlam, F., Aktaş, A.: Usability analysis of different distance measures on
TOPSIS. J. Fac. Eng. Archit. Gazi Univ. 32(1), 35–43 (2017)
8. Daim, T.U., Li, X., Kim, J., Simms, S.: Evaluation of energy storage technologies for
integration with renewable electricity: Quantifying expert opinions. Environ. Innov. Soc.
Transit. 3, 29–49 (2012)
9. Raza, S.S., Janajreh, I., Ghenai, C.: Sustainability index approach as a selection criteria for
energy storage system of an intermittent renewable energy source. Appl. Energy 136, 909–
920 (2014)
10. Özkan, B., Kaya, İ., Cebeci, U., Başlıgil, H.: A hybrid multicriteria decision making
methodology based on type-2 fuzzy sets for selection among energy storage alternatives. Int.
J. Comput. Intell. Syst. 8(5), 914–927 (2015)
11. Vo, T.T.Q., Xia, A., Rogan, F., Wall, D.M., Murphy, J.D.: Sustainability assessment of
large-scale storage technologies for surplus electricity using group multi-criteria decision
analysis. Clean Technol. Environ. Policy 19, 689–703 (2017)
12. Ren, J.: Sustainability prioritization of energy storage technologies for promoting the
development of renewable energy: a novel intuitionistic fuzzy combinative distance-based
assessment approach. Renew. Energy 121, 666–676 (2018)
13. Zhao, H., Guo, S., Zhao, H.: Comprehensive performance assessment on various battery
energy storage systems. Energies 11, 2841 (2018)
14. Acar, C., Beskese, A., Temur, G.T.: A novel multicriteria sustainability investigation of
energy storage systems. Int. J. Energy Res. 43, 6419–6441 (2019)
15. Zhang, C., Chen, C., Streimikiene, D., Balezentis, T.: Intuitionistic fuzzy MULTIMOORA
approach for multi-criteria assessment of the energy storage technologies. Appl. Soft
Comput. J. 79, 410–423 (2019)
16. Zhao, H., Guo, S., Zhao, H.: Comprehensive assessment for battery energy storage systems
based on fuzzy-MCDM considering risk preferences. Energy 168, 450–461 (2019)
17. Rodriguez, R.M., Martinez, L., Herrera, F.: Hesitant fuzzy linguistic term sets for decision
making. IEEE Trans. Fuzzy Syst. 20(1), 109–119 (2012)
18. Yavuz, M., Oztaysi, B., Onar, S.C., Kahraman, C.: Multi-criteria evaluation of alternative-
fuel vehicles via a hierarchical hesitant fuzzy linguistic model. Expert Syst. Appl. 42, 2835–
2848 (2015)
Solar Energy Power Plant Investment Selection
with Unbalanced Hesitant Fuzzy
Linguistic MULTIMOORA Method
Based Score-HeDLiSF
Veysel Çoban1(&) and Sezi Çevik Onar2

1
Industrial Engineering Department, Bilecik Seyh Edebali University,
Gülümbe, Bilecik, Turkey
[email protected]
2
Industrial Engineering Department, Istanbul Technical University, Macka,
Istanbul, Turkey
Abstract. Solar energy is the most important alternative and renewable energy
source in meeting the increasing energy demand. Despite technological advan-
ces, solar energy systems have a high initial cost. Therefore, the installation of
solar energy systems in the right place is of great importance for the return of high
investment costs. The Hesitant Fuzzy Linguistic Term Set (HFLTS) is an
important tool in integrating experts’ complex linguistic knowledge into
decision-making process. In this study, the psychological orientations of the
specialists are taken into consideration by using unbalanced HFLTS based on the
score function of hesitant fuzzy linguistic term set based on hesitant degrees and
linguistic scale functions (Score-HeDLiSF) method. Unbalanced HFLTSs based
Score-HeDLiSF is used in the Multi-Objective Optimization on the basis of a
Ratio Analysis plus the full MULTIplicative form (MULTIMOORA) method
developed by ORESTE method and alternative solar plants are evaluated.
Keywords: HFLTS Score-HeDLiSF MULTIMOORA Solar power plant
1 Introduction
Linguistic terms are an important tool used to reflect the knowledge and experience of
experts in decision-making problems. Fuzzy linguistic approaches are developed that
allow specialists to make their evaluations more realistic and comfortable under
uncertain and imprecise evaluation conditions. Hesitant Fuzzy Linguistic Term Set
(HFLTS) [1] and Hesitant Fuzzy Linguistic Elements (HFLEs) [2] is developed to
eliminate the ambiguity between linguistic terms used by experts in the evaluation
process. Experts with different personal and psychological characteristics perceive
linguistic terms differently and give different value [3]. Therefore, unbalanced HFLTS
is recommended to eliminate the complexity of the experts’ evaluation and to reflect
their complex information more accurately. The score function [4] is used to resolve the
computational complex in HFLTS and convert linguistic expressions to numerical
values. The Score Function of HFLTS based on Hesitant Degrees and Linguistic Scale
Solar Energy Power Plant Investment Selection 275
Functions (Score-HeDLiSF) is an important tool for resolving unbalanced HFLTS [3].

Since experts have different knowledge and experience characteristics, the impact
levels of their evaluations in the decision-making process are also different.
Solar energy is the most important and popular resource among renewable energy
systems. Investment decisions of solar energy systems are important decision making
problems since they include multiple objectives, especially high installation costs [5].
Sectoral and academic experts with different knowledge and experience are included in
the evaluation process of the solar energy power plant investment selection (SEPPIS)
decision problem. Determining the method that enables the experts to make their
evaluations more stable and reflects their effects on the evaluation most accurately is
the motivating aims of the study. In this study, a tool is developed to make a solar
power plant investment decision using the MULTIMOORA model based on Score-
HeDLiSF [6]. Important contributions are made to the literature with this original
feature.
The next sections of the study are organized as follows: Sect. 2 contains the
implementation steps, basic concepts and notations of the Unbalanced HFL-
MULTIMOORA method. Experts evaluate the alternatives and criteria that numeri-
cal expressions are insufficient for definition with hesitant linguistic expressions.
Linguistic evaluations are converted into HFLEs and the HFLE’s score values are used
to calculate expert and criterion weights. Criterion and expert weights are aggregated to
achieve subordinating and final rankings. Chapter 3 performs a sample application for
the evaluation of solar power plant investment alternatives. Four alternatives are
evaluated by four experts according to the four basic criteria (technical, economic,
environmental, social). In the conclusion part, the study is generally evaluated and
future studies are mentioned.
2 Unbalanced HFL-MULTIMOORA Method
The weight values of the experts in the evaluation process are determined based on the
pairwise comparisons of the evaluations. MULTIMOORA decision making method is
developed with unbalanced HFL and ORESTE method. The decision making problem
consists of alternatives as A ¼ fa1 ; a2 ; . . .; ai , evaluation criteria as
C ¼ fc1 ; c2 ; . . .; cj g, and experts as E ¼ fe1 ; e2 ; . . .; ek g. The linguistic term set is
defined as S ¼ fst ; . . .; 0; . . .; st g [7]. Linguistic evaluations by experts based on
context-free grammar are transformed into Hesitant Fuzzy Linguistic Elements

HFLEs; hmn
s
ð pÞ
with transformation functions ðEGH Þ as follows [1]:
• E ðG HÞðs iÞ ¼ fs ijs i 2 Sg
• EGH ðat least si Þ ¼ sj jsj 2 S and sj si
• E ðG HÞðat most s iÞ ¼ fs jjs j 2 S and s j s ig
• EGH ðless than si Þ ¼ sj jsj 2 S and sj \si
• E ðG HÞðmore than s iÞ ¼ fs jjs j 2 S and s j [ s ig
• E ðG HÞðbetween s i and s jÞ ¼ fs kjs k 2 S and s i s k s jg
276 V. Çoban and S. Ç. Onar
hmn
s
ð pÞ
means that HFLE is generated by expert ep to evaluate the alternative am
according to the criterion cn . Experts with different knowledge and experience perceive
alternatives and criteria from different perspectives and evaluate them at different
scales. Decision matrix, M p of expert ep is defined as;
2 3
hs11ð pÞ h1j
s
ð pÞ
p 6 . .. .. 7
M ¼ 4 .. . . 5
hi1
s
ð pÞ
hijs ð pÞ
A comprehensive hesitant degree (C) is calculated for each expert (ep ) to determine
the weights of the experts [3, 8, 9].
1 XX
j i
CHp d ¼ Hd ðhmnðpÞ
s Þ ð1Þ
i j m¼1 n¼1
Lðhs Þlnðhs Þ
Hd ðhs Þ ¼ ð2Þ
ð2t þ 1Þlnð2t þ 1Þ
where Hd ðhs Þ is a hesitant degree of hs and Lðhs Þ represents the number of linguistic
terms in hs . Expert weights are calculated by normalizing their comprehensive hesitant
degrees.
0:5 CHp d
w p ¼ Pk ð3Þ
p¼1 ð0:5 CHp d Þ
The experts’ evaluations are brought together under the collective decision matrix.
Xk
E hmn
s ¼ p¼1
wp
E h mnð pÞ
s ð4Þ
Differentiation of linguistic scale functions causes to change the semantic of the

linguistic term. The semantic differentiation of linguistic terms can be prevented by
defining the HFLE’s score values.
1 H d ð hs Þ X B
E ð hs Þ ¼ gðsab Þ ð5Þ
B b¼1
where hs ¼ fsab jsab 2 S; b ¼ 1; 2; . . .; B; sab 2 fn; . . .; 1; 0; 1; . . .; ng and gðsab Þ

represents the semantic of sab . Score functions based on semantics of linguistic terms
are defined as Score-HeDLiSF [10].
8 n
>
> d db
> ; b ¼ n; . . .; 0
< 2dn 2
g sb ¼ ð6Þ
>
> dn þ db 2
>
: ; b ¼ 0; . . .; n
2dn 2
where n means that threshold value and is defined as a value greater than one. In this
study, n value is accepted as 1.5 according to the literature applications [10]. A con-
sistency test is performed to measure the consistency of expert evaluations. The
pairwise comparison matrices of the criteria are converted to the HFLE matrix and the
score values of the HFLEs are calculated with the following equation [11, 12].

g sb ¼ ðn þ bÞ=2n ð7Þ
Subordinate criteria weights are calculated according to the HFLE score values
obtained as follows:
Xj
wpi ¼ Ppd ðci Þ= i¼1
Ppd ðci Þ ð8Þ
where Ppd ðci Þ shows the priority degree of criterion ci and calculated as [13];
Pj
r
E ðhps ðci ;cj ÞÞ
Ppd ðci Þ ¼ e j m¼1 ð9Þ
The experts’ weight set is calculated based on the individual preference matrix and
the comprehensive hesitant degree of HFLE.
Pi Pi p

p
0:5 iði1Þ
1
m¼1 n¼1 Hd ðhs ðcm cn ÞÞ
w ¼P Pi Pi ; m 6¼ n ð10Þ
P p
p¼1 0:5 1
iði1Þ m¼1 n¼1 H d ð h s ð c m c n Þ Þ
The criteria weights are calculated by evaluating the subordinate criteria weights
and the weight values of the experts.
PP p p
p¼1 w wi
wci ¼ P j PP ð11Þ
p p
i¼1 p¼1 w wi
The aggregation methods defined for the MULTIMOORA method [14] are adapted
to the unbalanced hesitant fuzzy linguistic terms. The methods called Unbalanced
Hesitant Fuzzy Linguistic Ratio System (UHFLRS), Unbalanced Hesitant Fuzzy
Linguistic Reference Point (UHFLRP) and Unbalanced Hesitant Fuzzy Linguistic Full
Multiplicative Form (UH- FLFMF) are as follows [3]:
X
e X
j

UHFLRSðai Þ ¼ wm En hmn
s wm En hmn
s ð12Þ
m¼1 m¼e þ 1

where En hmns represents the normalized score of HFLE, hmn s decision matrix, wm
defines the weight of criteria. The impact values of the cost criteria ðe þ 1; . . .; jÞ are
subtracted from the impact value of the benefit criteria (1; 2; ; . . .e) for each alternative.
The score values of the alternatives (ai ) according to the criteria are combined with the
arithmetic weighted aggregation method.

UHFLRPðai Þ ¼ max wn #j En hmn
s
ð13Þ
n

where #j represents the benefit and cost criterion as #j ¼ max En hmn
s and #j ¼
j
mn
min En hs respectively. UHFLRP aggregation method lists the worst performing
j
alternatives by criteria.
mn wn Qe
E n hs m¼1
UH FLFMF ðai Þ ¼ Q j wn ð14Þ
mn
m¼e þ 1 En hs
Geometric weights of normalized scores are calculated and compared with UH

FLFMF aggregation method. Also, the equation used in the normalization of score
values is calculated as follows:

E hmn s
En hmn
s ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pi ffi ð15Þ
Eðh mn Þ 2
l¼1 s
Alternatives listed by UHFLRS, UHFLRP, UH- FLFMF subordinating methods are

aggregate with the ORESTE method [15] and the general importance ranking of the
alternatives is performed. The order and weight of subordinating methods are deter-
mined by experts [16].
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Os ðam Þ ¼ 0:5ðvðTs ÞÞ2 þ 0:5ðvr ðam ÞÞ2 ð16Þ
where vðTs Þ refers the importance weights of the methods and vr ðam Þ refers the ranking
values of alternatives in each method. The alternatives are ranked using the global
Besson’s mean ranking method and the general alternative ranking is obtained.
X
F
G R ð am Þ ¼ GRf ðam Þ ð17Þ
f ¼1
where GRf ðam Þ defines the overall sort value of alternatives sorted by models.
3 SEPPIS with Unbalanced HFL-MULTIMOORA Method
In this section, the unbalanced HFL-MULTIMOORA method is applied to solve the

SEPPIS decision making problem. The investment decisions of the solar power plant are
based on technical, economic, environmental and social fundamental criteria. Although
these criteria have detailed features and sub-factors, this study evaluates alternative
investments only on general evaluation. The technical criterion takes into account the
location relative to the sun, the feature of the system (PV, CPV, CSP) and technical
infrastructure competence. The economic criterion addresses the return and cost factors
of the investment. The environmental criterion covers the geographical, climatic, resi-
dential and agricultural features of the installation site. The social criterion examines the
effects of investment on social behavior and social mobility in residential areas.
In the study, four alternative investments ðA ¼ fa1 ; a2 ; a3 ; a4 gÞ are evaluated by
four experts ðE ¼ fe1 ; e2 ; e3 ; e4 gÞ according to four criteria ðC ¼ fc1 ; c2 ; c3 ; c4 gÞ. The
linguistic term set used in the evaluation is S = {s3 = very bad (vb), s2 = bad (b),
s1 = little bad (lb), s0 = medium (m), s1 = good (g), s2 = little good (lg), s3 = very
good (vg)}. The set of linguistic terms used in the pairwise evaluation of the criteria is
S = {s3 = very less important (vli), s2 = less important (li), s1 = little less
important (lli), s0 = medium (m), s1 = little more important (lmi), s2 = more important
(mi), s3 = very more important (vg)}. Alternative investments are evaluated linguis-
tically by experts using context-free grammar (Table 1).
Table 1. Hesitant linguistic evaluations by expert 1 and their HFLEs.

c1 c2 c3 c4
a1 Betweenmg Betweenblg Betweenmlg less thanm
fs0 ; s1 ; s2 g fs2 ; s1 ; s0 ; s1 g fs0 ; s1 g fs3 ; s2 ; s1 g
a2 Betweenlgg less thanlb Betweenlbvg less thanb
fs1 ; s2 g fs3 ; s2 g fs1 ; s0 ; s1 ; s2 ; s3 g fs3 g
a3 at leastg at mostlb Betweenmg less thanvg
fs2 ; s3 g fs3 ; s2 ; s1 g fs0 ; s1 ; s2 g fs3 ; . . .; s2 g
a4 at leastg more thanlb less thanlb at leastlb
fs2 ; s3 g fs0 ; s1 ; s2 ; s3 g fs3 ; s2 g fs1 ; s0 ; s1 ; s2 ; s3 g
Linguistic evaluations of alternatives are transformed into HFLEs with transfor-

mation functions (EGH ) (Table 1). Hesitant degrees, which are converted from HFLEs
using Eq. (2), use in the Eq. (1) and Eq. (3) to calculate the weights of the experts as
w1 ¼ 0:216, w2 ¼ 0:215, w3 ¼ 0:283 and w4 ¼ 0:286. The score values of the HFLEs
are calculated (Eq. (5)) and a general decision matrix is created by bringing together all
expert evaluations as:
2 3
0:328 0:531 0:546 0:863
6 0:390 0:299 0:695 0:214 7
DE ¼ 4 0:598 0:282 0:433 0:370 5
0:618 0:740 0:177 0:268
Pairwise linguistic evaluations of the criteria are converted to HFLE and the score
values of HFLEs are calculated using Eq. (7). Subordinate weights of the criteria are
calculated using Eq. (8) as; w11 ¼ 0:259, w12 ¼ 0:234, w13 ¼ 0:225, w14 ¼ 0:282,
w21 ¼ 0:239, w22 ¼ 0:231, w23 ¼ 0:287, w24 ¼ 0:243, w31 ¼ 0:265, w32 ¼ 0:230,
w33 ¼ 0:246, w34 ¼ 0:259, w41 ¼ 0:244, w42 ¼ 0:268, w43 ¼ 0:252, w44 ¼ 0:237. The
weight set of experts based on the hesitant grades are calculated (Eq. 10) and the
weight of the criteria is calculated along with the subordinate weights of the criteria
(Eq. 11) as wc1 ¼ 0:253, wc2 ¼ 0:240, wc3 ¼ 0:252 and wc4 ¼ 0:256.
Table 2. Ranking alternatives with Unbalanced Hesitant Fuzzy Linguistic subordinate models
for unbalanced HFL-MULTIMOORA method
UHFLRS UHFLRP UH FLFMF T3 [ T1 [ T2
Utility value Rank Utility value Rank Utility value Rank GR New rank
a1 0.568 1 0.135 1 0.536 1 6.5 1
a2 0.400 4 0.166 3 0.363 4 14 3
a3 0.423 3 0.139 2 0.409 2 9 2
a4 0.447 2 0.173 4 0.381 3 15.5 4
Subordinate ranks of alternatives are found using the UHFLRS (Eq. (13)),
UHFLRP (Eq. (14)) and UH-FLFMF (Eq. (15)) methods (Table 2). Models are
weighted according to their evaluation power. In this study, the methods are weighted
as vðT1 Þ ¼ 2, vðT2 Þ ¼ 1 and vðT3 Þ ¼ 3. The evaluations of the alternatives based on the
rankings are made according to Eq. (16) and Eq. (17) and the final ranking is obtained
(Table 2). According to the final ranking, the first alternative is the most suitable
investment choice according to the defined criteria.
4 Conclusion
Linguistic expressions are used to evaluate problems that cannot be explained and inter-
preted with numerical values. HFLTS allows experts to make more comfortable evalua-
tions in conditions of lack of information or uncertainty. The Score-HeDLiSF method,
which allows scaling and grading linguistic expressions, performs the numerical trans-
formation of HFLTS. Linguistic expressions converted to numbers with Score-HeDLiSF
can be used in unbalanced hesitant fuzzy linguistic MULTIMOORA decision making
method. In this study, the investment decision problem of solar energy systems, which
have an important place in the renewable energy sector, is examined with the unbalanced
hesitant fuzzy linguistic MULTIMOORA method based Score-HeDLiSF method.
Experts make their evaluations easily with the wide set of linguistic terms offered
by HFLTS. Weights are defined based on hesitant degree and aggregate criteria and
expert weights to evaluate alternatives leads to more accurate results. In future studies,
alternative projects can be compared by detailing the criteria defined for the SEPPIS
problem and adding the cost criteria to the evaluation process. Also, by comparing the
unbalanced hesitant fuzzy linguistic MULTIMOORA method with other HFLTS based
decision making methods, the power and validity of the method can be questioned.
References
1. Rodriguez, R.M., Martinez, L., Herrera, F.: Hesitant fuzzy linguistic term sets for decision
2. Liao, H., Xu, Z., Zeng, X.J., Merigó, J.M.: Qualitative decision making with correlation
coefficients of hesitant fuzzy linguistic term sets. Knowl.-Based Syst. 76, 127–138 (2015)
3. Liao, H., Qin, R., Gao, C., Wu, X., Hafezalkotob, A., Herrera, F.: Score-HeDLiSF: a score
function of hesitant fuzzy linguistic term set based on hesitant degrees and linguistic scale
functions: an application to unbalanced hesitant fuzzy linguistic MULTIMOORA. Inf.
Fusion 48, 39–54 (2019)
4. Liao, H., Xu, Z., Zeng, X.-J.: Hesitant fuzzy linguistic VIKOR method and its application in
qualitative multiple criteria decision making. IEEE Trans. Fuzzy Syst. 23(5), 1343–1355
(2014)
5. Çoban, V., Onar, S.Ç.: Modelling solar energy usage with fuzzy cognitive maps. In:
Intelligence Systems in Environmental Management: Theory and Applications, pp. 159–187.
6. Brauers, W.K.M., Zavadskas, E.K.: Project management by MULTIMOORA as an
instrument for transition economies. Technol. Econ. Dev. Econ. 16(1), 5–24 (2010)
7. Xu, Z.: A note on linguistic hybrid arithmetic averaging operator in multiple attribute group
decision making with linguistic information. Group Decis. Negot. 15(6), 593–604 (2006)
8. Wei, C., Rodríguez, R.M., Martínez, L.: Uncertainty measures of extended hesitant fuzzy
linguistic term sets. IEEE Trans. Fuzzy Syst. 26(3), 1763–1768 (2017)
9. Wei, C., Zhao, N., Tang, X.: Operators and comparisons of hesitant fuzzy linguistic term
sets. IEEE Trans. Fuzzy Syst. 22(3), 575–585 (2013)
10. Bao, G.Y., Lian, X.L., He, M., Wang, L.L.: Improved two-tuple linguistic representation
model based on new linguistic evaluation scale. Control Decis. 25(5), 780–784 (2010)
11. Gou, X., Xu, Z.: Novel basic operational laws for linguistic terms, hesitant fuzzy linguistic
term sets and probabilistic linguistic term sets. Inf. Sci. 372, 407–427 (2016)
12. Wang, J.Q., Wu, J.T., Wang, J., Zhang, H.Y., Chen, X.H.: Interval-valued hesitant fuzzy
linguistic sets and their applications in multi-criteria decision-making problems. Inf. Sci.
288, 55–72 (2014)
13. Lootsma, F.A.: Scale sensitivity in the multiplicative AHP and SMART. J. Multi-Criteria
Decis. Anal. 2(2), 87–110 (1993)
14. Brauers, W.K., Zavadskas, E.K.: The MOORA method and its application to privatization in
a transition economy. Control Cybern. 35, 445–469 (2006)
15. Roubens, M.: Preference relations on actions and criteria in multicriteria decision making.
Eur. J. Oper. Res. 10(1), 51–55 (1982)
16. Wu, X., Liao, H.: An approach to quality function deployment based on probabilistic
linguistic term sets and ORESTE method for multi-expert multi-criteria decision making.
Inf. Fusion 43, 13–26 (2018)
Application of Linear Programming
Model in Multiple Criteria Decision
Making Under the Framework
of Interval-Valued Hesitant Fuzzy Sets
Tabasam Rashid(B) and M. Sarwar Sindhu
Department of Mathematics, University of Management and Technology,

Lahore 54770, Pakistan
Abstract. Hesitant fuzzy sets (HFSs) further extended into interval-

valued hesitant fuzzy sets (IVHFSs). IVHFSs comprises belonging degree
in terms of intervals. The weights of the criteria are the basic key to
decide the best option from the identical options. We introduce the linear
programming (LP) model to compute the criteria’s the unknown weights
in the present work. Furthermore, an extended TOPSIS is applied under
the INVHFs circumstances. Our proposed method is beneficial to reduce
the complexity of computation as compare to other existing MCDM
methods. The said model is implemented in a practical example regard-
ing the energy sector to explain the validity and strength.
Keywords: Hesitant fuzzy sets · Interval-valued hesitant fuzzy sets ·

Linear programming methodology · TOPSIS
1 Introduction
The idea of fuzzy sets (FSs) is presented by Zadeh [1] in the middle of the 20th
century. FSs explore the new horizons for the experts. The idea of fuzzy sets was
praised because it deals with ambiguity, vagueness and hesitation involved in the
information directly. Later on, Zadeh [2] introduced a novel concept of FSs called
the interval-valued fuzzy sets (IVFSs) which comprises a pair of membership level
in terms of interval to reduce the information loss during the decision making
process.
Hesitancy or uncertainty appear mostly everywhere in our daily life. It is dif-
ficult to decide in real life to choose one of the best options with alike structures.
In most of the decision making process, DMs feel confusion and hesitation to
assign the belonging degree to the alternatives. In order to handle such situa-
tions, Torra [3] offered a novel extension of FSs, the hesitant fuzzy sets (HFSs)
that contains the belonging degree in terms of set instead of single one. Since
HFSs have set of values as a membership function due to which it minimize the
loss of information. Rashid and Beg [4] extended convex fuzzy sets to convex
Application of LP Model in MCDM Under the Framework of IVHFSs 283
HFSs. Later on, Chen et al. [5] augmented the idea of HFSs into IVHFSs to
resolve such issues. IVHFSs comprises the belonging function in terms of set of
plausible interval-values in [0, 1] instead of set of real numbers that can deal hesi-
tancy, vagueness and unpredictability involved in the process of decision making
accurately. Bai [6] applied the distance measures of similarity for MCDM prob-
lems under the framework of IVHFSs. Moreover, Meng et al. [7] computed the
correlation coefficients to select the best alternative by using IVHFSs. Zhang
et al. [8] presented an interval-valued hesitant fuzzy multi granulation rough
set, a novel approach that developed from the IVHFSs. Moreover, Zhang et al.
[9] provided the analysis about MCDM by using IVHF linguistic numbers and
shapley fuzzy measures. The process in which DMs achieved an optimal decision
from the various identical criteria, called the multiple criteria decision making
(MCDM) [10]. In 1981, Hawang and Yoon [11] put forwarded a novel technique,
TOPSIS to handle the MCDM issues for crisp information. The center thought
behind the TOPSIS is that the perfect option has minimal separation from the
positive ideal solution (PIS) and have maximal separation from the negative
ideal solution (NIS).
The linear programming (LP) [12] technique, allow to minimize or maximize
an objective function inside the system of given situational limitations. LP is
a computational technique that enables DMs to solve the problems which they
face in decision-making model. It encourages the DMs to deal with constrained
ideal conditions which they need to make the best of their resources. Recently,
a LP technique is implemented to evaluate the criteria’s weights of an MCDM
problem under the probabilistic IVHFSs environment by Sindhu et al. [13].
Almost several extensions of FSs like HFSs have a discrete set of values in
its membership degree which cause the loss of information. However, IVHFSs
consist of continuous set values in its specification that minimize the loss of
information, thereby, we use the INVHFSs to select the best energy project in
this article.
The remaining part of this article is planned accordingly: We briefly throw
light on the basic notions of IVHFSs, the measure of similarity amongst IVHFEs
and a model based on LP in Sect. 2. A proposed MCDM model under the IVHFSs
framework, the LP model and the extended TOPSIS technique are presented in
Sect. 3. An explanatory example is provided in Sect. 4, we utilize the proposed
technique and investigate the experimental outcomes. Finally, the comparative
analysis and conclusions are debated in Sect. 5.
2 Fundamentals
The current section comprises brief fundamentals about IVHFSs, the measure
of similarity amongst IVHFEs and the LP model.
Definition 1 [3]. Let Y = {y1 , y2 , ..., yn } be a discourse set, the hesitant fuzzy
set A on Y is defined in the form of a function hA (y) that when applied to Y
returns a finite subset of [0, 1]. In order to understand easily, Xia andXu [14]
presented the mathematical notion of HFSs as: H = y, hH(y) : y ∈ y , where
284 T. Rashid and M. S. Sindhu
hH(y) is a set of some distinct real values in [0, 1] representing the plausible
belonging degree of the component y ∈ Y to the set H.
Definition 2 [15]. Let X = {x1 , x2 , ..., xn } be a universe

set, then,
interval-

valued hesitant fuzzy set (IVHFS ) on X is defined as: H̃ = x, h̃H̃(x) : x ∈ X ,
where h̃H̃(x) : X → [0, 1] represents all plausible interval membership degrees
of the element x ∈ X to the set H̃. Generally, h̃H̃(x) is called an interval-valued

hesitant fuzzy element (IVHFE) denoted by h̃ = γ̃ 1 , γ̃ 2 , γ̃ 3 , ..., γ̃ #h̃ , where

γ̃ λ λ = 1, 2, 3, ..., #h̃ is an interval number represented by γ̃ λ = γ̃ λL , γ̃ λU ∈

[0, 1]. The IVHFE become a HFE if γ̃ λL = γ̃ λU λ = 1, 2, 3, ..., #h̃ , the symbol
#h̃ indicate the number of components in h̃.

Definition 3 [16]. Let h̃ = γ̃ 1 , γ̃ 2 , γ̃ 3 , ..., γ̃ #h be an IVHFE with γ̃ + , the max-
imal and γ̃ − are the minimal interval values in h̃, then we can extend the interval
value of h̃ as, γ̃ ∗ = θγ̃ + + (1 − θ) γ̃ − , where θ (0 ≤ θ ≤ 1), is the constant eval-
uated according to the desire of DMs (the decision maker (DM) is risk-neutral,
risk-seeking and risk-averse). Accordingly, for any two IVHFEs h̃1 and h̃2 with
unequal cardinality that is |h̃1 | = |h̃2 |, we can equalize the cardinalities by adding
the distinct number of interval values in the lower IVHFE by using the values of
the parameter θ which depend upon the DM’s desire. The DM can be (a)risk-
neutral, when θ = 12 , the extended interval-value become, γ̃ ∗ = 12 (γ̃ + + γ̃ − );
(b) risk-seeking, when θ = 1, the extended interval-value become, γ̃ ∗ = γ̃ + ; (c)
risk-averse, when θ = 0, the extended interval-value become, γ̃ ∗ = γ̃ − .
Definition
4 [6]. Let X = {x1 , x2 , ..., xn } be a source set and, Ã =
{ xi , h̃Ã(xi ) : xi ∈ X}, B̃ = { xi , h̃B̃(xi ) : xi ∈ X}, be two IVHFSs in X. Con-

sider h̃Ã(xi ) (xi ) ∈ Ã and h̃B̃(xi ) (xi ) ∈ B̃ such that: h̃Ã(xi ) (xi ) = h̃− , h̃ +
Ã(xi ) Ã(xi )
−
∈ [0, 1] and h̃B̃(xi ) (xi ) = h̃B̃(x ) , h̃B̃(x ) ∈ [0, 1], where 1 ≤ i ≤ n.
+
i i
The Normalized Hamming distance between Ã and B̃ is as follows:

⎡ lx
⎤
n
1 1 i −

+

DH (Ã, B̃) = ⎣ h̃ (xi ) − h̃−
(xi ) + h̃ (xi ) − h̃+
(xi ) ⎦ (1)
n i=1 2lxi j=1 Ãσ(j ) B̃σ(j ) Ãσ(j ) B̃σ(j )
where lxi is the largest number of elements in l(h̃Ã (xi )) and l(h̃B̃ (xi )) for each
xi ∈ X. In general l(h̃Ã (xi )) = l(h̃B̃ (xi )), we can take l(h̃Ã (xi )) = l(h̃B̃ (xi )) in
two ways, risk-averse decision maker choose the shorter values while risk seeker
DM choose the largest values, h̃Ãσ(j ) (xi ) and h̃B̃σ(j ) (xi ) are the j th biggest values
of h̃Ã (xi ) and h̃B̃ (xi ), respectively.
Definition 5 [6]. Suppose that Cj are the criteria and wj the corresponding
m
weights with wj ∈ [0, 1], 1 ≤ j ≤ n and j=1 wj = 1. The generalized normalized
weighted distance similarity measures for IVHFSs is given below:
⎡ ⎛ ⎛ ⎞⎞⎤
−
m
1
lxi
h̃A∗ (Cj ) − h̃− (Cj )
Dn (A , Ai ) = 1 − ⎣
∗
wj ⎝ ⎝ (t)
Ai
⎠⎠⎦ (2)
j=1

2lCj t=1 + h̃ ∗ (Cj ) − h̃ (Cj )
+ +
A Ai (t)
where A∗ is an ideal IVHFS which defined for each criteria in the ideal
alternative as,

A∗ = h̃−ve
1 , h̃+ve
1 , h̃−ve
2 , h̃+ve
2 , h̃−ve
3 , h̃+ve
3 , ..., h̃−ve
n , h̃n
+ve
(3)
Measures of distance similarity Sd defined in Eq. 2 fulfil the given three condi-
tions: (i) Sd (A∗ , Ai ) ∈ [0, 1], (ii) Sd (A∗ , Ai ) = 1 iff A∗ = Ai , (iii) Sd (A∗ , Ai ) =
Sd (Ai , A∗ ).
Definition 6 [12]. A LP model is described as:
Maximize: S = c1 t1 + c2 t2 + c3 t3 + ... + cn tn
Subject to: b11 t1 + b12 t2 + b13 t3 + ... + b1n tn ≤ a1
b21 t1 + b22 t2 + b23 t3 + ... + b2n tn ≤ a2
..
.
bm1 t1 + bm2 t2 + bm3 t3 + ... + bmn tn ≤ am
t1 , t2 , ..., tn ≥ 0,
3 MCDM Model Based on LP Methodology and the

Extended TOPSIS Under the Framework of IVHFSs
The present section comprises a proposed MCDM method for the IVHFSs
based on the LP model and the extended TOPSIS. Suppose that P =
{P1 , P2 , ..., pm } (1 ≤ i ≤ m) and C = {C1 , C2 , ..., Cn } (1 ≤ j ≤ n) are the
sets of alternatives and criteria, respectively. The ranking of the alternative Pi ,
relative to the criteria Cj is denoted by an IVHFE h̃ij . The suggested extended
TOPSIS method contain the steps below:
Step 1. Based on [15], get the IVHF positive ideal solution (IVHFPIS) Ã+ and
the IVHF negative ideal solution (IVHFNIS) Ã− , shown below:

1
{maxγ̃ij
#h̃ij
, ..., maxγ̃ij } : Cj ∈ J1
+
Ã = h̃+ve
1 , h̃+ve
2 , h̃+ve
3 , ..., h̃+ve
n = i i
#h̃ij
(4)
1
{minγ̃ij , ..., minγ̃ij } : Cj ∈ J2
i i
⎛ ⎞
#h̃
1
minγ̃ij , ..., minγ̃ij ij : Cj ∈ J1
− −ve −ve −ve −ve
Ã = h̃1 , h̃2 , h̃3 , ..., h̃n = ⎝ i i
#h̃
⎠ (5)
1
maxγ̃ij , ..., maxγ̃ij ij : Cj ∈ J2
i i
where J1 is a subset of benefit criteria and J2 is a subset of cost criteria, and

J1 ∩ J2 = ∅.
Step 2. Based on Eq. 2, compute the degree of similarity Si+ amongst IVHFPIS
Ã+ obtained in Eq. 5 as well as the degree of similarity Si− amongst IVHFNIS
Ã− obtained in Eq. 6 with each alternative, respectively:
⎡ ⎛ ⎛ ⎞⎞⎤
lxi + −
n
1 h̃ (C j ) − h̃ (Cj )
Si = 1 −
+ ⎣ w̃j ⎝ ⎝ +
Ã (t) A i
⎠⎠⎦ (6)

j=1
2l C j t=1 + h̃ +
Ã+ (t)
(C j ) − h̃ +
Ai (C j )
⎡ ⎛ ⎛ ⎞⎞⎤
lxi
− −
n
1 h̃ (C j ) − h̃ (Cj )
−
Si = 1 − ⎣ w̃j ⎝ ⎝ −
Ã (t) A i
⎠⎠⎦ (7)

j=1
2l C j t=1 + h̃ +
Ã− (t)
(C j ) − h̃ +
Ai (C j )
n
where w̃j ∈ [0, 1] and j=1 w̃j = 1.
Step 3. Based on Eq. 6 and Eq. 7, an LP model is constructed, as shown below:

m

max S = Si+ − Si− (8)
i=1
Step 4. Calculate the weights w̃1 , w̃2 , w̃3 by solving the LP model shown in
Eq. 8 so that the value of the objective function S is maximized.
Step 5. By using the criteria’s weights attained in Step 4, compute the degree
of similarity Si+ and Si− amongst each alternative and the elements achieved in
the IVHFS ideal solutions Ã+ and Ã− , respectively.
Step 6. Determine the relative closeness R̃i (C) of alternative Ai with respect
to the IVHFSPIS Ã+
S+
R̃i (C) = + i − , (9)
Si + Si
here R̃i (C) ∈ [0, 1] and 1 ≤ i ≤ m. The greater the value of the relative close-
ness R̃i (C) of options Ai with respective to the IVHFPIS Ã+ the superior the
arrangement order of options Ai with 1 ≤ i ≤ m.
4 Explanatory Example
Suppose that H, S, N, R, and K are five energy projects, and these are calculated
under the criteria, C1 (sociopolitical), C2 (economic), C3 (technological) and C4
(environmental). The numeric data is collected form [16] and the weights w̃1 , w̃2 ,
w̃3 and w̃4 of the criteria C1 , C2 , C3 and C4 given by the DMs, respectively, are:
0.15 ≤ w̃1 ≤ 0.2;0.16 ≤ w̃2 ≤ 0.18; 0.3 ≤ w̃3 ≤ 0.35; 0.3 ≤ w̃4 ≤ 0.45; where
n
w̃j ∈ [0, 1] and j=1 w̃j = 1. The decision data given by the decision maker
represented in IVHFEs, as listed in the interval-valued hesitant fuzzy decision
Table 1: Table 1 shows that the cardinalities of each IVHFEs are unequal, to
make it to be equal we should extend the smaller one until all of them have the
Table 1. IVHFEs decision criteria
C1 C2
H {[0.2, 0.3], [0.3, 0.4], [0.3, 0.5]} {[0.1, 0.3], [0.6, 0.7], [0.7, 0.8], [0.7, 0.9]}
S {[0.2, 0.3], [0.2, 0.5]} {[0.1, 0.2], [0.4, 0.5], [0.4, 0.6], [0.6, 0.7], [0.7, 0.9]}
N {[0.5, 0.6], [0.5, 0.7]} {[0.5, 0.6], [0.7, 0.9]}
R {[0.2, 0.3], [0.3, 0.4], [0.6, 0.7], [0.7, 0.8]} {[0.1, 0.2], [0.3, 0.4], [0.6, 0.7]}
K {[0.1, 0.3], [0.2, 0.4], [0.5, 0.6], [0.6, 0.7], [0.7, 0.9]} {[0.3, 0.4], [0.5, 0.6], [0.6, 0.7], [0.6, 0.8]}
C3 C4
H {[0.1; 0.2], [0.3, 0.4], [0.3, 0.5]} {[0.2, 0.3], [0.4, 0.5], [0.4, 0.6], [0.7, 0.9]}
S {[0.1, 0.3], [0.4, 0.4], [0.4, 0.6], [0.6, 0.8]} {[0.2, 0.3], [0.3, 0.4], [0.6, 0.7]}
N {[0.2, 0.3], [0.4, 0.5], [0.6, 0.7]} {[0.3, 0.4], [0.5, 0.6]}
R {[0.1, 0.3], [0.7, 0.8]} {[0.5, 0.6], [0.7, 0.8], [0.7, 0.9]}
K {[0.6, 0.7], [0.7, 0.8], [0.7, 0.9]} {[0.2, 0.3], [0.5, 0.6], [0.5, 0.7], [0.7, 0.9]}
Table 2. The normalized results of IVHFEs decision criteria
C1 C2
H {[0.2, 0.3], [0.2, 0.3], [0.2, 0.3], [0.3, 0.4], [0.3, 0.5]} {[0.1, 0.3], [0.1, 0.3], [0.6, 0.7], [0.7, 0.8], [0.7, 0.9]}
S {[0.2, 0.3], [0.2, 0.3], [0.2, 0.3], [0.2, 0.3], [0.2, 0.5]} {[0.1, 0.2], [0.4, 0.5], [0.4, 0.6], [0.6, 0.7], [0.7, 0.9]}
N {[0.5, 0.6], [0.5, 0.6], [0.5, 0.6], [0.5, 0.6], [0.5, 0.7]} {[0.5, 0.6], [0.5, 0.6], [0.5, 0.6], [0.5, 0.6], [0.7, 0.9]}
R {[0.2, 0.3], [0.2, 0.3], [0.3, 0.4], [0.6, 0.7], [0.7, 0.8]} {[0.1, 0.2], [0.1, 0.2], [0.1, 0.2], [0.3, 0.4], [0.6, 0.7]}
K {[0.1, 0.3], [0.2, 0.4], [0.5, 0.6], [0.6, 0.7], [0.7, 0.9]} {[0.3, 0.4], [0.3, 0.4], [0.5, 0.6], [0.6, 0.7], [0.6, 0.8]}
C3 C4
H {[0.1, 0.2], [0.1, 0.2], [0.1, 0.2], [0.3, 0.4], [0.3, 0.5]} {[0.2, 0.3], [0.2, 0.3], [0.4, 0.5], [0.4, 0.6], [0.7, 0.9]}
S {[0.1, 0.3], [0.1, 0.3], [0.4, 0.4], [0.4, 0.6], [0.6, 0.8]} {[0.2, 0.3], [0.2, 0.3], [0.2, 0.3], [0.3, 0.4], [0.6, 0.7]}
N {[0.2, 0.3], [0.2, 0.3], [0.2, 0.3], [0.4, 0.5], [0.6, 0.7]} {[0.3, 0.4], [0.3, 0.4], [0.3, 0.4], [0.3, 0.4], [0.5, 0.6]}
R {[0.1, 0.3], [0.1, 0.3], [0.1, 0.3], [0.1, 0.3], [0.7, 0.8]} {[0.5, 0.6], [0.5, 0.6], [0.5, 0.6], [0.7, 0.8], [0.7, 0.9]}
K {[0.6, 0.7], [0.6, 0.7], [0.6, 0.7], [0.7, 0.8], [0.7, 0.9]} {[0.2, 0.3], [0.2, 0.3], [0.5, 0.6], [0.5, 0.7], [0.7, 0.9]}
same length. By using Definition 5, add the minimal values to make a normalize
interval valued hesitant fuzzy table. The normalized results are listed in Table 2:
Step 1. As C1 , C2 , C3 and C4 are beneficial criteria, i.e., J1 ={C1 , C2 , C3 , C4 } ,
based on Eq. 4 and Eq. 5, we can get, the IVHFPIS (Ã+ ) and the IVHFNIS (Ã− )
respectively:

+ {[0.5, 0.6], [0.5, 0.6], [0.5, 0.6][0.5, 0.6], [0.7, 0.9]} {[0.5, 0.6], [0.5, 0.6], [0.6, 0.7][0.7, 0.8], [0.7, 0.9]}
Ã =
{[0.6, 0.7], [0.6, 0.7], [0.6, 0.7][0.7, 0.8], [0.7, 0.9]} {[0.5, 0.6], [0.5, 0.6], [0.5, 0.6][0.7, 0.8], [0.7, 0.9]}

− {[0.1, 0.3] , [0.2, 0.3] , [0.2, 0.3] , [0.2, 0.3] , [0.2, 0.5]} {[0.1, 0.2] , [0.1, 0.2] , [0.1, 0.2] , [0.3, 0.4] , [0.6, 0.7]}
Ã =
{[0.1, 0.2] , [0.1, 0.2] , [0.1, 0.2] , [0.1, 0.3] , [0.3, 0.5]} {[0.2, 0.3] , [0.2, 0.3] , [0.2, 0.3] , [0.3, 0.4] , [0.5, 0.6]}
Step 2. Based on Eq. 6 and Eq. 7, evaluate the similarity degree Si+ amongst
IVHFPIS Ã+ and each alternative also the degree of similarity Si− amongst
IVHFNIS Ã− with each alternative respectively: Step 3. Based m on Eq. 6 and

Eq. 7, formulate the LP model, shown as follows: max S = i=1 S +
i − Si− ,
where 1 ≤ i ≤ 5. Step 4. LP model obtained in Step 3 is used to evaluate the
criteria’s weights, w̃1 , w̃2 , w̃3 and w̃4 for C1 , C2 , C3 and C4 respectively by
using the traits given. The weights of the criteria are: w̃1 = 0.1500, w̃2 = 0.1600,
w̃3 = 0.3500 and w̃4 = 0.3400. Step 5. By using the criteria’s weights attained
in Step 4, evaluate Si+ and Si− amongst each alternative and the elements
obtained in IVHFPIS Ã+ and IVHFNIS Ã− respectively. S1+ = 0.8335; S2+ =
0.8240; S3+ = 0.8935; S4+ = 0.6710; S5+ = 0.8572; S1− = 0.9385; S2− = 0.9296;
S3− = 0.8755; S4− = 0.7970; S5− = 0.6260.
Step 6. With the help of Eq. 9, the relative closeness R̃i (C) of alternatives Ai ,
S1+ S2+
where 1 ≤ i ≤ 5 are: R̃1 (C) = S1+ +S1−
= 0.4704; R̃2 (C) = S2+ +S2−
= 0.4699;
S3+ S4+ S+
R̃3 (C) = S3 +S3−
+ = 0.5051; R̃4 (C) = = 0.4571; R̃5 (C) = S + +S
S4 +S4−
+
5
− =
5 5
0.5779.
From the above results of the relative closeness R̃i (C), we get; As R̃5 (C)
R̃3 (C) R̃1 (C) R̃2 (C) R̃4 (C), that is, K N H S R, so the
desired alternative achieved by the proposed method is K, which coincides with
the result obtained by the method presented in [16,17].
Table 3. Results obtained by proposed method, Xu [16] and Alinezhad et al. [17]
Proposed method Alinezhad et al. method [17] Xu technique [16]

H 0.4704 0.2905 0.3233
S 0.4699 0.2653 0.3650
N 0.5051 0.5181 0.4581
R 0.4571 0.5132 0.4214
K 0.5779 0.7742 0.6934
KN HSR KN RHS KN RSH
Fig. 1. Preference arrangement. Fig. 2. Ranking order of the alternatives.

5 Comparative Analysis and Concluding Remarks

The LP model has been proposed to compute the criteria’s weights for MCDM.
The weights attained are then applied for selecting the best energy project. In
addition, the extended TOPSIS is implemented to line the alternatives. Lastly,
the efficiency and applicability of the proposed method have been demonstrated
with an energy police selection example presented in [16]. Our proposed tech-
nique is forthright and has fewer loss of data, can help the DMs to take the
decision in MCDM issues under the IVHF situations. From the Table 3, the
preference arrangement of the alternatives obtained by using the our approach
is K N H S R that is, K is the best choice. The technique pre-
sented by Alinezhad et al. [17], strengthen the proposed ranking order. The
ranking order of the alternatives attained by Xu [16] is, K N R S H,
which shows that, K is the desirable alternative. Figure 1 illustrate the pref-
erence arrangement of the alternatives taken from Table 3. Furthermore, Fig. 2
reveal the ranking order of the alternatives. In order to show the validity and
effectiveness of the proposed MCDM model, we will implement it for pattern
recognition under the environment of interval-valued picture fuzzy sets in the
future.
References
1. Zadeh, L.: Fuzzy sets. Inf. Control 8, 338–353 (1965)
2. Zadeh, L.: The concept of a linguistic variable and its applications to approximate
3. Torra, V.: Hesitant fuzzy sets. Int. J. Intell. Syst. 25, 529–539 (2010)
4. Rashid, T., Beg, I.: Convex hesitant fuzzy sets. J. Intell. Fuzzy Syst. 30(5), 2791–
2796 (2016)
5. Chen, N., Xu, Z., Xia, M.: Interval valued hesitant preference relations and their
applications to group decision making. Knowl.-Based Syst. 37, 528–540 (2013)
6. Bai, Z.: Distance similarity measures for interval-valued hesitant fuzzy sets and
their application in multicriteria decision making. J. Decis. Syst. 22(3), 190–201
(2013)
7. Meng, F., Wang, X.C., Zhang, Q.: Correlation coeficients of interval-valued hesitant
fuzzy sets and their application based on the shapely function. Int. J. Intell. Syst.
31, 17–43 (2016)
8. Zhang, C., Li, D., Mu, Y., Song, D.: An interval-valued hesitant fuzzy multigranu-
lation rough set over two universes model for steam turbine fault diagnosis. Appl.
Math. Model. 42, 693–704 (2017)
9. Zhang, W., Ju, Y., Liu, X.: Multiple criteria decision analysis based on shapley
fuzzy measures and interval-valued hesitant fuzzy linguistic numbers. Comput. Ind.
Eng. 105, 28–38 (2017)
10. Antucheviciene, J., Kala, Z., Marzouk, M., Vaidogas, E.R.: Solving civil engineering
problems by means of fuzzy and stochastic MCDM methods. Math. Probl. Eng.
2015, 1–16 (2015)
11. Hwang, C., Yoon, K.: Multiple Attributes Decision Making Methods and Applica-
tions. Springer, Heidelberg (1981)
12. Vanderbei, R.J.: Linear Programming: Foundations and Extensions, pp. 5–6.
13. Sindhu, M.S., Rashid, T., Kashif, A., Guirao, J.L.G.: Multiple criteria decision
making based on probabilistic interval-valued hesitant fuzzy sets by using LP
methodology. Discret. Dyn. Nat. Soc. Article ID 1527612, 12 p. (2019). https://
doi.org/10.1155/2019/1527612
14. Xia, M., Xu, Z.: Hesitant fuzzy information aggregation in decision making. Int.
J. Approx. Reason. 52(3), 395–407 (2011)
15. Zhang, X., Xu, Z.: Hesitant Fuzzy Methods for Multiple Criteria Decision Analysis.
16. Xu, Z., Zhang, X.: Hesitant fuzzy multi attribute decision making based on TOPSIS
with incomplete weight information. Knowl. Based Syst. 52, 53–64 (2013)
17. Alinezhad, A., Sarrafha, K., Amini, A.: Sensitivity analysis of saw technique, the
impact of changing the decision making matrix elements on the final ranking of
alternatives. Iran. J. Oper. Res. 5, 82–94 (2014)
Neutrosophic Sets
Evaluation of the Fourth Party Logistics
Provider Under Neutrosophic Environment
Serhat Aydın1(&), Mehmet Yörükoğlu1, and Mehmet Kabak2

1
National Defence University, 34149 Istanbul, Yeşilyurt, Turkey
[email protected]
2
Gazi University, 06560 Ankara, Yenimahalle, Turkey
Abstract. The fourth party logistics (4PLs) is an integrator that assembles the
resources, capabilities, and technology of the service provider and its client to
design, develop, and implement comprehensive supply chain solutions. A 4PLs
provider is a technological service provider who owns computer systems and
intellectual capital. Choosing the best 4PLs among many alternatives is critical
for the firms and it can be considered in multi-criteria decision making (MCDM)
problems and includes subjectivity and vagueness data. Fuzzy sets theory is a
useful tool to handle such subjectivity and vagueness in the evaluation process.
Recently, some extensions of fuzzy sets have been developed to define the
uncertainty more comprehensively. Neutrosophic sets are one of the extensions
of the fuzzy sets, which, unlike the other extensions, utilizes independent
indeterminacy-membership function, thereby extracts significant information
and enhances the accuracy of the decision-making process. In this paper, we
proposed a neutrosophic MCDM method to evaluate the 4PLs providers’ per-
formance. In the application section, three experts applied neutrosophic lin-
guistic scale to the providers’ performance. Then we calculated the closeness
coefficient of each provider and ranked them in descending order. The results
pointed out that the proposed neutrosophic MCDM method can be effectively
used in 4PLs providers’ performance evaluation problems.
Keywords: Supply chain management 4PLs Multi criteria decision

making Neutrosophic sets
1 Introduction
Supply chains (SCs) have a central act in the overall success of the organizations they
are involved in, especially the economic issues, for they have a fundamental respon-
sibility to convey the product or service to the final consumer. Managing such systems
is quite complex; SCs’ goals are often contradictory; assets within it have different
characteristics; the flow of the material and information is varied. Additionally, there
are many uncertainties in operations within the SC that increase its complexity [1].
SC can be represented as a connected network of all members involved in fulfilling
customer demand, either directly or indirectly. Customers, retailers, wholesalers/
distributors, manufacturers and suppliers of components/raw materials as different
layers form a typical SC [2]. Logistics is a basic and critical service network that
294 S. Aydın et al.
enables the physical flow of goods across the entire supply chain. It facilitates cross-
border trade through a range of activities including warehousing, shipping terminal
operations, brokerage, express delivery, and information management [3]. The main
objective of logistics in modern scientific literature is to solve potential problems
between producers and consumers by using effective and scientific methods in all
movements of goods in the supply chain processes. The dynamic market environment,
the intensity of competition and the customer’s expectation of quality lead to the
shaping of logistics as a science [4].
In today’s professional business life, there is a tendency to transfer side activities to
the companies which have specialized in these areas, thus gaining more competence
over the main activities. It is also possible for companies to transfer their logistics
activities to a company specializing in this field, and this practice is defined as “3PL”
services (third party logistics) or “3PL”. Although logistics is not connected to the most
important areas of expertise in an organization, it requires sufficient resources and high
care. More companies are transferring their logistics activities to logistics professionals,
so they can transfer their resources to their main capabilities, giving them an advantage
over their competitors in a competitive environment and gaining flexibility in their
activities [4]. 3PL is mainly concerned with the transportation and delivery of various
products; moreover, 3PL is a supply chain that includes a variety of services. 3PL
functions consist of storage, terminal processes, customs brokerage and much more. It
also uses analysis services and IT software to monitor and track the distribution status
of goods [5]. The concept of 3PL reminds of the concepts of 1PL and 2 PL. 1PL is a
firm or a person that does not outsource its/his/her shipping and logistics enterprise.
2PL has transport and transport vehicles. They have their own aircrafts, ships, and
trucks. They can offer services with a lease agreement [6].
Following the concept of 3PL, the concept of 4PLs is widely used today. 4PLs
firms are becoming progressively crucial in logistics management. 4PLs companies
integrate the activities of multiple 3PL companies to optimize the entire logistics
process, thus providing high levels of control and efficiency. The 4PLs notion was first
announced as a brand in 1996 by Anderson Consulting (now Accenture Consulting).
The definition of 4PLs was made by them as “an integrator that assembles its own
resources, capabilities, technology and other service providers to design and manage
complex supply chains”. “Cainiao Logistics,” for instance, which integrates more than
30 3PL firms, is the largest 4PLs firm in China as a subsidiary of Alibaba Group [7].
4PLs operations combine the resources of businesses and 3PL providers [8]. A 4PL
service provider is a supply chain integrator that manages and integrates the technol-
ogy, capabilities, and resources of the company it provides services to itself and its
suppliers [9]. 4PL service provider evaluation problem is a subject studied by many
researcher as follows; Chen and Su [10] proposed a decision-making method that uses
a modified particle swarm optimizations and fuzzy goal programming, they aimed to
solve operation allocation of 4PLs with preemptive form. Yao [11] used ant colony
optimization algorithm to solve operational bottleneck in 4PLs by analyzing the inte-
gration of resource in the supply chain. Liu et al. [12] proposed a study by using a
nondominant sorting genetic algorithm to reduce the cost of logistics in 4PLs. Huang
et al. [7] studied on a problem for risk management of logistics outsourcing within the
principal-agency structure for 4PLs firms. Krakovics et al. [13] created a performance
Evaluation of the Fourth Party Logistics Provider Under Neutrosophic Environment 295
evaluation system from the perspective of a 4PLs firm by their research. They devel-
oped an s-curve method to evaluate the performance indicators. Huang et al. [14]
presented a fuzzy duration time model for 4PLs RP (routing problem) to find a min-
imum cost rota with restraints under vogue environment and they, with Lee [15],
extended their study with a genetic algorithm model.
Evaluation of 4PLs service providers includes both tangible and intangible criteria,
so the evaluation of 4PLs service providers can be considered as a multi-criteria
decision making (MCDM) problem. In MCDM methods, experts assess the options
based on their knowledge and judgments and need natural language statements instead
of crisp numerical values in the assessment. The fuzzy set theory was developed by
Zadeh [16], is quite a useful method to meet this requirement. Neutrosophic sets are
one of these extensions, and their notable feature is that they have the first usage of
uncertainty membership. The neutrosophic set is represented by the usage of three
parameters, “Truthiness (T)”, “Indeterminacy (I)” and “Falsity (F)”. In the universe U.
x = x(T, I, F) 2 A, a neutrophic set is defined with T, I, and F; and these T, I, and F are
the standard or non-standard subsets of ]−0,1+[.
Throughout the rest of the study, the steps of the proposed method are given in the
second section. An application is given in the third section. Finally, the conclusion is
presented.
2 The Steps of the Proposed Method
A MCDM problem can be expressed as a decision matrix whose elements show the
evaluation values of all alternatives with respect to each criterion under spherical fuzzy
environment. Let A ¼ fA1 ; A2 ; . . .Am g (m 2) be a discrete set of m feasible alterna-
tives and C ¼ fC1 ; C2 ; . . .Cn g be a finite set of criteria and wP¼ fw1 ; w2 ; . . .wn g be the
weight vector of all criteria which satisfies 0 wj 1 and nj¼1 wj ¼ 1. The steps of
the proposed methodology can be expressed using the following steps.
Step 1: Experts establish pair wise comparison matrix by using Table 1 with nego-
tiation to evaluate criteria. Neutrosophic sets are one of the extensions of Pythagorean
fuzzy sets, so we adopted a Neutrosophic fuzzy entropy measure in this section. The
criteria weights are calculated by using entropy theory as follows:
(i) EN ð AÞ ¼ 0 if A is a crisp set
(ii) EN ð AÞ ¼ 1 if (TA ð xÞ; IA ð xÞ; FA ð xÞ ¼ ð0:5; 0:5; 0:5Þ8x 2 X
(iii) EN ð AÞ EN ðBÞ if A more uncertain than B,
i.e. TA ð xÞ þ FA ð xÞ TB ð xÞ þ FB ð xÞ and jIA ð xÞ þ IAC ð xÞj jIB ð xÞ þ IBC ð xÞj
(iv) EN ð AÞ EN ðAc Þ8A 2 N ð xÞ
Now notice that in a neutrosophic set the presence of uncertainty is due to two
factors: firstly, due to the partial belongingness and partial non-belongingness and
secondly due to the indeterminacy factor. Considering these two factors, we utilize an
entropy measure E1 of a single valued neutrosophic sets A as follows [17]:
1X
E1 ð AÞ ¼ 1 xi 2X
ðTA ðxi Þ þ FA ðxi ÞÞ:jIA ðxi Þ IAc ðxi Þj ð1Þ
n
di ¼ 1 E1 ð AÞ; 8i ð2Þ
di
w i ¼ Pn ; 8i ð3Þ
i¼1 di
Table 1. Linguistic terms and their corresponding neutrosophic numbers

Linguistic terms ðT; I; F Þ Linguistic terms ðT; I; F Þ
Absolutely More h0:9; 0:1; 0:1i Slightly Low Importance h0:4; 0:6; 0:4i
Importance (AMI) (SLI)
Very High Importance h0:8; 0:2; 0:2i Low Importance (LI) h0:3; 0:7; 0:3i
(VHI)
High Importance (HI) h0:7; 0:3; 0:3i Very Low Importance h0:2; 0:8; 0:2i
(VLI)
Slightly More Importance h0:6; 0:4; 0:4i Absolutely Low h0:1; 0:9; 0:1i
(SMI) Importance (ALI)
Equally Importance (EI) h0:5; 0:5; 0:5i
Step 2: Let experts fill in the decision and criteria evaluation matrices according to
their knowledge, background, and judgments. Experts are free to assign any spherical
number without using any linguistic term scale. Experts evaluate the alternatives with
respect to the criteria as if they were benefit criteria such that they assign a lower
linguistic term if it is a cost criterion.
Step 3: Aggregate the judgments of each expert using Neutrosophic Weighted
Arithmetic Mean (NWAM) as given in Eq. (4). Thus, an aggregated neutrosophic
decision matrix based on the opinions of decision makers has been
obtained.
For a
MCDM problem with neutrosophic numbers, decision matrix ~xsij ¼ Cj ð~xi Þ mxn should
be written as in Eq. (5).
D Yn Yn Yn E
~ ~ ~
Fw Av1 ; Av2 ; . . .; Avn ¼ 1 j¼1 ð1 TAi ð xÞÞwi ; 1 j¼1 ð1 IAi ð xÞÞwi ; 1 j¼1 ð1 FAi ð xÞÞwi
ð4Þ
Where W ¼ ðw1 ; w2 ; . . .wn Þ is weight vector and Aj ðj ¼ 1; 2; . . .; nÞ, wj 2 ½0; 1 and

Pn
j¼1 wj ¼ 1.
0 1
ðT11 ; I11 ; F11 Þ ðT12 ; I12 ; F12 Þ ... ðT1n ; I1n ; F1n Þ
B ðT21 ; I21 ; F21 Þ ðT22 ; Iv22 ; F22 Þ ... ðT2n ; I2n ; F2n Þ C
B C
~xNmxn ¼B
B : : : C
C ð5Þ
@ : : : A
ðTm1 ; Im1 ; Fm1 Þ ðTm2 ; Im2 ; Fm2 Þ ... ðTmn ; Imn ; Fmn Þ
þ þ T
Step 4: Define a~Nþ ¼ a~N1 þ T
; ~aN2 ; . . .; ~aNm a
and ~ N ¼ ~
aN1 ; ~ a
aN2 ; . . .; ~ Nm as the
neutrosophic ideal solution and the neutrophic negative ideal solution, respectively,
where ~aNþ ¼ h1; 0; 0i (i = 1, 2, …, m) are the m largest neutrosophic number, and
a
~ N ¼ h0; 0; 1i (i = 1, 2, …, m) are the m smallest neutrosophic number.
Step 5: Calculate the cosine similarity between the alternative Ai and the neutrosophic
ideal solution, and the cosine similarity between Ai and neutrosophic negative ideal
solution by using Eqs. (6–7) [18].

Ciþ ¼ Ai ; ~aNþ
Xn TA ðxi ÞT~a þ ðxi Þ þ IA ðxi ÞI~a þ ðxi Þ þ FA ðxi ÞF~a þ ðxi Þ
¼ i¼1 i
w pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffipffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð6Þ
TA2 ðxi Þ þ IA2 ðxi Þ þ FA2 ðxi Þ T~a2 þ ðxi Þ þ I~a2 þ ðxi Þ þ F~a2 þ ðxi Þ

Ci ¼ Ai ; ~a
N
Xn TA ðxi ÞT~a ðxi Þ þ IA ðxi ÞI~a ðxi Þ þ FA ðxi ÞF~a ðxi Þ
¼ i¼1 i
w pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffipffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð7Þ
TA ðxi Þ þ IA2 ðxi Þ þ FA2 ðxi Þ T~a2 ðxi Þ þ I~a2 ðxi Þ þ F~a2 ðxi Þ
2
Step 6: Calculate the similarity ratio by using Eq. (8) as follows:
Ciþ
Si ¼ ; i ¼ 1; 2; . . .; m ð8Þ
Ci þ Ciþ
Step 7: Rank the alternatives according to their similarity ratio in descending order.
3 Application
In the application section, a 4PLs service provider evaluation problem is handled and
the proposed neutrosophic method is used for the evaluation procedure. The studies in
the literature are examined and six main criteria are determined as follows;
C1: Ability to use information technology
C2: Ability to manage customer relationships
C3: International recognition
C4: Supply chain integration
C5: Cost
C6: Efficient improvement
Then three 4PLs service provider companies are evaluated by three experts from
logistics sector by using the proposed method according to the six different criteria.
Step 1: Experts establish pair wise comparison matrix by using Table with negotiation
to evaluate criteria. The comparison matrix is shown in Table 2.
By using Eq. (1–3) the weights of criteria are calculated as follows:
wC1 ¼ 0; 152; wC2 ¼ 0; 190; wC3 ¼ 0; 200; wC4 ¼ 0; 133; wC5 ¼ 0; 114; wC6 ¼ 0; 210
Table 2. Comparison matrix

C1 C2 C3 C4 C5 C6
C1 h0:5; 0:5; 0:5i h0:7; 0:3; 0:3i h0:8; 0:2; 0:2i h0:6; 0:4; 0:4i h0:6; 0:4; 0:4i h0:4; 0:6; 0:4i
C2 h0:3; 0:7; 0:3i h0:5; 0:5; 0:5i h0:6; 0:4; 0:4i h0:2; 0:8; 0:2i h0:4; 0:6; 0:6i h0:2; 0:8; 0:2i
C3 h0:2; 0:8; 0:2i h0:4; 0:6; 0:4i h0:5; 0:5; 0:5i h0:4; 0:6; 0:4i h0:3; 0:7; 0:3i h0:2; 0:8; 0:2i
C4 h0:4; 0:6; 0:4i h0:8; 0:2; 0:2i h0:6; 0:4; 0:4i h0:5; 0:5; 0:5i h0:5; 0:5; 0:5i h0:3; 0:7; 0:3i
C5 h0:4; 0:6; 0:4i h0:6; 0:4; 0:4i h0:7; 0:3; 0:3i h0:5; 0:5; 0:5i h0:5; 0:5; 0:5i h0:3; 0:7; 0:3i
C6 h0:6; 0:4; 0:4i h0:8; 0:2; 0:2i h0:8; 0:2; 0:2i h0:7; 0:3; 0:3i h0:7; 0:3; 0:3i h0:5; 0:5; 0:5i
Step 2: Each expert establishes decision matrix (Table 3, 4 and 5).
Table 3. Decision matrix by first expert

C1 C2 C3 C4 C5 C6
A1 h0:9; 0:1; 0:1i h0:7; 0:3; 0:3i h0:4; 0:5; 0:4i h0:7; 0:3; 0:2i h0:8; 0:2; 0:2i h0:7; 0:4; 0:2i
A2 h0:6; 0:4; 0:2i h0:7; 0:4; 0:4i h0:8; 0:2; 0:2i h0:1; 0:7; 0:2i h0:4; 0:6; 0:6i h0:2; 0:8; 0:2i
A3 h0:2; 0:8; 0:2i h0:4; 0:6; 0:4i h0:5; 0:5; 0:5i h0:4; 0:6; 0:4i h0:2; 0:8; 0:4i h0:4; 0:5; 0:5i
A4 h0:6; 0:5; 0:4i h0:5; 0:7; 0:4i h0:6; 0:5; 0:4i h0:8; 0:2; 0:1i h0:7; 0:2; 0:2i h0:9; 0:1; 0:1i
A5 h0:7; 0:4; 0:2i h0:6; 0:4; 0:3i h0:5; 0:4; 0:5i h0:2; 0:7; 0:7i h0:7; 0:1; 0:1i h0:8; 0:2; 0:2i
A6 h0:6; 0:4; 0:5i h0:7; 0:2; 0:2i h0:8; 0:2; 0:2i h0:4; 0:5; 0:5i h0:4; 0:6; 0:5i h0:3; 0:5; 0:7i
Table 4. Decision matrix by second expert

C1 C2 C3 C4 C5 C6
A1 h0:8; 0:2; 0:2i h0:6; 0:4; 0:4i h0:5; 0:4; 0:4i h0:6; 0:4; 0:3i h0:7; 0:3; 0:3i h0:5; 0:6; 0:3i
A2 h0:7; 0:3; 0:3i h0:8; 0:2; 0:2i h0:6; 0:4; 0:4i h0:2; 0:7; 0:3i h0:5; 0:5; 0:4i h0:4; 0:5; 0:4i
A3 h0:3; 0:7; 0:3i h0:5; 0:5; 0:3i h0:6; 0:4; 0:4i h0:3; 0:5; 0:5i h0:3; 0:7; 0:5i h0:3; 0:6; 0:5i
A4 h0:7; 0:4; 0:3i h0:6; 0:5; 0:2i h0:7; 0:4; 0:3i h0:7; 0:3; 0:3i h0:6; 0:4; 0:4i h0:7; 0:2; 0:2i
A5 h0:6; 0:5; 0:3i h0:7; 0:5; 0:4i h0:5; 0:5; 0:4i h0:3; 0:5; 0:6i h0:6; 0:3; 0:3i h0:7; 0:3; 0:3i
A6 h0:5; 0:6; 0:4i h0:8; 0:1; 0:1i h0:6; 0:4; 0:4i h0:5; 0:4; 0:4i h0:6; 0:4; 0:4i h0:2; 0:7; 0:2i
Table 5. Decision matrix by third expert

C1 C2 C3 C4 C5 C6
A1 h0:6; 0:4; 0:4i h0:7; 0:2; 0:2i h0:5; 0:4; 0:4i h0:8; 0:2; 0:2i h0:6; 0:2; 0:4i h0:6; 0:4; 0:4i
A2 h0:6; 0:4; 0:4i h0:7; 0:2; 0:3i h0:7; 0:5; 0:6i h0:4; 0:7; 0:6i h0:6; 0:7; 0:2i h0:5; 0:4; 0:3i
A3 h0:4; 0:6; 0:6i h0:6; 0:7; 0:4i h0:7; 0:4; 0:5i h0:4; 0:6; 0:7i h0:4; 0:5; 0:6i h0:6; 0:5; 0:4i
A4 h0:8; 0:1; 0:2i h0:7; 0:4; 0:4i h0:8; 0:2; 0:2i h0:7; 0:2; 0:2i h0:5; 0:3; 0:4i h0:6; 0:3; 0:6i
A5 h0:5; 0:4; 0:6i h0:6; 0:4; 0:5i h0:4; 0:5; 0:6i h0:2; 0:4; 0:7i h0:5; 0:4; 0:5i h0:7; 0:2; 0:2i
A6 h0:6; 0:2; 0:3i h0:7; 0:2; 0:5i h0:4; 0:5; 0:4i h0:6; 0:7; 0:2i h0:7; 0:3; 0:5i h0:4; 0:6; 0:4i
Step 3: Aggregate the judgments of each expert using Neutrosophic Weighted

Arithmetic Mean (NWAM) as given in Eq. (4). The aggregated matrix is shown in
Table 6.
Table 6. Aggregated matrix

C1 C2 C3 C4 C5 C6
A1 h0:80; 0:24; 0:24i h0:66; 0:31; 0:31i h0:47; 0:43; 0:40i h0:70; 0:31; 0:24i h0:71; 0:24; 0:30i h0:60; 0:49; 0:30i
A2 h0:64; 0:36; 0:30i h0:74; 0:27; 0:29i h0:70; 0:38; 0:42i h0:24; 0:70; 0:38i h0:51; 0:60; 0:42i h0:38; 0:60; 0:31i
A3 h0:30; 0:71; 0:38i h0:51; 0:60; 0:36i h0:61; 0:43; 0:46i h0:36; 0:56; 0:55i h0:30; 0:69; 0:51i h0:43; 0:54; 0:47i
A4 h0:71; 0:36; 0:30i h0:61; 0:55; 0:33i h0:71; 0:38; 0:30i h0:73; 0:24; 0:21i h0:61; 0:31; 0:35i h0:76; 0:20; 0:33i
A5 h0:61; 0:44; 0:38i h0:64; 0:44; 0:41i h0:47; 0:47; 0:50i h0:24; 0:55; 0:66i h0:61; 0:28; 0:32i h0:73; 0:24; 0:24i
A6 h0:56; 0:44; 0:41i h0:74; 0:16; 0:27i h0:63; 0:38; 0:35i h0:51; 0:54; 0:38i h0:59; 0:44; 0:46i h0:29; 0:62; 0:45i
Step 4: a~Nþ ¼ h1; 0; 0i (i = 1, 2, …, 6) are the 6 largest neutrosophic number, and

~a
N ¼ h1; 0; 0i (i = 1, 2, …, 6) are the 6 smallest neutrosophic number.
Step 5: Calculate the cosine similarity between the alternative Ai and the neutrosophic
ideal solution, and the cosine similarity.
CAþ1 ¼ 0:79; CAþ2 ¼ 1:27; C3þ ¼ 1:35; CAþ4 ¼ 2:90; CAþ5 ¼ 1:38; CAþ6 ¼ 1:68
CA1 ¼ 1:16; CA2 ¼ 2:90; C3 ¼ 5:02; CA4 ¼ 3:87; CA5 ¼ 2:86; CA6 ¼ 3:28
Step 6: Calculate the similarity ratio by using Eq. (8) as follows;
SA1 ¼ 0:404; SA2 ¼ 0:305; SA3 ¼ 0:212; SA1 ¼ 0:429; SA1 ¼ 0:326; SA6 ¼ 0:339
Step 7: Alternatives are ranked according to their similarity ratio by descending order.
A4 [ A1 [ A6 [ A5 [ A2 [ A3
4 Conclusion
In this study, a neutrosophic MCDM method was presented to assess the 4PLs service
providers’ performance. In the implementation section, three experts assessed the
performance of 4PLs service providers by using the neutrosophic linguistic scale. Then
4PLs service providers were arranged in descending order by calculating each service
provider’s closeness coefficient.
The results indicated that the proposed neutrosophic MCDM method can be taken
as an appropriate method in the evaluation of 4PLs service providers’ performance.
Prospective studies may apply new types of fuzzy extensions to the problem.
References
1. Barbosa-Povoa, A.P., Pinto, J.M.: Process supply chains: perspectives from academia and
industry. Comput. Chem. Eng. 132, 106606 (2020)
2. Chopra, S., Meindl, P.: Supply Chain Management: Strategy, Planning, and Operation.
Pearson, New Jersey (2016)
3. Ekici, Ş.Ö., Kabak, Ö., Ülengin, F.: Improving logistics performance by reforming the
pillars of global competitiveness index. Transp. Policy 81, 197–207 (2019)
4. Batarlienė, N., Jarašūnienė, A.: 3PL Service improvement opportunities in transport
companies. Procedia Eng. 187, 67–76 (2017)
5. 16 January 2019. https://ithinklogistics.com/blog/logistics-and-its-types-difference-between-
1pl-2pl-3pl-and-4PLs/. Accessed 10 Mar 2020
6. Jayshree, M.: Logistics: What is difference between 1PL, 2PL, 3PL, 4PL services? (2019).
https://www.quora.com/Logistics-What-is-difference-between-1PL-2PL-3PL-4PLs-services.
Accessed 10 Mar 2020
7. Huang, M., Tu, J., Chao, X., Jin, D.: Quality risk in logistics outsourcing: a fourth party
logistics perspective. Eur. J. Oper. Res. 276, 855–872 (2019)
8. Leung, L.C., Cheung, W., Hui, Y.V.: A framework for a logistics e-commerce community
network: The Hong Kong air cargo industry. IEEE Trans. Syst. Man Cybern.-Part-A: Syst.
Hum. 30(4), 446–455 (2000)
9. Mukhopadhyay, S.K., Setaputra, R.: The role of 4PLs as the reverse logistics integrator:
optimal pricing and return policies. Int. J. Phys. Distrib. Logist. Manag. 36(9), 716–729
(2006)
10. Chen, K.H., Su, C.T.: Activity assigning of fourth party logistics by particle swarm
optimization-based on optimization based on optimization-based preemptive fuzzy integer
goal programming. Expert Syst. Appl. 37, 3630–3637 (2010)
11. Yao, J.: Optimization analysis on supply chain resources integration in fourth party logistics.
J. Manuf. Syst. 29, 121–129 (2010)
12. Liu, Q., Zhang, C., Zhu, K., Rao, Y.: Novel multi-objective resource allocation and activity
scheduling for fourth party logistics. Comput. Oper. Res. 44, 42–51 (2014)
13. Krakovics, F., Leal, J.E., Mendes, P.J., Santos, R.L.: Defining and calibrating performance
indicators of a 4PLs in the chemical industry in Brazil. Int. J. Prod. Econ. 115, 502–514
(2008)
14. Huang, M., Cui, Y., Yang, S., Wang, X.: Fourth party logistics routing problem with fuzzy
duration time. Int. J. Prod. Econ. 145, 107–116 (2013)
15. Cui, Y., Huang, M., Yang, S., Lee, L.H., Wang, X.: Fourth party logistics routing problem
model with fuzzy duration time and cost discount. Knowl.-Based Syst. 50, 14–24 (2013)
17. Majumdar, P., Kumar, K.S.: On similarity and entropy of neutrosophic sets. J. Intell. Fuzzy
Syst. 26, 1245–1252 (2014)
18. Ye, J.: Vector similarity measures of simplified neutrosophic sets and their application in
multicriteria decision making. Int. J. Fuzzy Syst. 16(2), 204–211 (2014)
A* Algorithm Under Single-Valued
Neutrosophic Fuzzy Environment
Esra Çakır(&) and Ziya Ulukan
Department of Industrial Engineering,

Galatasaray University, 34349 Ortakoy/Istanbul, Turkey
{ecakir,zulukan}@gsu.edu.tr
Abstract. A* (A-star) is a computer algorithm that is widely used in

pathfinding and graph traversal to effectively compute optimal solutions. The
algorithm uses the best first search and finds the lowest cost path from a given
start node to a target node. It follows a path of the lowest known heuristic cost,
keeping a sorted priority queue of alternate path segments along the way. It is
more accurate than greedy best-first search algorithm because it takes into
account already moved nodes. Each iteration calculates the cost from the starting
node and adds the heuristic cost to the target node. The cost of paths can be
given based on truth, indeterminate and false membership functions. In the
graph, cost path and heuristic cost paths can be expressed with fuzzy numbers.
The contribution of this study to the existing literature comes from the usage of
A* algorithm to find the shortest path of a graph with arcs expressed in single
valued neutrosophic numbers. A numerical example is given to show the use-
fulness of proposed approach.
Keywords: A-star algorithm Fuzzy set theory Graph theory Shortest path
problem Single-valued neutrosophic fuzzy numbers
1 Introduction
The fuzzy set idea was originally introduced by Zadeh [1] in 1965 with a membership
degree of an element. In 1986, Atanassov [2] developed this concept with intuitionistic
fuzzy numbers by adding the degree of non-member values to express an imprecise
information. In addition, Samarandache [3] represented neutrosophic fuzzy sets to
express uncertain, incomplete and indeterminate information in 1998. Neutrosophic
fuzzy number consists of a truth-membership function (T), an indeterminate – mem-
bership function (I) and false-membership function (F) independently, which are within
the real standard or non-standard unit interval 0; 1 þ ½ [4]. Numerous recent researches
on single-valued neutrosophic fuzzy set were related on fuzzy graphs [5–7]. Some of
these researches focused on finding shortest paths of fuzzy graphs. Shortest path
problem can be solved by different crisp algorithms and these algorithms can be
adapted to fuzzy sets.
A* algorithm is a graph/tree search algorithm that finds a shortest path from an
initial node to end node [8]. This algorithm uses a “heuristic estimate” h(x) that gives
an estimate of the best route that goes through that node. It follows best first search and
A* Algorithm Under Single-Valued Neutrosophic Fuzzy Environment 303
visits the nodes in order of heuristic estimate. This algorithm is similar to greedy best-
first search but more accurate because A* takes into account the nodes that have
already been traversed. It combines features of uniform-cost search and pure heuristic
search to effectively compute optimal solutions. The time complexity of A* depends on
the heuristic. In the worst case of an unbounded search space, the number of nodes
expanded is exponential in the depth of the solution (the shortest path) d: H(bd ), where
b is the branching factor (the average number of successors per state) [9]. In the
literature, there are numerous researches on A* algorithm [10–12].
The contribution of this paper is to combine single-valued neutrosophic fuzzy sets and
A* algorithm. Thus, proposed algorithm can be used for shortest path between start node
and end node on a graph -or digraph- with single-valued neutrosophic fuzzy weights.
The organization of this paper is as follows. Section 2 introduces single-valued
neutrosophic fuzzy sets and its operations. Single-valued neutrosophic fuzzy A*
algorithm is given in Sect. 3. Proposed fuzzy A* algorithm is applied on a single-
valued neutrosophic fuzzy graph in Sect. 4. Conclusions are given in Sect. 5.
2 Single-Valued Neutrosophic Fuzzy Sets
In this section, the preliminaries and definitions of the proposed method with single-
valued neutrosophic fuzzy information [3, 13–15] are given.
Definition 1. Let X be a space of points (objects) and the generic element in X is
denoted by x; A = {<x: TA ð xÞ; IA ð xÞ; FA ð xÞ>, x 2 X} is the form of an object that is
the neutrosophic set A, where the functions T, I, F: x ! 0; 1 þ ½ define respectively
the truth-membership function, an indeterminacy-membership function and a falsity-
membership function of the element x 2 X to the set A with condition:

0 TA ð xÞ þ IA ð xÞ þ FA ð xÞ 3 þ ð1Þ
The function TA ð xÞ; IA ð xÞ and FA ð xÞ are real standard or non-standard subset of

0; 1 þ ½.
Definition 2: Let X be a space of points (objects) and the generic element in X is
denoted by x. A single-valued neutrosophic set A is characterized by truth-membership
function TA ð xÞ, an indeterminacy-membership function IA ð xÞ and a falsity-membership
function FA ð xÞ. TA ð xÞ; IA ð xÞ; FA ð xÞ 2 ½0; 1 for each point x in X. A single-valued
neutrosophic fuzzy number A is written as A = {<x: TA ð xÞ; IA ð xÞ; FA ð xÞ>, x 2 X}.
Definition 3: Let A ~ 2 ¼ ðT2 ; I2 ; F2 Þ be two single valued neu-
~ 1 ¼ ðT1 ; I1 ; F1 Þ and A
trosophic number. Then, the operations of two single valued neutrosophic numbers are
defined as follows:
~ 1 ¼ \1 ð1 T1 Þk ; ðI1 Þk ; ðF1 Þk [
kA ð2Þ
~ k ¼ ðT
A ~ k ; 1 ð1 I1 Þk ; ð1 ð1 F1 Þk Þ ð3Þ
1 1
304 E. Çakır and Z. Ulukan
~1 A
A ~ 2 ¼ \T1 þ T2 T1 T2 ; I1 I2 ; F1 F2 [ ð4Þ
A ~ 2 ¼ \T1 T2 ; I1 þ I2 I1 I2 ; F1 þ F2 F1 F2 [
~1 A ð5Þ
where k [ 0.
Definition 4: 0n may be defined as 0n = {<x, (0, 1, 1)>: x 2 X}. A useful approach for
comparing two single-valued neutrosophic numbers is the use of a score function.
Definition 5: Let A ~ 1 ¼ ðT1 ; I1 ; F1 Þ be a single-valued neutrosophic fuzzy number.

~ 1 , accuracy function a A
Then, the score function s A ~ 1 and certainty function c A
~1
~ 1 are defined as below:
of A

~ 1 ¼ 2 þ T 1 I 1 F1
s A ð6Þ
3

~ 1 ¼ T 1 F1
a A ð7Þ

~ 1 ¼ T1
c A ð8Þ
Definition 6: Let A ~ 1 ¼ ðT1 ; I1 ; F1 Þ and A ~ 2 ¼ ðT2 ; I2 ; F2 Þ are two single valued neu-
trosophic numbers. Then, we define a ranking method as follows:

– If s A~1 s A ~ 2 , then A ~ 1 is bigger than A ~ 2 and denoted by A ~1 A~ 2.

– If s A~1 ¼ s A ~ , then
2
– If a A~1 a A ~ 2 , then A~ 1 is bigger than A ~ 2 and denoted by A ~1 A ~ 2.

~ ~
– If a A1 ¼ a A2 , then

– If c A~1 c A ~ 2 , then A ~ 1 is bigger than A ~ 2 and denoted by A ~1 A ~ 2.

~ ~ ~ ~
– If c A1 ¼ c A2 , then A1 is equal to A2 and denoted by A1 ¼ A2 . ~ ~
3 Proposed Model
A* algorithm [9] consists of a “open list” and a “closed list”. The open list is a record
of all nodes immediately adjacent to areas that have already been explored. All nodes
which have been explored by the algorithm are recorded in closed list, as well.
Algorithm proceeds by calculating distance travelled from start node with estimated
distance to the end node.
Let w~ ij be a single-valued neutrosophic weight of edge (i, j), ~
g(j) be a single-valued
neutrosophic fuzzy distance from start node to node n and ~ h(j) be a single-valued
neutrosophic fuzzy heuristic function that estimates the distance of the shortest path
from node n to the end node. A* selects the path that minimizes ~f (j) = ~ g(j) + ~h(j). A*
terminates when the shortest path is found or if there are no paths eligible to be
extended. The heuristic function depends on the problem. If the heuristic function is
admissible, A* guarantees to find a shortest path from start node to end node.
A* uses a priority queue to perform the repeated selection of minimum distance
nodes to expand. This priority queue is called as the fringe or open set. At each step of
the algorithm, the node with the lowest ~f (j) value is removed from the queue, the ~f and
~g values of its neighbors are updated accordingly, and then these neighbors are added
to the queue. It continues until the queue is empty or until end node has a lower ~f value
than any node in the queue [16]. Since ~h at the end node is zero in an admissible
heuristic, the ~f value of the goal is then the distance of the shortest path.
The steps of single-values neutrosophic fuzzy A* algorithm are as follows:
Step 1: Select start node and add it to Open set.
Step 2: Add nodes reachable from selected node to Open set
Step 3: Remove selected node from Open and add it to Closed.
Step 4: Calculate ~f (n) = ~g(n) + ~h(n) of nodes in Open.
Step 5: Calculate scores of nodes in Open to compare single-valued neutrosophic
fuzzy numbers and select node with the smallest score.
Step 6: If end node is selected or Open is empty terminate the algorithm. Else,
return Step 2.
Step 7: Find the shortest path by tracing backward through the network.
Pseudocode of single-valued neutrosophic A* algorithm is shown in Table 1.
Table 1. Pseudocode of single-valued neutrosophic A* Algorithm.

4 Application
Assume a network with single-valued neutrosophic fuzzy weights as in Fig. 1. Single-

valued neutrosophic heuristic distances to end node are given in Table 2. According to
the proposed model, the shortest path from node 1 (start node) to node 7 (end node) is
calculated as follows:
Fig. 1. A graph with single-valued neutrosophic distance.
Table 2. Single-valued neutrosophic heuristic distances to end node (node 7).

Node ~h(n)
2 (0.6, 0.4, 0.3)
3 (0.6, 0.2, 0.1)
4 (0.6, 0.4, 0.5)
5 (0.1, 0.5, 0.4)
6 (0.1, 0.3, 0.6)
7 (0, 1, 1)
According to the proposed model, the shortest path between node 1 and end node 7
is calculated as follows:
In order to compare the ~f (2), ~f (3) and ~f (4), according to the Eq. (6) the scores are s
(~f (2)) = 0.82, s(~f (3)) = 0.73 and s(~f (4)) = 0.787. Since the rank of ~f (3) is less than
others, the algorithm proceeds from node 3.
The scores are s(~f (2)) = 0.82, s(~f (4)) = 0.787, s(~f (5)) = 0.876 and s(~f (6)) = 0.872.
Since the rank of ~f (4) is less than others, the algorithm proceeds from node 4.
The scores are s(~f (2)) = 0.82, s(~f (5)) = 0.876, s(~f (6)) = 0.872 (from node 3) and
s(~f (6)) = 0.841 (from node 4). Since the rank of ~f (2) is less than others, the algorithm
proceeds from node 2 and since the score of ~f (6) (from node 4) is less than ~f (6) (from
node 3), in the rest of the algorithm, score of ~f (6) (from node 4) is used for node 6.
The scores are s(~f (5)) = 0.876 (from node 3), s(~f (6)) = 0.841 (from node 4) and
s(f (5)) = 0.877 (from node 2). Since the rank of ~f (6) (from node 4) is less than others,
~
the algorithm proceeds from node 6 and since the score of ~f (5) (from node 3) is less
than ~f (5) (from node 2), in the rest of the algorithm, score of ~f (5) (from node 3) is used
for node 5.
The scores are s(~f (5)) = 0.876 (from node 3) and s(~f (7)) = 0.855 (from node 6).
Since the rank of ~f (7) (from node 6) is less than others and node 7 is the end node, the
algorithm is terminated by selecting end node. Thus, the shortest path is {node 1- node
4- node 6 - node 7} and it is shown in Fig. 2.
Fig. 2. Shortest path from node 1 to node 7 with node scores.

The end node (node 7) can also be reached from node 5. However, this path has not
been tried by the algorithm. Because the score of the node 5 is higher than the end
node. This means that each path to go from node 5 has a higher score than the shortest
path found.
5 Conclusion
A* algorithm can be used for single-valued neutrosophic fuzzy graphs to find shortest
path. The contribution of this paper is to combine single-valued neutrosophic fuzzy sets
and A* algorithm. By adapting fuzzy A* algorithm to the single-valued neutrosophic
numbers, weights of arcs and heuristic values of nodes which express estimated dis-
tance to end node are aggregated in every iteration. By comparing score values the
smallest path between start and end node is determined. A numerical example is given
to show usefulness of proposed model on a single-valued neutrosophic fuzzy graph.
For further research, we suggest to use A* algorithm on single-valued neutrosophic
fuzzy digraphs. We also recommend to apply this fuzzy algorithm with different fuzzy
sets such as intuitionistic, Pythagorean and spherical fuzzy sets.
References
1. Zadeh, L.A.: Fuzzy sets. Inform. Control 8, 338–353 (1965)
2. Atanassov, K.: Intuitionistic fuzzy sets. Fuzzy Sets Syst. 20, 87–96 (1986)
3. Smarandache, F.: A Unifying Field in Logics. Neutrosophy: Neutrosophic Probability, Set
and Logic. American Research Press, Reboboth (1998)
4. Smarandache, F.: A geometric interpretation of the neutrosophic set – a generalization of the
intuitionistic fuzzy set. In: 2011 International Conference on Granular Computing (GrC),
pp. 602–606 (2011)
5. Şahin, R.: An introduction to neutrosophic graph theory with applications. Soft. Comput. 23,
569–581 (2019)
6. Naz, S., Rashmanlou, H., Malik, M.A.: Operations on single valued neutrosophic graphs
with application. J. Intell. Fuzzy Syst. 32(3), 2137–2151 (2017)
7. Broumi, S., Talea, M., Smarandache, F.: Shortest path problem on single valued
neutrosophic graphs. In: 2017 International Symposium on Networks, Computers and
Communications (ISNCC), pp. 1–8 (2017)
8. Reddy, H.: PATH FINDING – Dijkstra’s and A* Algorithms, pp. 1–15 (2013)
9. Russell, S., Norvig, P.: Artificial Intelligence: A Modern Approach, 2nd edn, pp. 97–104.
Prentice Hall, Upper Saddle River (1995)
10. Duchon, F., Babinec, A.: Path planning with modified A star algorithm for a mobile robot.
Procedia Eng. 96, 59–69 (2014)
11. Liu, X., Daoxiong, G.: A comparative study of A-star algorithms for search and rescue in
perfect maze. In: 2011 International Conference on Electric Information and Control
Engineering (ICEICE). IEEE (2011)
12. Elisabete, F., Pedro, C., Jośe, L., Germano, V.: Towards an Orientation Enhanced A star
Algorithm for Robotic Navigation. IEEE (2015)
13. Wang, H., Smarandache, F., Zhang, Y., Sunderraman, R.: Single valued neutrosophic sets.
Multisspace Multistructure 4, 410–413 (2010)
14. Chi, P., Liu, P.: An extended TOPSIS method for the multiple attribute decision making
problems based on interval neutrosophic set. Neutrosophic Set Syst. 1, 1–8 (2013)
15. Jayagowri, P., Geetha Ramani, G.: Using trapezoidal intuitionistic fuzzy number to find
optimized path in a network. Adv. Fuzzy Syst. 2014, 6 p. (2014)
16. Dechter, R., Judea, P.: Generalized best-first search strategies and the optimality of A*.
J. ACM 32(3), 505–536 (1985)
Bipolar Neutrosophic Fuzzy Dijkstra
Algorithm and Its Application
Esra Çakır(&) and Ziya Ulukan
Department of Industrial Engineering, Galatasaray University,

34349 Ortakoy/Istanbul, Turkey
{ecakir,zulukan}@gsu.edu.tr
Abstract. Finding the shortest path of complex networks and transportation

problems is an important task. Many algorithms have been developed to solve
this problem and one of the most well-known is the Dijkstra algorithm, also
called “label algorithm”. The graph consists of directed or undirected arc with
positive weights as well and Dijkstra’s algorithm can solve both of them. This
classical problem in graph theory can be remodeled in a fuzzy environment. The
weights of arcs can be expressed with bipolar neutrosophic numbers based on
positive and negative effects and the fuzzy graph can be solved with the Dijkstra
algorithm for the shortest path problem. This study shows how the Dijkstra
algorithm is used to find the shortest path of arcs expressed in bipolar neutro-
sophic numbers in a single source network. Finally, a numerical example is
given to demonstrate the effectiveness of the method.
Keywords: Bipolar neutrosophic fuzzy numbers Dijkstra algorithm Fuzzy

set theory Graph theory Shortest path problem
1 Introduction
To express uncertainty, fuzzy set theory by Zadeh [1] and intuitionistic fuzzy sets by
Atanassov [2] originally introduced. Based on their ideas, Smarandache [3] proposed
neutrosophic fuzzy set which is generalization of fuzzy set theory and intuitionistic
fuzzy sets. The concept of the bipolar neutrosophic fuzzy sets has six components such
as truth-membership (T), indeterminacy membership (I), falsity membership (F) and
their respective negative membership degrees [4]. Some of the recent researches on
neutrosophic fuzzy sets has related neutrosophic sets with shortest path problem [5–9].
Finding the shortest path of a graph is a significant problem in graph theory and in the
real world, edges such as time and cost depend on uncertain conditions. Thus, fuzzy
edges graphs are used to represent real case. Numerous algorithms exist to solve this
problem and it is necessary to adapt these algorithms to different types of fuzzy sets.
Dijkstra algorithm is one of the most known algorithms in computer science. It is
developed in 1959 to find shortest path within a graph whose edges are all non-negative
values [10]. To solve the single source shortest path problem, it uses the greedy
approach. The algorithm repeatedly selects from the unselected vertices and declares the
distance to be actual shortest distance from start node. In every iteration “Permanent”
label is assigned to one node which has shortest path between “Temporary” nodes.
For general graphs its time complexity with Big-O notation is H((E + V) * log V) [11]
(E as edges and V as vertex).
The contribution of this paper is to combine bipolar neutrosophic fuzzy sets and
Dijkstra algorithm. Thus, proposed algorithm can find shortest paths to every node on a
graph –or a digraph- with bipolar neutrosophic fuzzy edges.
The rest of the paper is organized as follows. Bipolar neutrosophic fuzzy sets and
its operations are introduced in Sect. 2. Section 3 gives bipolar neutrosophic fuzzy
Dijkstra algorithm. Section 4 applies proposed Dijkstra algorithm on a bipolar neu-
trosophic fuzzy graph. Conclusions and further researches are given in Sect. 5.
2 Bipolar Neutrosophic Fuzzy Sets
This section gives the preliminaries and definitions of the proposed method with
bipolar neutrosophic fuzzy information [4].
Definition 1: Let X be a universe of discourse. Then, A = {<x; TA ð xÞ; IA ð xÞ; FA ð xÞ>:
x 2 X} is defined as a neutrosophic set, which is characterized by a truth-membership
function TA : X ! 0; 1 þ ½, an indeterminacy-membership function IA : X ! 0; 1 þ ½
and a falsity-membership function FA : X ! 0; 1 þ ½. The sum of TA ð xÞ; IA ð xÞ and
FA ð xÞ has no restrictions, so
0 sup TA ð xÞ þ sup IA ð xÞ þ sup FA ð xÞ 3 ð1Þ
Definition 2: A bipolar fuzzy set A in X defined as an object of the form

A = {<x; TAþ ð xÞ; IAþ ð xÞ; FAþ ð xÞ; TA ð xÞ; IA ð xÞ; FA ð xÞ>: x 2 X}, where T þ , I þ , F þ :
X ! ½1; 0 and T , I , F : X ! ½1; 0. The positive membership degrees T þ ðxÞ,
I þ ðxÞ and F þ ðxÞ denote the truth membership, indeterminate membership and false
membership of element x 2 X corresponding to a bipolar neutrosophic set A. The
negative membership degrees T ðxÞ, I ðxÞ and F ðxÞ denote the truth membership,
indeterminate membership and false membership of element x 2 X to some implicit
counter-property corresponding to a bipolar neutrosophic set A.
Theorem: A bipolar neutrosophic set is the generalization of a bipolar fuzzy set.
Proof: Assume that X is a bipolar neutrosophic fuzzy set. If the membership degrees
I þ , F þ ; I and F equal to zero then, X is reduced from the bipolar neutrosophic set
to the bipolar fuzzy set.
Definition 3: Let ~a1 ¼ ðT1þ ; I1þ ; F1þ ; T1 ; I1 ; F1 Þ and ~a2 ¼ ðT2þ ; I2þ ; F2þ ; T2 ; I2 ; F2 Þ
be two bipolar neutrosophic fuzzy numbers. Then, the operations for bipolar neutro-
sophic numbers are as follows:
k k
k~a1 ¼ \1 1 T1þ ; ðI1þ Þk ; F1þ ; ðT1 Þk ; ðI1 Þk ; ð1
k
1 ðF1 Þ [ ð2Þ
Bipolar Neutrosophic Fuzzy Dijkstra Algorithm and Its Application 313

þ k

þ k
~ak1 ¼ \ðT1þ Þk ; 1 1 I1 ; 1 1 F1 ; ð1
k
1 T1 ; ðI1 Þk ; ðF1 Þk [ ð3Þ
~a1 ~a2 ¼ \T1þ þ T2þ

T1þ T2þ ; I1þ I2þ ; F1þ F2þ ; T1 T2 ; I1 I2 I1 I2 ; ðF1 F2 ð4Þ
F1 F2 Þ [
~a1 ~a2 ¼ \T1þ T2þ ; I1þ þ I2þ I1þ I2þ ; F1þþ F2þ
F1þ F2þ ; T1 T2 I1 I2 ; I1 I2 ; F1 F2 [ ð5Þ
where k [ 0.
Definition 4: 0n may be defined as 0n = {<x, (0, 1, 1, −1, 0, 0)>: x 2 X}. A conve-
nient method for comparing bipolar neutrosophic number is by using of score function.
Definition 5: Let ~a1 ¼ ðT1þ ; I1þ ; F1þ ; T1 ; I1 ; F1 Þ be a bipolar neutrosophic fuzzy
number. Then, the score function sð~a1 Þ, accuracy function að~a1 Þ and the certainty
function cð~a1 Þ of an bipolar neutrosophic fuzzy number are defined as below:
T1þ þ 1 I1þ þ 1 F1þ þ 1 þ T1 I1 F1

sð~a1 Þ ¼ ð6Þ
6
að~a1 Þ ¼ T1þ F1þ þ T1 F1 ð7Þ
cð~a1 Þ ¼ T1þ F1 ð8Þ
Definition 6: Let ~a1 ¼ ðT1þ ; I1þ ; F1þ ; T1 ; I1 ; F1 Þ and ~a2 ¼ ðT2þ ; I2þ ; F2þ ; T2 ; I2 ; F2 Þ
be two bipolar neutrosophic numbers. The comparison method for two bipolar neu-
trosophic number can be defined as follows:
– If sð~a1 Þ [ sð~a2 Þ, then ~a1 is bigger than ~a2 and denoted by ~a1 [ ~a2 .
– If sð~a1 Þ ¼ sð~a2 Þ, then
– If að~a1 Þ [ að~a2 Þ, then ~a1 is bigger than ~a2 and denoted by ~a1 [ ~a2 .
– If að~a1 Þ ¼ að~a2 Þ, then
– If cð~a1 Þ [ cð~a2 Þ, then ~a1 is bigger than ~a2 and denoted by ~a1 [ ~a2 .
– If cð~a1 Þ ¼ cð~a2 Þ, then ~a1 is equal to ~a2 and denoted by ~a1 ¼ ~a2 .
3 Proposed Model
Dijkstra’s algorithm [12] solves the problem of finding the shortest path from a source
to a destination. This algorithm can be used the deliver the set of edges connecting all
vertices.
Let ~di be a bipolar neutrosophic shortest distance from start node to node i and w
~ ij
be a bipolar neutrosophic weight of edge (i, j). The bipolar neutrosophic labels of node
j coming from node i with a distance ~di from start node is:

~dj ; i ¼ ~di w
~ ij ; i ð9Þ

~ ij [ 0
s w ð10Þ
Note that the score of bipolar neutrosophic label should be positive, because
Dijkstra’s algorithm proceeds on graph with positive edge. In this algorithm, there is
two status of nodes as “Permanent” and “Temporary”. Temporary status is assigned to
a node reachable (node j) from a permanent node (node i). If there are two labels of a
node with temporary status, algorithm proceeds with the label with smallest score. In
each iteration of the algorithm, only one Permanent status is assigned to a node which
has a label with smallest score.
Pseudocode of bipolar neutrosophic Dijkstra label correcting algorithm is shown in
Table 1.
Table 1. Pseudocode of bipolar neutrosophic Dijkstra label correcting algorithm.

The steps of bipolar neutrosophic fuzzy Dijkstra’s algorithm are as follows:

Step 1 - Assign permanent status and label ½0n ; to start node.
Step 2 - Assign temporary status and calculate labels [~
di w~ ij ; i] to nodes reachable
(node j) from permanent nodes (node i). (If node j has already a label with tem-
porary status, compare the scores of labels and choose the label with smallest score.)
Step 3 - Calculate scores of labels to compare bipolar neutrosophic numbers.
Assign permanent status to node with the smallest score.
Step 4 - If all nodes have permanent status, terminate the algorithm. Else, return to
Step 2.
Step 5 - Find the shortest path by tracing backward through the network using the
label information.
4 Application
(0.7,0.7,0.5,-0.8,-0.7,-0.6)
(0.5,0.2,0.7,-0.5,-0.1,-0.9)
(0.9,0.7,0.2,-0.8,-0.6,-0.1)
Fig. 1. A network with bipolar neutrosophic fuzzy weight.
Consider a network with bipolar neutrosophic fuzzy weight as in Fig. 1. According to

the proposed model, the shortest path from node 1 to node 6 is calculated as follows:
Iteration 0: Permanent status and label ½0n ; are assigned to start node (node 1).
Iteration 1: Temporary status and labels are assigned to nodes reachable from per-
manent nodes.
According to the algorithm, the node with temporary status and the smallest label
should be given permanent status. Scores of labels are calculated to compare bipolar
neutrosophic numbers. Since the s(0.3, 0.4, 0.2, −0.6, −0.3, −0.7) = 0.35 and s(0.4,
0.4, 0.5, −0.7, −0.8, −0.4) = 0.33, permanent status is assigned to node 3.
Iteration 2: Temporary status and labels are assigned to nodes reachable from nodes
with permanent status.
Since the s(0.3, 0.4, 0.2, −0.6, −0.3, −0.7) = 0.35, s(0.7, 0.8, 0.35, −0.35, −0.82,
−0.94) = 0.61, s(0.82, 0.28, 0.25, −0.56, −0.94, −0.76) = 0.57 and s(0.88, 0.2, 0.15,
−0.28, −0.84, −0.52) = 0.6 permanent status is assigned to node 2 (from node 1).
Iteration 3: In this iteration there is no new nodes reachable from permanent nodes
but new label is assigned to nodes reachable from new permanent node.
Since the s(0.82, 0.28, 0.25, −0.56, −0.94, −0.76) = 0.57, s(0.88, 0.2, 0.15, −0.28,
−0.84, −0.52) = 0.6 and s(0.93, 0.28, 0.04, −0.48, −0.72, −0.73) = 0.596 permanent
status is assigned to node 4.
Iteration 4: Temporary status and labels are assigned to nodes reachable from nodes
with permanent status.
Since the s(0.88, 0.2, 0.15, −0.28, −0.84, −0.52) = 0.6, s(0.93, 0.28, 0.04, −0.48,
−0.72, −0.73) = 0.596 and s(0.892, 0.016, 0.2, −0.28, −0.97, −0.904) = 0.711 per-
manent status is assigned to node 5 (from node 2).
Iteration 5: In this iteration there is no new nodes reachable from permanent nodes
but new label is assigned to nodes reachable from new permanent node.
Since the s(0.892, 0.016, 0.2, −0.28, −0.97, −0.904) = 0.711 and s(0.951, 0.014,
0.008, −0.24, −0.86, −0.784) = 0.701 permanent status is assigned to node 6 (from
node 5).
(0.7,0.7,0.5,-0.8,-0.7,-0.6)
(0.5,0.2,0.7,-0.5,-0.1,-0.9)
(0.9,0.7,0.2,-0.8,-0.6,-0.1)
Fig. 2. Shortest path of each node from node 1.

Now all nodes have permanent status. Thus, the algorithm terminates. By tracing
backward through the network using the label information, shortest path from node 1 to
node 6 is “node 1 -> node 2 -> node 5 -> node 6”. Shortest path of each node from
node 1 is represented in Fig. 2.
5 Conclusions
Using fuzzy Dijkstra algorithm, the shortest path of a graph with bipolar neutrosophic
edges can be found. The contribution of this paper is to combine Dijkstra algorithm and
bipolar neutrosophic fuzzy sets. By adapting fuzzy Dijkstra algorithm to the bipolar
neutrosophic numbers, edges are compared with score value and permanent paths are
selected in each iteration. According to the classical algorithm, at the end of the
iterations, shortest paths for each node in the graph is calculated. A numerical example
is given to prove the validity of proposed model and show iterations of fuzzy Dijkstra
algorithm on bipolar neutrosophic graphs.
For further research, we recommend to use fuzzy Dijkstra algorithm with other
fuzzy sets such as intuitionistic fuzzy sets, Pythagorean fuzzy sets etc. We also suggest
to compare other shortest path algorithms with fuzzy sets to show the effectiveness of
algorithms on fuzzy graphs.
References
1. Zadeh, L.A.: Fuzzy sets. Inform. Control 8, 338–353 (1965)
3. Smarandache, F.: A Unifying Field in Logics. Neutrosophy: Neutrosophic Probability, Set
and Logic. American Research Press, Reboboth (1998)
4. Bosc, P., Pivert, O.: On a fuzzy bipolar relational algebra. Inf. Sci. 219, 1–16 (2013)
5. Broumi, S., Bakal, A., Talea, M., Smarandache, F., Vladareanu, L.: Applying Dijkstra
algorithm for solving neutrosophic shortest path problem. In: IEEE International Conference
on Advanced Mechatronic Systems (ICAMechS), pp. 412–416 (2016)
6. Şahin, R.: An introduction to neutrosophic graph theory with applications. Soft Comput. 23,
569–581 (2019)
7. Broumi, S., Talea, M., Bakali, A., Smarandache, F.: Application of Dijkstra algorithm for
solving interval valued neutrosophic shortest path problem. In: 2016 IEEE Symposium
Series on Computational Intelligence (SSCI), pp. 1–6 (2016)
8. Naz, S., Rashmanlou, H., Malik, M.A.: Operations on single valued neutrosophic graphs
with application. J. Intell. Fuzzy Syst. 32(3), 2137–2151 (2017)
9. Broumi, S., Talea, M., Smarandache, F.: Shortest path problem on single valued
neutrosophic graphs. In: 2017 International Symposium on Networks, Computers and
Communications (ISNCC), pp. 1–8 (2017)
10. Reddy, H.: Path Finding – Dijkstra’s and A* Algorithms, pp. 1–15 (2013)
11. Barbehenn, M.: A note on the complexity of Dijkstra’s algorithm for graphs with weighted
vertices. IEEE Trans. Comput. 47(2), 263 (1998)
12. Ahuja, R.K., Magnanti, T.L., Orlin, J.B.: Network Flow Theory, Algorithms, and
Applications. Prentice-Hall, Englewood-Cliffs (1993)
Physician Selection with a Neutrosophic
Multi-criteria Decision Making Method
Ahmet Sarucan , M. Emin Baysal , and Orhan Engin(&)
Industrial Engineering Department, Konya Technical University, Konya, Turkey

[email protected]
Abstract. In this study, it is done that the evaluation of neurology clinics of

two hospitals under the Ministry of Health and their physicians, considering
patient preferences. While effective service policies and procedures of hospitals
increase customer satisfaction, the performance of physicians ensures prefer-
ability. Therefore, measuring satisfaction and performance is crucial for man-
agement. However, measurement includes many challenges in connection with
many factors and uncertainty on these factors. Thus, neutrosophic multi-criteria
decision-making methods have been particularly chosen to overcome these
challenges in this study. The related literature on the subject shows that easy
access, cleaning and comfort, technological infrastructure, and equipment, total
service time, physicians’ communication skills, recognition of physicians,
treatment effectiveness, treatment rate, academic career are the main criteria. The
values of weights for these criteria and the physicians ranking based on the
values of those weights were found by using the Neutrosophic Analytical
Hierarchy Process (N-AHP). It has been validated that the ranking is fitting to
patient preferences in the Central Physicians Appointment System. As a result,
the recommended method can be effectively used in selecting physicians.
Keywords: Neutrosophic sets Physician selection Multi-criteria decision

making Neutrosophic AHP
1 Introduction
As in the world, the healthcare sector in Turkey is one of the areas where most works and
investments are made for the maintaining of human life, quality of life and well-being. To
ensure the protection and continuity of people’s health, the structure of all institutions and
organizations producing health-related goods and services is generally called the
“Healthcare Sector” [1]. The Covid-19 epidemic, the common agenda of the whole world,
has also unquestionably increased the importance of this sector to the top. Hospitals are
the most important unit of healthcare institutions. In addition to the services provided to
patients, hospitals have tasks such as providing medical education, training nurses and
medical personnel, doing medical studies, conducting research and development activi-
ties in the field of health, following rapidly developing technology and keeping hospitals
up to date [2]. Hospitals in Turkey are divided into 3 groups: University Hospitals, Health
Ministry Hospitals, and Private hospitals. Improving the quality and performance of the
320 A. Sarucan et al.
service provided in the health sector is possible by applying the correct diagnosis and
treatment methods by physicians to patients who come to receive services. While patients
evaluate the quality of the service they receive, they rate the hospital and physicians’s
performance together. Research on patient satisfaction shows that the medical proficiency
level of the healthcare institution (physicians quality, etc.) and its physical structure, the
behavior of the staff (communication, courtesy, etc.), patient waiting time, cleaning of the
facility affect the satisfaction level [3]. Also, hospital types and outpatient departments in
hospitals affect the number of patients who want to receive services. Besides, the number
of applications of many health institutions, the recognition of physicians and their
achievements in the field are also factors in the frequent preference of those healthcare
institutions [4]. Physicians preferences of patients within these and similar criteria are a
difficult problem. Naturally, multi-criteria decision-making (MCDM) methods are very
suitable to compare the performance of alternatives, as there are many alternatives and
criteria in the preference process. In fact, in a study in this field, the problem of hospital
selection of heart patients in hospitals operating in Ankara was discussed by AHP and
TOPSIS method which are well-known MCDM methods [5]. In this study, the evaluation
of physicians at the center of healthcare was examined by the AHP method, which is one
of the MCDM techniques based on neutrosophic sets. To the best of our knowledge, the
N-AHP is the first time applied for a physicians selection as a neutrosophic multi-criteria
decision-making method. The advantages of the applied method can be summarized as:
(1) The method represents the both uncertainty and indeterminacy; (2) This study can be
explanatory work for the researchers who intend to research on this topic. The rest of the
paper has been organized as follows: In Sect. 2 neutrosophic sets are briefly summarized.
In Sect. 3, the steps of N-AHP are briefly summarized. In Sect. 4, a healthcare application
is carried out and, the results are analyzed. The paper is completed with a section of
conclusion and, a suggestion for further studies.
2 Neutrosophic Sets
Some important definitions of neutrosophic sets are introduced as follow [6, 7].
Definition 1. Let S be a space of points and s 2 S. N neutrosophic set N in S is definite
by a truth-membership function TN(s), an indeterminacy-membership function
IN(s)
and a falsity-membership function FN(s). Also, TN(s):S ! 0; 1 þ ; IN(s):

S ! 0; 1 þ and FN(s):S ! 0; 1 þ . There is no restriction on the sum of
TN ðsÞ; IN ðsÞ and FN ðsÞ so 0 supTN ðsÞ þ supIN ðsÞ þ supFN ðsÞ 3 þ :
Definition 2. Let S be a universe of discourse. N single valued neutrosophic set N over
S is an object taking the form N = {hs; TN ðsÞ; IN ðsÞ; FN ðsÞ : s 2 Si}, where TN ðsÞ:
S ! ½0; 1; IN ðsÞ:S ! ½0; 1 and FN ðsÞ:S ! ½0; 1 with 0 TN ðsÞ þ IN ðsÞ þ FN ðsÞ
3 for all s 2 S: The intervals TN ðsÞ; IN ðsÞ and FN ðsÞ represent the truth-membership
degree, the indeterminacy membership degree and the falsity membership degree of
s to N, respectively. For convenience, a single valued neutrosophic number is repre-
sented by N = (t, y, z), where t, y, z 2 [0, 1] and t + y + z 3.
Physician Selection with a Neutrosophic Multi-criteria Decision Making Method 321
Definition 3. The single valued triangular neutrosophic number, ~t ¼ hðt1 ; t2 ; t3 Þ;

a~t ; h~t ; b~t i is a neutrosophic set on the real line set R, whose truth, indeterminacy and
falsity membership functions are as follows:
8 st1
>
> a~t ðt2 t1 Þ ðt1 s t2 Þ
>
< a
~t ðs ¼ t2 Þ
T~t ðsÞ ¼ t3 s
ð1Þ
>
> a ~tð Þ ð t2 s t3 Þ
>
: t3 t2
0 otherwise
8
> ðt2 s þ h~t ðst1 ÞÞ
>
> ðt2 t1 Þ ðt1 s t2 Þ
<
I~t ðsÞ ¼ h~t ðs ¼ t2 Þ ð2Þ
>
>
ðst2 þ h~t ðt3 sÞÞ
ð t s t3 Þ
>
: ðt3 t2 Þ 2
1 otherwise
8
> ðt2 s þ b~t ðst1 ÞÞ
>
> ðt2 t1 Þ ðt1 s t2 Þ
<
F~t ðsÞ ¼ b~t ðs ¼ t2 Þ ð3Þ
>
>
ðst2 þ bt ðt3 sÞÞ
ðt2 s t3 Þ
>
: ðt3 t2 Þ
1 otherwise
where a~t ; h~t ; b~t 2 [0, 1] and t1, t2, t3 2 R, t1 t2 t3.

Definition 4. Let ~t ¼ hðt1 ; t2 ; t3 Þ; a~t ; h~t ; b~t i and ~y ¼ hðy1 ; y2 ; y3 Þ; a~t ; h~t ; b~t i be two
single-valued triangular neutrosophic numbers and c 6¼ 0 be any real number. Then:
Addition of two triangular neutrosophic numbers
~t þ ~y ¼ hðt1 þ y1 ; t2 þ y2 ; t3 þ y3 Þ; a~t ^ a~y ; h~t _ h~y ; b~t _ b~y i
Subtraction of two triangular neutrosophic numbers
~t ~y ¼ hðt1 y3 ; t2 y2 ; t3 þ y1 Þ; a~t ^ a~y ; h~t _ h~y ; b~t _ b~y i
Inverse of a triangular neutrosophic number
1 1 1
~t1 ¼ h ; ; ; a~t ; h~t ; b~t i; where ð~t 6¼ 0Þ
t3 t2 t1
3 Neutrosophic AHP
Radwan et al. [8] proposed a neutrosophic AHP method and applied it to the selection
of the best learning management system. They stated that the traditional AHP method
considers the definite judgments of decision makers. While the neutrosophic set theory
makes the experts judgments more flexible. Bolturk and Kahraman [9] presented a new
AHP method with interval-valued neutrosophic sets and an interval-valued neutro-

sophic AHP based on cosine similarity measures. An application is given in energy
alternative selection by proposed method. Another paper which is related to neutro-
sophic AHP is published by Abdel-Basset et al. They developed a neutrosophic AHP-
Delphi group decision-making model based on trapezoidal neutrosophic numbers in
order to handle experts’ non-deterministic evaluation values [10]. Abdel-Basset et al.
[7] proposed a method with the group decision making based on N-AHP and solved a
real life problem structured by the experts from Zagazig Universityin Egypt. Ortega
et al. [11] focused on an environmental problem related to water quality from a river
basin by using N-AHP and TOPSIS linked to fuzzy cognitive map. Yücesan [12]
presented an application of Failure Mode Effect Analysis Integrated N-AHP. Abdel-
Bassed et al. [6] studied on the decision making process as an extension to N-AHP and
SWOT analysis for developing a strategy.
In this study, the steps of physicians’ selection study with N-AHP method are
shown below [8].
Step 1: Identifies the criteria, sub criteria and alternatives of the decision-making
problem. Then follows the constructing of the hierarchy of the considered problem.
Step 2: Defines the neutrosophic numbers that correspond to the 1–9 Saaty scale,
they are used to compare different criteria and physicians.
Step 3: Determines the neutrosophic preference via the pairwise comparison
between each criterion and sub criterion. Afterwards comes the comparing of the
alternatives under each criterion or sub criterion.
Step 4: Presents the calculation of the neutrosophic relative weight of each pref-
erence relation. The relative weight is calculated by the addition of each column in the
matrix, then each number in the matrix is divided on the sum of its column, with
averaging across the rows being the last step.
Step 5: Ranks the overall weights, and a choice is made of the best alternative, by
having the structure of the number of alternatives multiplied by the number of criteria.
4 Application
In this study, the evaluation of physicians and hospitals in the Neurology department of
2 hospitals operating under the Ministry of Health in Konya was carried out by a
decision-making group created. This group consists of academics, health managers and
people receiving services from the hospital. The purpose of choosing the neurology
department is that this problem has not been studied in the literature. All data of 2
public hospitals were last accessed by the Public Hospitals Statistics Report published
in 2018 by the General Directorate of Public Hospitals [4].
Step 1: There are a total of 5 physicians working in the Neurology department of
the designated hospitals. The names of these physicians and which hospital they served
were kept secret and encoded as P1, P2,…, P5. The information of physicians was
accessed from the Turkish Ministry of Health Central Physicians Appointment System
in January 2020 through the appointment portal. Besides, comments on forum sites
were used in the evaluation of physicians.
8 criteria were determined as a result of the literature screening. The criteria used in
the study are encoded as C1, C2,…, C8 as follows. The characteristics of both hospitals
and physicians were considered together when determining the criteria used during the
physician’s selection problem phase of the study. Hospital and physician concepts are
interconnected factors. All activities of physicians depend on hospital conditions and
facilities, regardless of their knowledge, skills and, experience. Since these factors
cannot be considered independent of each other, the 8 criteria selected were evaluated
together (C1: Ease of transportation, C2: Cleaning and comfort, C3: Technological
infrastructure and equipment, C4: Total service time, C5: physicians communication
skills, C6: physicians recognition, C7: Treatment effectiveness, treatment rate, C8:
Academic career). The first four of these criteria are of hospital characteristics and the
last four criteria belong to physicians’ characteristics [13].
The hierarchical structure of physicians selection is shown in Fig. 1.
Fig. 1. The hierarchy of a physician’s selection problem
Step 2: Defines the neutrosophic numbers that correspond to the 1–9 Saaty scale,
they are used to compare different criteria and physicians.
Structure the neutrosophic pair-wise comparison matrix of criteria and physicians,
through the linguistic terms which are shown in Table 1.
Table 1. Linguistic terms and the identical triangular neutrosophic numbers.

Saaty Scale Explanation Neutrosophic Triangular Scale
1 Equally influential ~1 ¼ hð1; 1; 1Þ; 0:50; 0:50; 0:50i
3 Slightly influential 3~ ¼ hð2; 3; 4Þ; 0:30; 0:75; 0:70i
5 Strongly influential ~5 ¼ hð4; 5; 6Þ; 0:80; 0:15; 0:20i
7 Very strongly influential ~7 ¼ hð6; 7; 8Þ; 0:90; 0:10; 0:10i
9 Absolutely influential ~9 ¼ hð9; 9; 9Þ; 1:00; 0:00; 0:00i
2 Intermediate values ~2 ¼ hð1; 2; 3Þ; 0:40; 0:65; 0:60i
4 4~ ¼ hð3; 4; 5Þ; 0:60; 0:35; 0:40i
6 6~ ¼ hð5; 6; 7Þ; 0:70; 0:25; 0:30i
8 e
8 ¼ hð7; 8; 9Þ; 0:85; 0:10; 0:15i
Step 3: Determines the neutrosophic preference via the pairwise comparison

between each criterion and sub criterion. Afterwards comes the comparing of the
alternatives under each criterion or sub criterion.
The values in Table 2 pertain to the decision maker group. The pair-wise com-
parison matrix of criteria is presented in Table 2. In table shows the linguistic scale and
their corresponding neutrosophic numbers that will be structure the neutrosophic pair-
wise comparison matrix of criteria and physicians, through the linguistic terms which
are shown in Table 1.
Table 2. The neutrosophic comparison matrix of criteria.

Criteria C1 C2 C3 C4 C5 C6 C7 C8
C1 ~1 ~2 5g g
1 61 4g g
1 71 1 ~
8g 3
C2 2g1 ~1 4g1 g
51 3g
1 g
61 7g
1 e
2
C3 ~5 ~4 ~1 3g1 ~ 2 4g
1 5g
1 ~
5
C4 ~6 ~5 ~3 1~ ~4 3g
1 4g
1 ~
6
C5 ~4 ~3 2g1 4 g1 1 ~ 5g1 6g1 ~
4
C6 ~7 ~6 ~4 ~3 ~5 ~1 3g1 ~
7
C7 ~8 ~7
~5 ~4 ~6 ~3 ~
1 ~
8
C8 1 g
3g 21 5g 1 6g1 4g g
1 7 g
1 81 ~
1
Step 4: Calculate the weight of the criteria and physicians.

The neutrosophic pair-wise comparison matrix, by transforming it to a deterministic
matrix using the following equations [6].
Let teij ¼ hðt1 ; y1 ; z1 Þ; a~t ; h~t ; b~t i be a single triangular neutrosophic number; then,
1
C ~tij ¼ ½t1 þ y1 þ z1 ð2 þ a~t h~t b~t Þ ð4Þ
8
and
1
N tij ¼ ½t1 þ y1 þ z1 ð2 þ a~t h~t b~t Þ ð5Þ
8
which are the score and accuracy degrees of ~tij respectively.
To get the score and the accuracy degree of ~tij , we use the following equations:
1
C ~tij ¼ ð6Þ
C ~tij
1
N ~tij ¼ ð7Þ
N ~tij
The crisp matrix is presented in Table 3 (by using Eq. (4)).
Table 3. The crisp comparison matrix of criteria.

Criteria C1 C2 C3 C4 C5 C6 C7 C8
C1 1 1 1/5 1/5 1/3 1/7 1/8 1
C2 1 1 1/3 1/5 1 1/5 1/7 1
C3 5 3 1 1 1 1/3 1/5 5
C4 5 5 1 1 3 1 1/3 5
C5 3 1 1 1/3 1 1/5 1/5 3
C6 7 5 3 1 5 1 1 7
C7 8 7 5 3 5 1 1 8
C8 1 1 1/5 1/5 1/3 1/7 1/3 1
The crisp matrix must be normalized as follows when it is created. (1) Calculate the
sum of each column of the crisp matrix. (2) Divide each matrix element into this total
value. (3) Calculate the average of the row elements of the normalized matrix.
The normalized comparison matrix of criteria and calculated the weight of the
criteria (Wcriteria) is presented in Table 4.
Table 4. The result matrix of criteria.

Criteria C1 C2 C3 C4 C5 C6 C7 C8 Wcriterion
C1 0.03 0.04 0.02 0.03 0.02 0.04 0.04 0.03 0.03
C2 0.03 0.04 0.03 0.03 0.06 0.05 0.05 0.03 0.04
C3 0.16 0.13 0.09 0.14 0.06 0.08 0.06 0.16 0.11
C4 0.16 0.21 0.09 0.14 0.18 0.25 0.11 0.16 0.16
C5 0.10 0.04 0.09 0.05 0.06 0.05 0.06 0.10 0.07
C6 0.23 0.21 0.26 0.14 0.30 0.25 0.32 0.23 0.24
C7 0.26 0.29 0.43 0.43 0.30 0.25 0.32 0.26 0.32
C8 0.03 0.04 0.02 0.03 0.02 0.04 0.04 0.03 0.03
Between step 2 and 4 are repeated and neutrosophic evaluation of physicians is

made for each criterion. Then next step is taken.
Step 5: Determine the total priority of each physicians (alternative) and the final
ranking.
The relative scores for each physician (alternative) is calculated as follows:
WP1 = 0:19 0:03 + 0:56 0:04 + 0:34 0:11 + 0:11 0:16 + 0:07 0:07 þ
0:09 0:24 + 0:15 0:32 + 0:23 0:03 ¼ 0:17
WP2 ¼ 0:15 0:03 + 0:16 0:04 + 0:11 0:11 + 0:15 0:16 + 0:44 0:07 +
0:06 0:24 + 0:14 0:32 + 0:31 0:03 = 0:15
WP3 ¼ 0:33 0:03 + 0:08 0:04 + 0:05 0:11 + 0:23 0:16 + 0:16 0:07 þ
0:41 0:24 + 0:19 0:32 + 0:15 0:03 = 0:23
WP4 ¼ 0:19 0:03 + 0:09 0:04 + 0:43 0:11 + 0:19 0:16 + 0:04 0:07 +
0:27 0:24 + 0:19 0:32 + 0:28 0:03 = 0:22
WP5 ¼ 0:14 0:03 + 0:10 0:04 + 0:08 0:11 + 0:32 0:16 + 0:29 0:07 +
0:16 0:24 + 0:33 0:32 + 0:03 0:03 = 0:23
Finally, the N-AHP ranking of physicians selection are shown in Table 5.
Table 5. The physicians selection ranking

Physicians Priority Ranking
P1 0.17 3
P2 0.15 4
P3 0.23 1
P4 0.22 2
P5 0.23 1
5 Conclusion
To look at the results of the study, the first place has been shared by the P3 and the P5.
The second place is the P4. It has been a remarkable result that these physicians served
in the same hospital. We have taken the criteria C6 and C7 with the highest criteria
weights to this result. P3 was in the first place according to the Criterion C6. According
to the criterion of C7, the P5 was the first. Since the total weight of these two criteria in
decision-making is 56%, these criteria have also determined the physician’s ranking. In
the physician selection problem, the weight of the hospital criteria was calculated by
34%, and the physician’s criteria were calculated by 66%.
This research has been limited to public hospitals. For this reason, it will be useful
to compare the results by applying them to the neurology departments of private
hospitals. Also, for further study, we suggest the solution of the physician selection
problem be compared with the solution of the interval-valued spherical fuzzy AHP.
References
1. Temür, Y., Bakırcı, F.: Türkiye’de sağlık kurumlarının performans analizi bir VZA
uygulaması. Sosyal Bilimler Dergisi. 10(3), 261–282 (2008)
2. Karahan, A., Özgür, E.: Hastanelerde Performans Yönetim Sistemi ve Veri Zarflama Analizi.
Nobel Yayın Dağıtım, Ankara (2011)
3. Kılıç, T., Topuz, R.: Hastalarla iletişimin, hasta memnuniyetine etkisi: özel, devlet ve
üniversite hastanesi karşılaştırması. Sağlıkta Performans Kalite Dergisi. 2015(9), 81–82
(2015)
4. Başara, B.B., Soytutan, Ç.İ.: T.C. Sağlık Bakanlığı Sağlık İstatistikleri Yıllığı 2017. Yayın
No: 1106 (2018)
5. Taş, C., Bedir, N., Eren, T., Alağaş, H.M., Çetin, S.: Policlinic evaluation with integrating
AHP-TOPSIS methods: an application in Ankara. Sağlık Yönetimi Dergisi 2(1), 1–17
(2018)
6. Abdel-Basset, M., Mohamed, M., Smarandache, F.: An extension of neutrosophic AHP–
SWOT analysis for strategic planning and decision-making. Symmetry 10(116), 1–18 (2018)
7. Abdel-Basset, M., Mohamed, M., Zhou, Y., Hezam, I.: Multi-criteria group decision making
based on neutrosophic analytic hierarchy process. J. Intell. Fuzzy Syst. 33, 4055–4066
(2017)
8. Radwan, N.M., Senousy, M.B., Riad, A.E.D.M.: Neutrosophic AHP multi criteria decision
making method applied on the selection of learning management system. Int. J. Adv.
Comput. Technol. (IJACT) 8(5), 95–105 (2016)
9. Bolturk, E., Kahraman, C.: A novel interval-valued neutrosophic AHP with cosine similarity
measure. Soft. Comput. 22(15), 4941–4958 (2018)
10. Abdel-Basset, M., Mohamed, M., Sangaiah, A.K.: Neutrosophic AHP-Delphi group
decision making model based on trapezoidal neutrosophic numbers. J Ambient Intell.
Hum. Comput. 9, 1–17 (2017)
11. Ortega, R.G., Vázquez, M.L., Figueiredo, J.A.S., Guijarro-Rodríguez, A.: Sinos river basin
social-environmental prospective assessment of water quality management using fuzzy
cognitive maps and neutrosophic AHP-TOPSIS. Neutrosophic Sets Syst. 23, 160–171
(2018)
12. Yücesan, M.: FMEA analysis in mechanical installation project based on best worst and
neutrosophic AHP integrated model. Electron. J. Soc. Sci. 19(73), 363–382 (2020)
13. Doğan, E.: Hizmet Performansının Çok Ölçütlü Karar Verme Yöntemleri ile Değer-
lendirilmesi: Sağlık Sektöründe Bir Uygulama. Yüksek Lisans Tezi, Konya Teknik
Üniversitesi Lisansüstü Eğitim Enstitüsü, Konya (2019)
Intuitionistic Fuzzy Sets
An Intuitionist Fuzzy Method for Discovering
Organizational Structures that Support Digital
Transformation
Zineb Besri1(&) and Azedine Boulmakoul2(&)

1
National School of Applied Sciences, Tetouan, Morocco
[email protected]
2
Department of Computer Science, FSTM, Hassan II University of Casablanca,
Casablanca, Morocco
[email protected]
Abstract. With recent developments in information technology and the advent

of new communication practices, organizations are currently facing enormous
changes in the design of their activities and their development strategy.
Unfortunately, these changes are not so easy to model and to insert into digital
transformation processes. The digital transformation (DT) is triggering changes
in operations and organizational processes. The results of this transformation are
not often predictable. Sometimes digital transformation leads to entirely new
solutions and has a deep impact on the organization. Understanding how
organizations operate when undergoing digital transformation is essential, as it
is the right way to overcome organizational obstacles and find the optimal path
for future-oriented development. By focusing on the processes and on the
organizational units resulting from an analytical approach, we show that the
development of digital strategies reveals new processes and new organizational
structures resulting from emerging strategies. More specifically, we find that a
DT is constantly evolving. This paper deals with a new approach to drive the
transformation of the organization regarding to digital transformation. Indeed,
we propose a fuzzy intuitionist approach to discover organizational structures
that support digital transformation. The theory of intuitionist fuzzy sets handles
uncertain situations, by extension of the classical fuzzy theory, with the inte-
gration of the concept of hesitation.
Keywords: Digital transformation Fuzzy logic Organizational

reengineering Intuitionistic fuzzy sets
1 Introduction
Business organizations throughout the world are experiencing a high level of compe-
tition and uncertainty in all type of business. As a result, most of the organizations are
desperately feeling the immediate and drastic need of change in their structure, pro-
cesses, strategies, people etc. so, that they remain relevant in the highly compact
market.
332 Z. Besri and A. Boulmakoul
Modern organizations have brought about sensational progress for humanity in less
than two centuries. The blink of an eye in the overall timeline of our species. None of
the recent advances in human history would have been possible without organizations
as vehicles for human collaboration.
The lack of a hierarchical structure does not prevent the existence of a structure of
another order, which is no longer based on the organization of positions to be filled but
on the articulation of roles to be filled. Thus, each employee fulfills operational,
technical roles, but also management roles, which are distributed between the different
members of the team instead of being concentrated in the hands of one person.
The digital transformation (DT) is triggering changes in operations and processes in
organizations. The results of this transformation are not often predictable. Sometimes
digital transformation leads to entirely new solutions and has a profound impact on the
organization. Understanding how organizations operate when undergoing digital
transformation is essential, as it is the right way to overcome organizational obstacles
and find the optimal path for future-oriented development. By focusing on the pro-
cesses and on the organizational units resulting from an analytical approach, we show
that the development of digital strategies reveals new processes and new organizational
structures resulting from emerging strategies. More specifically, we find that a DT is
constantly evolving. This paper deals with a new approach to drive the transformation
of the organization regarding to digital transformation. Indeed, we propose a fuzzy
intuitionist approach to discover organizational structures that support digital trans-
formation. The theory of intuitionist fuzzy sets handles uncertain situations, by
extension of the classical fuzzy theory, with the integration of the concept of hesitation.
The remain of the paper is as follow. Related works about organizational reengi-
neering and digital transformation are given in Sect. 2. Section 3 deals with the pro-
posed approach based on fuzzy intuitionist theory to discover organizational structure
that support digital transformation. Finally, we conclude in Sect. 4 the synthesis of our
approach and our future works on perspectives.
2 Related Works
2.1 Organizational Reengineering Using Structural Analyses
Reengineering is an approach for the reconstruction of the management and organi-
zational management literature emerged in the early 90s. Any organization or company
is a social institution that is based on objective and is consistent active and coordinate
systems and is associated with the external environment. In the past, when relatively
stable environment, most organizations take advantage of opportunities for ongoing,
satisfy with gradual and slight changes. But over time, across the world, organizations
have found that the gradual changes are not the solution of their cur-rent problem and
sometimes it is necessary for the survival of organization make basic changes in the
organization. Today around the world, these revolutionary changes know as reengi-
neering. Reengineering (BPR) is a process in which the current functions of the
organization change its place with main processes of business, therefore organization
moves from function-oriented to process-oriented.
An Intuitionist Fuzzy Method for Discovering Organizational Structures 333
Organizational reengineering is a technique used to redesign business processes in

order to take advantage of organizational strengths, called core competencies. Using
this technique, an organization can align itself for the future and increase its market
share and profitability.
The scope of an organizational overhaul is to achieve a mature level of the
enterprise information system’s governance. Thus, to reach this level several methods
and technics are used to diagnose, analyze, then propose some organizational re-
structure actions. In anterior works, we propose two approaches of structural analyses.
The first one is based on algebraic topology and the use of topological metrics to
analyze the organizational structure of the company. The second concerns, particularly,
the use of social network analysis techniques. Illustrate with the case for a re-design
business processes in harbor container terminal [1]. We used what we called a Topo-
Scopy for the combinatorial approached using simplicial analysis for the aim of per-
forming enterprise organizational structure redesign through structural analysis and
simplicial complexes framework [2]. And the scope we called an SNA-Scopy for the
social network analysis technics [3].
2.2 Digital Transformation

The digital conversion is anything but simple, it must be treated with caution. Digital is
indeed the new standard. It affects different areas of society and the economy. It allows
cooperation between different actors. It must be sufficiently agile from a technological,
structural and cultural point of view to constantly adapt to the rapidly changing
commercial and technological environments.
Enabling an organization to successfully deliver on digital business objectives, the
following key digital imperatives should be considered.
“Digital transformation - it’s really about how do you go from an organizational
structure that’s based on performance or hierarchies to an area where it’s based on
collaboration and co-creation. Leading means helping the organization along that way
to become open to different approaches” [4].
While it proved difficult to untangle the multifaceted issues related to digital
transformation, 9 major areas leaders are defined and must pay attention to, and which
are outlined as follow:
1. The legacy challenge;
2. The resources allocations;
3. The agility challenge;
4. The ambidexterity challenge;
5. The start-up challenge;
6. The connectivity challenge;
7. The governance challenge;
8. The functional identity challenge;
9. The people challenge.
The nine challenges are highly interdependent and feed on each other. While it is
impossible to isolate and untie them, the purpose of this paper is to discover organi-
zational structure that support digital transformation [7–9]. Then structure the arena
and provide some insights on key elements of the complex transformational journey
towards digital maturity.
The next section proposes a new approach, an intuitionist fuzzy method, for dis-
covering organizational structures that support digital transformation.
3 Organizational Structures Discovery by an Intuitionist

Fuzzy Method
Below, we present necessary reminders to understand our approach. In the same section
we develop the process of calculating collaboration intuitionistic matrices. Intuitionistic
fuzzy centrality degree will be evaluated on those collaboration matrices, to generate
communities corresponding to the emerging organizational structures.
3.1 Preliminaries
• pA : indicatrices
• k: instant, iteration
• t: time
• AðkÞ : Activities at the iteration k
• OðkÞ : Organizational Unit at the iteration k
• Xðk þ 1Þ ¼ AðkÞ [ Aðk þ 1Þ : Union of activities
• xi : activity of Xðk þ 1Þ
• Dðk þ 1Þ ¼ OðkÞ [ Oðk þ 1Þ : Union of organizational units
• KðkÞ ¼ AðkÞ OðkÞ : Collaboration’s relationship at the iteration k (Activity x
Organizational Unit adjacency matrix)
• Kðk þ 1Þ ¼ Xðk þ 1Þ Dðk þ 1Þ Collaboration’s relationship at the iteration k + 1
• l: degree of membership P
jCðkÞ ðxi Þ \ CðakÞ ðxj Þj PKk \ Kðk þ 1Þ ðxi ;xÞPKk \ Kðk þ 1Þ ðxj ;xÞ
• lðk þ 1Þ xi ; xj ¼ aðkÞ ¼ P x2AðkÞ
jCa ðxi Þ [ CðakÞ ðxj Þj ð Þ max ½PKk \ Kðk þ 1Þ ðxi ;xÞ;PKk \ Kðk þ 1Þ ðxj ;xÞ
n x2A k
x2A=KðkÞ ðxi ;xÞ6¼0

• Ca ðxi Þ ¼ andKðk þ 1Þ ðx ;xÞ6¼0
k
i
• m: degree of non-membership

Pðbk þ 1Þ ðxi Þ \ Pðbk þ 1Þ ðxj Þ
• m ð k þ 1Þ
xi ; xj ¼ ðkÞ þ 1
Pb
ð k þ 1Þ
ð x i Þ [ Pb ðxj Þ
. ..
ðxi Þ ¼ Cðk þ 1Þ ðxi Þ . . Kðk þ 1Þ Kðk þ 1Þ \ KðkÞ
ð k þ 1Þ
• Pb with . denotes
restrictions

• We have 0 lðk þ 1Þ xi ; xj þ mðk þ 1Þ xi ; xj 1

• Hesitation: Pij ¼ 1 lðk þ 1Þ xi ; xj þ mðk þ 1Þ xi ; xj
Definition 1. A fuzzy set A, on the universe of discourse of X can also be described a

set of ordered pairs as:
A ¼ fx; lA ð xÞjx 2 X g ð1Þ
In classical set theory, lA ð xÞ has only the tow membership values 0 (‘false’) and 1
(‘true’). Instead of considering only those tow truth values, fuzzy set can assign the
value of lA ð xÞ within the interval between values 0 (‘false’) and 1 (‘true’) for better
outcome [5].
Definition 2. Let A be an intuitionistic fuzzy set. A in X is defined by
A ¼ fx; lA ð xÞ; #A ð xÞjx 2 X g ð2Þ
Where lA ðxÞ 2 ½0; 1 and mA ð xÞ 2 ½0; 1 follows with the condition

0 lA ð xÞ þ mA ð xÞ 1. The lA ð xÞ is used to represent the degree of membership and
mA ð xÞ represents the degree of non-membership of the element x in the set A. for each
intuitionistic fuzzy set A in X, the intuitionistic fuzzy index is described as pA ð xÞ ¼
1 ðlA ð xÞ þ mA ð xÞÞ [6].
Definition 3. Let A and B be tow intuitionistic fuzzy set. The union, intersection and
complement are described as follow:
A [ B ¼ f x; maxðlA ð xÞ; lB ð xÞÞ; minðmA ð xÞ; mB ð xÞÞjx 2 X; y 2 Yg ð3Þ
A \ B ¼ f x; minðlA ð xÞ; lB ð xÞÞ; maxðmA ð xÞ; mB ð xÞÞjx 2 X; y 2 Yg ð4Þ
Ac ¼ fmA ð xÞ; lA ð xÞjx 2 Xg ð5Þ
3.2 Intuitionistic Fuzzy Degree Centrality

Social network can be represented by a directed or undirected weighted intuitionistic
fuzzy graph. Let Gdn ¼ ðV; E Þ be a directed intuitionistic fuzzy graph.
We can describe a directed intuitionistic fuzzy graph social network as an intu-

~ dn ¼ V; E
itionistic fuzzy relational structure: G ~ dn where V ¼ ft1 ; t2 ; . . .tn g denotes a
0 1
~e11 ~e1n
non-empty set if intuitionistic fuzzy nodes, and E ~ dn ¼ B
@ ..
. ..
.
.. C
. A denotes an
~en1 ~enm
undirected intuitionistic fuzzy relation on V.
The Intuitionistic Fuzzy In-degree Centrality
The sum of the lengths of the arcs that are adjacent to a social node ti is called the
intuitionistic fuzzy in-degree centrality of the node ti .
The intuitionistic in-degree centrality of node ti , d~I ðti Þ, is described as follows:

Xn
d~I ðti Þ ¼ ~e
j¼1;j6¼i ij
ð6Þ
The Intuitionistic Fuzzy Out-degree Centrality

The sum of the lengths of the arcs that are adjacent from a social node ti is called the
intuitionistic fuzzy out-degree centrality of the node ti .
The intuitionistic in-degree centrality of node ti , d~O ðti Þ, is described as follows:
Xn
d~O ðti Þ ¼ ~e
j¼1;j6¼i ij
ð7Þ
P
Here the symbol is an addition operation of intuitionistic fuzzy set and ~eij is an
intuitionistic fuzzy set associated with the arc (i, j).
The Intuitionistic Fuzzy Degree Centrality
The sum of intuitionistic in-degree centrality and intuitionistic out-degree centrality of
node ti is called intuitionistic degree centrality (IDC) of ti
d~ðti Þ ¼ d~I ðti Þ d~O ðti Þ ð8Þ
is an addition operation of intuitionistic fuzzy set.
3.3 The Process of Calculating Collaboration Intuitionistic Matrices

The process of calculating collaboration intuitionistic matrices is illustrated by the
Fig. 1. Intuitionistic fuzzy centrality degree will be evaluated on those collaboration
matrices, to generate communities corresponding to the emerging organizational
structures.
By this process we lead to, first, discover a possibility of a new organizational
structure after we explore the intuitionistic fuzzy collocational relationship matrices
defined by the relation between organizational unit (group of performers) and a set of
activities (the activities defined by the extracted processes to assess then to develop).
Those matrices will be the input of the proposed intuitionistic fuzzy analysis
approach. Once we result the k + 1 iteration previously defined. We can compute some
metrics such as degree centrality as presented in previous part. Besides, and the
important one is to compute the degree of hesitation to be or not to be part of the new
organizational structure.
Fig. 1. Intuitionistic fuzzy method for discovering organizational structure.
The aim is to help top management to approve then to generate the emerging
organizational structures that should be align and support their digital transformation
strategy.
4 Conclusion
Business process reengineering is a technique which is used to induce the radical

change in the organization set up. The goal of the technique is to achieve an optimum
level of effectiveness and efficiency. The paper aim to study a new approach to drive
the transformation of the organization regarding to digital transformation. Indeed, we
propose a fuzzy intuitionist approach to discover organizational structures that support
digital transformation. The theory of intuitionist fuzzy sets handles uncertain situations,
by extension of the classical fuzzy theory, with the integration of the concept of
hesitation. The paper presents such a process to conduct this method in order to
discover this new emergent organizational structure that support the digital transfor-
mation of the company. This approach will be more develop and a case study will be
handled in future works.
References
1. Besri, Z., Boulmakoul, A.: Framework for organizational structure re-design by assessing
logistics’ business processes in harbor container terminals. Transp. Res. Procedia 22, 164–173
(2017)
2. Boulmakoul, A., Besri, Z.: Performing enterprise organizational structure redesign through
structural analysis and simplicial complexes framework. Open Oper. Res. J. 7, 11–24 (2013)
3. Boulmakoul, A., Besri, Z.: Scoping enterprise organizational structure through topology
foundation and social network analysis. In: Innovation and New Trends in Information
Systems, 3rd edn, pp. 3–17. Ryad Mogador Tangier (2013)
4. Holt, T.: CEO Siemens Power Generation Services. Roland Deiser. Digital transformation
challenges in Large and complex organizations. Deiser (2018)
6. Atanassov, K.T.: Intuitionistic Fuzzy Sets. Springer, Berlin (1999)
7. Deiser, R.: Digital Transformation Challenges in Large and Complex Organizations. CFFO
Press (2018)
8. Laloux, F.: Reinventing organizations. Nelson Parker (2016). ISBN 978-2-9601335-6-1
9. Sommerfeld, B., Moise-Cheung, R.: The digitally-fit organization. Inside Magazine, issue 7,
Part 01. From Digital Perspective (2016)
Modeling Humanoid Robots Mimics
Using Intuitionistic Fuzzy Sets
Cengiz Kahraman1(&), Sezi Cevik Onar1, Basar Oztaysi1,

and Irem Otay2
1
Istanbul Technical University, 34367 Macka, Besiktas, Istanbul, Turkey
[email protected]
2
Istanbul Bilgi University, Istanbul, Turkey
Abstract. Intuitionistic fuzzy sets (IFS) have been often used in modeling the
problems under vagueness and impreciseness in order to better define the
problems together with the hesitancy of decision makers. IFS have been often
employed in modeling decision making problems in the literature. Human
decision making process can be used in humanoid robots by imitating the human
decisions and behaviors.
Keywords: Fuzzy sets Extensions Intuitionistic fuzzy sets Pythagorean

fuzzy sets Picture fuzzy sets Q-rung orthopair fuzzy sets Spherical fuzzy
sets
1 Introduction
Humanoid robots will be very usable in human society to improve the quality of life in
the future. Humanoid robots are being developed to perform human tasks like personal
assistance. They can assist the sick and elderly, and dirty or dangerous jobs. Humanoid
robots imitate human mechanisms of decision making and information processing. In
the literature, humanoid robots have been studied for biped walk control (Tawara et al.
2001), cooperative object transformation (Hawley and Suleiman 2019), gender repre-
sentation (Carpenter et al. 2009), emotion and sociable humanoid robots (Breazel
2003), healt assist (Robins et al. 2005), customer acceptance (Belanche et al. 2020),
speaker recognation (Ding and Shi 2017). Human behaviors are mostly emotional and
based on the degrees of feelings such as being strongly sad or being slightly happy.
Fuzzy sets can be very useful in modeling humanoid robots’ behaviors. In this paper,
we present an example of intuitionistic fuzzy modeling of humanoid robots.
Fuzzy sets theory has been extended to many new types of ordinary fuzzy sets after its
first introduction by Zadeh (1965). New extensions try to represent the thoughts of
humans more correctly. It is clear that classical logic is not sufficient to handle so complex
events. Intuitionistic fuzzy sets (IFS) have been introduced by Atanassov (1986), which
differ from ordinary fuzzy sets by an independent membership degree that is not neces-
sarily the complement of membership degree. The secondary extensions of IFS, which are
Pythagorean fuzzy sets and q-rung orthopair fuzzy sets aim at providing a larger domain
to experts than it is in IFS.
340 C. Kahraman et al.
Figure 1 illustrates a real humanoid robot on the left side, whose name is Sophia
developed by Hong Kong based company Hanson Robotics in 2016. On the right side,
a hypothetical humanoid robot is given.
(a) A real HR: Sophia (b) Hypothetical HR
Fig. 1. Humanoid robots of the future
In this paper, we try to model face mimics using intuitionistic fuzzy sets and fuzzy
functions in order to express the emotional feelings of HRs. In the literature, there are
very few works on fuzzy or intelligent modeling of humanoid robots. Kats and
Vukobratovi (2003) surveyed intelligent control techniques for humanoid robots.
Wong et al. (2008) studied fuzzy control of humanoid robots. Fang et al. (2019) studied
fuzzy brain emotional learning models.
The rest of the paper is organized as follows. Section 2 summarizes the basic
definitions of these sets together with graphical illustrations. These graphs provide a
comparative analysis among the extensions. Section 3 concludes the paper with future
directions and suggestions.
2 Intuitionistic Fuzzy Sets
Intuitionistic fuzzy set: Let X 6¼ ; be a given set. An intuitionistic fuzzy set in X is an

object A~ given by

~ ¼ \x; l ~ ð xÞ; v ~ ð xÞ [ ; xX ;
A ð1Þ
A A
where lA~ : X ! ½0; 1 and vA~ : X ! ½0; 1 satisfy the condition 0 lA~ ð xÞ þ vA~ ð xÞ 1;
for every x X. Hesitancy of an expert is calculated as pA~ ¼ 1 lA~ ð xÞ vA~ ð xÞ.
Modeling Humanoid Robots Mimics Using Intuitionistic Fuzzy Sets 341
Let D ½0; 1 be the set of all closed subintervals of the interval and X be a universe
of discourse. An interval-valued intuitionistic fuzzy set (IVIFS) in A ~ over X is an object
having the form (Otay et al. 2017; Kahraman et al. 2018).

~ ¼ \x; l ~ ð xÞ; v ~ ð xÞ [ jx X ;
A ð2Þ
A A
where lA~ ! D ½0; 1; vA~ ð xÞ ! D ½0; 1 with the condition 0 suplA~ ð xÞ þ
supvA~ ð xÞ 1; 8x 2 X:
The intervals lA~ ð xÞ and vA~ ð xÞ denote the membership function and the non-
membership function of the element x to the set A, ~ respectively. Thus for each x 2 X,
lA~ ð xÞ and vA~ ð xÞ are closed intervals and their starting and ending points are denoted by
l þ þ ~
A~ ð xÞ; lA
~ ð xÞ; tA ~ ð xÞ and tA
~ ð xÞ, respectively. IVIFS A is then denoted by
n h i h i o
~ ¼ \x; l~ ð xÞ; l ~þ ð xÞ ; t
A þ
~ ð xÞ; tA
~ ð xÞ [ jx 2 X ; ð3Þ
A A A
where 0 lA~þ ð xÞ þ tA~þ ð xÞ 1; l

~ ð xÞ 0; tA
A ~ ð xÞ 0:
For each element x, the unknown degree (hesitancy degree) of an IVIFS of x 2 X in
~ defined as follows;
A
h i
pA~ ð xÞ ¼ 1 lA~ ð xÞ vA~ ð xÞ ¼ 1 lA~þ ð xÞ tA~þ ð xÞ ; 1 l
~ ð xÞ tA
A ~ ð xÞ : ð4Þ
~ ¼ ð½a; b; ½c; d Þ can be defuzzified by using Eq. (5).

An IVIF number A
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
a þ b þ ð 1 cÞ þ ð 1 d Þ þ a b ð1 cÞ ð1 d Þ
~ ¼
Def A ð5Þ
4
In Eq. (5), the terms (1 − c) and (1 − d) convert non-membership degrees to

pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
membership degrees while the term ð1 cÞ ð1 d Þ decreases the defuzzified
value.
3 Face Mimics of HRs Using IFS
A HR must select the appropriate expression of these mimics when it meets an event
which necessitates it to change its face expression. Scientists discovered that humans
have 21 different facial expressions: Happy, Sad, Fearful Angry, Surprised, Disgusted,
Appalled, Happily surprised, Happily disgusted, Sadly fearful, Sadly angry, Sadly sur-
prised, Sadly disgusted, Fearfully angry, Fearfully surprised, Fearfully disgusted,
Angrily surprised, Angrily disgusted, Disgustedly surprised, Hatred, and Awed. Figure 2
presents some of these expressions on a woman (https://www.dailymail.co.uk/science
tech/article-2593459/Which-face-happily-disgusted-Scientists-discover-humans-21-
different-facial-expressions.html).
Using classical functions, the face mimics are tried to model as illustrated in Fig. 3
(https://www.desmos.com/calculator/xwa7eu2tc9).
Fig. 2. Different mimics of a woman
Fig. 3. Smile using mathematical functions
The crisp function fstrongsmile ¼ 0:5x2 3; 2\x\2 can be modeled by using IFS as
follows. Assume that the humanoid robot’s smile upon an event should be between strong
smile and medium strong smile. In this case, the coefficient k = 0.5 in the strong smile
function may take a value between 0.36 and 0.5. In this case, Fig. 3 turns to be Fig. 4.
k=0.36 k=0.40 k=0.45 k=0.5
Fig. 4. Smiles between medium strong smile and strong smile

Let us assume the membership function be defined between 0.40 and 0.46 and the
non-membership function be defined between 0.36 and 0.50. This can be represented
by an intuitionistic fuzzy number as given in Fig. 5:
Fig. 5. Intuitionistic fuzzy number for the coefficient k
The function in Fig. 5 can be represented by the following a-cuts:
lL ¼ 0:40 þ 0:03a; lR ¼ 0:46 0:03a; vL ¼ 0:43 0:07a; vR ¼ 0:43 þ 0:07a
The fuzzy function between strong smile and medium strong smile (ss&mss) then
becomes
fsm&mss ¼ ð½0:40 þ 0:03a; 0:46 0:03a; ½0:43 0:07a; 0:43 þ 0:07aÞx2

3; 2\x\2
Using Eq. (5), we obtain

pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi!
2 þ ð0:40 þ 0:03aÞ ð0:46 0:03aÞ ð0:57 þ 0:07aÞ ð0:57 0:07aÞ
fsm&mss ¼ x2 3;
4
where 2\x\2.
Table 1 gives us the defuzzified values of the IVIF number depending on the
various a values. Table 2 presents the results of the f sm&mss function with respect to the
defuzzified values given in Table 1.
Table 1. Defuzzified values of IVIF number

a Deffuzzified value a Deffuzzified value
0 0.4035 0.55 0.404005
0.05 0.403525 0.6 0.404076
0.1 0.403553 0.65 0.404152
0.15 0.403587 0.7 0.404232
0.2 0.403624 0.75 0.404317
0.25 0.403666 0.8 0.404405
0.3 0.403711 0.85 0.404498
0.35 0.403762 0.9 0.404596
0.4 0.403816 0.95 0.404698
0.45 0.403875 1 0.404804
0.5 0.403938
Table 2. Results of fsm&mss function with respect to the defuzzified values

x a
0 0.05 0.1 0.15 0.2 1
−1.98 −1.41812 −1.41802 −1.41791 −1.41778 −1.41763 −1.41301
−1.96 −1.44991 −1.44982 −1.44971 −1.44958 −1.44944 −1.44491
−1.94 −1.48139 −1.48129 −1.48119 −1.48106 −1.48092 −1.47648
−1.92 −1.51254 −1.51245 −1.51234 −1.51222 −1.51208 −1.50773
−1.9 −1.54337 −1.54328 −1.54317 −1.54305 −1.54292 −1.53866
−1.88 −1.57387 −1.57378 −1.57368 −1.57356 −1.57343 −1.56926

+1.98 −1.41812 −1.41802 −1.41791 −1.41778 −1.41763
Using these values, the graphs in Fig. 6 is obtained for a = 0.30 and a = 0.35.
Fig. 6. fsm&mss functions for a = 0.30 and a = 0.35
Figure 7 illustrates the smile functions depending on the different levels of a. For
larger membership degrees in IVIF numbers, the smile gets stronger.
Fig. 7. Smile functions depending on the different levels of a
4 Conclusions
New extensions of type-1 fuzzy sets try to define a more detailed membership function
with additional new parameters.. These parameters can be successfully used in mod-
eling the human-like behaviors of humanoid robots. For instance, the high degree of
hesitancy parameter can lead a humanoid robot to require additional information on the
problem. Equal degrees of membership, non-membership, and hesitancy parameters
may cause the robot to select any alternative with equal possibilities while a higher
degree of membership or non-membership than hesitancy degree can do it with a larger
possibility. The fuzzy extensions such as Pythagorean fuzzy sets or spherical fuzzy sets
can present larger domains for humanoid robots to judge about the decision making
problems they face. Every mimic and movement of humanoid robots can be directed by
these membership functions much easier than classical control mechanisms. For further
research, we suggest these new extensions to be used in modeling the humanoid robots’
emotional movements.
References
Atanassov, K.T.: Intuitionistic fuzzy sets. Fuzzy Sets Syst. 20(1), 87–96 (1986)
Belanche, D., Casaló, L.V., Flavián, C.: Customers acceptance of humanoid robots in services:
the moderating role of risk aversion. In: Marketing and Smart Technologies, pp. 449–458.
Springer, Cham (2020)
Breazeal, C.: Emotion and sociable humanoid robots. Int. J. Hum. Comput. Stud. 59(1–2), 119–
155 (2003)
Carpenter, J., Davis, J.M., Erwin-Stewart, N., Lee, T.R., Bransford, J.D., Vye, N.: Gender
representation and humanoid robots designed for domestic use. Int. J. Soc. Robot. 1(3), 261
(2009)
Ding, J., Shi, J.Y.: Kinect microphone array-based speech and speaker recognition for the
exhibition control of humanoid robots. Comput. Electr. Eng. 62, 719–729 (2017)
Fang, W., Chao, F., Lin, C.M., Yang, L., Shang, C., Zhou, C.: An improved fuzzy brain
emotional learning model network controller for humanoid robots. Front. Neurorobotics 13, 2
(2019)
Hawley, L., Suleiman, W.: Control framework for cooperative object transportation by two
humanoid robots. Robot. Auton. Syst. 115, 1–16 (2019)
Kahraman, C., Cebi, S., Onar, S.C., Oztaysi, B.: A novel trapezoidal intuitionistic fuzzy
information axiom approach: an application to multicriteria landfill site selection. Eng. Appl.
Artif. Intell. 67, 157–172 (2018)
Katic, D., Vukobratovic, M.: Survey of intelligent control techniques for humanoid robots.
J. Intell. Rob. Syst. 37(2), 117–141 (2003)
Otay, I., Oztaysi, B., Cevik Onar, S., Kahraman, C.: Multi-expert performance evaluation of
healthcare institutions using an integrated intuitionistic fuzzy AHP&DEA methodology.
Knowl.-Based Syst. 133, 90–106 (2017)
Robins, B., Dautenhahn, K., Te Boekhorst, R., Billard, A.: Robotic assistants in therapy and
education of children with autism: can a small humanoid robot help encourage social
interaction skills? Univ. Access Inf. Soc. 4(2), 105–120 (2005)
Tawara, T., Okumura, Y., Shimizu, M., Tomiyama, K.: Design and construction of a series of
compact humanoid robots and development of biped walk control strategies. Robot. Auton.
Syst. 37(2–3), 81–100 (2001)
Wong, C.C., Cheng, C.T., Huang, K.H., Yang, Y.T.: Fuzzy control of humanoid robot for
obstacle avoidance. Int. J. Fuzzy Syst. 10(1), 1–10 (2008)
Zadeh, L.A.: Fuzzy sets. Inf. Control 8(3), 338–353 (1965)
Understanding the Blockchain Technology
Adoption from Procurement Professionals’
Perspective - An Analysis of the Technology
Acceptance Model Using Intuitionistic
Fuzzy Cognitive Maps
Ayça Maden1(&) and Emre Alptekin2

1
Management Engineering Department, Beykent University,
34398 Maslak, İstanbul, Turkey
[email protected]
2
Industrial Engineering Department, Galatasaray University,
34357 Ortaköy, Istanbul, Turkey
[email protected]
Abstract. Nowadays, the trend towards new technologies has promoted fierce
competition in supply chains. Blockchain technology attracts widespread
attention in supply chain processes via its potential benefits. Efficiently managed
supply chain processes provide operational and organizational advantages, and
procurement is one of the critical processes to gain such advantages. With its
multi-participant nature, procurement process deserves particular attention in
terms of its potential to be transformed by blockchain technology. When eval-
uating a new technology, variables affecting behavior to use technology should
be analyzed carefully. In literature, additional research is needed to obtain a
broader understanding of blockchain technology acceptance. This study aims to
analyze procurement professionals’ adoption of the blockchain technology with
the Technology Acceptance Model (TAM) and Intuitionistic Fuzzy Cognitive
Maps (IFCM). The IFCM was used as it has the capacity to model the pro-
fessionals’ hesitation, and it also copes with incomplete or even conflicting
information. Our results indicate that the influences between variables are in line
with most of the other studies. However, the findings further strengthens that the
procurement professionals give Intention, Job Relevance and Output Quality
more importance.
Keywords: Blockchain technology Technology Acceptance Model

Intuitionistic Fuzzy Cognitive Maps
1 Introduction
Fast evolution of information and communication technologies has affected the adoption
and utilization of these technologies [1]. Blockchain technology attracts great attention in
both academia and industry, and it is predicted to be as revolutionary as the Internet [2].
Although there is not enough research on blockchain technology implementation in
348 A. Maden and E. Alptekin
logistics and supply chains, this does not reject the fact that blockchain will reshape the
relations in the logistics and supply chain management [1]. The evolutionary benefits of
the blockchain technology are not well understood, which may cause reduced adoption
rates of this technology [3]. Although researches on blockchain have acquired a relative
speed, there is not enough studies on behavior of the blockchain technology adoption of
individuals [1]. According to the literature, blockchain adoption behavior and the driving
forces of blockchain adoption are almost unexplored [1]. To achieve competitive
advantages, company managers need to figure out potential outcomes of blockchain
practices [3].
Technology Acceptance Model (TAM) is a significant tool for researching the
social mechanisms of technology adoption which has attracted great attention in the
literature [2]. Although experimental assistance for the model has changed relying on
specific situations, it continues to be a popular and beneficial framework for the
research of technology acceptance factors [2]. One of the core strengths of TAM is its
predictive ability [3]. When performing analysis via statistical procedures, it should be
considered that external variables of TAM are not dependent, and they do not have
important relations [4]. When the variables do not satisfy independency, the traditional
and modified TAM will not accurately investigate causal relations which may yield
incorrect results [4]. Beside this, some technology frameworks have high complications
as in the ERP [4]. If people do not fully understand these, and they have no experience
of using them, collected data can have inference bias [4]. The Cognitive Mapping does
not accept the independent assumption as the statistical procedures do. Unlike tradi-
tional Multi Criteria Decision Making procedures that necessitate full data, the intuitive
cognitive map can cope with the deficiency of information and uncertainty of decision
makers [5]. The main purpose of this study to better understand the behavior of the
procurement professionals in the context of blockchain adoption by providing answers
to the question: What are the main drivers of blockchain adoption in your company?
The results will help experts to figure out a variety of factors that affect blockchain
adoption, and will help to prepare a proper plan of application.
This paper is organised as follows: Sect. 2 describes the blockchain concept,
Sect. 3 describes the literature survey, Sect. 4 presents the methodology, Sect. 5
describes the case study, and Sect. 6 concludes.
2 Blockchain Concept
In 2008, a novel article was presented, Bitcoin: A Peer-to-Peer Electronic Cash System,
which put forward ‘chain of blocks’ using the nickname Satoshi Nakamoto [6].
Blockchain is a timestamped, organized, and immutable record of entire transactions in
the Bitcoin system [6]. In terms of arrangement and simplicity, various types of
blockchain activities are categorized, such as Blockchain 1.0, 2.0 and 3.0 [7]. The
Blockchain 1.0 deals with currency, spreading of cryptocurrencies in cash involved
implementations, such as money transfer and digital payment mechanisms [7]. The
Blockchain 2.0 deals with contracts spanning whole list of economic and market
applications that utilize blockchain, which is more inclusive than basic cash operations
[7]. The Blockchain 3.0 spans blockchain implementations especially in general-
purpose fields, such as government, health, science, and culture [7].
Understanding the Blockchain Technology Adoption 349
Blockchain technology has many areas of applications. Financial Technology,

abbreviated as FinTech, means organizations that use new technologies to make
financial tasks more effective [8]. Using these novel programs and internet technolo-
gies, these organizations intend to disturb conventional implementations, processes,
business types, and products in the field of finance [8]. Nowadays, there are 5 billion
Internet-of-Things (IoT) appliances that will increase by 2022 and become 29 billion
[9]. Blockchain characteristics such as immutability, accountability, auditability,
encoding, and flexibility can address almost all drawbacks of IoT [9]. As a decen-
tralized mechanism, blockchain has potential to facilitate data storage in healthcare
systems [8]. Smart contracts can ensure storing various health data like demographic
data, input and output data, and provided services etc. [8]. In e-commerce, there are a
giant number of listed counterfeit customers who wants to gain reputation [10].
Blockchain technology has potential to find beneficial solutions for this kind of issue,
and for reputation mechanisms as well [10]. Development of government regulations
can be seen as one of the most important aspects in showing whether the blockchain
will expand in a fully developed financial industry or not [7]. Blockchains and par-
ticularly cryptocurrencies are not accepted as a legal currency by governments [6].
Although it is categorized as money in some situations in the United States of America
and Germany, it is still not considered to be a normal currency [6]. Sometimes,
blockchain technology implementation can be disadvantageous. Blockchain technol-
ogy is based on a distributed consensus mechanism to build trust [11]. On the other
hand, consensus mechanism owns a 51% vulnerability that can be used by attackers to
operate overall chain [11]. Additionally, a great many of cryptocurrencies do not have a
high transaction ratio [12]. Bitcoin mechanism cannot be checked against the mecha-
nisms like VISA’s network that operates thousands of transactions [12]. With the wide
adoption of cryptocurrencies, the latency problem can be solved [12].
3 Literature Survey
In literature, there are a relatively small number of research papers on the blockchain
technology applications from the perspectives of procurement professionals. For
example, with a blockchain implementation study in [13], firm employees obtained
benefits: the blockchain created value in cryptocurrency creation, privacy, and disin-
termediation. In [14], the authors focused on encryption mechanism, distributed led-
gers, and smart contracts. Using the blockchain, payments can be settled directly by
subsidiaries and suppliers [14], which was actualised via a blockchain wallet mecha-
nism in [14]. Blockchain literature is mostly conceptual, and experimental evidence for
its adoption is rare [3]. Thus, it is extremely significant to analyze the factors that
influence blockchain acceptance [3]. Little has been studied on the adoption of
blockchain implementations [3]. As noted in the TAMs, different authors have made
important struggles to figure out how people behave when it comes to accept tech-
nology usage [1]. In literature, there are blockchain technology adoption studies that
use TAM models. Blockchain technology adoption studies with TAM models, and also
with various methodologies can be seen in Table 1.
Table 1. Blockchain technology adoption studies with TAM models.

Method Aim of the study
[15] TAM Aims to inspect blockchain technology
acceptance in for Bitcoin and financial
technology
[2] TAM Aims to investigate blockchain technology
acceptance by mining the collective
intelligence of users on Twitter
[16] TAM, Innovation Diffusion Aims to investigate the research and
Theory application views of blockchain technology
acceptance
[17] The Unified Theory of Acceptance Aims to investigate the blockchain technology
and Use of Technology (UTAUT) adoption in supply chain traceability
applications
[1] TAMs, UTAUT Aims to investigate the blockchain technology
adoption behavior in the logistics and supply
chain fields in India and the USA
[3] TAM, Theory of Planned Aims to provide an in-depth exploration of the
Behavior, Technology Readiness decision-making components that impact
Index behavioral intention of blockchain adoption in
supply chains
4 Methodology
4.1 TAM
TAM was introduced to figure out personal behaviors in IT adoption [1]. The aim of
the TAM model is to clarify or estimate usage of information technology [18].
The TAM provides a hypothetical base for realizing the influence of external variables
on internal opinions, behaviors and users’ purposes and implementation of information
technologies [18]. Impact variables for real system usage of the model contain: external
variables, Perceived Usefulness (PU), Perceived Ease of Use (PEU), Attitude toward
using (A) and Behavioral Intention to use (BI) [4]. The PEU is described as degree of
belief that a person will find it easy to use the system [15]. PEU involves items, such as
necessary physical and mental working, simplicity to understand use cases, simplicity
to learn different usages, operational, operational performance of use cases, and ease of
use for remembering and guidance [2]. Concerning the variables in the TAM model,
PU was determined as user attitudes towards improving performance by using a
specific application in a specific environment [18]. The more easier to use the system,
the more favorable attitude users demonstrate when using the system [18]. (A) is
defined as a personal’s entire efficient response to utilize technology [3]. External
variables are technology features, training, user participation in system design and
structure of system establishment process [18]. All external variables implicitly influ-
ence behavioral purposes of users and real system usage [18].
4.2 Theoretical Extension of TAM (TAM2)

The TAM2 model admitted a social influence process that involves Subjective Norm
(SN), Image (IM), Voluntariness (VO) and Experience (EX) [4]. In addition, the model
admitted a cognitive instrumental process that contains Job Relevance (JR), Output
Quality (OQ) and Result Demonstrability (RD) [4]. PU is determined by external
variables and PEU, which is different from the previous basic model [4]. JR is an extent
of which the technology is assumed practicable to the job [18]. TAM2 describes JR as a
cognitive judgment, rather than a SN which has an immediate impact on PU [18]. If
users understand business knowledge and tools, efficiency can be assured. In tech-
nology connected works, if users have difficulty in improving their performance, they
will not adopt the technology [18]. The TAM2 has more explanatory power than the
TAM model [4, 18]. The TAM2 can be seen in Fig. 1.
Experience Voluntariness
Subjective
Norm
Image
Perceived
Job
Usefulness
Relevance
Output Intention to Usage

Quality Use Behavior
Perceived
Result Ease of Use
Demonstrability
Fig. 1. Modified Technology Acceptance Model (TAM2).
4.3 Cognitive Mapping
Intuitionistic Fuzzy Sets (IFSs). The concept of IFS can be considered a tool that
describes a fuzzy set when existing data is not adequate to describe a vague concept
using a traditional fuzzy set [5]. The traditional fuzzy numbers indicate uncertainty of
information, but they ignore lack of information [5]. When decision-makers are hesitant
about data, the IFSs are appropriate to represent this data [5].
Fuzzy Cognitive Map (FCM) Methods. The FCMs are basic, but robust instruments
for modeling dynamic networks to define specific areas, while using fuzzy logic that
may help to determine hesitancy [19]. The primary benefit of FCMs over other decision-
making frameworks is that it has simple graphical representation [19]. In recent years,
main disadvantages of the FCM are debated and extensions of the FCM are recom-
mended [5]. In 2011, the Intuitionistic Fuzzy Cognitive Maps (IFCM) was proposed and
implemented in medicine [5]. The method was improved in 2013 and the IFCM II was
proposed [5].
Intuitionistic Cognitive Map (ICM) Methodology. An IFS I is identified with its
membership (lI) and non-membership (mI) degrees. In a finite universe U, an IFS I is in
the form [5]:
I ¼ fðx; lðxÞ; mðxÞÞjx 2 U g ð1Þ
where lI(x): U ! [0, 1], mI(x): U ! [0, 1] with the condition 0 lI(x)+ mI(x) 1.
In IFS, the sum of the degrees of membership and nonmembership must be less than 1.
The degree of indeterminacy membership pI is described as [5]:
pI ð xÞ ¼ 1 lI ð xÞ mI ðxÞ: ð2Þ
ICM involves six steps: definition of the concepts, signing causality links for
positive, negative or null categories, IFSs assignment, construction of weight matrix,
and calculation of intuitionistic cognitive mapping iterations using ICM equation [5];
ðk þ 1Þ
XN ðkÞ
Ai ¼ f ð j6¼i Aj ½lðwji Þ hs pðwji Þ ð3Þ
j¼i
where Aki þ 1 is the value of the concept Ci at iteration k + 1, f is the threshold function,
Aki is the value of the concept Ci at iteration k, l(wji) is the degree of membership of wji,
p(wji) is the degree of indeterminacy membership of wji, and hs is the coefficient of
synergy of the application field, hs 2 [0, 1]. Last step includes calculation of each
concept’s value [5].
5 Case Study
Blockchain technology has important elements in PEU and PU, based on its use cases
and characteristics. For the company experts, there are both benefits and risks to
Blockchain. Although blockchain has a strong encryption mechanism, they were still
worried about security risks. According to the experts, when it comes to handling larger
data volumes, scalability in Blockchain might causes problems. In particular, smart
contracts and cryptocurrencies were among their interests, as blockchain use cases in
the PEU part. In the PU part, they focused on performance, efficiency and effectiveness
for their company. Without third party intermediaries, transactions could be time and
cost saving. According to their PU of blockchain compared to existing technologies,
blockchain provides trust, transparency, traceability, and agility. The literature does not
focus on which blockchain characteristics are perceived as more beneficial for specific
blockchain applications. TAM2 variables, causal relationships and directions can be
seen in Fig. 1. Procurement professionals assigned IFSs to the weights of causal
relationships, and w matrix is constructed as can be seen in Table 2.
Table 2. W matrix.
SN IM JR OQ RD EX VO PU PEU INT UB
SN 0 M M M U U 0 M 0 U 0
IM M 0 M 0 0 0 0 M 0 0 0
JR 0 M 0 M 0 U 0 M I 0 0
OQ 0 0 M 0 0 0 0 I 0 0 0
RD 0 I M I 0 U U I 0 0 0
EX 0 0 0 0 0 0 0 M 0 M 0
VO U 0 0 0 0 U 0 0 0 M 0
PU 0 0 0 0 0 0 0 0 0 M 0
PEU 0 0 0 0 0 0 U I 0 I 0
INT 0 0 0 0 0 0 0 0 0 0 M
UB 0 0 0 0 0 0 0 0 0 0 0
ICM iteration values converged in 15 iterations. For the company professionals, the
variables affecting the Usage Behavior (UB) to adopt blockchain technology were, in
order of priority: PU, BI, JR, OQ, IM, EX, SN, PEU, UB, VO and RD. Our results
indicate that the influences between variables are in line with most of the other studies.
However, the finding further strengthens the view that the procurement professionals
give BI, JR and OQ more importance. In this study, the IFCM method avoided the
problem that experts do not have sufficient experience in blockchain technology.
6 Conclusion
In this article, we reviewed existing literature on blockchain technology adoption. In

literature, additional research is necessary to obtain a broader understanding of
blockchain technology adoption. As we reported, while there are studies in the existing
literature aiming to evaluate blockchain adoption behavior, very few efforts consider
procurement professionals’ perspectives. The findings further strengthen the view that
the procurement professionals give intention, job relevance and output quality more
importance. For further research, we suggest that professionals from different depart-
ments should participate in the evaluation process. In addition, a sensitivity analysis
should be applied for different professions.
References
1. Queiroz, M.M., Wamba, S.F.: Blockchain adoption challenges in supply chain: an empirical
investigation of the main drivers in India and the USA. Int. J. Inf. Manag. 46, 70–82 (2019)
2. Grover, P., Kar, A.K., Janssen, M., Ilavarasan, P.V.: Perceived usefulness, ease of use and
user acceptance of blockchain technology for digital transactions – insights from user-
generated content on Twitter. Enterp. Inf. Syst. 13(6), 771–800 (2019)
3. Kamble, S., Gunasekaran, A., Arha, H.: Understanding the Blockchain technology adoption
in supply chains-Indian context. Int. J. Prod. Res. 57(7), 2009–2033 (2019)
4. Lee, Y.C., Li, M.L., Yen, T.M., Huang, T.H.: Analysis of adopting an integrated decision
making trial and evaluation laboratory on a technology acceptance model. Expert Syst. Appl.
37, 1745–1754 (2010)
5. Dogu, E., Albayrak, Y.E.: Criteria evaluation for pricing decisions in strategic marketing
management using an intuitionistic cognitive map approach. Soft Comput. 22, 4989–5005
(2018)
6. Bashir, I.: Mastering Blockchain, 1st edn. Packt Publishing, Birmingham (2017)
7. Swan, M.: Blockchain: Blueprint for a New Economy, 1st edn. O’Reilly Media, Sebastopol
(2015)
8. Bahga, A., Madisetti, V.: Blockchain Applications: A Hands-on Approach. Arshdeep Bahga
& Vijay Madisetti (2017)
9. Panarello, A., Tapas, N., Merlino, G., Longo, F., Puliafito, A.: Blockchain and IoT
integration: a systematic survey. Sensors 18(8), 2575 (2018)
10. Zheng, Z., Xie, S., Dai, H.N., Chen, X., Wang, H.: Blockchain challenges and opportunities:
a survey. Int. J. Web Grid Serv. 14(4), 352–375 (2018)
11. Li, X., Jiang, P., Chen, T., Luo, X., Wen, Q.: A survey on the security of blockchain
systems. Future Gener. Comput. Syst. 107, 841–853 (2020)
12. Casino, F., Dasaklis, T.K., Patsakis, C.: A systematic literature review of blockchain-based
applications: current status, classification and open issues. Telematics Inform. 36, 55–81
(2019)
13. Ying, W., Jia, S., Du, W.: Digital Enablement of blockchain: evidence from HNA group. Int.
J. Inf. Manag. 39, 1–4 (2018)
14. Du, W., Pan, S.L., Leidner, D.E., Ying, W.: Affordances, experimentation and actualization
of FinTech: a blockchain implementation study. J. Strateg. Inf. Syst. 28, 50–65 (2019)
15. Folkinshteyn, D., Lennon, M.: Braving bitcoin: a technology acceptance model (TAM) anal-
ysis. J. Inf. Technol. Case Appl. Res. 18(4), 220–249 (2016)
16. Lou, A.T.F., Li, E.Y.: Integrating innovation diffusion theory and the technology acceptance
model: the adoption of blockchain technology from business managers’ perspective. In: 17th
International Conference on Electronic Business, Dubai, pp. 293–296 (2017)
17. Francisco, K., Swanson, D.: The supply chain has no clothes: technology adoption of
blockchain for supply chain transparency. Logistics 2(1), 2 (2018)
18. Lee, Y.C., Li, M.L., Yen, T.M., Huang, T.H.: Analysis of fuzzy decision making trial and
evaluation laboratory on technology acceptance model. Expert Syst. Appl. 38, 14407–14416
(2011)
19. Papageorgiou, E.I., Iakovidis, D.K.: Towards the construction of intuitionistic fuzzy
cognitive maps for medical decision making. In: 9th International Conference on
Information Technology and Applications in Biomedicine, Cyprus, pp. 1–4. IEEE (2009)
Distance and Similarity Measures of
Generalized Intuitionistic Fuzzy Soft Set
and Its Applications in Decision Support
System
Muhammad Jabir Khan1 and Poom Kumam1,2,3(B)

1
King Mongkut’s University of Technology, Thonburi, Bangkok 10140, Thailand
2
Center of Excellence in Theoretical and Computational Science (TaCS-CoE),
King Mongkut’s University of Technology Thonburi (KMUTT),
Bangkok 10140, Thailand
3
Department of Medical Research, China Medical University Hospital,
China Medical University, Taichung 40402, Taiwan
Abstract. The basic idea underneath the generalized intuitionistic

fuzzy soft set (GIFSS) is very constructive in decision making, since it
considers how to exploit an extra intuitionistic fuzzy input from the
director to make up for any distortion in the information provided by
the evaluation experts, which is redefined and clarified by F. Feng. In
this paper, we introduce the similarity measures based on the cosine
and cotangent function for GIFSSs and discuss their properties. At the
end, a numerical examples of decision making and pattern recognition is
discussed based on the proposed similarity measures.
Keywords: Generalized intuitionistic fuzzy soft set · Similarity

measures · Distance measures · Multi attribute decision making ·
Pattern recognition
1 Introduction
In fuzzy sets theory, the membership function is used to represents the infor-
mation [1]. Real life uncertainties handle effectively by fuzzy set theory. In Ref-
erence [2], Molodtsov defines the soft set which is a new logical instrument for
dealing uncertainties. Molodtsov soft set theory deals uncertainty effectively by
considering parametric point of view, that is, each element is judged by some
criteria of attributes. Atanassove defines intuitionistic fuzzy set (IFS) is the gen-
eralization of the fuzzy set theory [3]. The information in IFS is represented in
the form of membership (favour) and non-membership function (against). The
membership function and non-membership function assigns the values from the
Petchra Pra Jom Klao Ph.D. Research Scholarship, KMUTT and TaCS-CoE.
356 M. J. Khan and P. Kumam
unit interval [0, 1] with the condition that their sum is less than or equal to
one, i.e., if we represents the membership and non-membership functions by ξ
and ν, respectively, than 0 ≤ ξ + ν ≤ 1. This condition specifies the range of ξ
and ν. The range of membership and non-membership functions increases in
Yager’s Pythagorean fuzzy sets [4], i.e., the experts make their judgments more
freely in Pythagorean fuzzy environment. The condition 0 ≤ ξ 2 + ν 2 ≤ 1 specifies
the range of membership and non-membership functions. Further, improvement
continues by defining the more general environment, the q-rung orthopair fuzzy
sets by Yager [5]. The condition 0 ≤ ξ q + ν q ≤ 1, where q > 1 is any real number,
specifies the range of membership and non-membership functions.
Keeping in mind the importance of similarity measure and application in
data mining, medical diagnosis, decision making and pattern recognition many
authors work on this topic. A wide theory of similarity measures of fuzzy sets
and intuitionistic fuzzy sets is presented in the literature [7–12].
The idea of GIFSS was discussed by Feng et al. [13]. Further, soft decernibility
matrix for GIFSS was formulated by Khan et al. [14]. The generalized picture
fuzzy soft set was defined and applied to decision making by Khan et al. [15–17].
The generalized interval valued IFS was explained in [18].
Major contributions of our work is to define similarity measures for GIFSS
based on the cosine and cotangent functions. A method to solve decision making
problem and pattern recognition is proposed.
Rest of the paper is designed as follows: Sect. 2 contains the basic definitions.
In Sect. 3, the similarity measures based on the cosine and cotangent functions
are define. Numerical example of the decision making and pattern recognition is
discussed. Section 4 consist of the conclusion.
2 Preliminaries
The definitions of IFS, soft set, IFSS and GIFSS are written in this section.
Definition 1 ([3]). An IFS R on a universal set Ŷ is defined as
R = {(ξR (yi ), νR (yi )) | yi ∈ Ŷ},
where ξR and νR are the membership and non-membership functions from the
universal set Ŷ to the unit interval [0, 1], respectively. For IFSs, the sum of the
membership and non-membership functions is less than or equal to one, i.e.,
ξR (yi ) + νR (yi ) ≤ 1. The quantity πR (yi ) = 1 − (ξR (yi ) + νR (yi )) is called the hes-
itancy degree of the element yi ∈ Ŷ.
Molodtsov soft set theory deals uncertainty effectively by considering para-
metric point of view [2], that is, each element is judged by some criteria of
attributes (characteristics).
Definition 2 ([2]). Let universal space and parametric space are represented
by Ŷ and Ê, respectively. Let Â ⊂ Ê be a parametric set and power set of Ŷ is
displaced by P(Ŷ). A pair (F̂, Â) is called a soft set over Ŷ, where F̂ is a set valued
mapping given by F̂ : Â → P(Ŷ).
Distance and Similarity Measures of Generalized Intuitionistic Fuzzy Soft Set 357
In [6], Maji defines the IFSS as follows.

Definition 3 ([6]). Let universal space and parametric space are represented by
Ŷ and Ê, respectively. Let Â ⊂ Ê be a parametric set and IF(Ŷ) the set of all IFSS
of Ŷ. A pair (F̂, Â) is called an IFSS over Ŷ, where F̂ is a set valued mapping
given by F̂ : Â → IF (Ŷ).
The idea of GIFSS is constructive in decision-making since it considers how
to take advantage of an extra intuitionistic fuzzy input from the director to
decrease any possible distortion in the data provided by evaluating experts.
Definition 4 ([13]). Let universal space and parametric space are represented
by Ŷ and P̂, respectively. Let Â ⊂ P̂ be a parametric subset. We call (F̂, Â, ρ̂) a
GIFSS, where (F̂, Â) is an IFSS over Ŷ and ρ̂ : Â → IF(Â) is an IFS in Â.
where (F̂, Â) is called basic intuitionistic fuzzy soft set (BIFSS) and ρ̂ is called
the parametric intuitionistic fuzzy set (PIFS).
3 Distance and Similarity Measures
The similarity measures based on cosine and cotangent functions are defined
in this section. A numerical example is discussed here to support the proposed
similarity measures.
Definition 5. A distance measure between two GIFSSs Γ1 and Γ2 is a mapping
D̂ : GIFSS × GIFSS → [0, 1], which satisfies the following properties:
(D1) 0 ≤ D̂(Γ1 , Γ2 ) ≤ 1
(D2) D̂(Γ1 , Γ2 ) = 0 ⇐⇒ Γ1 = Γ2
(D3) D̂(Γ1 , Γ2 ) = D̂(Γ2 , Γ1 )
(D4) If Γ1 ⊆ Γ2 ⊆ Γ3 then D̂(Γ1 , Γ3 ) ≥ D̂(Γ1 , Γ2 ) and D̂(Γ1 , Γ3 ) ≥ D̂(Γ2 , Γ3 ).
Definition 6. A similarity measure between two GIFSSs Γ1 and Γ2 is a mapping

D̂ : GIFSS × GIFSS → [0, 1], which satisfies the following properties:
(S1) 0 ≤ Ŝ(Γ1 , Γ2 ) ≤ 1
(S2) Ŝ(Γ1 , Γ2 ) = 1 ⇐⇒ Γ1 = Γ2
(S3) Ŝ(Γ1 , Γ2 ) = Ŝ(Γ2 , Γ1 )
(S4) If Γ1 ⊆ Γ2 ⊆ Γ3 then Ŝ(Γ1 , Γ3 ) ≤ Ŝ(Γ1 , Γ2 ) and Ŝ(Γ1 , Γ3 ) ≤ Ŝ(Γ2 , Γ3 ).
Definition 7. Let Γ1 = (F̂, Â, ρ̂) and Γ2 = (Ĝ, B̂, σ̂) be two GIFSSs in Ŷ, the sim-
ilarity measures between Γ1 and Γ2 based on cosine functions are defined as
follows:

1 π
n m
S1 (Γ1 , Γ2 ) = cos ξF̂(ej ) (yi ) − ξĜ(ej ) (yi ) + νF̂(ej ) (yi ) − νĜ(ej ) (yi )
mn 8
j=1 i=1

+ ξρ̂(e ) − ξσ̂(e ) + νρ̂(e ) − νσ̂(e ) (1)
j j j j

1 π
n m
S2 (Γ1 , Γ2 ) = cos ξF̂(ej ) (yi ) − ξĜ(ej ) (yi ) + νF̂(ej ) (yi ) − νĜ(ej ) (yi )
mn 8
j=1 i=1

+ ξρ̂(ej ) − ξσ̂(ej ) + νρ̂(ej ) − νσ̂(ej )

+ (yi ) −
F̂(ej ) (yi ) + ρ̂(e ) − σ̂(e )
Ĝ(ej ) j j
(2)

1 π
n m
S3 (Γ1 , Γ2 ) = cos ξF̂(ej ) (yi ) − ξĜ(ej ) (yi ) ∨ νF̂(ej ) (yi ) − νĜ(ej ) (yi )
mn 2
j=1 i=1

∨ ξρ̂(e ) − ξσ̂(e ) ∨ νρ̂(e ) − νσ̂(e )
j j j j
(3)

1 π
n m
S4 (Γ1 , Γ2 ) = cos ξF̂(ej ) (yi ) − ξĜ(ej ) (yi ) ∨ νF̂(ej ) (yi ) − νĜ(ej ) (yi )
mn 2
j=1 i=1

∨ ξρ̂(ej ) − ξσ̂(ej ) ∨ νρ̂(ej ) − νσ̂(ej )

∨ F̂(ej ) (yi ) − Ĝ(ej ) (yi ) ∨ ρ̂(ej ) − σ̂(ej ) (4)
Definition 8. Let Γ1 = (F̂, Â, ρ̂) and Γ2 = (Ĝ, B̂, σ̂) be two GIFSSs in Ŷ, the sim-
ilarity measures between Γ1 and Γ2 based on cosine functions are defined as
follows:

1 π
n m
π
S5 (Γ1 , Γ2 ) = cot + ξF̂(ej ) (yi ) − ξĜ(ej ) (yi ) + νF̂(ej ) (yi ) − νĜ(ej ) (yi )
mn j=1 i=1 4 16

+ ξρ̂(e ) − ξσ̂(e ) + νρ̂(e ) − νσ̂(e )
j j j j
(5)

1 π
n m
π
S6 (Γ1 , Γ2 ) = cot + ξF̂(ej ) (yi ) − ξĜ(ej ) (yi ) + νF̂(ej ) (yi ) − νĜ(ej ) (yi )
mn j=1 i=1 4 16

+ ξρ̂(ej ) − ξσ̂(ej ) + νρ̂(ej ) − νσ̂(ej )

+ (yi ) −
F̂(ej ) (yi ) + ρ̂(e ) − σ̂(e )
Ĝ(ej ) j j
(6)

1
n m
π π
S7 (Γ1 , Γ2 ) = cot + ξ − ξĜ(ej ) (yi ) ∨ νF̂(ej ) (yi ) − νĜ(ej ) (yi )
F̂(ej ) (yi )
mn j=1 i=1 4 4

∨ ξρ̂(ej ) − ξσ̂(ej ) ∨ νρ̂(ej ) − νσ̂(ej ) (7)

1 π
n m
π
S8 (Γ1 , Γ2 ) = cot + ξF̂(ej ) (yi ) − ξĜ(ej ) (yi ) ∨ νF̂(ej ) (yi ) − νĜ(ej ) (yi )
mn j=1 i=1 4 4

∨ ξρ̂(ej ) − ξσ̂(ej ) ∨ νρ̂(ej ) − νσ̂(ej )

∨ F̂(ej ) (yi ) − Ĝ(ej ) (yi ) ∨ ρ̂(ej ) − σ̂(ej ) (8)
where ∨ is the maximum operation.
Theorem 1. Si , 0 ≤ i ≤ 8 satisfies the axioms of similarity measures (Defini-

tion 6).
Theorem 2. The distance measures for GIFSSs is obtained form the proposed
similarity measures by Di = 1 − Si , 0 ≤ i ≤ 8, i.e., Di satisfies the axioms of dis-
tance measures (Definition 5).
Example 1. Suppose two GIFSSs Γ1 = (F̂, Â, ρ̂) and Γ2 = (Ĝ, B̂, σ̂) be given in
Table 1.
We find the similarity between Γ1 and Γ2 by using above mentioned similarity
measures. The results of similarity measures are
S1 (Γ1 , Γ2 ) = 0.405329, S2 (Γ1 , Γ2 ) = 0.378678, S3 (Γ1 , Γ2 ) = 0.394212,

S4 (Γ1 , Γ2 ) = 0.394212, S5 (Γ1 , Γ2 ) = 0.271662, S6 (Γ1 , Γ2 ) = 0.242351,
S7 (Γ1 , Γ2 ) = 0.259162, S8 (Γ1 , Γ2 ) = 0.259162.
Table 1. Two GIFSSs Γ1 and Γ2
Ŷ e1 e2 e3
y1 (0.7, 0.2) (0.5, 0.4) (0.2, 0.7)
y2 (0.9, 0.0) (0.7, 0.2) (0.4, 0.6)
ρ̂ (0.3, 0.5) (0.5, 0.3) (0.3, 0.4)
y1 (0.3, 0.5) (0.7, 0.2) (0.5, 0.4)
y2 (0.4, 0.5) (0.6, 0.3) (0.4, 0.5)
σ̂ (0.5, 0.2) (0.7, 0.3) (0.8, 0.1)
3.1 Application in Decision Making and Pattern Recognition

A technological institute have some scholarships for their students. The scholar-
ships awarded to the best students. To specify a criteria for selecting students.
A committee is established, which decide the criteria for evaluation of student
performance in the institute. The director of the institute is the head of the
committee. For example, five students are shortlisted for further evaluation. The
committee set up the criteria for evaluation which consist of: e1 , e2 , e3 and e4
which stand for knowledge and understanding, practical applications of their
work, lab work and performance in the test, respectively. The committee eval-
uate the students according to the predefined criteria and their evaluation are
saved in the form of BIFSS. The head of the committee judged the evaluation
of the students by the committee and carefully monitored the whole procedure
of evaluation and give their preferences in the form of PIFS which completes the
formation of GIFSS. For comparison, an ideal student performance according to
the prescribed criteria is set by the committee and all the data is presented in
Table 2.
Table 2. The GIFSS
Ŷ e1 e2 e3 e4
y1 (0.7, 0.2) (0.5, 0.4) (0.2, 0.7) (0.8, 0.1)
y2 (0.9, 0.0) (0.7, 0.2) (0.4, 0.6) (0.6, 0.3)
y3 (0.5, 0.5) (0.8, 0.2) (0.5, 0.4) (0.5, 0.1)
y4 (0.4, 0.5) (0.6, 0.3) (0.4, 0.5) (0.7, 0.3)
ρ̂ (0.3, 0.5) (0.5, 0.3) (0.3, 0.4) (0.7, 0.1)
yideal (0.9, 0.1) (0.8, 0.1) (0.9, 0.0) (0.8, 0.2)
ρ̂ideal (0.9, 0.1) (0.8, 0.1) (0.9, 0.0) (0.8, 0.2)
To select best candidate for scholarship, the similarity measures is used. The
similarity between the evaluated values and the ideal performance set by the
committee is calculated by using above proposed similarity measures.
Table 3. Similarity Measures
S1 0.840373 0.882546 0.866388 0.841827 y2 y3 y4 y1

S2 0.802397 0.840732 0.819271 0.80089 y2 y3 y1 y4
S3 0.730118 0.754408 0.739396 0.763566 y4 y2 y3 y1
S4 0.72096 0.754408 0.718899 0.754408 y4 y2 y1 y3
S5 0.609948 0.652298 0.624671 0.600666 y2 y3 y1 y4
S6 0.548173 0.584508 0.538229 0.548589 y2 y4 y1 y3
S7 0.50797 0.497296 0.46886 0.52918 y4 y1 y2 y3
S8 0.476085 0.497296 0.443041 0.497296 y4 y2 y1 y3
The candidate y2 has the maximum similarity with the ideal candidate.
Therefore, y2 selected for scholarship from similarity measures results (Table 3).
4 Conclusion
In this paper, we have introduced the similarity measures based on the cosine
and cotangent function for GIFSSs. The Feng et al. [13] approach has considered
for GIFSSs. At the end, a numerical example of decision making and pattern
recognition is discussed based on the proposed similarity measures.
References
2. Molodtsov, D.: Soft set theory-first results. Comput. Math. Appl. 37, 19–31 (1999)
3. Atanassov, K.T.: Intuitionistic fuzzy sets. Fuzzy Sets Syst. 20, 87–96 (1986)
4. Yager, R.R.: Pythagorean fuzzy subsets. In: 2013 Joint IFSA World Congress and
NAFIPS Annual Meeting (IFSA/NAFIPS), Edmonton, AB, pp. 57–61 (2013)
5. Yager, R.R.: Generalized orthopair fuzzy sets. IEEE Trans. Fuzzy Syst. 25(5),
1222–1230 (2017)
6. Maji, P.K., Biswas, R., Roy, A.R.: Intuitionistic fuzzy soft sets. J. Fuzzy Math. 9,
677–692 (2001)
7. Wang, W., Xin, X.: Distance measure between intuitionistic fuzzy sets. Pattern
Recogn. Lett. 26, 2063–2069 (2005)
8. Ye, J.: Cosine similarity measures for intuitionistic fuzzy sets and their applications.
Math. Comput. Modell. 53(1–2), 91–97 (2011)
9. Garg, H., Kumar, K.: An advanced study on the similarity measures of intuitionis-
tic fuzzy sets based on the set pair analysis theory and their application in decision
making. Soft. Comput. 22(15), 4959–4970 (2018)
10. Hung, W.L., Yang, M.S.: Similarity measures of intuitionistic fuzzy sets based on
Hausdorff distance. Pattern Recogn. Lett. 25(14), 1603–1611 (2004)
11. Boran, F.E., Akay, D.: A biparametric similarity measure on intuitionistic fuzzy
sets with applications to pattern recognition. Inf. Sci. 255, 45–57 (2014)
12. Khan, M.J., Kumam, P., Deebani, W., Kumam, W., Shah, Z.: Distance and sim-
ilarity measures for spherical fuzzy sets and their applications in selecting mega
projects. Mathematics 8(4), 519 (2020). https://doi.org/10.3390/math8040519
13. Feng, F., Fujita, H., Ali, M.I., Yager, R.R.: Another view on generalized intuition-
istic fuzzy soft sets and related multi attribute decision making methods. IEEE
Trans. Fuzzy Syst. 27, 474–488 (2018)
14. Khan, M.J., Kumam, P., Liu, P., Kumam, W., Ashraf, S.: Ashraf a novel approach
to generalized intuitionistic fuzzy soft sets and its application in decision support
system. Mathematics 7(8), 742 (2019). https://doi.org/10.3390/math7080742
15. Khan, M.J., Kumam, P., Ashraf, S., Kumam, W.: Generalized picture fuzzy soft
sets and their application in decision support systems. Symmetry 11(3), 415 (2019).
https://doi.org/10.3390/sym11030415
16. Khan, M.J., Kumam, P., Liu, P., Kumam, W., Rehman, H.: An adjustable weighted
soft discernibility matrix based on generalized picture fuzzy soft set and its appli-
cations in decision making. J. Int. Fuzzy Syst. 38(2), 2103–2118 (2020). https://
doi.org/10.3233/JIFS-190812
17. Khan, M.J., Phiangsungnoen, S., Rehman, H., Kumam, W.: Applications of gen-
eralized picture fuzzy soft set in concept selection. Thai J. Math. 18(1), 296–314
(2020)
18. Khan, M.J., Kumam, P., Liu, P., Kumam, W.: Another view on generalized interval
valued intuitionistic fuzzy soft set and its applications in decision support system.
J. Int. Fuzzy Syst. 1–15 (2019, accepted). https://doi.org/10.3233/JIFS-190944
Intuitionistic Fuzzy Analysis of Variance
of Movie Ticket Sales
Velichka Traneva(B) and Stoyan Tranev
“Prof. Asen Zlatarov” University, “Prof. Yakimov” Blvd, 8000 Bourgas, Bulgaria
http://www.btu.bg
Abstract. This paper was provoked by the interest in investigating the

dependence of movie sales on the factors of “a day of the week” and
the “ticket price”. Analysis of variance (ANOVA) is a statistical method
concerned with comparing means of several samples. In the paper, a real
data set on daily ticket sales for the Bulgarian premiere movie “Cozi-
ness” (2-D, 2019) for a month from a Cinema City Bulgaria multiplex, is
analysed using one-way ANOVA. The observations including vagueness
can be expressed exactly using intuitionistic fuzzy set (IFS) member-
ship and non-membership functions. In order to analyze such data, it
is necessary to apply the proposed in the paper an intuitionistic fuzzy
ANOVA (IFANOVA). The proposed approach employs the apparatus of
IFSs and index matrices (IMs). The paper also explores an application
of IFANOVA to find out the dependencies of the ticket sales for this
premiere on a “day of the week” factor. A comparative analysis of the
rezults obtained by ANOVA and IFANOVA is performed. The main con-
tribution of the paper is that it shows the effectiveness of the proposed
IFANOVA for investigating the effects of factor levels.
Keywords: Index matrix · Intuitionistic fuzzy sets · Variation analysis
1 Introduction
The paper was provoked by the idea of determining the dependencies of the ticket
sales of a Bulgarian premiere “Coziness” (2-D, 2019) on the factors “ticket price”
and “day of the week”. There are situations where the data values are vaguely
specified in many cases due to the human errors in measuring, information noise,
data loss, etc. It is obvious that the traditional ANOVA cannot analyze these
unclear numbers, therefore it is necessary to use fuzzy sets theory for dealing
with the uncertain environment [22]. In the following we review some recent
works on ANOVA in fuzzy environment. A bootstrap approach to FANOVA has
Supported by the Ministry of Education and Science under the National Programme
“Young scientists and postdoctoral students”, approved by DCM # 577/17.08.2018
(V. Traneva) and by the Asen Zlatarov University under Project NIX-423/2019 “Inno-
vative methods for extracting knowledge management” (S. Tranev).
364 V. Traneva and S. Tranev
introduced in [11]. FANOVA was considered in [8] using a set of confidence inter-
vals for variance. FANOVA has proposed by [12] using the levels of pessimistic
and optimistic of the triangular fuzzy data. FANOVA has presented in [15,16]
based on Zadeh’s extension principle. IFSs [1,3] are an extension of the fuzzy
sets. IFANOVA by converting IFSs to fuzzy sets has proposed in [13].
In this study, are determined the dependencies of the sales of the premiere
“Coziness” on the factors “ticket price” and “day of the week” by one-factor
ANOVA. Also in the work is proposed one-way IFANOVA, based on the appa-
ratus of IFSs and IMs, to a case where observed data are intuitionistic fuzzy
rather than real numbers. The concept of IMs has been introduced in [2] and is
generalized in [4]. An application of IFANOVA is explored in the paper to ana-
lyze the dependencies of the daily movie sales of the premiere “Coziness” on a
factor “day of the week”. We used the real data set from a Cinema City Bulgaria
multiplex, part of Cineworld PLC Group. The main contributions of the paper
lie in its proposition of the IFANOVA and iis study of the effectiveness of the
proposed method on real daily movie sales data in an uncertain environment.
The rest of this paper is structured as follows: Sect. 2 describes some basic
definitions of the concepts of IMs and IF logic. Section 3 describes the classical
ANOVA and its application over the movie sales. In Sect. 4, we propose the
IFANOVA, based on the concepts of IMs and IFSs. The proposed approach
is applied over the IF data with the sales of the “Coziness” for investigating
the effect of the factor “day of the week”. The obtained results of IFANOVA
are compared with those obtained from classical ANOVA. Section 5 offers the
conclusions and outlines aspects for future research.
2 Basic Definitions on IMs and Intuitionistic Fuzzy Logic
Intuitionistic Fuzzy Pairs

The IFP is an object of the form a, b = μ(p), ν(p), where a, b ∈ [0, 1] and a+b ≤
1, that is used as an evaluation of a proposition p [6]. μ(p) and ν(p) respectively
determine the degrees of membership and non-membership. Let us have two IFPs
x = a, b and y = c, d. We recall some basic operations [3,6,7,18].
¬x = b, a; x ∧1 y = min(a, c), max(b, d);

x ∨1 y = max(a, c)), min(b, d); x ∧2 y = x + y = a + c − a.c, b.d;
x ∨2 y = x.y = a.c, b + d − b.d; α.x = 1 − (1 − a)α , bα (α ∈ R);
x@y = 2 , 2 ;
a+c b+d
x2 = a2 , 1 − (1 − b)2 ;
x−y = max(0, a − c), min(1, b + d, 1 − a + c);
x : y = min(1; a/c); min(max(0; 1 − a/c); max(0; (b − d)/(1 − d)))
and relations with IFPs

x ≥ y iff a ≥ c and b ≤ d; x ≥2 y iff a ≥ c;
(1)
x ≥ y iff b ≤ d; x ≥R y iff Ra,b ≤ Rc,d ;
where Ra,b = 0.5(1 + 1 − a − b)0.5(|1 − a| + |b| + |1 − a − b|) [18].

Intuitionistic Fuzzy Analysis of Variance 365
Short Notes over Intuitionistic Fuzzy Index Matrices

Let I be a set. By two-dimensional IF index matrix (2D-IFIM) A =
[K, L, {μki ,lj , νki ,lj }] with index sets K and L (K, L ⊂ I), we denote [4]:
l1 ... lj ... ln
k1 μk1 ,l1 , νk1 ,l1 . . . μk1 ,lj , νk1 ,lj . . . μk1 ,ln , νk1 ,ln
A≡ . .. .. .. .. .. ,
.. . . . . .
km μkm ,l1 , νkm ,l1 . . . μkm ,lj , νkm ,lj . . . μkm ,ln , νkm ,ln
where for every 1 ≤ i ≤ m, 1 ≤ j ≤ n: 0 ≤ μki ,lj , νki ,lj , μki ,lj + νki ,lj ≤ 1.
We use symbol “⊥” for lack of component in the definitions. The some oper-
ations over A = [K, L, {μki ,lj , νki ,lj }] and B = [P, Q, {ρpr ,qs , σpr ,qs }] are [4]:
Addition-(◦, ∗): A ⊕(◦,∗) B = [K ∪ P, L ∪ Q, {φtu ,vw , ψtu ,vw }], where ◦, ∗ ∈
{max, min, min, max, average, average}.
Termwise subtraction-(max,min): A −(max,min) B = A ⊕(max,min) ¬B.
Termwise multiplication: A ⊗(min,max) B = [K ∩ P, L ∩ Q, {φtu ,vw , ψtu ,vw }],
where φtu ,vw , ψtu ,vw = min(μki ,lj , ρpr ,qs ), max(νki ,lj , σpr ,qs ).
Reduction: The operations (k, ⊥)-reduction of an IM A is defined by: A(k,⊥) =
[K − {k}, L, {ctu ,vw }], where ctu ,vw = aki ,lj (tu = ki ∈ K − {k}, vw = lj ∈ L).
Projection: Let M ⊆ K and N ⊆ L. Then, prM,N A = [M, N, {bki ,lj }], where
for each ki ∈ M and each lj ∈ N , bki ,lj = aki ,lj .
Substitution: kp ; ⊥ A = [(K − {k}) ∪ {p}, L, {ak,l }], ⊥; ql A = [K, (L−
{l}) ∪ {q}, {ak,l }] .
Aggregation operation
Let x#@ y = average(a, c), average(b, d), where x = a, b and y = c, d are
IFPs. Let k0 ∈/ K be a fixed index. The aggregation operation by K is [4,21]:
l1 ... ln
m m
αK,#@ (A, k0 ) = . (2)
k0 #@ μki ,l1 , νki ,l1 . . . #@ μki ,ln , νki ,ln
i=1 i=1
Aggregate global internal operation [19]: AGIO⊕(max,min) (A) .

Internal subtraction of IMs’ components [19]:
IO−(max,min) (ki , lj , A , pr , qs , B) = [K, L, {γtu ,vw , δtu ,vw }].
3 Application of ANOVA to the Sales of “Coziness”
3.1 Classical One-Way ANOVA
The original ideas of ANOVA, concerned with analyzing variation in the

means of several independent populations, were developed by Fisher [10].
The simplest case is one-way ANOVA [9]. Let xki ,lj for i = 1, 2, ..., m and
j = i1 , i2 , ...iI (1 ≤ iI ≤ n) denote the data from the ki −th level and lj −th
observation. Let N is the number of observations. The ANOVA has been con-
template to accept/reject hypothesis H0 : μk1 = μk2 = ... = μkm , against H1 :
not all μki are equal, where μki are the factor level means The sum of squares
SST , sum of squares within groups SSE and the sum of squares between groups
SSC are calculated. The total mean sum of squares M ST , the mean sums of
squares for error M SE and the mean sums of squares for treatment M SC are [9]:
SST SSE SSC

M ST = , M SE = , M SC = .
N −1 N − km km − 1
If the test statistic
M SC 1 1
F = ≥ F(α,m−1,N −m) or ≤ = F(1−α,N −m,m−1) , (3)
M SE F F(α,m−1,N −m)
where F(α,m−1,N −m) is α−quantile of F −distribution with m − 1 and N − m

degrees of freedom, then H0 is rejected on significance level α [9,10]. The p−value
[14] is the probability of rejection the H0 hypotesis. In case p ≤ α, where α is
chosen significance level, H0 is rejected with probability greater than (1 − α).
3.2 Application of ANOVA to the Sales Revenue of “Coziness”
Let us apply one-way ANOVA using the statistical software SPSS [14] to the data
containing the daily sales of the premiere “Coziness”. There are three levels of
the “ticket price” factor in the studied cinema. The following ANOVA Table 1
and Table 2 present one-way ANOVA respectively by a factor “day of week” and
a factor “price of ticket” for the movie “Coziness” with α = 0, 05:
Table 1. ANOVA table by the factor “day of week” for “Coziness”.
Source SS df MS F p-value F crit

Between groups 116910 6 19485 2,82 0,03 2,508
Within groups 165783,5 24 6907,64
Table 2. ANOVA table by the factor “price of ticket” for “Avengers”.
Source SS df MS F p-value F crit

Between groups 107193 2 53596,6 8,55 0,01 3,34
Within groups 175500 28 6267,87
The conclusion of the ANOVA is that the levels of “day of the week” and
“price” factors affect the amount of movie ticket sales. The average ticket sales
are the highest on Thursday (the price is the lowest), then – on the weekend.
Fig. 1. Number of tickets on a day of a week for a month for “Coziness”
Table 3. Number of tickets sold per a week of a month for “Coziness”
First week Second week Third week Fourth week

“Coziness” 1634 1276 734 594
The Fig. 1 presents the comparison between the number of tickets sold per a day
of the week of a month: The Table 3 presents the number of tickets sold per a
week of a month for the premiere movie:
The conclusions of the statistical analysis are:
– Ticket sales were the highest for the movie “Coziness” on Thursday (the
price is the lowest), then – on the weekend. Ticket sales for the premiere were
highest in the first week of its screenings.
– The movie sales were the lowest on Monday and Wednesday. On Thursday,
increased movie ticket sales were due to the low ticket price of 2D-movie
“Coziness”. On Saturday and Sunday, attendance is the highest and is not
affected by the ticket price.
The proposed approach enables the decision-maker to increase the revenue

from ticket sales by increasing the price of tickets on Saturday and Sunday and
its reducing on Monday and Wednesday.
4 Intuitionistic Fuzzy One-Way ANOVA

We propose the IFANOVA of [20], based on IM concepts and fuzzy logic. Some of
the IMs operations are performed on IndMatCalc v0.9 [23] and a part of Visual
Studio.NET 2010 C project’s program code is used.
Step 1. If the data are fuzzy or intuitionistic fuzzy then go to Step 2, else {Let us
transform the data values for m levels of an investigated factor into intuitionistic
fuzzy as follows: let us have the set of intervals: [i1 , iI ] for 1 ≤ i ≤ m and
A = min xi,j < max xi,j = Bi .
let i For interval [i1 , iI ] we construct IFPs [3] as
i1 ≤j≤iI i1 ≤j≤iI
follows:
xi,j − Ai Bi − xi,j
μi,j = , νi,j = (4)
Bi − Ai Bi − Ai
that satisfy the conditions 0 ≤ μi,j , νi,j ≤ 1, ≤ 0 ≤ μi,j + νi,j ≤ 1; Go to Step 2 }

Step 2. The intuitionistic fuzzy IM (IFIM) X[K, L] is created, whose elements
are the measured values according to the different levels of the studied factor.
l1 ... ln Sri Sr
k1 μk1 ,l1 , νk1 ,l1 . . . μk1 ,ln , νk1 ,ln μk1 ,Sri , νk1 ,Sri μk1 ,Sr , νk1 ,Sr
X= . .. .. .. .. .. ,
.. . . . . .
km μkm ,l1 , νkm ,l1 . . . μkm ,ln , νkm ,ln μkm ,Sri , νkm ,Sri μkm ,Sr , νkm ,Sr
where K = {k1 , k2 , . . . , km }, L = {l1 , l2 , . . . , ln , Sri , Sr} and for 1 ≤ i ≤ m, 1 ≤ j ≤

n, {xki ,lj , xki ,Sri , xki ,Sr } are IFPs. We use symbol “⊥” for empty cells of IM X.
Let us define the auxiliary IM S = [K, L, {ski ,lj }], such that S = X i.e. (ski ,lj =
xki ,lj ∀ki ∈ K, ∀lj ∈ L). Then we define 2 − D IMs S1 [K, L/{Sr1 , Sr}] =
prK,L/{Sr1 ,Sr} S, S2 [K, {Sr1 }] = αL,#1 (S1 , Sr1 ). Using the operation “addition”
we obtained a new form of IM S1 as follows: S := S ⊕(max,min)) S2 , go to Step 3.
Step 3. We calculate the mean of the sample by the aggregation operation (2)
S3 [k0 , {Sr1 }] = αK,#1 (S2 , k0 )(Sr1 ∈
/ L/{Sr1 , Sr}).

for (int i = 0; i < m; i + +) {S := S ⊕(max,min)) kk0i ; Sr
Sr S3 }.
1
From each element of the matrix S, subtract the mean of the data in each
row of the IM S corresponding to a given factor level:
for (int i = 0; i < m; i + +)
for (int j = 0; j < n; j + +)

{S4 = IO−(max,min) ki , lj , prK,{L/{Sr,Sr1 } S , ki , Sr1 , prK,{Sr1 }S }.
1
M SE = AGIO⊕(max,min) ) S4 ⊗(min,max) S4 . Go to Step 4.
N − km
Step 4. We calculate the mean sums of squares M SC by the operations:
– S5 [K, {Sr1 }] = prK,{Sr1 } S4 and S6 [K,

{Sr}] = prK,{Sr} S4 ;
Let Dif ∈/ K ∪ L then S7 [K, Dif ] = ⊥; Dif Sr 1 S −
5 (max,min)) ⊥; Dif
Sr S6 ;
–
1
M SC = AGIO⊕(max,min) ) S7 ⊗(min,max) S7 ; Go to Step 5.
km − 1
Step 5. We calculate the reciprocal of the Fisher’s coefficient by:

1/F = M SE/M SC. Fuzzy estimator of ANOVA key statistic
Ff uzzy(1−α,N −km ,m−1) value is obtained using Pietraszek’s approach ([17], 2016).
The Fig. 2 presents the fuzzy estimators of the bootstrapped F statistics:
Fig. 2. Fuzzy estimator of the bootstrapped F statistics
If 1/F ≤ Ff uzzy(1−α,N −km ,km −1) in accordance with the relations (1) and (3),
then hypotesis H0 is rejected on significance level α, else H0 is accepted on level
α. End of algorithm.
The effectiveness of proposed IFANOVA is tested by its application to detect
dependencies between ticket sales as IFPs and the factor “day of a week”. At
the beginning of the algorithm we transform the data values with monthly sales
by a day of a week for the movie “Coziness” into IFPs as the elements of IFIM
X[K, L] using (4). The initial form of the IM X is:
l1 l2 l3 l4 l5 Sr1 Sr
M onday 0, 14; 0, 86 0, 37; 0, 63 0, 06; 0, 94 0, 04; 0, 96 ⊥; ⊥ ⊥; ⊥ ⊥; ⊥
T uesday 0, 41; 0, 59 0, 17; 0, 83 0, 14; 0, 86 0, 10; 0, 90 ⊥; ⊥ ⊥; ⊥ ⊥; ⊥
W ednesday 0, 13; 0, 87 0, 08; 0, 92 0, 03; 0, 97 0; 1 ⊥; ⊥ ⊥; ⊥ ⊥; ⊥
T hursday 1; 0 0, 90; 0, 10 0, 41; 0, 59 0, 25; 0, 75 ⊥; ⊥ ⊥; ⊥ ⊥; ⊥
F riday 0, 53; 0, 47 0, 42; 0, 58 0, 20; 0, 80 0, 076; 0, 93 0, 04; 0, 96 ⊥; ⊥ ⊥; ⊥
Saturday 0, 39; 0, 61 0, 64; 0, 36 0, 41; 0, 59 0, 18; 0, 82 0, 13; 0, 87 ⊥; ⊥ ⊥; ⊥
Sunday 0, 56; 0, 44 0, 66; 0, 34 0, 31; 0, 69 0, 11; 0, 89 0, 02; 0, 98 ⊥; ⊥ ⊥; ⊥
After application of the IFANOVA, presented in this Sect. 4, we found that:
0, 15; 0, 85 0, 41; 0, 59

M SC = = 0, 027; 0, 97, M SE = = 0, 022; 0, 98.
6 24
We apply Pietraszek’s approach ([17], 2016), which led to the obtaining of fuzzy
estimator of ANOVA key statistics F value. The classic value F (0, 95; 24; 6) =
0, 39 is related to the fuzzy assessment Ff uzzy(0,95;24;6) = 0, 93; 0. Therefore
0, 79; 0, 21 = M SE/M SC = 1/F ≤ Ff uzzy(0,95;24;6) = 0, 93; 0 in accordance
with the relation (1). It is found from (3) that the “day of a week” factor affects
the amount of ticket sales for the movie “Coziness”. A comparison of the results of
IFANOVA with those obtained by ANOVA shows that the results coincide.
5 Conclusion
The paper proposed one-way IFANOVA, based on the concepts of IMs and IFSs,
to analyze the dependencies of intuitionistic fuzzy data on single factor levels.
Decision maker may use the proposed ANOVA approach to vague real life prob-
lems to make effective decisions. In future, the outlined approach for IFANOVA,
will be extended to apply to the interval-valued intuitionistic fuzzy sets [5].
References
1. Atanassov, K.: Intuitionistic fuzzy sets. In: Proceedings of VII ITKR’s Session,
Sofia (1983). (in Bulgarian)
2. Atanassov, K.: Generalized index matrices. Comptes rendus de l’Academie Bulgare
des Sciences 40(11), 15–18 (1987)
3. Atanassov, K.: On Intuitionistic Fuzzy Sets Theory. STUDFUZZ, vol. 283.
4. Atanassov, K.: Index Matrices: Towards an Augmented Matrix Calculus. Studies
in Computational Intelligence, vol. 573. Springer, Cham (2014)
5. Atanassov, K., Gargov, G.: Interval valued intuitionistic fuzzy sets. Fuzzy Sets
Syst. 31(3), 343–349 (1989)
6. Atanassov, K., Szmidt, E., Kacprzyk, J.: On intuitionistic fuzzy pairs. Notes Intu-
itionistic Fuzzy Sets 19(3), 1–13 (2013)
7. Atanassov, K.: Remark on an intuitionistic fuzzy operation “division”. Ann. “Inf.”.
Section Union Scientists Bulgaria 10 (2019, in press)
8. Buckley, J.J.: Fuzzy Probability and Statistics. Springer, Heidelberg (2006)
9. Doane, D., Seward, L.: Applied Statistics in Business and Economics. McGraw-Hill
Education, New York (2016)
10. Fisher, R.: Statistical Methods for Research Workers, London (1925)
11. Gil, M.A., Montenegro, M., González-Rodrı́guez, G., Colubi, A., Casals, M.R.:
Bootstrap approach to the multi-sample test of means with imprecise data. Com-
put. Stat. Data Anal. 51, 148–162 (2006)
12. Kalpanapriya, D., Pandian, P.: Fuzzy hypotesis testing of ANOVA model with
fuzzy data. Int. J. Mod. Eng. Res. 2(4), 2951–2956 (2012)
13. Kalpanapriya, D., Unnissa, M.: Intuitionistic fuzzy ANOVA and its application
using different techniques. In: Madhu, V., Manimaran, A., Easwaramoorthy, D.,
Kalpanapriya, D., Mubashir Unnissa, M. (eds.) Advances in Algebra and Analysis.
Trends in Mathematics, pp. 457–468. Birkhäuser, Cham (2017)
14. Marques de Sá, J.: Applied statistics using SPSS, STATISTICA, MATLAB and
R. Springer, Heidelberg (2007)
15. Nourbakhsh, M.R., Parchami, A., Mashinchi, M.: Analysis of variance based on
fuzzy observations. Int. J. Syst. Sci. 44(4), 714–726 (2013)
16. Parchami, A., Nourbakhsh, M., Mashinchi, M.: Analysis of variance in uncertain
environments. Complex Intell. Syst. 3(3), 189–196 (2017)
17. Pietraszek J., Kolomycki M., et al.: The fuzzy approach to assessment of ANOVA
results. In: Nguyen, NT., et al. (eds.) Computational Collective Intelligence. ICCCI
2016. LNCS, vol. 9875, pp. 260–268. Springer, Cham (2016)
18. Szmidt, E., Kacprzyk, J.: Amount of information and its reliability in the ranking
of Atanassov intuitionistic fuzzy alternatives, In: Rakus-Andersson, (eds.) Recent
Advances in Decision Making, vol. 222, pp. 7–19. SCI, Springer, Heidelberg (2009)
19. Traneva, V.: Internal operations over 3-dimensional extended index matrices. Proc.
Jangjeon Math. Soc. 18(4), 547–569 (2015)
20. Traneva, V., Tranev, S.: Inuitionistic Fuzzy Anova Approach to the Management
of Movie Sales Revenue. Studies in Computational Intelligence (2020, in press)
21. Traneva, V., Tranev, S., Stoenchev, M., Atanassov, K.: Scaled aggregation opera-
tions over 2- and 3-dimensional IMs. Soft. Comput. 22(15), 5115–5120 (2018)
22. Zadeh, L.: Fuzzy Sets. Inf. Control 8(3), 338–353 (1965)
23. Software for index matrices. http://justmathbg.info/indmatcalc.html. Accessed 20
Mar 2020
Interval Valued Intuitionistic Fuzzy Gaussian
Membership Function: A Novel Extension
Janani Bharatraj(&)
Hindustan Institute of Technology and Science, Chennai 603103, India

[email protected]
Abstract. The Interval-valued intuitionistic fuzzy sets (IVIFSs) have long been
used to model vagueness and have been used in decision making, pattern
recognition, image processing, and other applications. In this study, IVIFSs are
defined using Gaussian membership functions (GMFs), and new measures of the
distance, the overlap, and the angle between two sets are developed. The pro-
posed methodology is used to determine the similarities between test subjects in
genetic brain profiling.
Keywords: Bhattacharya angle Bhattacharya distance Gaussian

membership function Genetic brain profiling Interval valued intuitionistic
fuzzy sets
1 Introduction
Fuzzy set theory and its extensions have gained popularity in various areas of science,
engineering, and technology since their inception in 1967. Fuzzy set theory and its
extensions are founded on three main concepts, namely, imprecision, uncertainty, and
vagueness. Uncertainty and vagueness form an integral part of type-1 and type-2 fuzzy
sets, interval-valued fuzzy sets (IVFSs), and intuitionistic fuzzy sets (IFSs). Atanassov
combined IVFSs and IFSs to create a more effective method of modelling vagueness
known as interval-valued intuitionistic fuzzy sets (IVIFSs). Since its inception, IVIFSs
have been applied to various areas of study like decision making and pattern recog-
nition techniques. Distance measures on IVIFSs based on all three degrees of freedom
have been studied for use in pattern recognition. The present study explores IVIFSs
using an alternative membership function. The remainder of the paper is organized as
follows. Section 2 provides a literature review of similarity measures on IVIFSs with
applications in pattern recognition. Section 3 identifies areas for improvement in the
existing methods. Section 4 defines certain properties required for developing an
extended version of the IVIFS with Gaussian membership functions (IVIFS-GMFs).
Section 5 introduces IVIFS-GMFs and defines a distance, or similarity, measure for
IVIFS-GMFs, and a simple example of pattern recognition using IVIFS-GMFs is
Interval Valued Intuitionistic Fuzzy Gaussian Membership Function 373
2 Literature Review
Yuan and Li [4] developed five-valued fuzzy sets that constituted the threshold sets of
IVIFSs. The properties of these sets were found to be consistent with the properties of
fuzzy sets and intuitionistic fuzzy sets. Yuan et al. [8] further developed cut sets on
IVIFSs and derived representation and decomposition theorems for interval-valued and
intuitionistic fuzzy sets. Adak et al. [1] introduced interval cut sets of generalized
IVIFSs and studied their properties using fuzzy arithmetic. Taking a geometric inter-
pretation of IVIFSs, Zeshui [9] defined distance measures such as the Hamming dis-
tance and the Euclidean distance with the use of a Hausdorff metric. Weighting factors
on the elements were defined, and weighted distance measures between IVIFSs were
developed. An example problem, the classification of building materials, was presented
to demonstrate the approach. Zhang et al. [11] constructed a novel pattern recognition
approach for measuring the degree of similarity between IVIFSs by choosing a dif-
ferent weight for each feature according to its dissimilarity with other features. Zhang
et al. [10] constructed a new pattern recognition approach with IVIFSs by defining
separate Hausdorff metrics for lower membership functions and upper membership
functions and deriving different similarity measures for each of the functions. The effect
of using both the measures in combination was then tested on a pattern recognition
problem. Wei et al. [6] proposed an entropy measure on IVIFSs that generalized the
entropy measures on IFSs. New similarity measures were constructed using the new
entropy measures, and the effectiveness of these measures were demonstrated using a
pattern recognition example. Singh [5] investigated a cosine similarity measure for
IVIFSs. An example of pattern recognition in medical diagnostics was used to test the
effectiveness of the cosine measure. It was found that the new cosine similarity measure
provided better results than distance-based similarity measures. Zhang et al. [12]
introduced the dynamic IVIFS, which was shown to be more accurate in a medical
diagnostics and decision-making example. The technique was more comprehensive and
more flexible than IFS and IVIFS methods. Meng and Chen [4] defined an entropy
measure using Shapley-weighted similarity measures that included the interdepen-
dencies of the elements in a set. This method was applied to a pattern recognition
problem with the multi-criteria decision-making technique.
3 Research Statement and Objectives
The literature review in the previous section shows that alternative methods are
required. One possibility is to define IVIFSs using Gaussian membership functions and
develop distance and similarity measures between IVIFS-GMFs.
Thus, the objectives of this research included the following:
• define IVIFSs using Gaussian membership functions (IVIFS-GMFs).
• derive distance measures for IVIFS-GMFs.
• define similarity measures for IVIFS-GMFs.
374 J. Bharatraj
4 Preliminaries
In this section, we provide theoretical background material for developing the new
distance measures.
Definition 4.1. Gaussian membership functions (GMFs).
1 xc 2
In fuzzy logic, the function f ðx; r; cÞ ¼ e2ð r Þ defines a Gaussian membership
function, where c is the centre point of the membership function and 2r indicates the
span of the curve.
Definition 4.2. Interval-valued intuitionistic fuzzy set (IVIFS).
Let X be a universal set and let ½I denote the set of all closed subintervals of ½0; 1.
A set A in X is said to be an IVIFS if A ¼ fhx; lA ð xÞ; tA ð xÞi=x 2 X g where
lA ; tA : X ! ½I , lA ð xÞ ¼ ½lL ð xÞ; lU ð xÞ, tA ð xÞ ¼ ½tL ð xÞ; tU ð xÞ, And 0 lU ð xÞ þ
tU ð xÞ 1 for any x 2 X [2].
4.1 Bhattacharya Distance Between Two Distributions

The Bhattacharya distance gives a measure of the similarity between two probability
distributions, and the Bhattacharya coefficient gives a measure of the overlap between
two Gaussian distributions.
Definition 4.3. Bhattacharya distance.
The Bhattacharya distance is defined as
!! 2 !
1 rp rq lp lq
2 2
1 1
DB ðp; qÞ ¼ ln þ þ2 þ
4 4 r2q r2p 4 r2p þ r2q
where p and q are two distributions, r2p and r2q are the variances of the distributions p
and q, respectively, and lp and lq are the means of the distributions p and q,
respectively [3].
Definition 4.4. Bhattacharya coefficient (BC).
The Bhattacharya coefficient is defined using the Bhattacharya distance and is given
by
DB ðp; qÞ ¼ lnðBC ðp; qÞÞ;
BC ðp; qÞ ¼ eDB ðp;qÞ :
It can be shown that 0 BC 1 and 0 DB 1.

The Bhattacharya distance does not satisfy the triangle inequality.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Hence, the Hellinger distance 1 eDB ðp;qÞ , which does satisfy the triangle
inequality, was chosen as the distance metric [3].
Definition 4.5. Bhattacharya angle [3].

The Bhattacharya angle gives the angle between two probability measures and is
defined as
Dðp; qÞ ¼ arccos eDB ðp;qÞ :
5 Interval-Valued Intuitionistic Gaussian Membership

Functions
In this section, we present the IVIFS with Gaussian membership functions. The fol-
lowing distance measures were developed for IVIFS-GMFs:
• the Bhattacharya distance for IVIFS-GMFs and cut sets on an IVIGMF.
• an overlap measure for IVIFS-GMFs.
• a measure of the angle between two IVIFS-GMFs.
Definition 5.1. Interval-valued intuitionistic Gaussian membership function (Fig. 1).

The membership, non-membership, and hesitancy functions are, respectively,
h i
lA~ G ð xÞ ¼ lLA~ G ð xÞ; lU~
AG
ð x Þ ;
h i
mA~ G ð xÞ ¼ mLA~ G ð xÞ; mU
~ G ð xÞ ;
A
h i
pA~ G ð xÞ ¼ pLA~ G ð xÞ; pU~ G ð xÞ :
A
and satisfy the following conditions:

Membership:
1 xc 2 1 xc 2
lLA~ G ð xÞ ¼ u e2ð r Þ ; 0\u\1; lU þ 2ð r Þ
~ G ð xÞ ¼ u e
A
;
Non-Membership:
mLA~ G ð xÞ ¼ v ð1 lAL Þ; mU þ
~ G ð xÞ ¼ v ð1 lAU Þ; 0\v\u \1;
A
Hesitancy:
pLA~ G ð xÞ ¼ 1 lLA~ G ð xÞ mLA~ G ð xÞ; pU U

~ G ð x Þ ¼ 1 lA
A
U
~ G ð xÞ mA
~ G ð xÞ:
376 J. Bharatraj
Definition 5.2. Distance, overlap, and angle measures for IVIFS-GMFs.

The Bhattacharya distance for IVIFS-GMFs can be defined as
2 n 2 o 3
r r2
ln 14 rA2 þ rB2 þ 2 þ r2 þ1 r2
62 B A ð A BÞ 37
6 ðc c Þ2 7
~ 1 6 A B 7
dBh ðA; BÞðxi Þ ¼ 6 6 277
8 6 6 þ fðlAL ðxi Þ lBL ðxi ÞÞ þ ðlAU ðxi Þ lBU ðxi ÞÞg 7 7
4 4 þ fðt ðx Þ t ðx ÞÞ þ ðt ðx Þ t ðx ÞÞg2 5 5
AL i BL i AU i BU i
þ fðpAL ðxi Þ pBL ðxi ÞÞ þ ðpAU ðxi Þ pBU ðxi ÞÞg2
X
n
~ Bh ðA; BÞ ¼ 1
D d~Bh ðA; BÞðxi Þ
8 i¼1
The overlap of two IVIFS-GMFs can be defined as
CBh ðp; qÞ ¼ eDBh ðp;qÞ
The Bhattacharya angle between two IVIFS-GMFs is given by
DðA; BÞ ¼ arccos CBh ðA; BÞ
Fig. 1. IVIFS with Gaussian membership functions for c ¼ 0; r ¼ 2
The Bhattacharya distance for IVIFS-GMFs clearly satisfies the requirements for a
distance metric.
6 Example: Similarities in Two Individuals Using Genetic

Brain Profiling and IVIFS-GMF
Genetic Brain Profiling (GBP) is a unique assessment tool that provides individuals
with greater knowledge of their traits and their most beneficial environment. This tool
identifies innate talents, which assists in developing the full potential of an individual
and identifies the most beneficial educational system and environment for an indi-
vidual. GBP identifies possible behavioural and learning difficulties and academic
strengths and weaknesses. In addition, GBP provides information on how an individual
receives social and educational information. Furthermore, GBP identifies interpersonal
communication styles, the causes of stress, and an individual’s tendencies for handling
stress. GBP is based on the principles of the behavioural sciences and applied psy-
chology. This approach provides a procedure to map the learning, cognitive, and action
preferences of an individual and their effects on an individual’s behaviour. GBP
analyses fingerprints to identify an individual’s natural abilities, talents, and personality
traits. Learning style, personality, and aptitude can be identified in an individual as
early as age two with GBP, and this information can give parents a better understanding
of their child.
7 Materials, Method and Result
To test IVIFS-GMFs with the Bhattacharya distance using GBP analysis as an

example, the fingerprints of two individuals were obtained and the profiles were cre-
ated. Educational performance data were chosen as test samples.
Table 1 lists the performance in three levels of formal learning, namely, primary,
secondary, and tertiary, or university, education. The performance data are comfort
levels expressed as a percentage at each level.
Table 1. Performance at three levels of formal learning

Education level Primary Secondary University
Individual 1 30.51 36.5 32.99
Individual 2 33.49 31.86 34.65
Table 2 gives the scores for the primary sensory learning styles. These scores
indicate the comfort levels of an individual in using their various physical senses.
Scores are obtained for three styles of learning: kinaesthetic, auditory, and visual. An
analysis of these scores indicates the degree to which an individual is reflective,
affective, and cognitive; the analysis results are included in Table 2.
Table 3 provides the intelligence profiles for the two test samples, where several
types of intelligence such as logic, verbal, nature, and music were assessed.
Table 4 provide the four DISC factors, which summarize an individual’s profile.
DISC is an acronym for dominance, influence, steadiness, and compliance.
378 J. Bharatraj
Table 2. Primary sensory learning styles.

Learning Kinaesthetic Auditory Visual
Individual 1 26.98 29.89 43.13
Individual 2 32.77 33.45 33.78
Domains Reflective Affective Cognitive
Individual 1 40 30 30
Individual 2 20 70 10
Table 3. Genetic Intelligence profiles of test samples.

Intelligence category Logic Verbal Self People Body
Individual 1 12.5 12.7 11.6 9.4 9.7
Individual 2 10.2 8.9 10.2 10.7 10.7
Intelligence category Nature Music Picture Cosmic
Individual 1 12.6 9.7 9.9 11.8
Individual 2 12.2 13.1 14.7 9.1
Table 4. Genetic behaviour profiles.

GBP Persistence Sensitivity Self Thoughtfulness Enthusiasm Self-
confidence motivation
Individual 1 1.16 0.87 0.91 1.1 1.05 1.17
Individual 2 1.39 0.72 0.97 1.03 1.35 0.81
Intelligence Patience Cooperativeness Independence Friendliness Accuracy Efficiency
category
Individual 1 0.86 0.74 1.35 0.78 0.95 1.28
Individual 2 1.23 0.89 1.13 1.2 0.75 0.83
The DISC factors provide a useful technique for extracting information that describes
the relationships between various pairs of factors.
Experimental Results
The mean and the variance of the data were calculated, and the Bhattacharya distance,
coefficient, and angle were obtained. The results are shown in Table 5.
Table 5. Similarity evaluation of GBPs.

Profiles Bhattacharya distance Overlap Bhattacharya
angle
Radians Degrees
Education 0.4952 0.6094 0.9154 52.45
Learning 4.9977 0.0068 1.564 89.61
Domains 2.1155 0.1206 1.4499 83.07
Intelligence profile 0.1055 0.89 0.4513 25.86
GBP 0.1396 0.8697 0.5161 29.57
Table 5 clearly shows that the two individuals are most similar in their intelligence
and behaviour profiles and least similar in their learning styles and domains.
The results were found to be consistent with the profiler’s interpretation of the
results.
8 Conclusion and Future Direction
Interval-valued intuitionistic fuzzy sets have been applied to problems in distance-

based pattern recognition and in evaluating similarities between samples. In this
research, we defined interval-valued intuitionistic Gaussian membership functions and
defined new measures of the distance, the overlap, and the angle between two IVIFS-
GMFs. The proposed method was used to analyse samples of Genetic Brain Profiling
(GBP) to determine the similarity between two individuals. IVIFS-GMFs can be
applied in image segmentation, pattern recognition, and MCDM problems where
Bhattacharya angle and distance are widely used. Further, this formula can be used to
study the Gaussian distribution curve of the COVID 19 infected countries and the
uncertainty areas can be addressed.
Acknowledgements. The author is sincerely grateful to Mrs. Rekha Hariprawin, SARE Image
Consultancy, Chennai, for providing the GBP samples, which were provided in collaboration
with MIDNA Global (www.midnaglobal.com), India.
References
1. Adak, A.K., Bhowmik, M., Pal, M.: Interval cut-set of generalised interval-valued
intuitionistic fuzzy sets. Int. J. Fuzzy Syst. Appl. 2, 35–50 (2012)
2. Atanassov, K., Gargov, G.: Interval-valued intuitionistic fuzzy sets. Fuzzy Sets Syst. 31,
343–349 (1989)
3. Bhattacharyya, A.: On a measure of divergence between two statistical populations defined
by their probability distributions. Bull. Calcutta Math. Soc. 35, 99–109 (1983)
4. Meng, F., Chen, X.: Entropy and similarity measure for Atanassov’s interval valued
intuitionistic fuzzy sets and their application. Fuzzy Optim. Decis. Making 15, 75–101
(2016)
5. Singh, P.: A new method on measure of similarity between interval-valued intuitionistic
fuzzy sets for pattern recognition. J. Appl. Comput. Math. 1, 1–5 (2012)
6. Wei, C.P., Wang, P., Zhang, Y.Z.: Entropy, Similarity measure of interval valued
intuitionistic fuzzy sets and their applications. Inf. Sci. 181, 4273–4286 (2011)
7. Yuan, X.H., Li, H.X.: Cut sets on interval-valued intuitionistic fuzzy sets. In: Proceedings of
the 6th International Conference on Fuzzy Systems and Knowledge Discovery, pp. 167–171
(2009)
8. Yuan, X., Li, H., Sun, K.: The cut sets, decomposition theorems and representation theorems
on intuitionistic fuzzy sets and interval-valued fuzzy sets. Sci. China Inf. Sci. 54, 91–110
(2010)
9. Zeshui, X.: On Similarity measures of interval-valued intuitionistic fuzzy sets and their
applications to pattern recognition. J. Southeast Univ. 23, 139–143 (2007)
380 J. Bharatraj
10. Zhang, Q.S., Yao, H.X., Zhang, Z.H.: Some similarity measures of interval valued
intuitionistic fuzzy sets and application to pattern recognition. Appl. Mech. Mater. 44–47,
3888–3892 (2011)
11. Zhang, Y., Ma, P., Su, X.: Pattern recognition using interval-valued intuitionistic fuzzy set
and its similarity degree. In: Proceedings of the IEEE International Conference on Intelligent
Computing and Intelligent Systems, Shanghai, China (2009)
12. Zhang, Z., Wang, M., Hu, Y., Ye, Y., Li, Y.: A dynamic interval-valued intuitionistic fuzzy
sets applied to pattern recognition. Math. Probl. Eng. 2013, 1–16 (2013)
An Integrated IVIF-DEMATEL
and IVIF-TOPSIS Methodology for Hotel
Information System Selection
Gizem Erkal1,2 , Huseyin Selcuk Kilic1 ,

Zeynep Tugce Kalender1 , Ahmet Selcuk Yalcin2 ,
and Gulfem Tuzkaya1(&)
1
Industrial Engineering Department, Marmara University,
[email protected], {huseyin.kilic,tugce.simsit,
gulfem.tuzkaya}@marmara.edu.tr
2
Industrial Engineering Department, Okan University, 34959 Istanbul, Turkey
[email protected]
Abstract. The importance of information technologies in hotels has increased

more in recent years due to the variety of required processes and the level of
competitiveness in the sector. Depending on their contribution to creating value
and increasing efficiency, hotel enterprises make important investments in
information technologies. However, there are various criteria and alternatives in
the selection process of the most suitable hotel information system. Moreover,
the criteria affect each other in a fuzzy decision-making environment. Hence,
two techniques are utilized synergistically in this study. IVIF-DEMATEL is
initially performed to obtain the importance degrees of the criteria. Afterward,
IVIF-TOPSIS is applied to reach the final rankings. An application is performed
considering the case in the Turkish tourism sector, and the alternatives are
ranked.
Keywords: Hotel information system Channel manager Interval-valued

intuitionistic fuzzy sets DEMATEL TOPSIS
1 Introduction
Depending on the increasing trend in the tourism sector, the number of academic
studies has increased in the last years. There are several topics covered in the literature
dealing with the hotel industry, including hotel service quality evaluation, hotel loca-
tion selection, determination of hotel staff, and so on. In the evaluation and selection of
specific alternatives in the literature, many multi-criteria decision-making techniques
were applied as follows. To solve a hotel selection problem, Isik and Adali [1] pro-
posed a novel integrated approach based on SWARA and OCRA methods. Vujoševic
and Popovic [2] used an integrated PROMETHEE and BORDA method for the
selection of the most appropriate alternative of the hotel building with the atrium. Yu
et al. [3] used an extended VIKOR approach for the determination of hotels on their
382 G. Erkal et al.
websites. Peng et al. [4] introduced the Heronian mean tool for addressing hotel
selection issues on TripAdvisor and recommended two information fusion operators to
establish a multi-criteria group decision-making method. Kwok and Lau [5] proposed a
methodology called Vague Set TOPSIS to help tourists to sort hotel choices. Akincilar
and Dagdeviren [6] proposed a method that assesses the quality of hotel management
websites and used AHP and PROMETHEE. Chou et al. [7] developed a multi-criteria
decision-making approach based on fuzzy AHP method for the selection of an inter-
national tourist hotel location. Ji et al. [8] investigated the hotel location selection
problem, and multi-hesitant fuzzy linguistic term sets was proposed to indicate cog-
nitive knowledge. Afterward, an ELECTRE III based approach is used for hotel
location selection. Mahalik and Satpathy [9] used fuzzy AHP to help DMs of the hotel
industry to choose an appropriate supplier. Pahari et al. [10] used intuitionistic fuzzy
sets with TOPSIS method for hotel selection based on online reviews on the tourism
website. Roy et al. [11] used weighted interval rough number with the COPRAS
method. Various criteria were evaluated by the WIRN method and weighted, and then
used to rank hotels. Sin et al. [12] proposed an integrated DEMATEL and ANP
approach to determine the most appropriate total quality management practices in the
hotel industry. Kumru and Kumru [13] proposed a hybrid MCDM approach by using
fuzzy AHP and fuzzy TOPSIS for the evaluation of a set of thermal hotel alternatives.
Popovic et al. [14] proposed an integrated multi-criteria decision-making methodology
for the selection of hotel locations. The adapted SWARA method was employed to
calculate the criteria weights. The weighted-sum method was used for the assessment
and selection of the optimum location.
The results of the literature survey are given as follows. Although several issues are
addressed about hotels, to the best of authors’ knowledge, there is no study about the
selection and evaluation of hotel information systems, which play an important role in
the success of the hotels. In this study, channel manager, which is an essential part of
hotel information systems, is the focus on specifically. Channel managers are tools that
enable to update several reservation channels simultaneously and save a lot of time. In
the absence of them, it takes hours to make the required changes concerning price and
availability. Hence, “channel manager” is a must for hotels to survive on online
channels. Moreover, it has connections with international channels, therefore, provides
access to foreign guests. Considering the critical role of hotel information systems in
the success of hotels, the selection of proper “channel manager” becomes very
important. Hence, this study will fulfill the gap in that area with a robust integrated
methodology, including IVIF-DEMATEL and IVIF-TOPSIS techniques. The rationale
behind the integration of DEMATEL and TOPSIS methodologies in this study is to
provide a practical method in which interactions between criteria are handled. It has
been proven in the literature that integrating and using DEMATEL and TOPSIS
methodologies together give successful results in multi-criteria decision-making
problems, especially in fuzzy environments [15–20].
The remaining of this paper is organized as follows. Section 2 gives the funda-
mentals of integrated IVIF-DEMATEL and IVIF-TOPSIS techniques. Turkish tourism
sector application is given in Sect. 3. Finally, in Sect. 4, concluding remarks are given.
An Integrated IVIF-DEMATEL and IVIF-TOPSIS Methodology 383
2 The Proposed Methodology
In real-life applications, it is a difficult process for the DMs to pinpoint the degree of
membership, uncertainty, or hesitation. Hence, over time, the scope of the fuzzy theory
has been expanded. Atanassov and Gargov [21] propose an expansion of intuitionistic
fuzzy sets and develop interval valued intuitionistic fuzzy set (IVIFs) theory. For
details of the operational laws of IVIF sets, Xu [22] can be referenced. In this study,
integrated IVIF-DEMATEL and IVIF-TOPSIS methods are used. The DEMATEL
methodology was developed in the 1970s [15] to solve complex problems, and have
been applied in various papers [23–28]. In [16, 29], the DEMATEL is used to find the
weights of the criteria. Steps (adapted from [28] and [30]) are given below.
Step 1. Compute the direct-relation matrix (Z)

Direct-Relation Matrix Z is an n n matrix obtained by pairwise comparison to
determine criteria relations. For instance, zij value is assigned as the affect degree of ith
criterion on the jth criterion.
Step 2. Normalize the direct-relation
( matrix (D) !
n
n
)
P P
EðAÞ
D ¼ s where s ¼ max max1 i n zij ; max1 i n zij and
j¼1 i¼1

1 Xn 2 lLi ð xÞ vLi ð xÞ lU U L U
i ð xÞ vi ð xÞ þ pi ð xÞ þ pi ð xÞ
E ð AÞ ¼ i¼1 2 þ jlL ð xÞ vL ð xÞj þ jlU ð xÞ vU ð xÞj þ pL ð xÞ þ pU ð xÞ
ð1Þ
n i i i i i i
Step 3. Compute the total-relation matrix (T)

T ¼ tij nxn i; j ¼ 1; 2; . . .:; n and T ¼ D þ D2 þ D3 þ D4 þ . . . þ Dm
¼ DðI DÞ1 when m ! 1 ð2Þ
Step 4. Find out row and column totals of Matrix T

The sum of rowi indicates the total influence of ith criterion on other criteria and
hXn i
D ¼ ½di nx1 ¼ t
j¼1 ij
ð3Þ
nx1
The sum of columnj indicates the total influence of other criteria on the jth criterion and
hXn i
R ¼ rj nx1 ¼ i¼1
t ij ð4Þ
1xn
Step 5. Determine criteria weights ðWi Þ
xi
1=2
W i ¼ Pn where xi ¼ ðDi þ Ri Þ2 þ ðDi Ri Þ2 ð5Þ
i¼1 xi
384 G. Erkal et al.
Main goal of TOPSIS is to find the best alternative which has the shortest distance
to the positive ideal solution and farthest distance to the negative ideal solution [31]. In
this study, the IVIF-TOPSIS method is used to handle vagueness inherent to the
decision environment as in [32–34]. Steps of the IVIF-TOPSIS methodology are given
below (Adapted from Ye [35])
Step 1: Determine the problem n 0 o

0
X ¼ Xi i ¼ 1; 2. . .; n be a set of candidates and A ¼ Aj j ¼ 1; 2. . .; m be a set
0 0
of attributes.
Step 2: Determine positive ideal solution (PIS) and negative ideal solution (NIS)
For each decision-maker, determine (k = 1, 2, …, K);
~k ~k þ ~k þ

~k þ
Dh i h iE
~
Rk þ ¼ r1 ; r2 ; . . .; rm where rj maxi akij ; maxi bkij ; mini ckij ; mini dijk
þ
~k ~k ~k

~k
Dh i h iE ð6Þ
~
Rk ¼ r1 ; r2 ; . . .; rm where rj ¼ mini akij ; mini bkij ; maxi ckij ; mini dijk
Step 3: Calculate the separation measures

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n o
1 Xm ~ ~ ~ ~ ~ ~k þ 2
Dki þ ¼ J¼1 j
w ðakij ajk þ Þ2 þ ðbkij bkj þ Þ2 þ ðckij ckj þ Þ2 þ ðdijk djk þ Þ2 þ ðpkijL pkjLþ Þ2 þ ðpkijU pjU Þ
2
ð7Þ
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n o
1 Xm ~ ~ ~ ~ ~k 2 ~k 2
Dki ¼ J¼1
wj ðakij ajk Þ2 þ ðbkij bjk Þ2 þ ðckij cjk Þ2 þ ðdijk djk Þ2 þ ðpkijL pjL Þ þ ðpkijU pjU Þ
2
ð8Þ
Step 4: Aggregate separation measures

XK XK
Diþ ¼ k Dk þ and D
k¼1 k i i ¼ k¼1
kk Dki ð9Þ
where i = 1, 2, … n; k = 1, 2, … K and kk is the weight of decision-maker k.

Step 5: Compute the closeness coefficient
D
i
Ui ¼ where 0 Ui 1 ð10Þ
Diþ þ D
i
3 Application of the Proposed Methodology
Online travel agencies are controlling the travel market and the majority of hotel
reservations made from these channels. From the hoteliers’ point of view, being visible
on many channels is a vital issue to survive among the rivals. On the other hand, having
connections in many channels cause some problems and makes it difficult to control
information such as price and availability. In this regard, a tool called channel manager,
which is used for updating and managing all online channels where a hotel operates,
has become a must. Steps of the methodology and the application are given below:
Step 1. Determine the problem and alternatives: This study focused on the selection of
the most appropriate channel manager tool for the hotel information system. Five
alternatives are determined as Hotellinkage Pro (A1), Synexis (A2), SiteMinder (A3),
Travelclick (A4), and Vertical (A5). The evaluation criteria are Stability of Infras-
tructure (C1), Number of Connected Channels (C2), Integration with PMS System
(C3), Appearance (C4), Reliability (C5), User Friendliness (C6), Speed (C7), Cost
(C8), B2B and B2C Separation (C9), Reporting System (C10), GDS Connection (C11),
Supporting Local Channels (C12), Working Structure (C13), Support (C14), Access
from Mobile Devices (C15).
Step 2. Collect DMs’ opinions for criteria: Initially, an evaluation form is designed,
and opinions of five DMs are collected. In the evaluation form, DMs are asked to
specify the relation between criteria by using the scale given in [28]. In this study, it is
assumed that weights of DMs’ are equal. After collecting DMs’ opinions, collected
evaluations are aggregated by using a weighted averagingP operator defined by Xu [22].
Let w ¼ ðw1 ; w2 ; . . .; wn Þ be the weight vector n where nj¼1 wj ¼ 1 and wj 2 ½0; 1.
The aggregation operator IIFWA is given in Eq. 11.
h Yn Yn i hYn Yn w i
w
IIFWAwð~a1 ; ~a2 ; . . .; ~an Þ ¼ 1 j¼1
aj Þwj ; 1
ð1 ~ j¼1
ð1 ~
bj Þwj ; ~c j ; j¼1 d~j j
j¼1 j
ð11Þ
Step 3. Compute criteria weights by using IVIF-DEMATEL

Step 3.1. Compute the direct-relation matrix (Z): Direct-relation matrix is obtained
at Step 2 of the proposed methodology.
Step 3.2. Normalize the direct-relation matrix (D): Entropy of the matrix Z is
calculated by using Eq. 1, then, s value is determined, and finally, normalized direct-
relation matrix D is obtained and partially presented in Table 1.
Table 1. Normalized direct-relation matrix

C1 C2 C3 C4 C5 C6 … C10 C11 C12 C13 C14 C15
C1 0.000 0.047 0.051 0.046 0.048 0.043 … 0.040 0.042 0.040 0.042 0.042 0.040
C2 0.066 0.000 0.048 0.043 0.054 0.043 … 0.045 0.046 0.041 0.055 0.049 0.050
… … … … … … … … … … … … … …
C15 0.089 0.047 0.080 0.056 0.063 0.052 … 0.058 0.046 0.047 0.077 0.089 0.000
Step 3.3. Calculate total-relation matrix (T): After matrix D is obtained, by using
Eq. 2, T is calculated (Table 2).
386 G. Erkal et al.
Table 2. Total relation matrix

C1 C2 C3 C4 C5 C6 … C10 C11 C12 C13 C14 C15
C1 0.000 0.009 0.010 0.007 0.009 0.007 … 0.006 0.007 0.006 0.008 0.009 0.006
C2 0.017 0.000 0.011 0.007 0.011 0.007 … 0.008 0.008 0.007 0.012 0.011 0.009
… … … … … … … … … … … … … …
C15 0.029 0.012 0.024 0.013 0.016 0.012 … 0.013 0.010 0.010 0.022 0.028 0.000
Table 3. Final weights of each criterion

D+R D−R Weight
C1 0.394 0.181 0.079
C2 0.332 0.050 0.061
… … … …
C15 0.367 −0.119 0.070
Step 3.4. Find out row and column total of Matrix T: After matrix T is obtained, by
using Eq. 3 and 4, D and R values are calculated (Table 4).
Step 3.5. Determine criteria weights: Finally, criteria weights are found by
applying step 5 of IVIF-DEMATEL (Eq. 5) and presented in Table 3.
Step 4. Collect DMs’ opinions for alternatives: A new evaluation form is designed, and
opinions of five DMs are collected for alternatives. The evaluation scale given in [28] is
used.
Step 5. Rank alternatives by using IVIF-TOPSIS: After criteria weights are found in
Step 3 and opinions of DMs are collected in Step 4, IVIF-TOPSIS steps are applied.
Step 5.1. Determine the problem: In this study, the problem is defined in Step 1 of
the proposed methodology.
Step 5.2. Determine positive ideal solution (PIS) and negative ideal solution (NIS):
PIS and NIS are calculated for each DM, as explained in step 2 of IVIF-TOPSIS
(Eq. 6). An example of the first DM is partially presented in Table 4.
Table 4. Example of PIS and NIS calculation for 1st decision-maker

PIS NIS
lL lU mL mU lL lU mL mU
C1 0.725 0.775 0.175 0.225 0.175 0.175 0.525 0.675
C2 0.725 0.775 0.175 0.225 0.175 0.325 0.525 0.675
… … … … … … … … …
C15 0.45 0.55 0.35 0.45 0.175 0.175 0.525 0.675
Step 5.3. Calculate separation measures: Separation measures are calculated for
each alternative for each DM by using Eq. 7 and 8.
Step 5.4. Aggregate separation measures: Separation measure are used to find
aggregated separation measures by using Eq. 9 (Table 5).
Step 5.5. Calculate closeness coefficient: Aggregated separation measures are used
to calculate closeness coefficients by using Eq. 10 (Table 5).
Table 5. Final results according to IVIF-TOPSIS methodology

A1 A2 A3 A4 A5
D (+) 0.356 0.459 0.437 0.426 0.468
D (−) 0.553 0.433 0.454 0.470 0.428
U 0.608 0.486 0.510 0.525 0.478
Step 6. Choose the best alternative: According to the final results found in Step 5,
alternatives are ranked based on the closeness coefficient values in descending order.
Therefore, final ranking is obtained as A1 > A4 > A3 > A2 > A5.
4 Conclusions
For the organization and management of hotels, information technology (IT) is one of
the most important tools. The success of hotel management depends on the integration
of the various processes, and the required level of integration can be achieved with the
help of proper ITs. As a result of this reality, investments in ITs in the tourism industry
has increased in recent years. For a proper investment decision for the managers,
multiple criteria should be considered to evaluate multiple alternatives. Also, in gen-
eral, it is very hard to obtain exact values for evaluation processes, and a level of
vagueness may exist in the decision environment. Hence, in this study, an integrated
IVIF-DEMATEL and IVIF-TOPSIS methodology is applied. For future studies, the
results of the IVIF-TOPSIS can be compared with IVIF-VIKOR method, since, both
techniques are distance-based. Also, with a given budget limit and other investment
requirements, this problem can be considered as a resource allocation problem.
Moreover, the proposed methodology can be applied in various multi-attribute
decision-making problems where decision-makers face vagueness and imprecision.
References
1. Isik, A.T., Adali, E.A.: A new integrated decision-making approach based on SWARA and
OCRA methods for the hotel selection problem. Int. J. Adv. Oper. Manag. 8(2), 140–151
(2016)
2. Vujosevic, M.L., Popovic, M.J.: The comparison of the energy performance of hotel
buildings using PROMETHEE decision-making method. Thermal Sci. 20(1), 197–208
(2016)
388 G. Erkal et al.
3. Yu, S.M., Wang, J., Wang, J.Q., Li, L.: A multi-criteria decision-making model for hotel
selection with linguistic distribution assessments. Appl. Soft Comput. 67, 741–755 (2018)
4. Peng, H.G., Zhang, H.Y., Wang, J.Q.: Cloud decision support model for selecting hotels on
TripAdvisor.com with probabilistic linguistic information. Int. J. Hosp. Manag. 68, 124–138
(2018)
5. Kwok, P.K., Lau, H.Y.: Hotel selection using a modified TOPSIS-based decision support
algorithm. Decis. Support Syst. 120, 95–105 (2019)
6. Akincilar, A., Dagdeviren, M.: A hybrid multi-criteria decision-making model to evaluate
hotel websites. Int. J. Hosp. Manag. 36, 263–271 (2014)
7. Chou, T.Y., Hsu, C.L., Chen, M.C.: A fuzzy multi-criteria decision model for international
tourist hotels location selection. Int. J. Hosp. Manag. 27(2), 293–301 (2008)
8. Ji, P., Zhang, H.Y., Wang, J.Q.: A projection-based outranking method with multi-hesitant
fuzzy linguistic term sets for hotel location selection. Cogn. Comput. 10(5), 737–751 (2018)
9. Mahalik, D.K., Satpathy, B.: Prioritization of outsourcing in hotel industry: a fuzzy AHP
multi criteria decision making approach. South Asia J. Tourism Herit. 4(2), 111–119 (2011)
10. Pahari, S., Ghosh, D., Pal, A.: An online review-based hotel selection process using
intuitionistic fuzzy TOPSIS method. In: Progress in Computing, Analytics and Networking,
pp. 203–214. Springer, Singapore (2018)
11. Roy, J., Sharma, K.H., Kar, S., Zavadskas, K.E., Saparauskas, J.: An extended COPRAS
model for multi-criteria decision-making problems and its application in web-based hotel
evaluation and selection. Econ. Res.-Ekonomska Istraživanja 32(1), 219–253 (2019)
12. Sin, K.Y., Jusoh, M.S., Mardani, A.: Proposing an integrated multi-criteria decision making
approach to evaluate total quality management best practices in Malaysia hotel industry. In:
Journal of Physics: Conference Series, vol. 1432, no. 1, p. 012062. IOP Publishing, January
2020
13. Kumru, M., Kumru, P.Y.: An integrated approach to thermal hotel selection. Revısta
Cıentífıca Da Academıa Mılıtar Série VIII 2(193) (2019)
14. Popovic, G., Stanujkic, D., Brzakovic, M., Karabasevic, D.: A multiple-criteria decision-
making model for the selection of a hotel location. Land Use Policy 84, 49–58 (2019)
15. Uyanik, C., Tuzkaya, G., Kalender, Z.T., Oguztimur, S.: An integrated DEMATEL–IF-
TOPSIS methodology for logistics centers’ location selection problem: an application for
Istanbul Metropolitan area. Transport (article in press)
16. Baykasoglu, A., Kaplanoglu, V., Durmusoglu, Z.D.U., Sahin, C.: Integrating fuzzy
DEMATEL and fuzzy hierarchical TOPSIS methods for truck selection. Expert Syst. Appl.
40, 899–907 (2013)
17. Sangaiah, A.K., Subramaniam, P.R., Zheng, X.: A combined fuzzy DEMATEL and fuzzy
TOPSIS approach for evaluating GSD project outcome factors. Neural Comput. Appl. 26(5),
1025–1040 (2015)
18. Uygun, Ö., Dede, A.: Performance evaluation of green supply chain management using
integrated fuzzy multi-criteria decision-making techniques. Comput. Ind. Eng. 102, 502–511
(2016)
19. Baykasoğlu, A., Gölcük, İ.: Development of an interval type-2 fuzzy sets based hierarchical
MADM model by combining DEMATEL and TOPSIS. Expert Syst. Appl. 70, 37–51 (2017)
20. Zhang, X., Su, J.: A combined fuzzy DEMATEL and TOPSIS approach for estimating
participants in knowledge-intensive crowdsourcing. Comput. Ind. Eng. 137, 106085 (2019)
21. Atanassov, K., Gargov, G.: Interval valued intuitionistic fuzzy sets. Fuzzy Sets Syst. 31,
343–349 (1989)
22. Xu, Z.S.: Methods for aggregating interval-valued intuitionistic fuzzy information and their
application to decision making. Control Decis. 22(2), 215–219 (2007)
23. Chen, C.H., Tzeng, G.H.: Creating the aspired intelligent assessment systems for teaching
materials. Expert Syst. Appl. 38, 12168–12179 (2011)
24. Yalcin, A.S., Kilic, H.S., Guler, E.: Research and development project selection via IF-
DEMATEL and IF-TOPSIS. In: International Conference on Intelligent and Fuzzy Systems,
pp. 625–633. Springer, Cham (2019)
25. Lee, H.S., Tzeng, G.H., Yeih, W., Wang, Y.J., Yang, S.C.: Revised DEMATEL: resolving
the infeasibility of DEMATEL. Appl. Math. Model. 37, 6746–6757 (2013)
26. Kilic, H.S., Yalcin, A.S.: Comparison of municipalities considering environmental
sustainability via neutrosophic DEMATEL based TOPSIS. Socio-Econ. Plann. Sci.
100827 (2020). (in Press)
27. Yasmin, M., Tatoglu, E., Kilic, H.S., Zaim, S., Delen, D.: Big data analytics capabilities and
firm performance: an integrated MCDM approach. J. Bus. Res. 114, 1–15 (2020)
28. Abdullah, L., Zulkifli, N., Liao, H., Herrera-Viedma, E., Al-Barakati, A.: An interval-valued
intuitionistic fuzzy DEMATEL method combined with Choquet integral for sustainable solid
waste management. Eng. Appl. Artif. Intell. 82, 207–215 (2019)
29. Dalalah, D., Hajaneh, M., Batieha, F.: A fuzzy multi-criteria decision making model for
supplier selection. Expert Syst. Appl. 38, 8384–8391 (2011)
30. Kobryń, A.: Dematel as a weighting method in multi-criteria decision analysis. Multiple
Criteria Decis. Mak. 12, 153–167 (2017)
31. Tuzkaya, G., Sennaroglu, B., Kalender, Z.T., Mutlu, M.: Hospital service quality evaluation
with IVIF-PROMETHEE and a case study. Socio-Econ. Plann. Sci. 68, 100705 (2019)
32. Kahraman, C., Öztayşi, B., Onar, S.Ç.: Intuitionistic fuzzy multi-criteria evaluation of
outsource manufacturers. IFAC 49(12), 1844–1849 (2016)
33. Ren, H., Wang, G.: An interval-valued intuitionistic fuzzy MADM method based on a new
similarity measure. Information 6, 880–894 (2015)
34. Zhang, J., Hedge, G.G., Shang, J., Qi, X.: Evaluating emergency response solutions for
sustainable community development by using fuzzy multi-criteria group decision making
approaches: IVDHF-TOPSIS and IVDHF-VIKOR. Sustainability 8(291), 1–28 (2016)
35. Ye, F.: An extended TOPSIS method with interval-valued intuitionistic fuzzy numbers for
virtual enterprise partner selection. Expert Syst. Appl. 37, 7050–7055 (2010)
Decision Making Using New Distances
of Intuitionistic Fuzzy Sets and Study
Their Application in the Universities
Shuker Mahmood Khalil1(B) and Mohanad Abdulkareem Hasan Hasab2

1
Department of Mathematics, College of Science,
Basrah University, Basrah 61004, Iraq
[email protected]
2
Department of Psychological Counseling, College of Education for Human Sciences,
University of Basrah, Basrah, Iraq
[email protected]
Abstract. Most students in the universities experience big problem to

select the suitable scope for jobbing that will be suitable with their cre-
ativity. In this work, we will investigate new two types of distances, these
types of intuitionistic fuzzy set in decision making like an absolute nor-
malize Euclidean distance and square hamming distance. Furthermore,
we study their application to help these students to select the right scope.
Furthermore, our work in this paper is examined.
Keywords: Decision making · Hamming distance · Normalize

Euclidean distance · Intuitionistic fuzzy set
1 Introduction
There are many problems of decision making are discussed on different non-
classical sets like fuzzy sets (FS), soft sets (SS) and others, see [1–6]. In 1986 [7],
the degree of non-membership is added to (FS) and hence the new type is con-
sider, it is called intuitionistic fuzzy set (IFS). This type looks more accurate to
determine provides the occasion to completely model the problem established on
surveillances the existing familiarity there are many application on non-classical
sets,like fuzzy, soft, nano, permutation sets, see [8–25].
The main purpose of providing appropriate acquaintance to students for suit-
able career nominee cannot be overemphasized. This is important since the myr-
iad problems of scarcity substantial career guide notable by students are of
imposing consequence on their qualification and career nominee.
Thus it is dossier that students be presented full acquaintance on career deter-
mination or nominee to enhance appropriate planning, arranger and skillfulness.
Among the career determining operators such as interest, academic achievement,
personality industrialize etc.; the first mentioned exemplifies to be fundamental.
An (IFS) is utilized as tool since here the membership (non-membership) degree
Decision Making Using New Distances 391
will symbolize the marks of the true (fouls) answered by the student and the hes-
itation degree will symbolize the mark of the questions that are disregarded by
student. In this work, new distances of intuitionistic fuzzy set in decision making
like an absolute normalize Euclidean distance and square hamming distance are
investigated and their application is shown.
2 Definitions and Notations

In this section, some basic definition of (IFS) are recalled.
Definition 2.1 [7]: We say Y = {(x, μY (x), νY (x)), x ∈ W } is intuitionistic
fuzzy set (IFS) where μY (x) : W −→ [0, 1], νY (x) : W −→ [0, 1] with the
property 0 ≤ μY (x) + νY (x)≤ 1, ∀x ∈ W . The values μY (x) and νY (x) represent
the degree of membership and non-membership of x to Y respectively.
Operations on (IFS) and Their Basic Relations 2.2 [7]: Assume that
Y = {(x, μY (x), νY (x)), x ∈ W } and F = {(x, μF (x), νF (x)), x ∈ W } are two
(IFSs) of W . Then some relations are considered as following:
(1) Y ⊆ F if and only if μY (x) ≤ μF (x) and νY (x) ≥ νF (x) for all ∀x ∈ W ,
(2) Y = F if and only if Y ⊆ F and F ⊆ Y ,
(3) Y ∪ F = {(x, max{μY (x), μF (x)}, min{νY (x), νF (x)}) : x ∈ W },
(4) Y ∩ F = {(x, min{μY (x), μF (x)}, max{νY (x), νF (x)}) : x ∈ W },
(5) Y c = {(x, νY (x), μY (x)), x ∈ W }.
These above relations are called (inclusion, equality, union, intersection, comple-
ment) respectively.
Definition 2.3 [7]: Let πY (x) = 1 − μY (x) − νY (x) be the (IFS) index or
hesitation margin of x in Y is the degree of indeterminateness of x ∈ W to the
IFS Y and πY (x) ∈ [0, 1]. i.e., πY (x) : W −→ [0, 1] and 0 ≤ πY ≤ 1 for every
x ∈ W . πY (x) expresses the lack of knowledge of whether x belongs to (IFS) Y
or not. For instance, if Y is an (IFS) with μY (x) = 0.7 and νY (x) = 0.2, then
πY (x) = 1 − (0.7 + 0.2) = 0.1. It can be interpreted as “the degree that the
object x belongs to IFS Y is 0.7, the degree that the object x does not belong
to IFS Y is 0.2 and the degree of hesitancy is 0.1”.
Definition 2.4 [7]: Let Y = {(x, μY (x), νY (x)), x ∈ W }and F = {(x, μF (x),
νF (x)), x ∈ W } be IFS in W . Then,
(1) The normalize Euclidean distance between Y and F is defined as:

n
d(Y, F ) = (1/2n) [(μY (xi ) − μF (xi ))2 + (νY (xi ) − νF (xi ))2 + (πY (xi ) − πF (xi ))2 ]
i=1
where xi ∈ W , for i = 1, 2, . . . , n.
(2) The Hamming distance between Y and F is defined as:

n
d(Y, F ) = (1/2n) [|μY (xi ) − μF (xi )| + |νY (xi ) − νF (xi )| + |πY (xi ) − πF (xi )|]
i=1
where xi ∈ W , for i = 1, 2, . . . , n.
392 S. M. Khalil and M. A. H. Hasab
3 New Class of Intuitionistic Fuzzy Sets

In this work, we will introduce new category of intuitionistic fuzzy set in decision
making and study its application to help the students to select the right scope.
Definition 3.1. Let Y = {(x, μY (x), νY (x)), x ∈ W } and F = {(x, μF (x),
νF (x)), x ∈ W } be IFS in W . An absolute normalize Euclidean distance between
Y and F is denoted by dY (Y, F ) and define as:
n

dY (Y, F ) = (1/2n) [(μY (xi ) − μF (xi ))2 + (νY (xi ) − νF (xi ))2
i=1 (1)
+ |πY (xi ) − πF (xi )|]
where xi ∈ W , for i = 1, 2, . . . , n.
Definition 3.2. Let Y = {(x, μY (x), νY (x)), x ∈ W } and F = {(x, μF (x),
νF (x)), x ∈ W } be IFS in W . A square Hamming distance between Y and
F is denoted by ds (Y, F ) and define as:
n

ds (Y, F ) = (1/2n) [|μY (xi ) − μF (xi )| + |νY (xi ) − νF (xi )|
i=1 (2)
+ (πY (xi ) − πF (xi )) ] 2
where xi ∈ W , for i = 1, 2, . . . , n.
4 Applications on Intuitionistic Fuzzy Sets

The main purpose of providing appropriate acquaintance to students for suit-
able career nominee cannot be overemphasized. This is important since the
myriad problems of scarcity substantial career guide notable by students are
of imposing consequence on their qualification and career nominee. So, it is
dossier that students be presented full acquaintance on career determination
or nominee to enhance appropriate planning, arranger and skillfulness. Among
the career determining operators such as interest, academic achievement, per-
sonality industrialize etc.; the first mentioned exemplifies to be fundamen-
tal. An (IFS) is utilized as tool since here the membership (non-membership)
degree will symbolize the marks of the true (fouls) answered by the stu-
dent and the hesitation degree will symbolize the mark of the questions that
are disregarded by student. Assume that these sets U = {u1 , u2 , u3 , u4 },
H = {Surgery, P harmacy, M edicine, Anatomy} = {Sur., P har., M ed., Anat.}
and K = {Biology, M athematics, EnglishLanguage, P hysics, Chemistry} =
{Bi., M ath., En.L., P hys., Chem.} are the sets of students, careers and subjects,
respectively. Suppose the members of U sit for examinations, where the total
degree is 100 marks on the above aforesaid subjects to limited their career deploy-
ment and nominees. The related between subjects requirements and careers is
shown in Table 1.
Table 1. Careers vs Subjects
Bi. Math. En.L. Phys. Chem.

Med. (0.9,0.0,0.1) (0.7,0.2,0.1) (0.8,0.1,0.1) (0.6,0.3,0.1) (0.8,0.1,0.1)
Phar. (0.8,0.1,0.1) (0.8,0.1,0.1) (0.9,0.1,0.0) (0.5,0.3,0.2) (0.7,0.2,0.1)
Sur. (0.9,0.0,0.1) (0.5,0.2,0.3) (0.5,0.3,0.2) (0.5,0.4,0.1) (0.7,0.1,0.2)
Ana. (0.9,0.1,0.0) (0.5,0.4,0.1) (0.7,0.2,0.1) (0.6,0.3,01) (0.8,0.0,0.2)
Table 2. Students vs Subjects
Bi. Math. En.L. Phys. Chem.

u1 (0.6,0.2,0.2) (0.5,0.4,0.1) (0.6,0.3,0.1) (0.5,0.3,0.2) (0.5,0.5,0)
u2 (0.5,0.3,0.2) (0.6,0.2,0.2) (0.5,0.3,0.2) (0.4,0.5,0.1) (0.7,0.2,0.1)
u3 (0.7,01,0.2) (0.6,0.3,0.1) (0.7,0.1,0.2) (0.5,0.4,0.1) (0.4,0.5,0.1)
u4 (0.6,0.0,0.4) (0.8,0.1,0.1) (0.6,0.4,0.0) (0.6,0.3,01) (0.5,0.3,0.2)
Table 3. Students vs Careers
Sur. Phar. Med. Ana.

u1 0.085 0.079 0.062 0.087
u2 0.079 0.093 0.072 0.079
u3 0.109 0.068 0.08 0.101
u4 0.097 0.093 0.09 0.101
There are three maps μ, ν and π describe each performance they are member-
ship, non-membership and hesitation margin, respectively. The following marks
for mummers (students) in set U after the different examinations which are
shown in Table 2. Now, by Eq. (1) we will find the distance between each mem-
ber (student) in set U and each member (career) in set H with reference to the
subjects. That is explained in Table 3. We consider the following from Table 3,
the lest distance provides the proper career assigned as flows:
(1) – u1 is to mention surgery (surgeon),
(2) – u2 is to mention surgery (surgeon),
(3) – u3 is to mention pharmacy (pharmacist),
(4) – u4 is to mention surgery (surgeon).
Moreover, there are many applications. For example, in Basrah university college
of science in Iraq. For selecting the appropriate department to each member
(student) in set U , we need to know students degree of each object. In another
side, each department requires for them to be superior in determined objects as
follows:
(1) In Mathematics Department: The student need to be superior in Mathe-
matical and Computer.
Table 4. Departments vs. Subjects
Math. Phys. Com. Chem. En.L. Hum.Bi.

Dep.M ath. (0.9,0.1,0) (0.5,0.4,0.1) (0.8,0.1,0.1) (0.5,0.3,0.2) (0.5,0.5,0) (0.5,0.2,0.3)
Dep.P hys. (0.8,0.1,01) (0.9,0.1,0) (0.7,0.3,0) (0.6,0.2,0.2) (0.5,0.4,0.1) (0.5,0.4,0.1)
Dep.Com.Sci. (0.7,01,0.2) (0.5,0.2,0.3) (0.9,0.1,0) (0.5,0.1,0.4) (0.8,0.1,0.1) (0.5,0.2,0.3)
Dep.Chem. (0.5,0.1,0.4) (0.5,0.5,0) (0.8,0.2,0) (0.9,0.1,0) (0.8,0.1,0.1) (0.5,0.1,0.4)
Dep.Bi. (0.5,0.2,0.3) (0.5,0.4,0.1) (0.5,0.4,0.1) (0.5,0.3,0.2) (0.8,0.2,0) (0.9,0.1,0)
Table 5. Students vs. Subjects
Math. Phys. Com. Chem. En.L. Hum. Bi.

t1 (0.9,0.1,0) (0.6,0.1,0.3) (0.8,0.1,0.1) (0.5,0.3,0.2) (0.9,0.1,0) (0.7,0.1,0.2)
t2 (0.5,0.1,0.4) (0.6,0.1,0.3) (0.7,0.3,0) (0.6,0.2,0.2) (0.8,0.2,0) (0.9,0.1,0)
t3 (0.8,0.1,0.1) (0.9,0.1,0) (0.9,0.1,0) (0.5,0.2,0.3) (0.7,0.1,0.2) (0.5,0.1,0.4)
t4 (0.9,0.1,0) (0.5,0.5,0) (0.8,0.2,0) (0.8,0.1,0.1) (0.6,0.3,0.1) (0.5,0.3,0.2)
t5 (0.5,0.2,0.3) (0.6,0.3,0.1) (0.9,0.1,0) (0.9,0.1,0) (0.6,0.4,0) (0.7,0.1,0.2)
(2) In Physics Department: The number u (student) need to be superior in

Physic, Mathematical and Computer.
(3) In Computer Science Department: The number u (student) need to be supe-
rior in Computer, Mathematical and English.
(4) In Chemistry Department: The number u (student) need to be superior in
Chemistry, Computer, and English.
(5) In Biology Department: The number u (student) need to be superior in
Human Biology, and English.
We use (IFS) as tool because it is integrate the degree of membership (non-

membership), it symbolize the marks of the true (fouls) answered by the student
and the hesitation degree will symbolize the mark of the questions that are
disregarded by student.
Assume that these sets T = {t1 , t2 , t3 , t4 , t5 , }, D = {Department of Mathe-
matics, Department of Physics, Department of Computer Science,Department of
Chemistry, Department of Biology} = {Dep.M ath., Dep.P hys., Dep.Com.Sci.,
Dep.Bi.} and K = {Computer, Mathematical, Chemistry, Physics, Human Biol-
ogy, English Language} = {Com., M ath., Chem., P hys., Hum.Bi., En.L., } are
the sets of students, departments and subjects, respectively. Suppose the mem-
bers of T sit for examinations, where the total degree is 100 marks on the above
aforesaid subjects to limited their department deployment and nominees. The
related between subjects requirements and departments is shown in Table 4.
There are three maps μ, ν and π describe each performance they are member-
ship, non-membership and hesitation margin, respectively. The following marks
for mummers (students) in set T after the different examinations which are
shown in Table 5.
Table 6. Students vs. Departments
Dep. Math. Dep. Phys. Dep. Com. Sci. Dep. Chem. Dep. Bi.
t1 0.129 0.225 0.149 0.096 0.175
t2 0.186 0.186 0.118 0.144 0.222
t3 0.101 0.148 0.086 0.144 0.168
t4 0.188 0.158 0.174 0.109 0.125
t5 0.173 0.088 0.231 0.223 0.192
Now, by Eq. (2) we will find the distance between each member (student) in
set T and each member (department) in set D with reference to the subjects.
That is explained in Table 6.
The decision making to select the right department to get suitable scope or
right career for students after they graduated from the University can be consider
from Table 6, the lest distance provides the proper department assigned as flows:
(1) – t1 is to mention (Dep. Chemistry),
(2) – t2 is to mention (Dep. Computer Science),
(3) – t3 is to mention (Dep. Computer Science),
(4) – t4 is to mention (Dep. Chemistry),
(5) – t5 is to mention (Dep. Physics).
5 Conclusion
In this research, new distances of (IFS) in decision making like an absolute
normalize Euclidean distance and square hamming distance are investigated and
their application is shown. As planned research, we will study and discuss new
notions of (IFS) in decision making and we will apply more applications.
References
1. Mahmood, S.: Decision making using algebraic operations on soft effect matrix
as new category of similarity measures and study their application in medical
diagnosis problems. J. Intell. Fuzzy Syst. 37(2), 1865–1877 (2019). https://doi.
org/10.3233/JIFS-179249
2. Mahmood, S.: Dissimilarity fuzzy soft points and their applications. Fuzzy Inf.
Eng. 8(3), 281–294 (2016). https://doi.org/10.1016/j.fiae.2016.11.003
3. Peneva, V., Popchev, I.: Aggregation of fuzzy preference relations to multicriteria
decision making. Fuzzy Optim. Decis. Making 6, 351–365 (2007)
4. Xu, Z., Da, Q.: An approach to improving consistency of fuzzy preference matrix.
Fuzzy Optim. Decis. Making 2, 3–12 (2003)
5. Bezdek, J.C., Spillman, B., Spillman, R.: A fuzzy relation space for group decision
theory. Fuzzy Sets Syst. 1(4), 255–268 (1978)
6. Tanino, T.: Fuzzy preference orderings in group decision making. Fuzzy Sets Syst.
12(2), 117–131 (1984)
8. Mahmood, S.: New category of the fuzzy d-algebras. J. Taibah Univ. Sci. 12(2),
143–149 (2018)
9. Mahmood, S., Ulrazaq, M., Abdul-Ghani, S., Al-Musawi, A.: σ-algebra and σ-baire
in fuzzy soft setting. Adv. Fuzzy Syst. 2018, Article ID 5731682 (2018)
10. Mahmood, S., Hameed, F.: Applications of fuzzy ρ-ideals in ρ-ideals. Soft Comput.
2020 (2020). https://doi.org/10.1007/s00500-020-04773-3
11. Abdul-Ghani, S., Mahmood, S., Ulrazaq, M., Al-Musawi, A.: New branch of intu-
itionistic fuzzification in algebras with their applications. Int. J. Math. Math. Sci.
2018, Article ID 5712676 (2018)
12. Coker, D.: An introduction to intuitionistic fuzzy topological spaces. Fuzzy Sets
Syst. 88, 81–89 (1997)
13. Mahmood, S., Hameed, F.: Applications on cyclic soft symmetric groups. J. Phys.:
Conf. Ser. (2020, to appear )
14. Mahmood, S.: The permutation topological spaces and their bases. Basrah J. Sci.
(A) 32(1), 28–42 (2014)
15. Mahmood, S., Rajah, A.: Solving class equation xd = β in an alternating group
for each β ∈ H ∩ C α and n ∈ / θ. J. Assoc. Arab Univ. Basic Appl. Sci. 10(1), 42–50
(2011)
16. Mahmood, S., Rajah, A.: Solving class equation xd = β in an alternating group
for all n ∈ θ & β ∈ Hn ∩ C α . J. Assoc. Arab Univ. Basic Appl. Sci. 16(1), 38–45
(2014)
17. Mahmood, S., Abbas, N.M.: Characteristics of the number of conjugacy classes and
p-regular classes in finite symmetric groups. In: IOP Conference Series: Materials
Science and Engineering, vol. 571, p. 012007. IOP Publishing (2019). https://doi.
org/10.1088/1757-899X/571/1/012007
18. Torki, M.M., Mahmood, S.: New types of finite groups and generated algorithm to
determine the integer factorization by excel. AIP (2020, to appear)
19. Al-Musawi, A.M., Mahmood, S., Ulrazaq, M.A.: Soft (1,2)-strongly open maps
in bi-topological spaces. In: IOP Conference Series: Materials Science and Engi-
neering, vol. 571, p. 012002. IOP Publishing (2019). https://doi.org/10.1088/1757-
899X/571/1/012002
20. Mahmood, S., Abbas, N.M.: Applications on new category of the symmetric groups.
AIP (2020, to appear)
21. Abbas, N.M.A., Mahmood, S.: On α -open sets in topological spaces. In: IOP
Conference Series: Materials Science and Engineering, vol. 571, p. 012021. IOP
Publishing (2019). https://doi.org/10.1088/1757-899X/571/1/012021
22. Mahmood, S., Abdul-Ghani, S.: Soft m-ideals and soft s-ideals in soft s-algebras. J.
Phys.: Conf. Ser. 1234, 012100 (2019). https://doi.org/10.1088/1742-6596/1234/
1/012100
23. Mahmood, S., Abbas, N.M.A.: On nano with their applications in medical field.
AIP (2020, to appear)
24. Mahmood, S., Hameed, F.: An algorithm for generating permutations in symmetric
groups using soft spaces with general study and basic properties of permutations
spaces. J. Theor. Appl. Inform. Technol. 96(9), 2445–2457 (2018)
25. Mahmood, S., Hameed, F.: An algorithm for generating permutation algebras using
soft spaces. J. Taibah Univ. Sci. 12(3), 299–308 (2018)
Some New Ordered Semi-linear Spaces
of Intuitionistic Fuzzy Processes
and the Pair of Closely Related States
Nguyen Dinh Phu1(B) , Nguyen Nhut Hung1 , and Le Thi Ngoc Quynh2
1
Faculty of Engineering Technology, Quang Trung University, Qui Nho’n, Vietnam
{ndphu,nnhung}@qtu.edu.vn
2
Faculty of Economics and Tourism, Quang Trung University, Qui Nho’n, Vietnam
[email protected]
Abstract. In this paper by classification in the ordering of intuition-

istic fuzzy functions, we have some new ordered semi-linear spaces of
intuitionistic fuzzy processes L∗ . Modeling the intuitionistic fuzzy differ-
ential equations will be the ideal model to represent the transformation
of the pair of closely related states. For the applications, we have consid-
ered the model that describes the pair of states are closely related, for
example, the oriental philosophy and economic competitiveness models.
Keywords: Intuitionistic fuzzy numbers (IFNs) · The ordered

semi-linear space of IFNs · Intuitionistic fuzzy differential equation
(IFDE) · The pair of closely related states
1 Introduction
In 1983, Atanassov introduced the concept of “Intuitionistic fuzzy sets - IFSs” [2–
4]. It is a generalization of FSs that it could be an important idea when describing a
problem with a variable language (fuzzy) and was pretty useful in situations when
describing a problem. Because of the flexibility of Intuitionistic fuzzy sets in han-
dling uncertainty, they are a tool for a more human consistent reasoning under the
undefined event perfect and vague. In [12] the author have used the same termi-
nology “intuitionistic fuzzy set” as Atanassov but different in meaning to build the
concept of intuitionistic fuzzy logic and IFSs. In the present time, the IFS theory
has been applied to many different fields, for example, in [6] the author discussed
intuitionistic fuzzy medical diagnosis consisting of three major steps: symptom
identification, formulation of medical knowledge. In [11] the author had discussed
an application of intuitionistic fuzzy multiset in medical diagnosis, in [4] proposed
method of many measurement tools and multi-criteria decision making.
In the present approach, we will study them through L-fuzzy sets, which
means that we describe them as special cases of Goguen’s L-fuzzy sets [7] with
elements taking values in a complete lattice L∗G = (L∗ , G ) [5], where

L∗ = x = (x1 , x2 ) ∈ [0, 1]2 |0 x1 + x2 1
y = (y1 , y2 ) G x = (x1 , x2 ) ⇔ y1 ≤ x1 and y2 ≥ x2 .
398 N. D. Phu et al.
Based on the ideas of [4] Xu and Yager [14] defined the intuitionistic fuzzy
numbers which are considered as the basic elements of the intuitionistic fuzzy
sets. Most recently, In [1] the authors gave the definition of derivative opera-
tions for IFNs and their limited character analysis. Recently, the authors Lei,
Xu, Yager ... (see [8,13,14]) studied the intuitionistic fuzzy functions (IFFs) by
regarding the IFFs as the basic elements ϕ(α, β), where α, β ∈ L∗ .
1.1 Approach and Objective
The main objective of this study is the intuitionistic fuzzy numbers (IFNs),
which is related to the intuitionistic fuzzy sets (IFSs) [2,3].
In Nature and in Society there are always pairs of processes which are con-
tradictory, interrelated, there is one that is in the other, so they never accept
negative values, for example, the rate of healthy and diseased cells in a living
organism, ... or the ratio of capital and profit to one commodity of the econ-
omy. Such the pairs of states are closely related, have positive signs, a state
of decreasing and other increasing, or rising together, or decreasing together,
etc., they are geometric intuitionistic fuzzy functions. Modeling the geometric
intuitionistic fuzzy differential equations will be the ideal model to represent the
variation of this closely related states of the processes in L∗ .
1.2 Outline
This paper includes: In Sect. 2, we collect the fundamental notions to use in the
next section. In Sect. 3 we will build some new ordered semi-linear spaces L∗ of
intuitionistic fuzzy processes and sort the order of nature monotony, continu-
ity, differentiable and integrable properties of the intuitionistic fuzzy functions
(IFFs). We introduce about initial value problem of intuitionistic fuzzy differ-
ential equations (IVP for IFDEs). In the last section, we will simulate the laws
of the Oriental Phylosophy (Yin-Yang Theory) and Economic Competitiveness
models.
2 Preliminaries
We recall some notations and concepts presented in detail in series of researches

about fuzzy sets and its extended (see [2,3,15]).
Definition 1 ([15]). Let U be a universe of discourse, then a fuzzy set
A = {(u, μA (u)) |u ∈ U }
that is characterized by membership function:
μA : U → [0, 1] , (1)
that means membership function μA (u) is continuous on [0, 1].

Some New Ordered Semi-linear Spaces of Intuitionistic Fuzzy Processes 399
In [2,3] the author introduced the concept of intuitionistic fuzzy sets (IFSs),
characterized by a membership function and a non-membership function, which
is a generalization of fuzzy sets (FSs).
Definition 2 ([2,3]). Let U be a universe of discourse, then an intuitionistic
fuzzy set A = {(u, μA (u), νA (u))|u ∈ U } that is characterized by membership
function:
μA : U → [0, 1] , (2)
and by non-membership function:
νA : U → [0, 1] , (3)
that means membership function μA (u), and non-membership function νA (u) =
μĀ (u) are continuous on [0, 1] and satisfy 0 μA (u) + νA (u) 1.
In [1,8,13,14] the authors have studied the operations (calculus) of the intu-
itionistic fuzzy numbers, but the authors examined mathematical operations
(calculations) for intuitionistic fuzzy sets, although the operations were addi-
tion, multiplication, and subtraction logic. These logic operations have limita-
tions that prevent us from expanding on metric space, in mathematical analysis,
for example, they defined the IFNs as follows α = (μα , να ), β = (μβ , νβ ).
The authors also proposed (addition)
α ⊕ β = (μα + μβ − μα μβ , να νβ )
is a logic addition operation of the values of the logical proposition, where
O = (0, 1) is a neutral element of addition, α− a reverse element of α. This
⊕ operation is a binary operation and it has the following properties
(i) α ⊕ β = β ⊕ α (commutative law of addition);
(ii) (α ⊕ β) ⊕ γ = α ⊕ (β ⊕ γ) (associativity);
(iii) α ⊕ O = O ⊕ α = α;
(iv) α ⊕ α = O with α = (μα , να ),
and the authors also proposed (multiplication)
α ⊗ β = (να νβ , μα + μβ − μα μβ )
is a logic multiplication operation of the values of the logical proposition, E =
(1, 0). If so, we will have O = (1, 0) = E, and E = (0, 1) = O. While the neutral
O elements of addition and E are the neutral elements of multiplication, they
must be O = O and E = E, obviously, neutralizing elements of mathematical
operations are not well-defined, not suitable. Further to the operations of the
propositional value, we obviously have α ⊕ β = α ⊗ β, the result will be complex
when calculating IFNs.
Based on this ideas, in [1,8,13,14] the authors defined the intuitionistic fuzzy
functions depending on the IFNs variable, which are a ϕ(α) = (ϕ(μα ), ϕ(να )).
In [9] the authors have reiterated the definition of semi-linear spaces, it will
be useful for the next section of this study. The authors have buildt the metric
semi-linear space of lattice intuitionistic fuzzy numbers L∗ .
Definition 3 ([9]).
L∗ = {x = (x1 , x2 ) ∈ [0, 1] × [0, 1]|0 x1 + x2 1} , (4)
and classifying G order in L∗ for ∀x, y ∈ L∗ :
y = (y1 , y2 ) G x = (x1 , x2 ) ⇔ y1 ≤ x1 and y2 ≥ x2 .
where L∗G = (L∗ , G ) is a complete lattice, every x ∈ L∗G is called a lattice

intuitionistic fuzzy number (LIFN) (Fig. 1).
Fig. 1. A geometrical interpretation of LIFNs y G x in L∗G .
Definition 4. In [9] denoted that:
(i) θ = (0, 0) ∈ L∗G is called the zero element of L∗G ;

(ii) if x = (x1 , x2 ) ∈ L∗G then x = (x2 , x1 ) ∈ L∗G is called the reverse element of
x ∈ L∗G .
In [9] denoted too that: L∗G = (L∗ , G ) is a complete lattice, and L∗G is called
a set of the lattice intuitionistic fuzzy numbers (LIFNs), that satisfy order G
in L∗ .
Definition 5 [9]. The geometric distance between intuitionistic fuzzy numbers
x = (x1 , x2 ), y = (y1 , y2 ) ∈ L∗G denote by
HL∗G [x, y] = sup {|x1 − y1 | , |x2 − y2 |} .
Remark 1. In [9], the authors have proved that L∗ is semi-linear space of intu-
itionistic fuzzy numbers. Along with geometric distance between intuitionis-
tic fuzzy numbers (HL∗G (x, y)), it becomes a metric semi-linear space of intu-
itionistic fuzzy numbers and geometric lattice intuitionistic fuzzy function x(t)
depending on the time variable t that is a mapping x : [t0 , T ] → L∗G , with
x(t) = (x1 (t), x2 (t)) ∈ L∗G and t ∈ [t0 , T ], such that 0 ≤ x1 (t) + x2 (t) ≤ 1.
Based on this ideas, in [10] the authors defined two new binary operations,
which are addition and scalar-multiplication, for intuitionistic fuzzy numbers
(IFNs). Thereafter, they introduce a semi-linear space for IFNs that it is called
the narrow metric semi-linear space - L∗ . At the same time, we present a new
type of intuitionistic fuzzy functions (IFFs) with a real domain, and propose a
number of concepts and properties for these functions such as geometric differ-
ence, geometric differentiability, derivative and integral. In addition, we give a
model of initial value problem (IVP) for intuitionistic fuzzy differential equations
(IFDEs) and present its application to an AIDS model.
3 Main Results
3.1 The New Ordered Semi-linear Spaces of Monotone and
Continuity of the Fuzzy Functions
Different from the order in [9,10], here we have the other classifying S order
in L∗ for ∀x, y ∈ L∗ as follows
Definition 6.
L∗S = {x = (x1 , x2 ) ∈ [0, 1] × [0, 1]|0 x1 + x2 1} , (5)
and S order in L∗ for ∀x, y ∈ L∗ :
y = (y1 , y2 ) S x = (x1 , x2 ) ⇔ y1 ≤ x1 and y2 ≤ x2 , (6)
where L∗S = (L∗ , S ) (Fig. 2).
Fig. 2. A geometrical interpretation of GIFNs y S x in L∗S .
Remark 2. Purpose consider two order structures (L∗G and L∗S ) of L∗ is to per-
form separate difference operations for the elements on it. Putting L∗ = L∗G ∪L∗S ,
that is called a set of intuitionistic fuzzy numbers (IFNs), that L∗ is semi-linear
space of IFNs.
Definition 7. We say that there is a geometric intuitionistic fuzzy function

x(t) ∈ L∗ depending on the real time variable t if there is a mapping x : [t0 , T ] →
L∗ , with x(t) = (x1 (t), x2 (t)) ∈ L∗ , such that 0 ≤ x1 (t) + x2 (t) ≤ 1.
Definition 8. Let x, y : [t0 , T ] → L∗ , with x(t) = (x1 (t), x2 (t)) ∈ L∗ , y(t) =
(y1 (t), y2 (t)) ∈ L∗ and t ∈ [t0 , T ]. We say that y(t) is less than x(t) type G
for all t ∈ [t0 , T ]:
y(t) G x(t) for all t ∈ [t0 , T ] if and only if

y1 (t) ≤ x1 (t) and y2 (t) ≥ x2 (t) for all t ∈ [t0 , T ].
We say that y(t) is less than x(t) type S for all t ∈ [t0 , T ]:
y(t) S x(t) for all t ∈ [t0 , T ] if and only if

y1 (t) ≤ x1 (t) and y2 (t) ≤ x2 (t) for all t ∈ [t0 , T ].
Definition 9. We say that

a the intuitionistic fuzzy function with real domain x(t) ∈ L∗ is strictly mono-
tonic increasing by t iff x(t) G x(t + h) in L∗ ;
b the intuitionistic fuzzy function with real domain x(t) ∈ L∗ is strictly mono-
tonic decreasing by t iff x(t + h) G x(t) in L∗ ;
c the intuitionistic fuzzy function with real domain x(t) ∈ L∗ is regular mono-
tonic increasing by t iff x(t) S x(t + h) in L∗
d the intuitionistic fuzzy function with real domain x(t) ∈ L∗ is regular mono-
tonic decreasing by t iff x(t + h) S x(t) in L∗ .
e the distance between the intuitionistic fuzzy function with real domain
x(t), y(t) ∈ L∗ is
H[x(t), y(t)] = sup {|x1 (t) − y1 (t)| , |x2 (t) − y2 (t)|} .
Definition 10. Let x : [t0 , T ] → L∗ , with x(t) = (x1 (t), x2 (t)) ∈ L∗ and t, t +
h ∈ [t0 , T ]. We say that the intuitionistic fuzzy function x(t) is continuous at t
iff
lim xk (t + h) = xk (t), k = 1, 2.
h→0
3.2 The Differential Calculus of the Geometric Intuitionistic Fuzzy

Functions
Definition 11. Let x : [t0 , T ] → L∗ . We say that a geometric intuitionistic
fuzzy function x(t) is geometric differentiable at t ∈ (t0 , T ), if there are exist the
geometric differences x(t + h) G x(t) and DG x(t) ∈ L∗ , such that
x(t + h) G x(t) dx1 dx2
lim+ = , = DG x(t);
h→0 h dt dt
dx dx
1 2
where x(t + h) G x(t) is positive and DG x(t) = , is geometric deriva-
∗
dt dt
tives in L .
Theorem 1. Assume that the geometric intuitionistic fuzzy function x(t) ∈ L∗

is strictly monotonic then a geometric derivative DG x(t) in L∗ , then it’s geo-
metric derivative in L∗ will be:
dx dx2
1
a/ DG x(t) = ,− if the geometric intuitionistic fuzzy function x(t) is
dt dt
strictly monotonic increasing by t;
dx dx
1 2
b/ DG x(t) = − , , if the geometric intuitionistic fuzzy function x(t)
dt dt
is strictly monotonic decreasing by t.
Proof. a/ Because the geometric intuitionistic fuzzy function x(t) ∈ L∗ is strictly

monotonic increasing by t, then exists x(t + h) GS x(t) ∈ L∗GS and implies that
x(t) G x(t + h) or it’s components x1 (t), x2 (t) satisfy x1 (t) x1 (t + h), x2 (t +
h) x2 (t). By Definition 9: x(t+h)GS x(t) = (x1 (t+h)−x1 (t), x2 (t+h)−x2 (t))
and by Definition 10, we have:
x(t + h) GS x(t)
lim+
h
x2 (t) − x2 (t + h)
h→0
x1 (t + h) − x1 (t)
= lim , lim
h→0 x (t + h)h− x (t) h→0 x (t + hh) − x (t)
+ +
1 1 2 2
= lim , − lim
h→0+ h
dx h→0+ h
1 dx2
= ,− = DG x(t).
dt dt
b/ Because the geometric intuitionistic fuzzy function x(t) is strictly monotonic
decreasing by t then exists x(t + h) GS x(t) ∈ L∗G and implies that x(t + h) G
x(t) or it’s components x1 (t), x2 (t) satisfy x1 (t + h) x1 (t), x2 (t) x2 (t + h).
By Definition 9: x(t + h) G2 x(t) = (x1 (t) − x1 (t + h), x2 (t + h) − x2 (t)) and by
Definition 10, we have:
x(t+h)GS x(t)
lim h
h→0+
x1 (t) − x1 (t +
h) x2 (t + h) − x2 (t)
= lim , lim
h→0 x (t + hh) − x (t) h→0 x (t + hh) − x (t)
+ +
1 1 2 2
= − lim+ , lim+
h→0 h
dx dx h→0 h
1 2
= − , = DG x(t).
dt dt
Theorem 2. Assume that the geometric intuitionistic fuzzy function x(t) ∈
L∗GS is regular monotonic then a geometric derivative DG x(t) in L∗GS , then it’s
geometric derivative in L∗GS will be:
dx dx
1 2
a/ DG x(t) = , if the components x1 (t), x2 (t) of the geometric intu-
dt dt
itionistic fuzzy function x(t) ∈ L∗GS are regular monotone increase by t;
dx dx2
1
b/ DG x(t) = − ,− , if the components x1 (t), x2 (t) of the geometric
dt dt
intuitionistic fuzzy function x(t) ∈ L∗GS are functions decrease by t;
Proof. a/ Because the geometric intuitionistic fuzzy function x(t) ∈ L∗GS are
regular monotonic increasing by t, then exists x(t + h) GS x(t) ∈ L∗GS and
implies that x(t) GS x(t + h) or x1 (t) x1 (t + h), x2 (t + h) x2 (t). By
Definition 10: x(t + h) GS x(t) = (x1 (t + h) − x1 (t), x2 (t + h) − x2 (t)) and by
Definition 11, we have:
x(t+h)GS x(t)
lim h
h→0+
x1 (t + h) − x1 (t) x2 (t) − x2 (t + h)
= lim+ , lim+
h→0 x (t + h)h− x (t) h→0 x (t + hh) − x (t)
1 1 2 2
= lim , − lim
h→0+ h
dx dx h→0+ h
1 2
= , = DG x(t).
dt dt
b/ Because the geometric intuitionistic fuzzy function x(t) is regular monotonic
decreasing by t then exists x(t + h) GS x(t) ∈ L∗G and implies that x(t) GS
x(t+h) or x1 (t+h) x1 (t), x2 (t) x2 (t+h). By Definition 10: x(t+h)S x(t) =
(x1 (t) − x1 (t + h), x2 (t + h) − x2 (t)) and by Definition 11, we have:
x(t+h)GS x(t)
lim h
h→0+
x1 (t) − x1 (t +
h) x2 (t + h) − x2 (t)
= lim+ , lim+
h→0 x (t + hh) − x (t) h→0 x (t + hh) − x (t)
1 1 2 2
= − lim , lim
h→0+ h
dx h→0+ h
1 dx2
= − ,− = DG x(t).
dt dt
3.3 The Initial Valued Problem for Intuitionistic Fuzzy Differential

Equations
Theorem 3. (Some kinds of IVP for IFDEs)
a/ If the intuitionistic fuzzy function with real random x(t) ∈ L∗ is a strictly

monotonic by t, then the initial value problems for the intuitionistic fuzzy
differential equation with real random (IVP for IFDE) will have one of the
form:
DG x(t) ⊕ g(t, x(t)) = h(t, x(t)),
(7)
x(0) = x0 = (a1 , a2 ) ∈ L∗ , t ∈ [0, T ].
b/ If the intuitionistic fuzzy function with real random x(t) ∈ L∗ is a regular
monotonic increasing by t, then the initial value problems for the intuition-
istic fuzzy differential equation with real random (IVP for IFDE) will have
one of the form:

DG x(t) = f (t, x(t)),
(8)
x(t0 ) = (x1 (t0 ), x2 (t0 )) = (a1 , a2 ) ∈ L∗ , t ∈ [t0 , T ].
c/ If the intuitionistic fuzzy function with real random x(t) ∈ L∗ is a regular

monotonic decreasing by t, then the initial value problems for the intuition-
istic fuzzy differential equation with real random (IVP for IFDE) will have
one of the form:

−DG x(t) = f (t, x(t)),
(9)
x(t0 ) = (x1 (t0 ), x2 (t0 )) = (a1 , a2 ) ∈ L∗ , t ∈ [t0 , T ].
where x(t), g(t, x(t)), h(t, x(t)) ∈ L∗ and DG x(t)− geometric differentiability of
x(t).
Proof. a/ Because x(t) ∈ L∗ is a strictly monotonic function by t, then the real

function x1 (t) ≤ x1 (t+h) and x2 (t+h) ≤ x2 (t) a derivative of real function x2 (t)
is negative, then the initial value problems for intuitionistic fuzzy differential
equation with real random (IVP for IFDE) is equivalent to:
⎧ dx (t)
⎪ 1
⎨ dt = f1 (t, x(t)),
dx2 (t)
= −f2 (t, x(t)), (10)
⎪
⎩
dt
∗
x(0) = (x1 (0), x2 (0)) = (a1 , a2 ) ∈ LG , t ∈ [0, T ].
/ L∗ . We can rewrite (9) under type:

but f (t, x) = (f1 (t, x(t)), −f2 (t, x(t))) ∈
⎧ dx1 (t)
⎪
⎪ f1 (t, x(t))
⎪
⎪
dt
= ,
⎨ 2 2
dx2 (t)
+ f2 (t, x(t)) (11)
⎪
⎪
dt
= 0,
⎪
⎪ 2
⎩
x(0) = (x1 (0), x2 (0)) ∈ L∗ , t ∈ [0, T ].
or ⎧
⎨ dx1 (t) dx2 (t) f1 (t, x(t))
( , ) ⊕ (f2 (t, x(t)), 0) = ( , 0),
dt dt 2 (12)
⎩x(0) = (x (0), x (0)) = (a , a ) ∈ L∗ , t ∈ [0, T ].
1 2 1 2
f1 (t, x(t))
Putting g(t, x(t)) = (0, f2 (t, x(t))) ∈ L∗ and h(t, x(t)) = ( , 0) ∈ L∗ ,
2
(12) is equivalent to (8).
Because x(t) ∈ L∗ are functions increase by t then x(t + h) the real function
x1 (t) decreases and a derivative of real function x1 (t) is negative, then the initial
value problems for intuitionistic fuzzy differential equation with real random
(IVP for IFDE) is equivalent to:
⎧ dx (t)
⎪
⎨ dt = −f1 (t, x(t)),
1
dx2 (t)
= f2 (t, x(t)), (13)
⎪
⎩
dt
∗
x(0) = (x1 (0), x2 (0)) = (a1 , a2 ) ∈ L , t ∈ [0, T ].
but f (t, x) = (−f1 (t, x(t)), f2 (t, x(t))) ∈/ L∗ . We can rewrite (13) under type:
⎧ dx1 (t)
⎪
⎪ + f1 (t, x(t))
⎪
⎪
dt
= 0,
⎨ 2
dx2 (t)
f2 (t, x(t)) (14)
⎪
⎪
dt
= ,
⎪
⎪
⎩ 2 2
x(0) = (x1 (0), x2 (0)) ∈ L∗ , t ∈ [0, T ].
or ⎧
⎨ dx1 (t) dx2 (t) f2 (t, x(t))
( , ) ⊕ (f1 (t, x(t)), 0) = (0, ),
dt dt 2 (15)
⎩x(0) = (x (0), x (0)) = (a , a ) ∈ L∗ , t ∈ [0, T ].
1 2 1 2
f2 (t, x(t))
Putting g(t, x(t)) = (f1 (t, x(t)), 0) ∈ L∗ and h(t, x(t)) = (0, ) ∈ L∗ ,
2
(15) is equivalent to (8).
b/ Because the real function x1 (t), x2 (t) are decreasing and the geometric deriva-
dx1 dx2
tives of real function x1 (t), x2 (t) are negative, that means DG x(t) = ( , ),
dt dt
then the initial value problems for intuitionistic fuzzy differential equation with
real random (IVP for IFDE) is equivalent to:
⎧ dx (t)
⎪ 1
⎨ dt = f1 (t, x(t)),
dx2 (t)
= f2 (t, x(t)), (16)
⎪
⎩
dt
∗
x(0) = (x1 (0), x2 (0)) = (a1 , a2 ) ∈ L , t ∈ [0, T ].
where f (t, x) = (f1 (t, x(t)), f2 (t, x(t))) ∈ L∗G . We can rewrite (16) under type
(9). c/ Because the real function x1 (t), x2 (t) are decreasing and the geomet-
ric derivatives of real function x1 (t), x2 (t) are negative, that means DG x(t) =
dx1 dx2
(− ,− ), then the initial value problems for intuitionistic fuzzy differential
dt dt
equation with real random (IVP for IFDE) is equivalent to:
⎧ dx (t)
⎪
⎨− dt = f1 (t, x(t)),
1
dx2 (t)
− dt = f2 (t, x(t)), (17)
⎪
⎩ ∗
x(0) = (x1 (0), x2 (0)) = (a1 , a2 ) ∈ L , t ∈ [0, T ].
where f (t, x) = (f1 (t, x(t)), f2 (t, x(t))) ∈ L∗ . We can rewrite (17) under type
(10).
Lemma 1. (Dependence of solution)
The solution of the initial value problems for intuitionistic fuzzy differential equa-
tion with real random (IVP for IFDE) depends continuously on the function on
the right side of the equation.
Proof. We consider the initial value problems for intuitionistic fuzzy differential
equation with real random (IVP for IFDE) (9):

−DG x(t) = f (t, x(t)),
(18)
x(t0 ) = (x1 (t0 ), x2 (t0 )) = (a1 , a2 ) ∈ L∗ , t ∈ [t0 , T ].
with
∀ > 0, ∃γ, η(ε, x0 ) : ∀(t, x) ∈ G : H[f¯(t, x̄), f (t, x)] < η
We need prove that:
H[x̄0 , x0 ] < γ ⇒ H[x̄(t), x(t)] < ε, ∀x ∈ L∗ .
Indeed, because the function f (t, x(t)) ∈ L∗ then
t t
x(t) = x0 + f (s, x(s))ds, and x̄(t) = x̄0 + f¯(s, x̄(s))ds.
t0 t0
We have
t
H[x̄(t), x(t)] H[x̄0 , x0 ] + H[f¯(s, x̄(s)), f (s, x(s))]ds ε.
t0
Theorem 4. (Existence and uniqueness)

If the functions g(t, x(t)), h(t, x(t)) ∈ L∗ are the strictly monotonic by t, and
f1 (t, x(t))
g(t, x(t)) = (0, f2 (t, x(t))) ∈ L∗ and h(t, x(t)) = ( , 0) ∈ L∗ , then for
2
the IVP for IFDE:

DG x(t) ⊕ g(t, x(t)) = h(t, x(t)),
(19)
x(0) = x0 = (a1 , a2 ) ∈ L∗ , t ∈ [0, T ].
exists an uniqueness strictly monotonic by t intuitionistic fuzzy solution x(t) ∈
L∗ .
Proof. Because the functions g(t, x(t)), h(t, x(t)) ∈ L∗ are the strictly monotonic
functions by t, and assume that g(t, x(t)) = (0, f2 (t, x(t))) ∈ L∗ and h(t, x(t)) =
f1 (t, x(t))
( , 0) ∈ L∗ , then the IVP for IFDE (19) is equivalent to:
2
⎧ dx (t)
⎪ 1
⎨ dt = f1 (t, x(t)),
dx2 (t)
= −f2 (t, x(t)), (20)
⎪
⎩
dt
x(0) = (x1 (0), x2 (0)) = (a1 , a2 ) ∈ L∗G , t ∈ [0, T ].
/ L∗ imlies that an uniqueness intu-
where f (t, x) = (f1 (t, x(t)), −f2 (t, x(t))) ∈
∗
itionistic fuzzy solution x(t) ∈ L . By Lemma 1 this solution depends continu-
ously on the function f (t, x) on the right side of the equation in IVP IFDE (20).
Because the given functions g(t, x(t)), h(t, x(t)) ∈ L∗ are the strictly monotonic
by t then the real solutions x1 (t) ≤ x1 (t+h) and x2 (t+h) ≤ x2 (t) a derivative of
real function x2 (t) is negative, that means exists uniqueness a strictly monotonic
by t intuitionistic fuzzy function x(t) ∈ L∗ .
If the functions f (t, x(t)) ∈ L∗ are the regular monotonic increasing by t, then
for the IVP for IFDE:

DG x(t) = f (t, x(t)),
(21)
x(t0 ) = (x1 (t0 ), x2 (t0 )) = (a1 , a2 ) ∈ L∗ , t ∈ [t0 , T ].
exists an uniqueness regular monotonic increasing by t intuitionistic fuzzy solu-

tion x(t) ∈ L∗ .
Proof. Proof of Theorem 5 is similar proof of Theorem 4.

If the functions f (t, x(t)) ∈ L∗ are the regular monotonic decreasing by t, then
for the IVP for IFDE:

−DG x(t) = f (t, x(t)),
(22)
x(t0 ) = (x1 (t0 ), x2 (t0 )) = (a1 , a2 ) ∈ L∗ , t ∈ [t0 , T ].
exists an uniqueness regular monotonic decreasing by t intuitionistic fuzzy solu-
tion x(t) ∈ L∗ .
Proof. Proof of Theorem 6 is similar proof of Theorem 4.
4 Discussion on Application Directions for the Pair of

Closely Related Variable States of the Processes
4.1 Application to the Oriental Philosophy Models
The universe always contains in it the contradictory laws of self-contradiction:
Cold- Hot; Short-Tall; Dark- Light; Bad- Beautiful; Harm- Good; Cost- Benefit, ...
Oriental philosophy is the discovery of the most fundamental law of the universe,
the discipline of all things, the source of the evolution and destruction of material.
Yin - Yang is the postulate (axiom) of the Universe. Oriental philosophy has a
rule: Yin and Yang contrast each other, but rely on each other, there is nothing
entirely positive (Yang) and nothing is purely negative (Yin). In the Yin existed
Yang and vice versa. Therefore, it is possible to apply the model geometric lattice
fuzzy functions (GLFFs) to the Philosophy of Yin- Yang. The GLFFs theory basi-
cally dees the claim that from the fact that an element x “belong” to a given degree
to a fuzzy set A, naturally follows that x should “not belong” to A. Assume that the
degree of negative membership (Yin) of x is x1 , and the degree of positive member-
ship (Yang) of x is x2 , then implies that 0 x1 + x2 1, then the geometric lattice
fuzzy functions (GLFFs) x = (x1 , x2 ) belong to L∗GS . Let’s consider the cases when
dx1 dx2
Yin x1 (t) is decreasing and Yang x2 (t) is increasing DG x(t) = (− , ), such
dt dt
that they satisfy the ORIENTAL PHYLOSOPHY’S models:

DG x(t) = (1 − x1 (t), exp(−t)),
(23)
x(0) = (x1 (0), x2 (0)) = (0, 1) ∈ L∗ , t ∈ [0, 4],
with x(t) is a strict monotone increasing function by t. We find the solutions of
(IVP for IFDE) (23). Because x(t) is a strict monotone increasing function, thus
(IVP for IFDE) (23) be written as follows:
⎧ dx (t)
⎪
⎨ dt = 1 − x1 (t),
1
− dxdt
2 (t)
= exp(−t), (24)
⎪
⎩
x1 (t0 ) = 0, x2 (t0 ) = 1.
Solving (24), we have one of the solutions of (IVP for IFDE) (23) that is
(x1 (t), x2 (t)) = (1 − exp(−t), exp(−t)), t ∈ [0, 4].
This solution is shown in Fig. 3 and 4.
Fig. 3. The relationship between Yin Fig. 4. The relationship between Yin
x1 (t) and Yang x2 (t), when Yin x1 (t) is x1 (t) and Yang x2 (t), when Yin x1 (t) is
decreasing and Yang x2 (t) is increasing decreasing and Yang x2 (t) is increasing
for the Oriental philosophy’s models for the Oriental philosophy’s models
(23) belong to C([0, 4], [0, 1]). (23) in L∗GS .
4.2 Application to the Economic Competitiveness
The globalization market is both an opportunity and a challenge for each com-
pany and for each country. The index of economic competition often depends on
the management of the company and its development strategy.
Back Office: In the 1970s and 80s of the last century, some companies really
created a governance revolution. They are pioneers in implementing quality man-
agement systems, the factories have produced higher quality goods, prices are
cheaper for decades. It can be said that theire goods successfully occupied the
Western market thanks to the effectiveness of governance. This companies have
become the world’s leading brands. A large organization run by a strict man-
agement system (Back Office) is not dynamic enough in the current context.
Frontier: It is an innovation strategy. Looking back over the last two decades,
we have seen little businesses appearing on the list of the world’s most successful,
new companies like new companies today must be able to respond to fast changes
of customers, technology, the market; not only constantly benchmarking with the
best practices as before but there must always be new innovations to go ahead
of the opponent (Frontier). In order to survive and grow in new conditions,
companies need a high level Frontier.
In this paper we consider the relationship between back office and frontier
levels in economic competitiveness. Assume that x1 (t) is back office level, and
x2 (t) is frontier level - two states of the economic competitiveness- x(t) of a
company or county. If we compare the economic competitiveness y(t)- and the
other economic competitiveness x(t) we always x(t) G y(t) by means:
x = (x1 , x2 ) G y = (y1 , y2 ) ⇔ y1 ≤ x1 and y2 ≥ x2 .
According to [9], in L∗ we can write the economic model in following type:

DG x(t) = f (t, x(t)), t ∈ [t0 , T ],
(25)
x(t0 ) = (0, 1) ∈ L∗ ,
where x(t) = (x1 (t), x2 (t)), f (t, x(t)) = (f1 (.), f2 (.)) ∈ L∗ . Let consider the rela-
tionship between back office and frontier levels in economic competitiveness,
that means the economic model:

DG x(t) = (1 − x1 (t), exp(−t)), t ∈ [0, 12],
(26)
x(0) = (x1 (0), x2 (0)) = (0.5, 0.5) ∈ L∗G ,
with x(t) is a strict monotone increasing function by t. The results and illustra-
tions for IVP (26) are presented in Fig. 5.
Fig. 5. The relationship between the back office level x1 (t), and frontier level x2 (t) in
economic competitiveness by the economic model (26) belong to C([0, 12], [0, 1]).
5 Conclusions
In this paper we consider the new ordered semi-linear spaces of intuitionistic
fuzzy processes. We presented the existence and uniqueness theorems for solu-
tions of IVP for IFDEs.
For the applications, we have considered the model that describes the devel-
opmental processes in Oriental Philosophy and the relationship between back
office and frontier levels in Economic Competitiveness by IFDEs in L∗ .
References
1. Ai, Z., Xu, Z., Lei, Q.: Limit properties and derivative operations in the metric
space of intuitionistic fuzzy numbers. Fuzzy Optim. Decis. Making 16(1), 71–87
(2017). https://doi.org/10.1007/s10700-016-9239-7
2. Atanassov, K.T.: Intuitionistic fuzzy sets. VII ITKR’s Session, Sofia (1983)
4. Atanassov, K.T., Pasi, G., Yager, R.: Intuitionistic fuzzy interpretations of multi-
criteria multi-person and multi-measurement tool decision making. Int. J. Syst.
Sci. 36, 859–868 (2005)
5. Cornelis, C., Deschrijver, G., Kerre, E.E.: Implication in intuitionistic fuzzy and
interval-valued fuzzy set theory: construction, classication, application. Int. J.
Approx. Reason. 35, 55–95 (2004)
6. De, S.K., Biswas, R., Roy, A.R.: An application of intuitionistic fuzzy sets in med-
ical diagnostic. Fuzzy Sets Syst. 117, 209–213 (2001)
7. Goguen, J.A.: L-fuzzy sets. J. Math. Anal. Appl. 18, 145–174 (1967)
8. Lei, Q., Xu, Z.: A unification of intuitionistic fuzzy calculus theories based on
subtraction derivatives and division derivatives. IEEE Trans. Fuzzy Syst. 25, 1023–
1040 (2016)
9. Phu, N.D., Hung, N.N.: The geometric lattice intuitionistic fuzzy functions and
applications. J. Intell. Fuzzy Syst. 35, 3347–3358 (2018). https://doi.org/10.3233/
IFS-172027
10. Phu, N.D., Ahmadian, A., Hung, N.N., Salahshour, S., Senu, N.: Narrow metric
semi-linear space of intuitionistic fuzzy numbers: application to AIDS model. Int.
J. Fuzzy Syst. 36 (2019). https://doi.org/10.3233/IFS-172027
11. Shinoj, T.K., Sunil, J.J.: Intuitionistic fuzzy multisets and its application in med-
ical diagnosis. Int. J. Math. Comput. Phys. Electr. Comput. Eng. 6, 34–38 (2012)
12. Takeuti, G., Titani, S.: Intuitionistic fuzzy logic and intuitionistic fuzzy set theory.
J. Symb. Logic 49, 851–866 (1984)
13. Xu, Z.: Methods for aggregating interval-valued intuitionistic fuzzy information
and their application to decision making. Control Decis. 22, 215–219 (2007)
14. Xu, Z., Yager, R.R.: Some geometric aggregation operators based on intuitionistic
fuzzy sets. Int. J. Gen. Syst. 35, 417–433 (2006)
Pythagorean Fuzzy Sets and Q-Rung
Orthopair Fuzzy Sets
Emergency Decision Making Problem of Power
Cut in Turkey Using Pythagorean Fuzzy
Thermodynamic Approach with Prospect
Theory
Veysel Çoban1(&) and Sezi Çevik Onar2

1
Industrial Engineering Department, Bilecik Seyh Edebali University,
Gülümbe, Bilecik, Turkey
[email protected]
2
Istanbul, Turkey
Abstract. Evaluating alternatives and making the right decisions in case of an

emergency is very important for the normalization of life as soon as possible
after an emergency. The uncertainty of the data and the different responses of the
decision makers to the events in the emergency situation make it difficult to
make a quick and accurate decision. The cut of energy, which is very important
for the existence and maintenance of life, in unexpected situations is considered
as an emergency problem. The Pythagorean fuzzy thermodynamic approach is
used to eliminate uncertainties in emergency situations and the emotional effects
of decision-makers. The Pythagorean fuzzy thermodynamic approach allows
decision makers to make their evaluations more realistic and the resulting
decision data is evaluated in a holistic way. In this study, unusual circumstances
(earthquake, flood, storm, etc.) would cause energy bottleneck problem in
Turkey is developed a solution proposals by emergency decision making
approach based on prospect theory.
Keywords: Emergency decision making Pythagorean fuzzy set Multicriteria

decision making (MCDM) Prospect theory Thermodynamics
1 Introduction
Multicriteria decision making (MCDM) method is an important decision making tool

that determines the most suitable alternative among the alternative options. The failure
of classical decision making methods (AHP, TOPSIS, ELECTRE, PROMETEE) to
reflect the knowledge and experience of the experts leads to the development of new
methods based on classical methods. The new methods also aim to make the evalua-
tions of the experts under pressure more realistic. Prospect theory [1] is an important
tool in reflecting expert assessments in pressure and risk-making decision making
problems. Prospect theory explains that the experts perform their evaluations under the
expectation of potential loss and gain, and the theory also accepts that decision makers
are more sensitive to potential losses. The first study, which is made by explaining the
dominance among the alternatives with the prospect theory, is called TODIM [2].
Prospect theory, which is successful in reflecting the real life ideas and behaviors of
decision makers, is used in classical decision making methods and new decision
making methods are developed [3, 4].
Fuzzy logic based approaches are used to describe the complexity and uncertainty
in real life decision making problems. Pythagorean fuzzy sets (PFSs), which are
extensions of the intuitionistic fuzzy sets, are used to identify uncertainties in the
evaluation process [5]. PFSs define non-membership and membership values whose
sum of squares is less than one. PFSs have the ability to represent uncertainty more
strongly than intuitionistic fuzzy sets with this feature.
Unexpected situations affecting people, the environment and the economy are
called emergency events. Fast and effective response to emergencies are the most
important steps to minimize negative effects. Prospect theory and PFSs are important
methods to allow experts to make rational evaluations under time pressure and
uncertainties. The properties of thermodynamics are used to determine the importance
and degree of expert evaluations in complex and unstable conditions. Psychological
responses of specialists in emergency events need to be correctly included in the
evaluation and decision making process. In this study, an unexpected power cut
problems in Turkey are discussed, and these problems are evaluated by novel method
called as Pythagorean fuzzy thermodynamic decision-making method based on
Prospector theory.
The study is organized as follows. In section two, PFSs and its basic calculation
methods are mentioned. Section three contains the basic concepts and application steps
of the Pythagorean fuzzy thermodynamic method based on the prospect theory. In
section four, a sample emergency application is carried out for the proposed new
method. In section five, general evaluations are made about the proposed method and
suggested new studies for future.
2 Pythagorean Fuzzy Sets
PFSs provide decision makers to make more comfortable evaluations by increasing the
definition power of membership degrees [6]. PFS allows decision makers to easily
assess uncertain situations by offering the rule of being less than or equal to the sum of
the squares of membership and non-membership degrees [7, 8]. Membership and non-
membership function of PFS P ~ ¼ fðx; lP~ ð xÞ; vP~ ð xÞÞjx 2 Xg can be defined as
lP~ ð xÞ: X ! ½0; 1 vP~ ð xÞ: X ! ½0; 1 for each x 2 X and the following condition should
be met.
0 ðlP~ ð xÞÞ2 þ ðvP~ ð xÞÞ2 1 ð1Þ

Emergency Decision Making Problem of Power Cut in Turkey 417
~ is defined as follows:
The degree of uncertainty of x 2 X in P
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pP~ ð xÞ ¼ 1 lP~ ð xÞ2 vP~ ð xÞ2 ð2Þ

~ 1 ¼ f x; lP~ ð xÞ; vP~ ð xÞ jx 2 Xg and P
P ~ 2 ¼ f x; lP~ ð xÞ; vP~ ð xÞ jx 2 Xg are PFSs
1 1 2 2
and the basic operations of PFSs are defined as follows [9]:

• Distance between P ~ 1 and P ~2, d P~1; P
~2

d P ~ 2 ¼ 1=2ðj lP~ 2 lP~ 2 j þ j vP~ 2 vP~ 2 j þ jðpP~ Þ2 pP~ 2 jÞ ð3Þ
~1; P
1 2 1 2 1 2

~1, S P
• Score function of P ~ 1 2 ½1; 1
2 2
~ 1 ¼ lP~ mP~
S P ð4Þ
1 1

~1, A P
• Accuracy function of P ~ 1 2 ½1; 1
2 2
~ 1 ¼ lP~ þ vP~
A P ð5Þ
1 1
• Arithmetic operations [9, 10]

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~1 P
P ~2 ¼ ðlP~ 1 ð xÞ2 þ lP~ 2 ð xÞ2 lP~ 1 ð xÞ2 lP~ 2 ð xÞ2 ; vP~ 1 ð xÞvP~ 2 ð xÞ ð6Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P ~2 ¼
~1 P lP~ 1 ð xÞlP~ 2 ð xÞ; vP~ 1 ð xÞ2 þ vP~ 2 ð xÞ2 vP~ 1 ð xÞ2 vP~ 2 ð xÞ2 ð7Þ
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi !
k
~1 ¼
kP 2
1 1 lP~ 1 ð xÞ ; vP~ 1 ð xÞ ; k [ 0 k
ð8Þ
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi!
k
~ k1
P ¼ k
lP~ 1 ð xÞ ; 1 1 vP~ 1 ð xÞ2 ; k [ 0: ð9Þ
0v u
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1
ulP~ ð xÞ2 lP~ ð xÞ2 vP~ ð xÞ
~1 P
P ~ 2 ¼ @t 1 2
; 1 A; if lP~ 1 ð xÞ
1 lP~ 2 ð xÞ2 vP~ 2 ð xÞ
( ) ð10Þ
vP~ 2 ð xÞpP~ 1 ð xÞ
lP~ 2 ð xÞ; vP~ 1 ð xÞmin vP~ 2 ð xÞ;
pP~ 2 ð xÞ
ð11Þ
Pythagorean fuzzy weighted geometric

average (PFWGA) aggregation operator is
~
defined for Pk ¼ x; lP~ k ð xÞ; vP~ k ð xÞ for k ¼ 1; 2; . . .; K and the weight vector is w ¼
P
ðw1 ; w2 ; . . .; wK ÞT with Kk¼1 wk ¼ 1 as follows [11]:
YK YK
~k ¼
PFWGA P lP
wk
~ k ð xÞ ; v ~ k ð xÞ
wk
ð12Þ
k¼1 k¼1 A
3 Pythagorean Fuzzy Thermodynamic with Prospect Theory
Prospect theory [1] examines the factors that affect decision makers’ evaluation under
pressure and stress. The value function reflects experts’ loss and gain perspectives on
events.

xa ; x 0
# ð xÞ ¼ ð13Þ
kðxÞb ; x\0
where gains and losses are represented with x and risk aversion is shown with kð2:25Þ.
The parameters (a; ¼ b ¼ 0:88) determine the shape of the value function (convex,
concave). Thermodynamics based on the basic concepts of exergy and entropy explain
the transformation of energy [12, 13]. Thermodynamics integrates the knowledge of the
experts qualitatively and quantitatively into the decision-making process and con-
tributes to the more effective execution of the decision-making process.
Criteria and decision makers’ expectations from alternatives cause different degrees
of importance for alternatives. If the expectation levels of decision makers are accepted
as reference points, gain and loss are defined according to the level of the decision
value. Pythagorean fuzzy prospect decision matrix (PFPDM) can be created by cal-
culating the Pythagorean fuzzy prospect decision value (PFPDV) of alternatives
according to the corresponding criteria [7, 8, 12].
Emergency decision making problems consist of criteria (Cn ; n ¼ 1; 2; ::; N),
alternatives (Am ; m ¼ 1; 2; . . .; M) and expert (Ek ; k ¼ 1; 2; . . .; K) parts. Pythagorean
fuzzy preference relation (PFPR, Pk ) is obtained by pairwise comparisons of criteria
and consists of PFEs (pkmn ) that express the degree of superiority of criteria. The priority
weights (xkn ) of the criteria from PFPR are calculated to be used in the decision making
process. The Pythagorean fuzzy decision matrix (PFDM, PkD ) defined using PFS is
converted to Pythagorean fuzzy prospect decision matrix (PFPDM, P~ k ) based on loss

and gain assessments for each expert.
Pythagorean fuzzy energy (PFEN, E), which defines the capacities of alternatives,
is calculated as follows using Pythagorean Fuzzy potential (PFP, pp ) and PFF
(Pythagorean Fuzzy force, pf ):
E ¼ pp pf ð14Þ
The quality level of PFP (Q) defined in the range 0–1 is calculated as follows [14]:
1 XS
rðsÞ
rðsÞ
Q¼1
p
s¼1 ps

p ps
ð15Þ
S
where pps represents the mean potentials and calculated by using PFWGA. Pythagorean
Fuzzy Exergy (PFEX) which corresponds to the highest effectiveness level of PFP is
Ex ¼ EQ ð16Þ
Pythagorean fuzzy entropy (PFET) is calculated to define the inequality of the

alternative PFP as follow:
Et ¼ E Ex ð17Þ
Alternatives are sorted by Et values and the alternative with the smallest value is
chosen as the most suitable alternative.
4 Application for Energy Cut in Turkey
Energy, which is the main element of daily life and economic activities, has an
important place in national and international dimensions. Bottleneck or outage in
energy sources can lead to the emergence and growth of important problems locally
and globally. In this study, an effective decision making method is applied to the
energy cut emergency problem that may occur on a national scale. In the emergency
process of Turkey’s energy bottleneck problems are taken into account the following
considerations.
• Determining and recording the amount of current energy: The current situation is
evaluated to prevent energy losses.
• Distribution of existing energy to priority sectors: The normal flow of life is ensured
by maintaining the vital sectors.
• Creation of alternative energy sources: Energy dependency is reduced by providing
diversity of energy sources.
Three projects (Pra ; a ¼ 1; 2; 3) proposed for emergency response for power cut are
evaluated by three experts (Ek ; k ¼ 1; 2). The project that reflects the best performance
is determined by considering the considerations (Cb ; b ¼ 1; 2; 3). The Pythagorean

fuzzy thermodynamic method based on the prospect theory is used to solve the power
outage emergency problem. Considering that experts have equal weights and PFPR is
defined as:
0 1
0:5 0:55 0:65
P1 ¼ P2 ¼ @ 0:45 0:5 0:6 A
0:35 0:4 0:5
Expectation levels of the considerations defined by the experts are as

EL1 ¼ ð0:55; 0:45; 0:65Þ, EL2 ¼ ð0:65; 0:55; 0:45Þ and the original PFDMs are
obtained as follows.
0 1
ð0:6; 0:4Þ ð0:8; 0:3Þ ð0:6; 0:3Þ
P1D ¼ @ ð0:8; 0:5Þ ð0:6; 0:4Þ ð0:5; 0:5Þ A
ð0:5; 0:1Þ ð0:7; 0:5Þ ð0:8; 0:4Þ
0 1
ð0:7; 0:3Þ ð0:5; 0:2Þ ð0:8; 0:4Þ
P2D ¼ @ ð0:8; 0:3Þ ð0:7; 0:4Þ ð0:8; 0:5Þ A
ð0:6; 0:4Þ ð0:7; 0:4Þ ð0:5; 0:6Þ
The priority vectors of the criteria are as x1 ¼ x2 ¼ ð0:45; 0:35; 0:30ÞT [15].
PFPDM can be calculated based on the PFDMs and corresponding expectation-level as
follows [12]:
0 1
ð0:145; 0:203Þ ð0:672; 0:215Þ ð0:248; 0:126Þ
~ 1D ¼ @ ð0:596; 0:257Þ ð0:319; 0:264Þ ð0:187; 0:187Þ A
P
ð0:257; 0:126Þ ð0:500; 0:121Þ ð0:474; 0:150Þ
0 1
ð0:180; 0:126Þ ð0:257; 0:144Þ ð0:672; 0:264Þ
~ 2D ¼ @ ð0:474; 0:126Þ ð0:380; 0:203Þ ð0:672; 0:121Þ A
P
ð0:248; 0:150Þ ð0:380; 0:203Þ ð0:121; 0:319Þ
are calculated according to the defined

PFEN (E) and mean PFP for each PFEN (E)
PFPDMs for each expert as follows:
0 1
ð0:065; 0:091Þ ð0:235; 0:075Þ ð0:074; 0:038Þ
E1 ¼ @ ð0:268; 0:116Þ ð0:112; 0:092Þ ð0:056; 0:056Þ A
ð0:116; 0:057Þ ð0:175; 0:042Þ ð0:142; 0:045Þ
0 1
ð0:081; 0:057Þ ð0:090; 0:050Þ ð0:202; 0:079Þ
E2 ¼ @ ð0:213; 0:057Þ ð0:133; 0:071Þ ð0:202; 0:036Þ A
ð0:111; 0:068Þ ð0:133; 0:071Þ ð0:036; 0:096Þ
1 ¼ ðð0:137; 0:079Þ; ð0:181; 0:101Þ; ð0:155; 0:054ÞÞT

E
2 ¼ ðð0:130; 0:067Þ; ð0:202; 0:061Þ; ð0:108; 0:084ÞÞT

E
and quality matrices are obtained as follows:

0 1
0:958 0:946 0:948
Q1 ¼ @ 0:949 0:961 0:915 A
0:979 0:984 0:989
0 1
0:971 0:972 0:958
Q2 ¼ @ 0:993 0:961 0:988 A
0:991 0:981 0:958
Experts’ PFEX matrices are obtained using PFEN and quality matrices.
0 1
ð0:062; 0:087Þ ð0:222; 0:071Þ ð0:070; 0:036Þ
Ex1 ¼ @ ð0:254; 0:110Þ ð0:108; 0:088Þ ð0:051; 0:051Þ A
ð0:114; 0:056Þ ð0:172; 0:041Þ ð0:140; 0:045Þ
0 1
ð0:079; 0:055Þ ð0:087; 0:049Þ ð0:194; 0:076Þ
Ex2 ¼ @ ð0:212; 0:057Þ ð0:128; 0:068Þ ð0:200; 0:036Þ A
ð0:110; 0:067Þ ð0:130; 0:070Þ ð0:034; 0:092Þ
Projects’ PFETs are calculated and their score values are defined as
Et1 ¼ f0:0062; 0:122g, Et2 ¼ f0:0057; 0:118g, and Et3 ¼ f0:0124; 0; 138g and their
scores are obtained as Sc ðEt1 Þ ¼ 0:092, Sc ðEt2 Þ ¼ 0:088, Sc ðEt1 Þ ¼ 0:131. The score
values of the projects submitted to the power cut emergency are listed from small to
large as Pr2 \Pr1 \Pr3 . The project (Pr2 ) with the smallest score value is chosen as the
most appropriate emergency response.
5 Conclusions
The extent of the negative effects of emergency events is narrowed by the rapid and
effective execution of the decision making process. The decision makers’ evaluations
of alternative actions and their expectations from alternatives differ according to the
personal characteristics, knowledge and experience of the decision maker. The new
method, proposed by name Pythagorean fuzzy thermodynamic approach with prospect
Theory, gives decision makers the ability to make comfortable and consistent decisions
in uncertain and complex conditions. The proposed method is applied in the analysis of
power outages in Turkey emergency decision making problems and the results show
the consistency and applicability of the method. In future researches, it is planned to
use the proposed method in different emergency scenarios and develop it using dif-
ferent methods in the calculation steps.
References
1. Tversky, A., Kahneman, D.: Prospect theory: an analysis of decision under risk.
Econometrica 47(2), 263–291 (1979)
2. Gomes, L., Lima, M.: TODIM: basics and application to multicriteria ranking of projects
with environmental impacts. Found. Comput. Decis. Sci. 16(4), 113–127 (1992)
3. Fan, Z.P., et al.: Multiple attribute decision making considering aspiration-levels: a method
based on prospect theory. Comput. Ind. Eng. 65(2), 341–350 (2013)
4. Liu, P., et al.: Research on the multi-attribute decision-making under risk with interval
probability based on prospect theory and the uncertain linguistic variables. Knowl. Based
Syst. 24(4), 554–561 (2011)
5. Çoban, V., Onar, S.Ç.: Pythagorean fuzzy engineering economic analysis of solar power
plants. Soft. Comput. 22(15), 5007–5020 (2018)
6. Yager, R.R., Abbasov, A.M.: Pythagorean membership grades, complex numbers, and
decision making. Int. J. Intell. Syst. 28(5), 436–452 (2013)
7. Peng, X., Yuan, H., Yang, Y.: Pythagorean fuzzy information measures and their
applications. Int. J. Intell. Syst. 32(10), 991–1029 (2017)
8. Yager, R.R.: Pythagorean fuzzy subsets. In: 2013 Joint IFSA World Congress and NAFIPS
Annual Meeting (IFSA/NAFIPS). IEEE (2013)
9. Zhang, X., Xu, Z.: Extension of TOPSIS to multiple criteria decision making with
Pythagorean fuzzy sets. Int. J. Intell. Syst. 29(12), 1061–1078 (2014)
10. Peng, X., Yang, Y.: Some results for Pythagorean fuzzy sets. Int. J. Intell. Syst. 30(11),
1133–1160 (2015)
11. Yager, R.R.: Pythagorean membership grades in multicriteria decision making. IEEE Trans.
Fuzzy Syst. 22(4), 958–965 (2013)
12. Ren, P., Xu, Z., Hao, Z.: Hesitant fuzzy thermodynamic method for emergency decision
making based on prospect theory. IEEE Trans. Cybern. 47(9), 2531–2543 (2016)
13. Bhattacherjee, A.: Social Science Research: Principles, Methods, and Practices (2012)
14. Xu, Z., Xia, M.: Distance and similarity measures for hesitant fuzzy sets. Inf. Sci. 181(11),
2128–2138 (2011)
15. Dong, Y., Zhang, H., Herrera-Viedma, E.: Integrating experts’ weights generated
dynamically into the consensus reaching process and its applications in managing non-
cooperative behaviors. Decis. Support Syst. 84, 1–15 (2016)
Multi-criteria Cloud Computing Service
Provider Selection Employing Pythagorean
Fuzzy AHP and VIKOR
Irem Otay1 and Tuğba Yıldız2(&)

1
Faculty of Engineering and Natural Sciences,
Department of Industrial Engineering, Istanbul Bilgi University,
Eski Silahtarağa Elektrik Santrali, 34060 Eyüpsultan, Istanbul, Turkey
[email protected]
2
Faculty of Engineering and Natural Sciences,
Departmentof Computer Engineering, Istanbul Bilgi University,
Eski Silahtarağa Elektrik Santrali, 34060 Eyüpsultan, Istanbul, Turkey
[email protected]
Abstract. With the wide usage of Cloud Computing Services (CCS), finding
the most suitable cloud service by matching criteria provided by different CCS
providers have gained considerable importance in various commercial and sci-
entific tasks. In this paper, we analyze cloud service provider evaluation and
selection problem under fuzzy environment. To solve the problem, a novel fuzzy
integrated multicriteria decision making method is proposed. In the proposed
model, a hierarchical structure with six criteria which are financial, performance,
security and privacy, assurance, agility and usability, and four alternatives, is
designed based on comphrehensive literature review and consulting to the
experts. The evaluations regarding to the cloud alternatives are expressed using
Pythagorean fuzzy sets being an extension of intuitionistic fuzzy sets where the
square sum of membership and non-membership degrees is at most “1”. In the
analysis, the weights of the criteria are obtained using fuzzy Analytical Hier-
archy Process (AHP) based on single-valued Pythagorean fuzzy sets. Then, to
rank the alternatives, the calculated weights from the previous analysis are used
as inputs in single-valued Pythagorean fuzzy VlseKriterijumska Optimizacija I
Kompromisno Resenje, meaning of multi-criteria optimization and compromise
solution (VIKOR) method. Finally, sensitivity analysis is performed to check
the robustness and the applicability of the proposed model.
Keywords: Cloud service provider Pythagorean fuzzy sets Pythagorean

fuzzy AHP Pythagorean fuzzy VIKOR Fuzzy multi-criteria decision making
1 Introduction
In recent years, Cloud Computing Services (CCS) has gained considerable importance
as a new paradigm for computing, storage and especially service solution. The popular
service providers such as Google, Microsoft, and IBM started to offer similar profes-
sional technology-based services with high scalability and efficiency but also with
424 I. Otay and T. Yıldız
different quality of service (QoS) and prices to the customers. The diversity in the cloud
services led to competition between service providers. On the customer side, provider
selection (PS) attained a crucial role to choose the best service provider/s among
different service providers based on multiple qualitative and quantitative criteria. In this
manner, CCS PS problem can be seen as a Multi-Criteria Decision-Making (MCDM)
problem.
Recently, MCDM techniques have played a strategical role in consideration of criteria
and yielded the suitable alternative among others. A variety of MCDM techniques have
been applied to address different decision making problems. These techniques can be
listed as follows: Analytic Hierarchy Process (AHP) [1], Analytic Network Process [2],
Complex Proportional Assessment (COPRAS) [3], The Preference Ranking Organiza-
tion METHod for Enrichment of Evaluations (PROMETHEE) [4], Elimination and
Choice Expressing Reality (ELECTRE) [5], Technique for Order of Preference by
Similarity to Ideal Solution (TOPSIS) [6] and VIKOR [7]. Traditional MCDM techniques
use crisp values for judgments. However, these values are insufficient to cope with real life
decision making problems. Thus, the fuzzy systems have gained importance using fuzzy
sets to deal with uncertainty and ambiguity. Fuzzy sets have been extended to Intu-
itionistic fuzzy sets, Type-2 fuzzy sets, and Hesitant fuzzy sets etc. Recently, Pythagorean
Fuzzy Sets (PFSs) [8] are developed as an extension of intuitionistic fuzzy sets where the
square sum of membership and non-membership degrees is at most “1”. According to the
researchers [9], PFSs handle uncertainty in a better way by providing a wide range of
freedom and covering larger domain.
A variety of studies has been proposed to address CCS PS problem. In study [10],
the authors proposed a combination method involving 6 criteria and 27 sub-criteria to
rank CCS PS alternatives based on Interval-valued intuitionistic fuzzy (IVIF) AHP,
IVIF CORPAS, IVIF MOORA, IVIF TOPSIS and IVIF VIKOR. The studies [11, 12]
proposed a framework to address the problem of QoS-based CCS PS using AHP and
TOPSIS methods. The study [13] proposed an MCDM method for CCS considering the
changing priorities of user preferences (UPs). A Markov chain is utilized to extract the
patterns on changes on UPs. Then, the authors used Best Worst Method (BWM) [14] to
rank the services that are selected by patterns. The results showed that proposed
method outperformed AHP-based approaches for CCS. [15] proposed a hybrid MCDM
framework using extended Grey TOPSIS with AHP to rank CCS based on QoS
parameters. In [16], the authors proposed a decision model consisting of seven criteria
and 21 sub-criteria for CCS PS problem. Single-valued PFSs are used for decision
makers’ judgments and Pythagorean fuzzy (PF) TOPSIS technique is utilized for an
MCDM problem. In this study, we proposed a Pythagorean fuzzy model to address
CCS PS problem by integrating AHP and VIKOR methods considering the advantages
of pairwise comparison evaluations, providing maximum group utility and minimum
individual regret [17]. In the analysis, the weights of 6 criteria are calculated using
AHP based on single-valued PFSs. Then, PF-VIKOR is utilized to rank the four
alternatives. This study is considered to be the first attempt that integrates PF-AHP and
PF-VIKOR for a CCS PS problem.
The rest of this paper is organized as follows: Sect. 2 presents preliminaries on
PFSs. Methodology part is proposed in Sect. 3. Application is explained in Sect. 4.
Conclusion and future research directions are given in Sect. 5.
Multi-criteria Cloud Computing Service 425
2 Preliminaries on Pythagorean Fuzzy Sets (PFSs)
PFSs or intuitionistic Type-2 fuzzy sets [8] are defined by membership ðlðxÞÞ and non-
membership ðvðxÞÞ degrees where the sum of these degrees may be greater than “1”;
~ be a single-valued
however, the sum of their squares are less than or equal to “1”. Let P
Pythagorean fuzzy set (SVPFS); then,

~ffi
P x; lP~ ðxÞ; vP~ ðxÞ ; x 2 X where 0 lP~ ðxÞ2 þ vP~ ðxÞ2 1 and
ð1Þ
lP~ ðxÞ : X ! ½0; 1; vP~ ðxÞ : X ! ½0; 1 for 8x 2 X
~ is defined as in Eq. (2):

The degree of hesitancy for a Pythagorean fuzzy set, P,
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pP~ ðxÞ ¼ 1 lP~ ðxÞ2 vP~ ðxÞ2 ð2Þ
Definition 1. Assume that A ~ ¼ hl1 ; v1 i and B

~ ¼ hl2 ; v2 i are two single-valued
Pythagorean fuzzy numbers (SVPFNs). Some of the arithmetic operations for these
SVPFNs are as follows [18–20]:
Addition and Multiplication:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~B
A ~¼ð l21 þ l22 l21 l22 ; v1 v2 Þ ð3Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~B
A ~ ¼ ðl1 l2 ; v21 þ v22 v21 v22 Þ ð4Þ
~
Multiplication with a scalar k(k [ 0) and Power of A:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~¼ð
kA 1 ð1 l2 Þk ; vk Þ ð5Þ
~ k ¼ ðlk ;
A 1 ð1 v2 Þk Þ ð6Þ
Division:
8 rffiffiffiffiffiffiffiffiffi
< v21 v22
~ ~
A%B ¼ l1 =l2 ; 1v22
; if l1 l2 and v1 v2 and l1 p2 l2 p1 ð7Þ
:
0 Otherwise
Definition 2. Let A~ 1 ¼ hl1 ; v1 ; p1 i and B

~ 2 ¼ hl2 ; v2 ; p2 i be two SVPFNs, then the
Euclidean distance between these SVPFNs can be calculated as in Eq. (8) [20].
1
2
~1 B
dðA ~2Þ ¼ l1 l22 þ v21 v22 þ p21 p22 ð8Þ
2
~ i ¼ hli ; vi ; pi i be a collection of SVPFN and w ¼ ðw1 ; w2 ; . . .; wn ÞT

Pn
be the weight vector of A ~ i ðwi 0; wi ¼ 1Þ. Then, Pythagorean fuzzy weighted
i¼1
geometric (PFWG) operator is defined in Eq. (9) [21]:
" # " #
Yn Y
n
~1; A
PFWGðA ~ 2 ; . . .; A
~nÞ ¼ lwi i ; vwi i ð9Þ
i¼1 i¼1
~ ¼ hl1 ; v1 ; p1 i) is as [22]:
Definition 4. The defuzzification formula for a SVPFN (A
~ ¼ ð1 v21 Þ
def ðAÞ ð10Þ
ð2 v21 l21 Þ
~i ¼
Definition 5. The score and accuracy functions for a Pythagorean fuzzy set A
hli ; vi ; pi i are presented in Eq. (11) [21]:
ScðAi Þ ¼ l2i v2i ; AccðAi Þ ¼ l2i þ v2i ð11Þ
3 Proposed Pythagorean Fuzzy AHP and VIKOR

Methodology
The proposed integrated Pythagorean fuzzy method is composed of Pythagorean fuzzy

AHP and Pythagorean fuzzy VIKOR methods based on single-valued Pythagorean
fuzzy sets. The steps of the proposed methodology are presented below [9, 22]:
1. Construct the hierarchical structure representing a goal, a finite set of criteria C =
{C1, C2, …, Cn}, and a discrete set of alternatives X = {x1, x2, …, xm}.
2. Determine the linguistic terms with corresponding single-valued Pythagorean fuzzy
numbers [9].
3. Collect pairwise evaluations from decision maker/s using the linguistic scale.
4. Check the consistency of pairwise comparison matrix of criteria employing Saaty’s
classical consistency analysis [1].
5. Obtain the weights of criteria by Pythagorean fuzzy AHP method.
5:1. Calculate the geometric mean of each row. The operation is realized in two
steps. Primarily, Pythagorean values for each criteria are multiplied employing
Eq. (4) and then root values are obtained using Eq. (6).
5:2. Divide the geometric mean of each row by the sum of the geometric means by
means of Eq. (7) and defuzzify the evaluations using Eq. (10).
6. Evaluate & rank the alternatives through Pythagorean fuzzy VIKOR method [9].
6:1. Construct the decision matrix including Pythagorean fuzzy ratings of alterna-
tives based on a set of criteria as in Eq. (12).
2 C1 Cn 3
x1 Pðu11 ; v11 Þ Pðu1n ; v1n Þ
~¼ . 6
D .. .. .. 7 ð12Þ
.. 4 . . . 5
xm Pðum1 ; vm1 Þ Pðumn ; vmn Þ
where Pðuij ; vij Þ is the Pythagorean fuzzy rating of alternative xi with respect to
criterion Cj
6:2. Determine Pythagorean fuzzy positive and negative ideal solutions [PF-PIS
þ
ð~pPIS;j Þ and PF-NIS ð~p NIS;j Þ], respectively.
( (
max Pðuij ; vij Þ for benefit criteria min Pðuij ; vij Þ for benefit criteria
þ
pPIS;j ¼ i
min Pðuij ; vij Þ for cost criteria
~p
NIS;j ¼ i
max Pðuij ; vij Þ for cost criteria
ð13Þ
i i
where j = 1, 2,. . .; n.
6:3. Calculate Si and Ri values referring to the separation measure of alternative Xi
from PF-PIS and PF-NIS using Pythagorean fuzzy ordered weighted distance
operator [9, 23]:
X
n
Si ¼ ð wj djk Þ1=k ; Ri ¼ ðmaxðwj djk ÞÞ1=k where i = 1; 2; . . .; m
j ð14Þ
j¼1
and wj is the weight of criteria
In Eq. (14), dk points out the largest Pythagorean fuzzy distance between
Pðuij ; vij Þ and PF-PIS divided by distance between PF-PIS and PF-NIS.
6:4. Obtain Qi value for each alternative using Eq. (15). In the equation, a weight
value of the maximum group utility (v) is set to “0.5”.
Si S
Ri R
Qi ¼ v
þ ð1 vÞ i = 1; 2; . . .; m: ð15Þ
S S R R
where S
¼ min Si ; S ¼ max Si ; R
¼ min Ri ; R ¼ max Ri
i i i i
6:5. Rank the alternatives in increasing order based on Si, Ri, and Qi values.
4 Application
In this study, we have concentrated on solving a real life CCS PS problem by applying
the proposed integrated Pythagorean fuzzy MCDM approach combining Pythagorean
fuzzy AHP and Pythagorean fuzzy VIKOR methods using single-valued Pythagorean
fuzzy numbers. Based on the comprehensive literature review and meetings with an
expert who is a senior manager at one of the leading global provider of CCS, the three-
layer hierarchical structure of the problem is designed. At the first layer, the goal of the
study is presented while the set of criteria which are Financial (C1), Performance (C2),
Security and Privacy (C3), Assurance (C4), Agility (C5) and Usability (C6) are
demonstrated at the second layer. At the bottom layer, the four alternatives (CSP1,
CSP2, CSP3, and CSP4) are presented. The name of the cloud service providers are
kept secret because of the privacy policies of the company. At the meetings, first the
pairwise comparisons of criteria are asked from the expert; then, the pairwise com-
parison matrix is constructed as presented in Table 1.
Table 1. Pairwise comparison matrix of criteria.

Criteria C1 C2 C3 C4 C5 C6
C1 EE LI LI MLI 1/MI MLI
C2 EE LI MLI 1/MI MLI
C3 EE MI MI HI
C4 EE MI MHI
C5 EE MI
C6 EE
The consistency of pairwise comparison matrix of criteria is checked by using the

classical consistency analysis. Once the matrix is found to be consistent, the steps of
Pythagorean fuzzy AHP are applied. Initially, the geometric mean is obtained for each
criteria by applying Eqs. (4) and (6). The geometric mean of each row is divided by the
sum of the geometric means using division operator given in Eq. (7). In the following
step, the Pythagorean fuzzy weights are defuzzified using Eq. (10), and the weights of
the criteria are calculated as 0.029, 0.065, 0.358, 0.271, 0.182 and 0.095, respectively.
After obtaining the weights of criteria, the steps of the proposed methodology are
followed up to evaluate the alternative cloud service providers. Herein, Pythagorean
fuzzy VIKOR is implemented using Python libraries. After conducting several meet-
ings with the expert, the decision matrix illustrating the ratings of the alternatives based
on the criteria is collected as in Table 2. When the linguistic values are converted into
their corresponding Pythagorean fuzzy values, the Pythagorean fuzzy positive and
negative ideal solutions [PF-PIS and PF-NIS] are calculated as in Table 3.
Table 2. Decision matrix.

Criteria/Alternatives CSP1 CSP2 CSP3 CSP4
C1 ML MHI MLI MLI
C2 HI HI MHI MHI
C3 HI VHI HI MHI
C4 V HI MHI MLI
C5 VHI VHI HI HI
C6 VHI VHI HI MHI
þ
Table 3. The values of p~PIS;j ~
and pNIS;j .
þ
~pPIS;j =~p
NIS;j C1 C2 C3
þ
~pPIS;j (0.65, 0.35, 0.67) (0.75, 0.25, 0.61) (0.85, 0.15, 0.5)
~p NIS;j (0.5, 0.45, 0.74) (0.65, 0.35, 0.67) (0.65, 0.35, 0.67)
þ
~pPIS;j =~p
NIS;j C4 C5 C6
þ
~pPIS;j (0.85, 0.15, 0.5) (0.85, 0.15, 0.5) (0.85, 0.15, 0.5)
~p
NIS;j (0.5, 0.45, 0.74) (0.75, 0.25, 0.61) (0.65, 0.35, 0.67)
Then, Si, Ri and Qi values for each alternative are calculated using Eqs. (14)–(15) as
listed in Table 4. The alternatives are ranked in ascending order based on the values of Si,
Ri and Qi. The results show that “CSP2” is the leading cloud service provider while it is
followed by “CSP1” and “CSP3” with the Q values of “0.26” and “0.52”, respectively.
The least desired alternative is found as “CSP4”. The results are illustrated in Fig. 1.
Table 4. The values of Si, Ri and Qi.

Alternatives Si Ri Qi Rank
CSP1 0.22 0.19 0.26 2
CSP2 0.09 0.09 0.00 1
CSP3 0.69 0.19 0.52 3
CSP4 1 0.358 1.00 4
In this paper, the sensitivity analysis is also conducted for checking the robustness
and applicability of the proposed method. The model is solved for different values of
“v”. The results are illustrated in Fig. 2. In the figure, “q0, q1, …, q10” represent “v”
values ranging from “0” and “1.0” with increment of “0.1”. The results demonstrate
that for all values of “v”, “CSP2” is always found as the leading cloud service provider
while “CSP4” is the least desired alternative. In addition to that, for all the values of “v”
except “v = 0”, the same ranking of the alternatives is obtained.
Fig. 1. The results of the proposed Fig. 2. The results of sensitivity analysis
method
5 Conclusion
In this study, a novel integrated Pythagorean fuzzy AHP-VIKOR method based on

single-valued Pythagorean fuzzy numbers is proposed to solve cloud service provider
evaluation and selection problem. In the study, a real life problem is analyzed with
respect to six criteria such as financial, performance, security and privacy, assurance,
agility and usability, and four alternatives. In the paper, sensitivity analysis is also
performed for the validity of the developed method.
In the following studies, we aimed to extend the proposed hierarchical structure by
adding sub-criteria i.e. CPU performance, data privacy, and portability. In addition, for
further studies the results can be compared with other MCDM methods using Pytha-
gorean, Spherical, q-Rung Orthopair, Neutrosophic, Fermatean and/or Hexagonal
fuzzy sets by considering the judgments of multiple experts. Also, the researchers may
consider constructing a fuzzy mathematical model integrated with the proposed
Pythagorean fuzzy MCDM method.
References
2. Saaty, T.L.: Decision Making with Dependence and Feedback: The Analytic Network
Process. PWS Publications, Pittsburgh (1996)
3. Zavadskas, E.K., Kaklauskas, A., Sarka, V.: The new method of multicriteria complex
proportional assessment of projects. Technol. Econ. Dev. 1, 131–139 (1994)
4. Brans J.P.: L’ingénièrie de la décision; Elaboration d’instruments d’aide à la décision. La
méthode PROMETHEE. In: Nadeau, R., Landry, M. (eds.) L’aide à la décision: Nature,
Instruments et Perspectives d’Avenir, pp. 183–213. Presses de l’Université Laval, Québec
(1982)
5. Benayoun R., Roy B., Sussman B.: ELECTRE: Une Méthode Pour Guider Le Choix En
Présence De Points De Vue Multiples, Note De Travail 49. SEMA-METRA International,
Direction Scientifique (1996)
6. Hwang, C.L., Yoon, K.: Multiple Attribute Decision Making-Methods and Application.
Springer, New York (1981)
7. Opricovic, S.: Multicriteria optimization of civil engineering systems. Faculty of Civil
Engineering, Belgrade (1998)
8. Yager, R.R.: Pythagorean fuzzy subsets. In: Joint IFSA World Congress and NAFIPS
Annual Meeting (IFSA/NAFIPS), pp. 57–61. IEEE (2013)
9. Mete, S., Serin, F., Oz, N.E., Gul, M.: A decision-support system based on Pythagorean
fuzzy VIKOR for occupational risk assessment of a natural gas pipeline construction. J. Nat.
Gas Sci. Eng. 71, 102979 (2019)
10. Büyüközkan, G., Göçer, F., Feyzioğlu, O.: Cloud computing technology selection based on
interval-valued intuitionistic fuzzy MCDM methods. Soft. Comput. 22, 5091–5114 (2018)
11. Kumar, R.R., Mishra, S., Kumar, C.: A novel framework for cloud service evaluation and
selection using hybrid MCDM methods. Arab. J. Sci. Eng. 43, 7015–7030 (2018)
12. Kumar, R.R., Mishra, S., Kumar, C.: Prioritizing the solution of cloud service selection using
integrated MCDM methods under Fuzzy environment. J. Supercomput. 73, 4652–4682
(2017)
13. Nawaz, F., Asadabadi, M., Janjua, N., Hussain, O., Chang, E., Saberi, M.: An MCDM
method for cloud service selection using a Markov chain and the best-worst method.
Knowl.-Based Syst. 159, 120–131 (2018)
14. Rezaei, J.: Best-worst multi-criteria decision-making method. Omega 53, 49–57 (2015)
15. Jatoth, C., Gangadharan, G.R., Fiore, U., Buyya, R.: SELCLOUD: a hybrid multi-criteria
decision-making model for selection of cloud services. Soft. Comput. 23(13), 4701–4715
(2018)
16. Onar, S.Ç., Öztaysi, B., Kahraman, C.: Multicriteria evaluation of cloud service providers
using Pythagorean fuzzy TOPSIS. Multiple-Valued Log. Soft Comput. 30, 263–283 (2018)
17. Wang, T.C., Liang, J.L., Ho, C.Y.: Multi-criteria decision analysis by using fuzzy VIKOR.
In: International Conference on Service Systems and Service Management, pp. 1–6. IEEE
(2006)
18. Karasan, A., Ilbahar, E., Kahraman, C.: A novel Pythagorean fuzzy AHP and its application
to landfill site selection problem. Soft. Comput. 23, 10953–10968 (2019)
19. Gou, X., Xu, Z., Ren, P.: The properties of continuous Pythagorean fuzzy information. Int.
J. Intell. Syst. 31, 401–424 (2016)
20. Zhang, X., Xu, Z.: Extension of TOPSIS to MCDM with Pythagorean fuzzy sets. Inte.
J. Intell. Syst. 29(12), 1061–1078 (2014)
21. Peng, X., Yang, Y.: Some results for pythagorean fuzzy sets. Int. J. Intell. Syst. 30, 1133–
1160 (2015)
22. Oztaysi, B., Cevik Onar, S., Seker, S., Kahraman, C.: Water treatment technology selection
using hesitant Pythagorean fuzzy hierachical decision making. J. Intell. Fuzzy Syst. 37, 867–
884 (2019)
23. Cui, F.B., You, X.Y., Shi, H., Liu, H.C.: Optimal siting of electric vehicle charging stations
using Pythagorean fuzzy VIKOR approach. Math. Prob. Eng. 2, 1–12 (2018)
Extension of Classical Analytic Hierarchy
Process Using q-Rung Orthopair Fuzzy Sets
and Its Application to Disaster Logistics
Location Center Selection
Seyed Amin Seyfi Shishavan(&) , Yaser Donyatalab ,

and Elmira Farrokhizadeh

[email protected]
Abstract. The subjectivity and vagueness in the analytic hierarchy process

(AHP) method have been handled by the various extensions of ordinary fuzzy
sets. The q-Rung Orthopair Fuzzy Sets (q-ROFs) are superior to the intuition-
istic and the Pythagorean fuzzy sets to express uncertain information. Their
eminent characteristic is that the space of uncertain information they can
describe is broader. The aim of this paper is to extend the classical AHP to
q-Rung Orthopair Fuzzy sets. The proposed method has successfully evaluated
the disaster logistics location center selection in Istanbul with respect to the
considered attributes. The results are compared with Chang fuzzy AHP method,
and revealed that the proposed method better represents the uncertainty of
decision-making environment.
Keywords: q-Rung Orthopair Fuzzy Sets MADM AHP Disaster logistic

location center selection
1 Introduction
Among important branches of modern decision-making theory, MADM has been

widely used in economics, management, medical, social, and the other fields in recent
years and plays a vital role in modern decision science [1]. The imprecision and
uncertainty are two critical problems in the MADM. The fuzzy MADM is a decision-
making technique having the potential of using the fuzzy theory to measure the
vagueness and the uncertainties [2]. Analytic hierarchy process (AHP) is one of the
most prevalent MADM methods to evaluate, prioritize, and rank the decision choices.
Using the judgments of decision-makers, AHP forms the decomposition of problems
into hierarchies [3, 4]. Assessments of experts over alternatives are always imprecise
and uncertain due to the subjective nature of human thinking. To deal with the
uncertainty nature of this type of problems, using fuzzy logic can be an appropriate
approach [4]. The fuzzy set theory or ordinary fuzzy sets, which developed by has an
essential impact on the development of decision-making problems in an uncertain
Extension of Classical Analytic Hierarchy Process 433
environment [5]. In recent years, several extensions of ordinary fuzzy sets have
developed by different researchers. Type-2 fuzzy sets [6]; intuitionistic fuzzy sets [7];
neutrosophic fuzzy sets [8]; hesitant fuzzy sets [9]; Pythagorean fuzzy sets [10]; Pic-
ture fuzzy set [11] and Spherical fuzzy sets [12]. A new concept of fuzzy sets, the q-
Rung Orthopair Fuzzy Sets (q-ROFSs), was presented by Yager [13], in which the sum
of the qth power of the membership and the non-membership degrees have to be less
than or equal one. In this paper, in order to aggregate the comparison matrices infor-
mation, we use three different aggregation operators, q-Rung orthopair fuzzy weighted
arithmetic mean (q ROFWAMÞ, q-Rung orthopair fuzzy weighted geometric mean
(q ROFWGMÞ and q-Rung orthopair fuzzy weighted power geometric mean
(q ROFWPGMÞ, that are most usable aggregation operators in the literature. This
research has three main contributions. The first is to extend the classical analytic
hierarchy process (AHP) using q-Rung orthopair fuzzy sets for selecting the disaster
logistics location center. The second lies in the comparison among the results of the
three mentioned aggregation operators. The final is to demonstrate the application,
practicality, and effectiveness of the proposed AHP method for selecting the disaster
logistics location center in Istanbul. The novelty of this paper comes from its pre-
sentation of extending AHP process using q-Rung orthopair fuzzy sets and the
application of the proposed method in the disaster logistics location center selection.
The structure of the paper is as follows. Section 2 briefly recalls basic concepts and
preliminaries of the q-ROFS. Section 3, includes our proposed MADM technique,
q-Rung orthopair fuzzy AHP method (q-ROF-AHP). Section 4 applies q-ROF_AHP
method to the disaster logistics location center selection problem. In Sect. 5, includes a
comparative analysis of q-ROF-AHP and Chang’s fuzzy AHP method. Finally, the
study is concluded in Sect. 6.
2 Preliminaries
In this section, we give the definition of q-ROFS and summarize some basic concepts
including mathematical operations, aggregation operators, and score and accuracy
functions.
~ defined on X is given
Definition 1 [13]. Let X be an ordinary fixed set, a q-ROFS A
by:
A ¼ fx; lA ðxÞ; #A ðxÞjx 2 X g ð1Þ
Where lA ðxÞ and #A ðxÞ represent the membership degree and non-membership
degree respectively, satisfying lA ð xÞ 2 ½0; 1, #A ðxÞ 2 ½0; 1 and 0 lA ð xÞq þ
#A ðxÞq 1, ðq 1Þ. The indeterminacy degree is defined as pA ð xÞ ¼ ðlA ð xÞq þ
#A ð xÞq lA ð xÞq #A ð xÞq Þ1=q : In this paper, for convenience, we call ðlA ð xÞ; #A ðxÞÞ a
q-ROFS denoted by A ¼ ðlA ; #A Þ:
Definition 2 [13]. Let A~ ¼ ðl; #Þ, A ~ 2 ¼ ðl2 ; #2 Þ be three q-ROFNs,
~ 1 ¼ ðl1 ; #1 Þ and A
and k be a positive real number, then:
434 S. A. Seyfi Shishavan et al.
h i
~1 A
A ~ 2 ¼ ðlq þ lq lq lq Þ1=q ; #1 #2 ð2Þ
1 2 1 2
h i
A ~ 2 ¼ l1 l2 ; ð#q þ #q #q #q Þ1=q
~1 A ð3Þ
1 2 1 2
h i
~ ¼ ð1 ð1 lq Þk Þ1=q ; #k
kA ð4Þ
h i
~ k ¼ ðlk ; ð1 ð1 #q Þk Þ1=q
A ð5Þ
Yager [13], proposed the score and accuracy functions of q-ROFS to compare two
q-ROFs.
Definition 3 [13]. Let A ~ ¼ ðl; #Þ be a q-ROFN, then the score function of A
~ is defined
~ q q ~ ~
as SðAÞ ¼ l # , and the accuracy of A is defined as HðAÞ ¼ l þ # . For any two
q q
qROFNs, A ~ 2 ¼ ðl2 ; #2 Þ:
~ 1 ¼ ðl1 ; #1 Þ and A
1. If SðA ~ 2 Þ, then A
~ 1 Þ [ SðA ~1 [ A
~2;
~ ~
2. If SðA1 Þ ¼ SðA2 Þ, then
If HðA ~ 2 Þ, then A
~ 1 Þ [ HðA ~1 [ A
~2;
~ ~ ~ ~
If HðA1 Þ ¼ HðA2 Þ, then A1 ¼ A2 .
Definition 4. Let A ~ i ¼ ðli ; #i Þ, i ¼ ð1; 2; . . .; nÞ be a collection of q-ROFN and w ¼
Pn
ðw1 ; w2 ; . . .; wn Þ be the weight vector of A ~ i with wi ¼ 1, then the q-rung orthopair
i¼1
fuzzy weighted arithmetic mean (q-ROFWAM) operator is:
D Y 1 Y
E
~1; A
q ROFWAMðA ~ 2 ; . . .; A
~ n Þ ¼ ð1 n ð1 lq Þwi Þq ; n #wi ð6Þ
i¼1 i i¼1 i

Pn
i¼1
fuzzy weighted geometric mean (q-ROFWGM) operator is (Liu and Wang 2017):
DYn Yn E
~1; A
q ROFWGMðA ~ 2 ; . . .; A
~nÞ ¼ lwi ; ð1
w 1
ð1 #qi Þ i Þq ð7Þ
i¼1 i i¼1

Pn
i¼1
fuzzy weighted power geometric mean (q-ROFWPGM) operator is:
D Y Y E
~1; A
~ 2 ; . . .; A
~ n Þ ¼ ð1 n ð1 lq Þwi Þq ; ð1 n ð1 #q Þwi Þq
1 1
q ROFWPGMðA i¼1 i i¼1 i
ð8Þ
Definition 7. Since the score value of a q-ROFS, may be influenced by the hesitancy
degree; therefore a new score function defined by considering the hesitancy degree as
follow:
2 þ lq # q
~ ¼
S A ð9Þ
ð 2 lq þ # q Þ ð 2 l q # q Þ

~ Monotonically increases and decreases regarding l and #, respectively.
Where S A
3 Extension of AHP with Q-Rung Orthopair Fuzzy Sets
The proposed q-rung orthopair fuzzy AHP method in this research is composed of
several steps.
Step 1: Construct the hierarchical structure. In this step, a hierarchical structure
consisting of at least three levels is developed. Level 1 represents a goal or an
objective (selecting the best alternative) based on the score index. The scoring index
is estimated based on a finite set of criteria C ¼ fC1 ; C2 ; . . .; Cn g, which are shown
at Level 2. There are many sub-criteria defined for any criterion C in this hierar-
chical structure. Therefore, at level 3, a discrete set of m feasible alternatives X ¼
fx1 ; x2 ; . . .; xn g (m 2Þ is defined.
Step 2: Constitute pairwise comparison matrices using q-rung orthopair fuzzy

judgment matrices based on the linguistic terms given in Table 1.
Step 3: Normalize pairwise comparison matrices based on positive and negative

criteria. If we have a negative criteria (cost, production waste, tax etc.), in order to
normalize them, we have to change the membership and non-membership degrees.
Step 4: Check for the consistency of each pairwise comparison matrix. To do so,
using the rating value of each linguistic terms and apply the classical consistency
check method, the consistency ratio of each pairwise comparison matrix is calcu-
lated. The threshold of the CR is 10%.
Step 5: Calculate the q-rung orthopair fuzzy weights of criteria and alternatives.
Determine the weight of each alternative using q-ROFWAM, q-ROFWGM, and
q-ROFWPGM operators given in Eq. (6–8) with respect to each criterion where
w ¼ 1n.
Step 6: Compute the q-rung orthopair fuzzy global preference weights using
q-ROFS multiplication operation given in Eq. (3).
Step 7: Calculate the final score of each alternative using q-ROFS addition oper-
ation given in Eq. (2).
Step 8: Defuzzify the final score of each alternative using the score function given
in Eq. (9).
Step 9: Rank the alternatives with respect to the defuzzified final scores. The largest
value indicates the best alternative.
Disaster Logistics Center Location
C2: C4: C5:

C1: Cost C3: Infrastructure Geographic
Transportation locations Suitability
C41: Closeness to city center
C42: Closeness to potential

C32: Energy Infrastructure
C33: Water Infrastructure

C12: Maintenance Cost
C13: Operational Cost
C31: IT Infrastructure
C11: Investment Cost
C21: Land Route
Disaster location
C22: Sea Route
C23: Airway
A1: Kartal A2: Halkali A3: Tuzla
Fig. 1. Hierarchical structure for the problem
4 Application to Disaster Logistics Location Center Selection
The aim of this section is to illustrate the proposed method through a disaster logistics
location center selection problem. We first present the problem definition, followed by
evaluation criteria, problem solution, and comparative analyses, respectively.
4.1 Problem Definition

“Earthquakes are among the most destructive of natural disasters. Through human
history, people living in countries on seismic belts (like Turkey) have faced the
devastating impact of earthquakes many times. Each time, the vital importance of
disaster logistics centers and their location has been realized by the authorities. But,
more is the pity, realizing the importance after disasters claim many lives” [14]. Our
proposed methodology is applied for ranking and selecting the optimal location of
disaster logistics center for a massive city like Istanbul. After a comprehensive liter-
ature review, five criteria and 11 sub-criteria have been determined. Criteria are (C1)
cost, (C2) transportation, (C3) Infrastructure, (C4) Geographic locations and (C5)
suitability of climate. The possible locations (goals) for the disaster logistics were
Halkalı, Kartal, and Tuzla. Figure 1, illustrates this hierarchy which consists of all
criteria and sub-criteria. In this structure, while ‘‘cost’’ is a non-beneficial criterion, the
rest of them are beneficial. First of all, the assessments for the criteria and sub-criteria
are collected from a decision-makers group with respect to the goal, using the linguistic
terms given in Table 1.
Table 1. Weighting scale for the q-rung orthopair fuzzy AHP method
Linguistic terms q-ROFN equivalents numbers Rating value
(l; #)
Absolutely Weak (AW) (0.1, 0.9) 1/9
Very Weak (VW) (0.2, 0.8) 1/7
Weak (W) (0.3, 0.7) 1/5
Slightly Weak (SW) (0.4, 0.6) 1/3
Equal (E) (0.5, 0.5) 1
Slightly Strong (SS) (0.6, 0.4) 3
Strong (S) (0.7, 0.3) 5
Very Strong (VS) (0.8, 0.2) 7
Absolutely Strong (AS) (0.9, 0.1) 9
4.2 Solution Process

The solution process of the problem is as follows:
1. The consistency ratios of the pairwise comparison matrices are calculated based on
the corresponding numerical values in the classical AHP method for the linguistic
scale given in Table 1.
2. Pairwise comparisons and the q-rung orthopair fuzzy priority weights (~ w) are cal-
culated using three different aggregation operators (q-ROFWAM, q-ROFWGM, q-
ROFWPGM). Also, the consistency ratios (CR) of each pairwise matrix is calcu-
lated using the classical consistency check method. Due to space constraints, we do
not provide these tables.
3. This step presents the q-rung orthopair fuzzy weights of the alternatives with respect
to the evaluation criteria as given in Eq. (3). It is necessary to mention that, in this
step, if we have more than two numbers, we have to use the q-rung orthopair fuzzy
weighted geometric mean (q-ROFWGM) instead of Eq. (3) in order to compute the
weights of alternatives.
4. Table 2 gives the final results of the q-ROFs_AHP method based on three different
aggregation operators using addition operation Eq. (2) and score function Eq. (9) in
order to compute the total weight and score value of each alternative. It is necessary
to mention that, in this step, if we have more than two numbers, we have to use the
q-rung orthopair fuzzy weighted arithmetic mean (q-ROFWAM) instead of Eq. (2)
in order to compute the total weight of alternatives.
5. Based on the score value of each alternative, we rank the alternatives.
Table 2. Score values and ranking obtained from q-ROFs-AHP method

q-ROFWAM q-ROFWGM q-ROFWPGM
Total Score Ranking Total Score Ranking Total Score Ranking
A1 (0.47, 0.01) 0.585 1 (0.33, 0.018) 0.528 2 (0.47, 0.02) 0.586 1
A2 (0.46, 0.00) 0.578 2 (0.34, 0.016) 0.531 1 (0.46, 0.01) 0.578 2
A3 (0.34, 0.02) 0.531 3 (0.24, 0.05) 0.510 3 (0.34, 0.05) 0.531 3
As seen in Table 2, both q-ROFWAM and q-ROFWPGM aggregation operators

give the same ranking result (A1 [ A2 [ A3Þ, and the best location for a disaster
logistics center is Kartal. But the result of q-ROFWGM aggregation operator is dif-
ferent (A2 [ A1 [ A3Þ; and the best location for a disaster logistics center is Halkali.
This difference is due to the optimistic and pessimistic nature of the aggregation
operators.
5 Comparison with Chang Fuzzy AHP
Turgut et al. [14] proposed Chang fuzzy AHP method in order to determine the
appropriate location for a disaster logistics center in Istanbul. They used ordinary
triangular fuzzy numbers in their research. In this section, we compare the results of our
proposed q-ROFs-AHP method with the results of this research. Here, we use the same
pairwise matrices for a reliable comparison. Because of the space constraints, we only
present the obtained results of the Chang method. Based on the results obtained from
Chang fuzzy AHP method, the ranking is the same in both Chang fuzzy AHP and q-
ROFs-AHP (q-ROFWAM, q-ROFWPGM) methods. In both methods, the best alter-
native is Kartal. However, the results of the proposed q-ROFs-AHP are more reliable
and precise because the presence of non-membership degrees in the q-ROFs-AHP
method causes to consider the uncertainty degree of decision-makers.
6 Conclusion
q-ROFs are an extension of IFSs aiming that experts are not forced to assign mem-
bership and non-membership degrees whose sum must be at most 1. This provides a
larger range of freedom to experts in expressing their opinions about the vagueness and
impreciseness of the considered problem.
In this paper, the disaster logistics center location selection problem has been
successfully solved by q-ROFs-AHP and compared with Chang’s fuzzy AHP method.
The ranking results in both methods are different since different assumptions and scales
are used. For further research, we suggest q-ROFs-AHP to be compared with other
extensions of ordinary fuzzy sets such as intuitionistic fuzzy AHP, hesitant fuzzy AHP,
and Pythagorean fuzzy AHP. Alternatively, different MADM methods such as fuzzy
ELECTRE, fuzzy TOPSIS, fuzzy axiomatic design, and ANFIS based systems could
be performed in order to develop the proposed models.
References
1. Xing, Y., Zhang, R., Wang, J., Bai, K., Xue, J.: A new multi-criteria group decision-making
approach based on q-rung orthopair fuzzy interaction Hamy mean operators. Neural Comput.
Appl. 8, 1–24 (2019)
2. Sun, L., Dong, H., Liu, A.X.: Aggregation functions considering criteria interrelationships in
fuzzy multi-criteria decision making: State-of-the-art. IEEE Access 6, 68104–68136 (2018)
3. Kutlu Gündoğdu, Fatma, Kahraman, Cengiz: A novel spherical fuzzy analytic hierarchy
process and its renewable energy application. Soft. Comput. 24(6), 4607–4621 (2019)
4. Farrokhizadeh, E., Seyfi Shishavan, S.A., Donyatalab, Y., Kutlu Gündoğdu, F., Kahraman,
C.: Spherical fuzzy bonferroni mean aggregation operators and their applications to multiple-
attribute decision making. In: Kutlu Gündoğdu, F., Kahraman, C. (eds.) Decision Making
with Spherical Fuzzy Sets-Theory and Applications. Springer (2020). https://doi.org/10.
1007/978-3-030-45461-6_5
5. Donyatalab, Y., Farrokhizadeh, E., Seyed Garmroodi, S.D., Seyfi Shishavan, S.A.:
Harmonic mean aggregation operators in spherical fuzzy environment and their group
decision making applications. J. Mult.-Valued Log. Soft Comput. 33(6), 565–592 (2019)
reasoning. Inf. Sci. 8(3), 199–249 (1975)
8. Smarandache, F.: Neutrosophy: neutrosophic probability, set, and logic: analytic synthesis &
synthetic analysis. ISBN 1-879585-76-6 Contents: Preface by Charles T. Le: 3 (2000)
10. Yager, R.R.: Pythagorean fuzzy subsets. In: Joint IFSA World Congress on NAFIPS Annual
Meetind, vol. 2, no. x, pp. 57–61 (2013)
11. Cường, B.C.: Picture fuzzy sets. J. Comput. Sci. Cybern. 30(4), 409–420 (2015)
12. Gündoǧdu, F.K., Kahraman, C.: Spherical fuzzy sets and spherical fuzzy TOPSIS method.
(2017)
14. Turǧut, B.T., Taş, G., Herekoǧlu, A., Tozan, H., Vayvay, O.: A fuzzy AHP based decision
support system for disaster center location selection and a case study for Istanbul. Disaster
Prev. Manag. 20(5), 499–520 (2011)
Interval Valued q-Rung Orthopair Fuzzy
Prioritized Dual Muirhead Mean
Operator and Their Application in Group
Decision Making
Salih Berkan Aydemir(B) and Sevcan Yilmaz Gündüz
Eskisehir Technical University, 26470 Eskisehir, Turkey

{sbaydemir,sevcany}@eskisehir.edu.tr
Abstract. q-rung orthopair fuzzy set (q-ROFS) is a generalization of

various fuzzy consepts. As fuzzy set (FS), intuitionistic FS, Pytagorean
FS, Fermatean FS. Interval q-rung orthopair fuzzy set (IVq-ROFS) is a
more powerful mathematic tool to tackle, inconsistent and vague infor-
mation than the above sets. The aim of this study is to introduce a new
hybrid aggregation method on IVq-ROFS. Attributes and experts often
have different priority level in multiple criteria decision making (MCDM)
problems. Therefore, the prioritized aggregation (PA) operator plays an
important role in aggregation operators (AO). On the other hand, with
the help of a variable vector, a Muirhed mean (MM) operator is taken
into account, which takes into account the interrelationship between any
number of arguments. In the suggested method, the important cases
of both aggregation methods are combined. For multi-attribute group
decision making (MAGDM), prioritized dual Muirhead mean operators
(PDMM) have been proposed under the (IVq-ROFS) environment. In
addition, some covetable properties of the proposed operators are men-
tioned. Finally, a practical example shows the effect and superiority of
the proposed method. Its advantages are detailed when compared to
other existing methods.
Keywords: Multi-criteria group decision making · Muirhead mean

aggregation operators · Interval valued q-rung orthopair fuzzy
information · Prioritized aggregation operators
1 Introduction
The aim of MAGDM methods is to obtain the most appropriate alternative and
criteria-based ranking based on expert knowledge. Some problems encountered in
recent years are related to MCDM. Such as, service quality of commercial banks
Supported by organization scientific research project of Eskisehir Technical University
for project topic named “Using generalized fuzzy sets in multiple criteria decision
making systems” (Project Number: 20DRP041).
q-IVROFPDMM Operator and Their Application in GDM 441
[10], maintenance assessment in the hospitals [5], renewable energy [2]. Fuzzy
set types are used in multiple criteria problems in order to the uncertain and
vague data more accurately [4]. However, only the membership degree (MD) in
fuzzy sets proposed by Zadeh [17] restricts fuzzy set. Hence, Atanassov proposed
intuitionistic fuzzy sets (IFS) [1]. IFSs have the both membership and non-
membership degree (NMD). In addition, they have a hesitant degree. Therefore,
it was seen that it gave more accurate results in expressing incomplete and vague
information. Since multiple criteria problems are often based on individual or
expert opinions, they contain imprecise and incomplete judgments. Sum of the
MD and NMD in IFSs is 1. However, this limits the choice of MD and NMD. Let
the MD be μ and the NMD be ν. For example, IFS is not provided when μ = 0.6
and ν = 0.5. Hence IFSs were also expanded and Pythagorean fuzzy set(PFSs)
were obtained [14]. In the PFSs, the sum of squares of degrees is taken into
account. For example, μ = 0.62 and ν = 0.52 . The sum of squares of degrees is
seen to be less than 1. In this case, when the decision matrix is created in multiple
criteria problems, the assigned value range differs for each fuzzy set. For example,
if the μ and ν values are chosen equal, μ = 0.5 and ν = 0.5 in IFS, μ = 0.7
and ν = 0.7 in PFS. On the other hand, given the extension of fuzzy sets, Yager
introduced q-rung orthopair fuzzy sets (q-ROFSs) [15]. Thus, the μq + ν q ≤ 1
condition is provided for a more general structure. The decision matrix created
by individual or experts is created as a result of a broader evaluation. When
creating the decision matrix, interval valued q-rung orthopair fuzzy (IVq-ROF)
are introduced for situations where μ and ν values are expressed at certain
intervals [3]. Aggregation functions are an important component for information
aggregation and fusion systems. Many studies have been proposed that solve
MCDM problems through AO using q-ROFS information.
The Muirhead mean operator also includes BM and MSM operators due to
the vector parameter. In this study, DMM operator was used due to both its
general structure and the vector parameter that provides flexibility. On the other
hand, it is possible that experts and attributes have different priority level. Con-
sidering this situation, prioritized aggregate function introduced by Yager [13].
The motivation of this study is that the MM operator and the priority operator
are hybridized in a IVq-ROF environment. Thus, both interrelationships between
arguments and attributes of different importance with the priority operator were
taken into account. The advantages of proposed operator can be listed as follows.
1. Attribute and decision makers may have different priorities. Therefore, the
necessity of the prioritized operator is considered.
2. With IV q-ROF, experts have a wider evaluation capability.
3. The PDMM operator includes many aggregation operators. Such as dual
Maclaurin symmetric mean, geometric Bonferroni mean, geometric mean,
4. Proposed method(PDMM) is flexible to consider the interrelationship among
any numbers of attributes, making it more suitable and powerful to
deal with different kinds of complicated MAGDM problems. Moreover, IV
q-ROFPDMM operator provides idempotency and boundednes features.
442 S. B. Aydemir and S. Yilmaz Gündüz
Other parts of the study are organized as follows. In the second section, the
structures used for the proposed method are introduced. In the third section,
the method proposed and its advantages are mentioned. In section fourth, the
proposed aggregation method is tested on the multiple criteria decision making
problem. In the last part,conclusion and future studies are mentioned.
2 Basic Concepts
2.1 Interval q-Rung Orthopair Fuzzy Set
In this chapter, main definitions are emphasized. Also, the score function used
in the study is mentioned.
Definition 1. Let S be a nonempty, The IVq-ROF on the set s is expressed as
follows [3].
A = {< s, ([μL U L U
a (s), μa (s)], [νa (s), νa (s)]) > |s ∈ S} (1)
where [μL U L U
a (s), μa (s)] and [νa (s), νa (s)] denote the MD of NMD, respectively.
Also, the 0 ≤ (μa (s)) + (νa (s))q ≤ 1 condition is valid. The degree
U q U
of indeterminacy is [πaL (s), πaU (s)] = [(1 − (μU q U q 1/q

a (s)) − (νa (s)) ) , (1 −
L q L q 1/q
(μa (s)) − (νa (s)) ) ]. In terms of simplicity, it can be written as A =<
L U L U
[μ , μ ], [ν , ν ] >.
Score Function. Let A =< [μL , μU ], [ν L , ν U ] > be an IV q-ROF number, then

score function is defined as follows [3]:
1 + (μL )q − (ν L )q + 1 + (μU )q − (ν U )q
S(A) = (2)
4
L q L q U q U q
and accuracy function is H(A) = (μ ) +(ν ) +(μ
2
) +(ν )
.
The following theorem is used to compare two q-ROF numbers.
Theorem 1. Let A1 =< [μL U L U L U L U
1 , μ1 ], [ν1 , ν1 ] > and A2 =< [μ2 , μ2 ], [ν2 , ν2 ] > be
two IV q-ROF numbers, then comparision rules can be defined as [3]: If S(A1 ) >
S(A2 ), then A1 > A2 ; If S(A1 ) = S(A2 ), then; If H(A1 ) > H(A2 ), then H1 >
H2 ; If H(A1 ) = H(A2 ), then H1 = H2 ;
2.2 Prioritized Aggregation

The PA operator is proposed by Yager [13]. Prioritized aggregation operators
note that attributes and decision-makers are of different priority level [13]. The
definition of prioritized aggregation operator is given following.
Definition 2. Hi have higher priority than Hj . The value of Hi (x) is the perfor-
mance of any alternative x under attribute Hi and Hi (x) ∈ [0, 1]. The following
equality refers to the PA [13].
t i−1
P A (Hk (x)) = tχi
u=1 ζs Hs (x) Where ζi = i=1 χi
, χi = j=1 Hj (x) (i =
2, 3, . . . t), χ1 = 1.
2.3 Dual Muirhead Mean Operator

The DMM operator takes into account the relationship between any number
of attributers. It also includes several aggregation operators with its parameter
vector [9].
Definition 3. Let αj (j = 1, 2, ..., n) be a set of positive real numbers. Let’s take
[p] = [p1 , p2 , ..., pn ] ∈ Rn as a parameter vector [9].
⎛ ⎞1/n!
n

1 ⎝
DM M [p] (α1 , α2 , ...αn ) = n pj ασ(j) ⎠ (3)
j=1 pj
σ∈Sn j=1
DM M [p] is called Dual Muirhead Mean operator. Where σj (j = 1, 2, ...n) is any

a permutation {1, 2, ...n} and Sn is the set of all permutations of {1, 2, ...n}.
3 Prioritized Dual Muirhead Mean Operator

The proposed method was presented at this section. Using the IVq-ROFS num-
bers, the DMM operator is more sensitivity with the priority operator.
Definition 4. Let αj (j = 1, 2, ..., n) be a set of positive real numbers. Let’s take
[p] = [p1 , p2 , ..., pn ] ∈ Rn as a parameter vector
⎛ ⎞1/n!
n χ
1 n n j
P DM M [p] (α1 , α2 , ...αn ) = n ⎝ (pj ασ(j) ) j=1 j )⎠
χ
(4)
j=1 pj j=1 σ∈Sn
Let’s explicitly consider the aggregation method proposed using basic opera-
tions. The summation operator used here is the basic addition operator defined
on IVq-ROFS.
Theorem 2. Let αj =< [μL U L U
αj , μαj ], [ναj , ναj ] > be a set of IV q-ROF numbers.
Aggregated value of PDMM operator is also an IV q-ROF.
[p]
P DM M (α1 , α2 , ...αn )
⎡⎛ ⎞
⎛ ⎛ ⎛ ⎛ ⎛ χj ⎞q ⎞p ⎞⎞1/n! ⎞ n1
j q j=1 pj
⎢⎜ n L
n n
χj ⎟
⎢ ⎜ ⎝
= ⎣⎝1 − 1 − ⎝ ⎝1− ⎝ ⎝
1 − (μσ(j) ) j=1 ⎠ ⎠ ⎠⎠ ⎠ ⎟,
σ∈Sn j=1 ⎠
⎛ ⎞⎤
⎛ ⎛ ⎛ ⎛ ⎛ χj ⎞q ⎞p ⎞⎞1/n! ⎞ n1
j q j=1 pj
⎜ n n n ⎟⎥
⎜1 − ⎝1 − ⎝ ⎝1 − ⎝1 − ⎝(μU j=1 χj ⎠ ⎠ ⎠⎠ ⎠ ⎟⎥ ,
⎝ σ∈Sn j=1 σ(j) ) ⎠⎦
⎡
⎛ ⎛ ⎛ ⎛ ⎛ χj ⎞q ⎞p ⎞⎞1/n! ⎞ n1
j q j=1 pj
⎢ n n n
⎢⎝1 − ⎝ ⎝1 − L
⎝1 − ⎝(1 − νσ(j) ) j=1 χj ⎠ ⎠ ⎠⎠ ⎠ ,
⎣ σ∈Sn j=1
⎤
⎛ ⎛ ⎛ ⎛ ⎛ χj ⎞q ⎞p ⎞⎞1/n! ⎞ 1
n
j q j=1 pj
n n n ⎥
⎝1 − ⎝ ⎝1 − ⎝1 − ⎝(1 − U
νσ(j) ) j=1 χj ⎠ ⎠ ⎠⎠ ⎠ ⎥
σ∈Sn j=1 ⎦
In Theorem 2, the expansion of PDMM operator in IV q-ROF environment is

given. The proof of Theorem 2 can be easily demonstrated by the mathematical
induction principle. Since there are not enough pages, the proof is not mentioned
here.
4 Practical Example
This section focuses on a real decision-making problem to demonstrate the valid-
ity of the proposed method. For the effectiveness of the suggested procedure, we
touch on the advantages of our method and analyze the advantages of our sug-
gested method comparatively.
Example (Adapted from [16]). The example in the study [16] was used.
The school management and a group of specialists create decision matrices
to advance academic education. The decision matrix consists of 5 alternative
and 4 attribute for each decision maker. There are 5 candidates for evaluation
(A1 , A2 , A3 , A4 , A5 ). According to the following four attributes: 1 C1 is the
morality; 2 C2 is the research capability; 3 C3 is the teaching skill;
4 C4 is
the education background. The priority relationship between decision makers
and criteria has been determined. C1 C2 C3 C4 . The decision makers
q
also show as R(q) = (rij )5x4 , (q = 1, 2, 3). The decision matrices for the three
decision makers to evaluate the Ai (i = 1, 2, 3, 4, 5) candidates according to the
attributes are as follows (see Table 1).
Table 1. Decision matrix created by decision makers
C1 C2 C3 C4
DM1 A1 ([0.6, 0.8], [0.1, 0.2]) ([0.2, 0.4], [0.4, 0.5]) ([0.6, 0.7], [0.2, 0.3]) ([0.4, 0.5], [0.2, 0.4])
A2 ([0.4, 0.7], [0.0, 0.1]) ([0.5, 0.7], [0.1, 0.2]) ([0.7, 0.8], [0.1, 0.2]) ([0.7, 0.8], [0.1, 0.2])
A3 ([0.3, 0.7], [0.2, 0.3]) ([0.2, 0.4], [0.4, 0.5]) ([0.1, 0.4], [0.4, 0.5]) ([0.3, 0.4], [0.4, 0.6])
A4 ([0.7, 0.8], [0.1, 0.2]) ([0.2, 0.3], [0.4, 0.6]) ([0.6, 0.8], [0.0, 0.2]) ([0.6, 0.8], [0.0, 0.2])
A5 ([0.5, 0.6], [0.3, 0.4]) ([0.7, 0.8], [0.0, 0.1]) ([0.2, 0.4], [0.4, 0.5]) ([0.1, 0.3], [0.4, 0.6])
DM2 A1 ([0.2, 0.4], [0.4, 0.5]) ([0.6, 0.7], [0.1, 0.2]) ([0.5, 0.7], [0.1, 0.2]) ([0.5, 0.7], [0.1, 0.2])
A2 ([0.6, 0.8], [0.0, 0.2]) ([0.2, 0.3], [0.4, 0.6]) ([0.7, 0.8], [0.1, 0.2]) ([0.2, 0.4], [0.4, 0.5])
A3 ([0.1, 0.4], [0.4, 0.5]) ([0.8, 0.9], [0.0, 0.1]) ([0.1, 0.4], [0.2, 0.5]) ([0.4, 0.7], [0.2, 0.3])
A4 ([0.6, 0.8], [0.0, 0.2]) ([0.3, 0.8], [0.0, 0.1]) ([0.2, 0.3], [0.4, 0.6]) ([0.6, 0.7], [0.2, 0.3])
A5 ([0.2, 0.4], [0.5, 0.6]) ([0.6, 0.7], [0.2, 0.3]) ([0.6, 0.8], [0.0, 0.2]) ([0.1, 0.4], [0.3, 0.5])
DM3 A1 ([0.2, 0.4], [0.4, 0.5]) ([0.6, 0.7], [0.1, 0.2]) ([0.5, 0.7], [0.1, 0.2]) ([0.5, 0.7], [0.1, 0.2])
A2 ([0.6, 0.8], [0.0, 0.2]) ([0.2, 0.3], [0.4, 0.6]) ([0.7, 0.8], [0.1, 0.2]) ([0.2, 0.4], [0.4, 0.5])
A3 ([0.1, 0.4], [0.4, 0.5]) ([0.8, 0.9], [0.0, 0.1]) ([0.1, 0.4], [0.2, 0.5]) ([0.4, 0.7], [0.2, 0.3])
A4 ([0.6, 0.8], [0.0, 0.2]) ([0.3, 0.8], [0.0, 0.1]) ([0.2, 0.3], [0.4, 0.6]) ([0.6, 0.7], [0.2, 0.3])
A5 ([0.2, 0.4], [0.5, 0.6]) ([0.6, 0.7], [0.2, 0.3]) ([0.6, 0.8], [0.0, 0.2]) ([0.1, 0.4], [0.3, 0.5])
All attributes are of benefit type. Therefore, no normalization was performed.

Based on IV q-ROFPDMM operator the main steps are as follows:
Step 1. There are three priority decision matrices for the decision makers. The
priority matrix is calculated below only for the 3rd decision maker.
⎛ ⎞
0.3294 0.3188 0.5075 0.4169
⎜0.6000 0.2719 0.6400 0.3400⎟
⎜ ⎟
Tij = ⎜ ⎟
(3)
⎜0.2500 0.3825 0.1800 0.2763⎟
⎝0.6400 0.2813 0.3000 0.5600⎠
0.2250 0.6375 0.3400 0.1488
Step 2. Three decision makers are aggreted with IV q-ROFWG (Interval

q-rung orthopair fuzzy weighted geometric) [8]. Also, experiment steps have been
applied for q = 1 (Table 2).
Table 2. Aggregated q-ROF decision matrix
C1 C2 C3 C4
A1 ([0.3371, 0.5560], ([0.2614, 0.4585], ([0.5159, 0.7000], ([0.4796, 0.6231],
[0.2725, 0.3749]) [0.3377, 0.4393]) [0.1258, 0.2263]) [0.1328, 0.2903])
A2 ([0.3814, 0.5884], ([0.3165, 0.4586], ([0.6723, 0.7725], ([0.4154, 0.6044],
[0.1223, 0.2953]) [0.2649, 0.4340]) [0.1274, 0.2275]) [0.2266, 0.3289])
A3 ([0.2329, 0.6007], ([0.3019, 0.4840], ([0.1000, 0.3871], ([0.3018, 0.4600],
[0.2512, 0.3529]) [0.3234, 0.4259]) [0.3432, 0.5000]) [0.3553, 0.5230])
A4 ([0.5329, 0.8000], ([0.1949, 0.3497], ([0.3375, 0.4786], ([0.5090, 0.6486],
[0.0684, 0.2000]) [0.3264, 0.5194]) [0.2348, 0.4256]) [0.1787, 0.3514])
A5 ([0.3856, 0.5441], ([0.5345, 0.7643], ([0.2978, 0.5265], ([0.1195, 0.3537],
[0.3451, 0.4559]) [0.0734, 0.1000]) [0.2828, 0.4026]) [0.3457, 0.5486])
Step 3. The priority matrix for aggreted values is as follows.

⎛ ⎞
1 0.6345 0.3383 0.2557
⎜1 0.7271 0.4297 0.3343⎟
⎜ ⎟
Tij = ⎜
⎜1 0.6716 0.4537 0.1888⎟
⎟
⎝1 0.7971 0.4521 0.3044⎠
1 0.6532 0.5322 0.3450
Step 4. In this step, the proposed IV q-ROFPDMM aggregation method was

applied on the aggregated matrix in Table-2 (for q = 1 and P = (1, 1, 1, 1)).
r1 = ([0.5303, 0.6908], [0.0803, 0.3623]) r2 = ([0.5683, 0.7086], [0.2041, 0.3234])
r3 = ([0.3882, 0.6091], [0.3529, 0.5181]) r4 = ([0.5143, 0.6984], [0.0000, 0.3827])
r5 = ([0.4270, 0.6301], [0.0000, 0.4142])
Step 5. Score values are given in this step (for q = 1 and P = (1, 1, 1, 1)).
A1 = 0.6946 A2 = 0.6874 A3 = 0.5316 A4 = 0.7075 A5 = 0.6607 the ranking is as
follows. A4 > A1 > A2 > A5 > A3 It seems that the best alternative is A4 . Since
P = (1, 1, 1, 1) is taken, all interaction between the criteria is taken into consid-
eration. In below gives results in case of interaction between different criteria.
For P = (1, 0, 0, 0)S(A1 ) = 0.5615, S(A2 ) = 0.5757, S(A3 ) = 0.5408, S(A4 ) =
0.5649, S(A5 ) = 0.5608 A2 > A4 > A1 > A5 > A3 . For P = (1, 1, 0, 0)S(A1 ) =
0.6228, S(A2 ) = 0.6192, S(A3 ) = 0.5711, S(A4 ) = 0.6256, S(A5 ) = 0.6043
A4 > A1 > A2 > A5 > A3 . For P = (1, 1, 1, 0)S(A1 ) = 0.6630, S(A2 ) =
0.6603, S(A3 ) = 0.5368, S(A4 ) = 0.6717, S(A5 ) = 0.6368 A4 > A1 > A2 >
A5 > A3 . For P = (1, 1, 1, 1)S(A1 ) = 0.6496, S(A2 ) = 0.6874, S(A3 ) = 0.5316,
S(A4 ) = 0.7075, S(A5 ) = 0.6607 A4 > A2 > A5 > A1 > A3 .
When the criteria interact with each other, it is seen that the best alterna-
tive is A4 . If there is no relationship between the criteria, it is said to be in the
P = (p1 , 0, 0, 0), p1 > 0 state. However, in the case of P = (1, 1, 0, 0), there is a
interrelationship between any two criteria. Same way, for P = (1, 1, 1, 0), inter-
relationship between any three criteria and for P = (1, 1, 1, 1), interrelationship
between all criteria. If there is a relationship between more than one criterion, the
best alternative is A4 . If there is no relation between the criteria, ranking results
similar to [11] (A2 > A4 > A5 > A1 > A3 ) and [16] (A2 > A4 > A5 > A1 > A3 )
studies are obtained. However, considering the other criteria, the A4 alternative
is in the first place.
5 Conclusions
The weight matrix related to the priority operator is created according to the
priority between the criteria. Therefore, the priority operator plays an important
role in an aggregation structure. In this study, the advantages of aggregation
operator and MM operator can be combined and uncertain and incomplete data
can be expressed better. In multiple criteria problems, more sensitive results are
obtained according to the relationship between the criteria. In addition, validity
and reliability tests can be performed with the test criteria used in the [12] study.
In future studies, the proposed methods can be combined with archimedian
T-norm and T-conorm operators. Such as, Bonferroni Mean [7], Heronian Mean
[6]. On the other hand, muirhead mean and dual muirhead mean can be used
in ratio system and full multiplicative form parts of Multi-Moora structure,
respectively. In the [18] study, a similar situation was proposed for arithmetic
weighted average and geometric weighted average operators on IF.
References
1. Atanassov, K.T.: Intuitionistic fuzzy sets. In: Intuitionistic Fuzzy Sets, pp. 1–137.
2. Ilbahar, E., Cebi, S., Kahraman, C.: A state-of-the-art review on multi-attribute
renewable energy decision making. Energy Strategy Rev. 25, 18–33 (2019)
3. Joshi, B.P., Singh, A., Bhatt, P.K., Vaisla, K.S.: Interval valued q-rung orthopair
fuzzy sets and their properties. J. Intell. Fuzzy Syst. 35(5), 5225–5230 (2018)
4. Kahraman, C.: Fuzzy Multi-Criteria Decision Making: Theory and Applications
with Recent Developments, vol. 16. Springer, Heidelberg (2008)
5. Karimi, H., Sadeghi-Dastaki, M., Javan, M.: A fully fuzzy best-worst multi
attribute decision making method with triangular fuzzy number: a case study of
maintenance assessment in the hospitals. Appl. Soft Comput. 86, 105882 (2020)
6. Liu, P.D., Chen, S.M.: Heronian aggregation operators of intuitionistic fuzzy num-
bers based on the archimedean t-norm and t-conorm. In: 2016 International Con-
ference on Machine Learning and Cybernetics (ICMLC), vol. 2, pp. 686–691. IEEE
(2016)
7. Liu, P., Wang, P.: Multiple-attribute decision-making based on archimedean bon-
ferroni operators of q-rung orthopair fuzzy numbers. IEEE Trans. Fuzzy Syst.
27(5), 834–848 (2018)
8. Liu, P., Wang, P.: Some q-rung orthopair fuzzy aggregation operators and their
applications to multiple-attribute decision making. Int. J. Intell. Syst. 33(2), 259–
280 (2018)
9. Qin, J., Liu, X.: 2-tuple linguistic muirhead mean operators for multiple attribute
group decision making and its application to supplier selection. Kybernetes (2016)
10. Tang, G., Chiclana, F., Lin, X., Liu, P.: Interval type-2 fuzzy multi-attribute
decision-making approaches for evaluating the service quality of Chinese commer-
cial banks. Knowl.-Based Syst. 193, 105438 (2020)
11. Wang, W., Mendel, J.: Interval-valued intuitionistic fuzzy aggregation methodology
for decision making with a prioritization of criteria. Iran. J. Fuzzy Syst. 16(4), 115–
127 (2019)
12. Wang, X., Triantaphyllou, E.: Ranking irregularities when evaluating alternatives
by using some electre methods. Omega 36(1), 45–63 (2008)
13. Yager, R.R.: Prioritized aggregation operators. Int. J. Approx. Reason. 48(1), 263–
274 (2008)
14. Yager, R.R.: Pythagorean fuzzy subsets. In: 2013 Joint IFSA World Congress and
NAFIPS Annual Meeting (IFSA/NAFIPS), pp. 57–61. IEEE (2013)
15. Yager, R.R.: Generalized orthopair fuzzy sets. IEEE Trans. Fuzzy Syst. 25(5),
1222–1230 (2016)
16. Yu, D., Wu, Y., Lu, T.: Interval-valued intuitionistic fuzzy prioritized operators and
their application in group decision making. Knowl.-Based Syst. 30, 57–66 (2012)
18. Zhang, C., Chen, C., Streimikiene, D., Balezentis, T.: Intuitionistic fuzzy multi-
moora approach for multi-criteria assessment of the energy storage technologies.
Appl. Soft Comput. 79, 410–423 (2019)
Fuzzy Capital Budgeting Using Fermatean
Fuzzy Sets
Duygu Sergi1(&) and Irem Ucal Sari2

1
Department of Industrial Engineering, MEF University, Ayazağa Cad. No. 4,
Maslak, Sarıyer, 34396 İstanbul, Turkey
[email protected]
2
Istanbul, Turkey
[email protected]
Abstract. Investment projects are mostly evaluated by capital budgeting

techniques to measure their profitability. The parameters used in capital bud-
geting such as future cash flows, interest rate and useful life involves high
uncertainty due to the lack of information for the future environment. Since the
uncertainty involved in forecasting the parameters is mostly in high levels, fuzzy
set theory could be used in the determination of capital budgeting parameters to
handle uncertain information in the analyses. Fermatean fuzzy sets are one of the
most recent extensions of fuzzy set theory which are capable to handle higher
levels of uncertainties by assigning fuzzy parameters from a larger domain. In
this paper, fuzzy capital budgeting techniques that are fuzzy net present worth,
fuzzy net future worth and fuzzy net annual worth are extended using fermatean
fuzzy sets. An illustration for the calculations is also presented.
Keywords: Fermatean fuzzy sets Fuzzy net present worth Fuzzy annual
worth
1 Introduction
In today, global economies contain high uncertainty which forces businesses that are
one of the elements of the economic system to make their investment decisions under
high uncertain conditions to survive against competitors [1]. Under increasing and
lethal competition conditions, the sustainability of businesses depends on making a
profit, and in order to achieve this goal, it is necessary to make correct and reliable
decisions on investments.
Within this economic system, the responsibilities of decision-makers can be quite
high in order to carry out actions that will ensure companies’ survival, such as keeping
up with the developing technology, forming the right strategies, and outperforming the
competition. Uncertainties about the future force businesses to make a plan and take
quick decisions. Effective and efficient implementation of the decisions are taken as a
management issue too. In an environment where rapid fluctuations and competition are
experienced, for the companies whose first and main goal is to make profit, it is
Fuzzy Capital Budgeting Using Fermatean Fuzzy Sets 449
possible to achieve this goal with proper investment planning and the harmony and
coordination of the factors that determine the profit [2].
The profitability of businesses that are difficult to predict includes uncertainty due
to fluctuations in global economies. Using fuzzy logic theory in investment analysis
enables decision makers to approximately evaluate the investments planned by the
companies before operating activities are concluded and to estimate their financial
returns especially in uncertain economies. Thus, businesses have the opportunity to
make plans for the future [3]. Fermatean fuzzy sets are one of the most recent
extensions of fuzzy set theory which are capable to handle higher levels of uncertainties
by assigning fuzzy parameters from a larger domain.
In this paper, fermatean fuzzy net present worth formulas are proposed for the first
time to evaluate investment projects in which interest rates and cash flows are deter-
mined using fermatean fuzzy sets. An illustration is also done for numerical example.
The rest of the paper is organized as follows: In Sect. 2, a detailed literature review
on fuzzy capital budgeting techniques is given. The preliminaries of fermatean fuzzy
sets are determined in Sect. 3. In Sect. 4, fermatean fuzzy net present worth formulas
are proposed and applied on an illustrative example. The paper lasts with the con-
clusion and future research suggestions.
2 Literature Review
When the literature on fuzzy capital budgeting is analysed, many academic studies
have been found. Buckley [4] developed a fuzzy equivalent for compound and simple
interest rate problems used in mathematics of finance and formulated the fuzzy present
value and fuzzy future value using this interest rate. Pohjola and Turunen [5] expanded
the profitability analysis for fuzzy data by expressing the internal rate of return with
type 2 fuzzy numbers. Chiu and Park [6] cultivated the fuzzy present value analysis by
expressing fuzzy interest rates and fuzzy cash flows in their studies by triangular fuzzy
numbers. Wang and Liang [7] proposed two algorithms that give the fuzzy benefit cost
ratio and the fuzzy benefit cost increase rate in order to perform benefit cost analysis in
the fuzzy environment. Chiu and Park [8] developed the fuzzy present value analysis
for triangular fuzzy numbers and proposed the dominant project determination method
for prioritizing fuzzy projects using the present time criterion.
Kahraman et al. [9] used the fuzzy benefit-cost ratio method which is based on the
new fuzzy budgeting technique. Kuchta [10] defined fuzzy payback time, fuzzy net
present value, fuzzy net future value, and fuzzy internal rate of return formulas and
examined fuzzy cash flows, fuzzy interest rate and fuzzy project life by evaluating
investments. Karsak and Tolga [11] used fuzzy discounted cash-flow analysis (DCF) in
their fuzzy decision algorithm to evaluate economic criteria. Kahraman et al. [12]
developed the fuzzy net present value, fuzzy equivalent uniform annual value, fuzzy
future value, fuzzy benefit-cost ratio and fuzzy payback period formulas for geometric
and trigonometric cash flows in discrete and continuous discounting situations.
Since then several fuzzy extensions of capital budgeting methods have been pro-
posed. Kahraman et al. [13] proposed fuzzy net present value-based fuzzy engineering
economy decision models in which uncertain cash flow and interest rate are expressed
450 D. Sergi and I. U. Sari
in triangular fuzzy numbers. Ammar and Khalifa [14] proposed an algorithm aiming to
determine the best profit of an investment problem by using trapezoidal fuzzy numbers.
Sari and Kahraman [15] used both triangular and trapezoidal interval type-2 fuzzy sets
to develop interval type-2 fuzzy capital budgeting techniques which are interval type-2
fuzzy net present value analysis, future value analysis and equivalent uniform annual
value analysis. Kahraman et al. [16] developed interval-valued intuitionistic fuzzy
present worth analysis for the evaluation of CNC (computer numerical control) lathe
investments and more than one expert was evaluated through aggregation operators in
line with these analyses. Kahraman et al. [17] developed the pythagorean fuzzy present
worth analysis (PWA) method using pythagorean fuzzy sets that have a stronger ability
than intuitionistic fuzzy sets to model uncertainty in investment decision-making
problems to address fuzzy parameters of investments. Aydın and Kabak [18] provided
a new multiplication operator and a new term under the name of “neutrosophic
equivalent” using present value and future value analysis techniques with single-valued
neutrosophic set. Proposed techniques were used to analyze and evaluate alternatives
by experts.
3 Fermatean Fuzzy Sets
Fermatean fuzzy sets, which are proposed by Senapati and Yager [19] are one of the
most recent extensions of fuzzy set theory that are capable to handle higher levels of
uncertainties by assigning fuzzy parameters from a larger domain.
A fermatean fuzzy set (FFS) F can be represented on X which is the universe of
discourse, having the form as
F ¼ fhx; aFðxÞ; bFðxÞi : x 2 Xg; ð1Þ
where each element of X is independently mapped to a value between 0 and 1 as aF (x):

X ! [0, 1] and bF (x): X ! [0, 1] [19]. For all x 2 X includes the following condition,
0 ðaF ð xÞÞ3 + ðbF ð xÞÞ3 1 ð2Þ
The functions aF(x) and bF(x) respectively, indicates the degree of membership and
non-membership of x to in the set F. Thus, FFS is defined as the set where each
element of the universe has a degree of membership and non-membership and the sum
of these degrees’ cubes can be at most equal to 1.
Furthermore, the degree of indeterminacy of x to F is stated as
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pF ðxÞ ¼ 1 ðaF ð xÞÞ3 ðbF ð xÞÞ3 ð3Þ
3
for any FFS F and x 2 X.

When there is only one element in set X, this set ⟨aF(x), bF(x)⟩ called as a Fer-
matean fuzzy number (FFN); denoted by F = (aF, bF) [20]. In short, Fermatean fuzzy
numbers (FFNs) are FFS components.
Comparative representations of Intuitionistic Fuzzy Set (IFS), Pythagorean Fuzzy

Set (PFS) and Fermatean Fuzzy Set (FFS) membership degrees are given in Fig. 1
where x refers a and y refers b. Here, it shows that FFS allows for a higher body of a
degree of membership than IFS and PFS. Therefore, FFS can be applied more widely in
multi-criteria decision-making problems.
Fig. 1. The comparison of membership degrees of IFS, PFS and FFS
3.1 Arithmetic Operations

Let F = (aF , bF ), F1 = (aF1 , bF1 ) and F2 = (aF2 , bF2 ) be three FFNs, and k is a scalar
that is greater than 0. The arithmetic operations are given in Eqs. 4–10 [20, 21]:
Summation:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
F1 F2 ¼ ð 3 a3F1 þ a3F2 a3F1 a3a2 ; bF1 bF2 Þ ð4Þ
Subtraction:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
3 aF1 aF2 bF1 b p1
3 3
F1 F2 ¼ ð ; Þ if aF1 aF2 ; bF1 minfbF2 ; F2 g ð5Þ
1 aF2 F2 3 b p2
Multiplication:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
F1 F2 ¼ ðaF1 aF2 ; b3F1 þ b3F2 b3F1 b3F2 Þ ð6Þ
3
Scalar Product:
k F = ð 1 ð1 a3F Þk ; bkF Þ ð7Þ
3
Exponentiation:
F k = ðakF ; 1 ð1 b3F Þk Þ ð8Þ
3
Division:
αF1 3 βF 3 - βF 3 αF2 π1
1 2
F1 F2 = ( , ) if αF1 ≤ min { αF2 , } , βF ≥ βF ð9Þ
αF2 1- βF 3 π2 1 2
2
3.2 Ranking of Fermatean Fuzzy Sets

Senapati and Yager [19, 20] proposed the score function and the accuracy function to
provide a comparison between FFSs. The score function of F = (aF , bF ), can be
represented as follows for any FFN:
scoreðF Þ ¼ a3F b3F where scoreðF Þ 2 ½1; 1 ð10Þ
The accuracy function of F = (aF , bF ), can be described as follows for any FFN:
accðF Þ ¼ a3F þ b3F ð11Þ
where accðFÞ 2 ½0; 1.

The ranking technique for any two FFNs can be defined as follows depending on
the score and accuracy functions of the FFNs [20]:
Let F1 = (aF1 , bF1 ) and F2 = (aF2 , bF2 ) be two FFNs. score(F1), score(F2), acc(F1)
and acc(F2) are the score values and accuracy values of F1 and F2, then
scoreðF1 Þ\scoreðF2 Þ; then F1 \F2 ; ð12Þ
scoreðF1Þ [ scoreðF2Þ; then F1 [ F2; ð13Þ
accðF1Þ\accðF2Þ; then F1\F2

scoreðF1Þ ¼ scoreðF2Þ; then If accðF1Þ [ accðF2Þ; then F1 [ F2 ð14Þ
accðF1Þ ¼ accðF2Þ; then F1 ¼ F2
4 Fermatean Fuzzy Net Present Worth
Net present worth is one of the most used discounted cash flow methods in the eval-
uation of investment projects that can be described as the summation of the equivalent
values of all cash flow of a project on present time. The equivalent values of cash flows
which are occurring on different time periods is calculated by interest formulas that
represents the time effect on the cash flow amount. There are several assumptions that
are made in the net present worth analysis. One of the critical assumptions is that the
future cash flows, future interest rates and other factors such as useful lives and timing
of the cash receipts can be forecasted with certainty [22]. Most of the future forecasts
involve high uncertainty due to the unpredictable manner of the factors affecting the
forecasted parameter. For this reason, as it is summarized in Sect. 2, fuzzy extensions
of net present worth analysis are proposed
by various authors for different uncertainty
levels. Fuzzy net present worth NPV g can be calculated using Eq. 15 where CF fj
represents fuzzy net cash flow on time j, ~i represents fuzzy interest rate and n represents
study period of the project:
Xn
g ¼
NPV f j 1 þ ~i n
CF ð15Þ
j¼0
If the cash flows of an investment project have annual uniform values, than NPVg can
~
be formulated by Eq. 16 where I denotes initial investment cost which occurs at the
~ denotes annual uniform net cash flow starting at the end of
beginning of the project, A
1 period and lasts at the end of nth period, and SV
st f denotes salvage value/residual
value of the project:

g ¼ ~I þ A
NPV f 1 þ ~i n
~ 1 þ ~i n 1 ~i 1 þ ~i n þ SV ð16Þ
Fermatean fuzzy sets are capable to handle higher levels of uncertainties by assigning
fuzzy parameters from a larger domain. Equation 17 is proposed to calculate fermatean
fuzzy net present worth (FFNPV) as follows:
Xn
FFNPVðaNPV ; bNPV Þ ¼ f j ðaCFj ; bCF Þ 1 þ ~iðai ; bi Þ n
CF ð17Þ
j¼0 j
When the parameters are determined using fermatean fuzzy sets, Eq. 16 can be written
as given in Eq. 18 as follows:
n
~ 1 þ ~iðai ; bi Þ 1 f ðaSV ; bSV Þ
SV
~
FFNPVðaNPV ; bNPV Þ ¼ IðaI ; bI Þ þ AðaA ; bA Þ þ n
~iðai ; bi Þ 1 þ ~iðai ; bi Þ n 1 þ ~iðai ; bi Þ
ð18Þ
When the conditions of subtraction and division operations are satisfied, Eq. 18 can be
rewritten as in Eq. 19 using the arithmetic operations that are given in Eqs. 4–9.
n !
1 þ ~i 1 f
SV
FFNPVðaNPV ; bNPV Þ ¼ ~
~I þ A þ n ðaNPV ; bNPV Þ ð19Þ
~i 1 þ ~i n 1 þ ~i
where;
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ua3 a3 a3SV a3A
u A
þ aSV 3n 3n þ 3 aI
3
t
3
a3i ai
aNPV ¼ i
ð20Þ
1 a3I
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
u 3 3 3
u 1 b 3
b b 3
þ b 1 b b 3
bNPV ¼t i A i SV A i
ð21Þ
3

3 nþ1
bI 1 bi
3
If the conditions for subtraction and division operations are not satisfied in Eq. 19 then
Eq. 22 can be used to calculate FFNPV:
FFNPV ðaNPV ; bNPV Þ

n !
1 þ ~i 1 f
SV
¼ ~I þ A~ þ n ðminðaA ; aSV ; ai ; aI Þ; maxðbA ; bSV ; bi ; bI ÞÞ
~i 1 þ ~i n 1 þ ~i
ð22Þ
5 An Illustration
Let Project I and Project II are two mutually exclusive investment projects which have
the cash flows that are determined using fermatean fuzzy sets and shown in Table 1.
Fuzzy fermatean interest rate is determined as 5% (0.7, 0.3) per year.
Table 1. Cash flows of the projects.

Project I Project II
Initial investment cost 100,000 (0.9, 0.3) 130,000 (0.8, 0.4)
Net annual income 45,000 (0.7, 0.4) 50,000 (0.9, 0.6)
Salvage value 15,000 (0.6, 0.5) 30,000 (0.8, 0.4)
Interest rate 0.05 (0.7, 0.3) 0.05 (0.7, 0.3)
Useful life 5 5
Using Eqs. 19–22, FFNPV of Project I can be calculates as $106,579 (0.6, 0.5) and
FFNPV of Project II can be calculated as $109.979 (0.7, 0.6). Since Project II has a
greater FFNPV, it is selected for the investment.
6 Conclusion
Capital budgeting techniques are used to determine an equivalent value for the cash
flows of an investment project to decide whether it is profitable or not. Since the
estimations of future cash flows and the parameters such as interest rate, useful life are
involves high degrees of uncertainty, fuzzy logic and its extensions are used widely in
economic analysis of investment projects. In this paper, present worth analysis is done
using single valued fermatean fuzzy sets to increase the uncertainty taken into account
in the economic analysis.
For further researches, it is suggested to extend the analysis for several fuzzy
membership functions such as interval valued fermatean fuzzy sets and triangular
fermatean fuzzy sets. Also, fermatean fuzzy sets could be applied to additional capital
budgeting methods such as benefit cost ratio, internal rate of return.
References
1. Baral, G.: Cost–value–profit analysis and target costing with fuzzy logic theory. Mediter.
J. Soc. Sci. 7(2), 21 (2016)
2. Baral, G.: Bulanık Mantık Kuramı Kullanarak Belirsizlik Şartlarında Maliyet-Hacim-Kar
Analizleri. Sakarya Üniversitesi Sosyal Bilimler Enstitüsü, Ph.D. Thesis (2011)
3. Oke, S.A., Johnson, A.O., Popoola, I.O., Charles-Owaba, O.E., Oyawale, F.A.: Application
of fuzzy logic concept to profitability quantification in plastic recycling. Pac. J. Sci. Technol.
7(2), 163–175 (2006)
4. Buckley, J.J.: The fuzzy mathematics of finance. Fuzzy Sets Syst. 21(3), 257–273 (1987)
5. Pohjola, V.J., Turunen, I.: Estimating the internal rate of return from fuzzy data. Eng.Costs
Prod. Econ. 18(3), 215–221 (1990)
6. Chiu, C.Y., Park, C.S.: Fuzzy cash flow analysis using present worth criterion. Eng. Econ.
39(2), 113–138 (1994)
7. Wang, M.J., Liang, G.S.: Benefit/cost analysis using fuzzy concept. Eng. Econ. 40(4), 359–
376 (1995)
8. Chiu, C.Y., Park, C.S.: Capital budgeting decisions with fuzzy projects. Eng. Econ. 43(2),
125–150 (1998)
9. Kahraman, C., Tolga, E., Ulukan, Z.: Justification of manufacturing technologies using
fuzzy benefit/cost ratio analysis. Int. J. Prod. Econ. 66(1), 45–52 (2000)
10. Kuchta, D.: Fuzzy capital budgeting. Fuzzy Sets Syst. 111(3), 367–385 (2000)
11. Karsak, E.E., Tolga, E.: Fuzzy multi-criteria decision-making procedure for evaluating
advanced manufacturing system investments. Int. J. Prod. Econ. 69(1), 49–64 (2001)
12. Kahraman, C., Ruan, D., Tolga, E.: Capital budgeting techniques using discounted fuzzy
versus probabilistic cash flows. Inf. Sci. 142(1–4), 57–76 (2002)
13. Kahraman, C., Beskese, A., Ruan, D.: Measuring flexibility of computer integrated
manufacturing systems using fuzzy cash flow analysis. Inf. Sci. 168(1–4), 77–94 (2004)
14. Ammar, E., Khalifa, H.A.: Characterization of optimal solutions of uncertainty investment
problem. Appl. Math. Comput. 160(1), 111–124 (2005)
15. Sari, I.U., Kahraman, C.: Interval type-2 fuzzy capital budgeting. Int. J. Fuzzy Syst. 17(4),
635–646 (2015)
16. Kahraman, C., Onar, S.Ç., Öztayşi, B.: Interval valued intuitionistic fuzzy investment
analysis: application to CNC lathe selection. IFAC-PapersOnLine 49(12), 1323–1328 (2016)
17. Kahraman, C., Onar, S.C., Oztaysi, B.: Present worth analysis using pythagorean fuzzy sets.
In: Advances in Fuzzy Logic and Technology 2017, pp. 336–342. Springer, Cham (2017)
18. Aydın, S., Kabak, M.: Investment analysis using neutrosophic present and future worth
techniques. J. Intell. Fuzzy Syst. 38(1), 627–637 (2020)
19. Senapati, T., Yager, R.R.: Fermatean fuzzy weighted averaging/geometric operators and its
application in multi-criteria decision-making methods. Eng. Appl. Artif. Intell. 85, 112–121
(2019)
20. Senapati, T., Yager, R.R.: Some new operations over Fermatean fuzzy numbers and
application of Fermatean fuzzy WPM in multiple criteria decision making. Informatica
30(2), 391–412 (2019)
21. Senapati, Tapan, Yager, Ronald R.: Fermatean fuzzy sets. J. Amb. Intell. Hum. Comput. 11
(2), 663–674 (2019). https://doi.org/10.1007/s12652-019-01377-0
22. Sullivan, W.G., Wicks, E.M., Luxhoj, J.T.: Engineering Economy, vol. 12. Prentice Hall,
Upper Saddle River (2003)
Interval Valued q-Rung Orthopair
Fuzzy EDAS Method and Its Application
to Supplier Selection
Elmira Farrokhizadeh1(&) , Seyed Amin Seyfi Shishavan2 ,

Yaser Donyatalab3 , and Seyyed Hadi Seifi4
1
Industrial Engineering Department,
Urmia University of Technology, Urmia, Iran
[email protected]
2
Industrial Engineering Department, PNU University of Urmia, Urmia, Iran
[email protected]
3
University of Moghadas Ardabili, Ardabil, Iran
[email protected]
4
Industrial and Systems Engineering Department, Mississippi State University,
Starkville, MS 39762, USA
[email protected]
Abstract. Nowadays, choosing the best supplier among the various suppliers
has high importance for all enterprises, factories, and purchasers, which require
high accuracy of the decision-making process by using specific methods and
tools for the analysis of relevant factors. To reflect the uncertainty and
changeable environment, fuzzy sets are used frequently by researchers. One of
the extensions of ordinary fuzzy sets is Interval Valued q-Rung Orthopair
Fuzzy Sets (IVq-ROFSs) which is an important way to express uncertain
information. The objective of this paper is supplier assessment using EDAS
method in q-Rung Interval-Valued Orthopair Fuzzy environment in Multiple-
Attribute Group Decision Making (MAGDM) problems. Due to aggregate the
comments of decision-makers, Interval Valued q-Rung Orthopair Fuzzy
Weighted Power Mean aggregation operator (IVq-ROFWPM) is presented and
used. Finally, to show the validity and applicability of the proposed method, a
case study of supplier evaluation in a transformer company is implemented.
Keywords: IVq-ROFSs MAGDM IVq-ROFWPM EDAS Supplier

evaluation
1 Introduction
Theory of fuzzy set introduced by Zadeh (1965) has achieved great success to handle
inexact and imprecise data in various fields [1]. To handle the more ambiguity and
complexity in the real world problem, Interval Valued q-Rung Orthopair Fuzzy Sets
which allow decision-makers to provide their membership and non-membership
degrees to given set of alternatives by an interval value have defined as a useful tool for
458 E. Farrokhizadeh et al.
characterizing uncertainty and complexity of the MAGDM problems [2]. Multi-

Attribute Group Decision Making (MAGDM) which plays an important role in modern
decision science, and widely used in economics, management and the other fields in
recent years, is a process that can rank the finite alternatives according to the more than
one attribute values which can be tangible or intangible with assessing the different
decision maker’s comments [3]. Due to the uncertainty of the decision-making envi-
ronment, usually in many decision-making problems be difficult for DMs to exactly
describe their views with an exact number. This denotes the necessity of developing the
Interval Valued q-Rung Orthopair Fuzzy Sets (IVq-ROFSs) which the degrees of
membership and non-membership are given by an interval value between 0 and 1. In
recent studies, many researchers extended IVq-ROFSs in decision-making problems
[4–7]. The Evaluation based on Distance from Average Solution (EDAS) method was
introduced in 2015 for the first time to solve the multi-criteria inventory classification
[8]. This method played an important role when the conflicting attributes exist in the
MADM problem. The efficiency of the EDAS method is extended in a different type of
fuzzy sets such as IFSs [9], Interval Type-2 Fuzzy Sets [10], Neutrosophic Fuzzy Sets
[11], Interval Fuzzy Sets [12], Linguistic Hesitant Fuzzy Sets [13] and q-ROFSs [14].
Literature analysis has shown that EDAS method with Interval Valued q-Rung
Orthopair Fuzzy Sets (IVq-ROFSs) does not exist to solve the MADM problems. The
purpose of this paper is to Extend the EDAS method for IVq-ROFSs according to the
definition of the traditional EDAS method and IVq-ROFSs to study the MADM
problem with weighted power mean (WPM) aggregation operator which is proposed
for the first time in this paper. It is an important type of means aggregation operator
which is a generalization of the arithmetic and geometric means and acts as a com-
pensative operator for fuzzy set theory [15]. The nature of compensability makes WPM
a flexible and powerful operator that is used in many applications. The WPM has been
proposed in MADM within intuitionistic fuzzy [16], pythagorean fuzzy [17], and q-
rung orthopair fuzzy [1]. This paper proposed EDAS method with IVq-ROFSs to study
the supplier selection problem with weighted power mean (WPM) aggregation operator
to aggregate the comments of decision-makers. In Sect. 2 some basic definitions are
briefly introduced. In Sects. 3 and 4 new aggregation operator based on weighted
power mean and extended EDAS mothode for IVq-ROFSs are proposed and appli-
cability and validity of the new model illustrate in Sect. 5. Finally, in the last section
results are discussed and conclusions of the paper are given.
2 Preliminaries
Definition 1 [5]. An Interval-Valued q-Rung Orthopair Fuzzy Sets (IVq-ROFSs) on

the universe X can be given as follow:
nD h i h iE o
e¼
A l
x; le u
ð xÞ; le l
ð xÞ ; #e u
ð xÞ; #e ð xÞ jx 2 X ð1Þ
A A A A
Interval Valued q-Rung Orthopair Fuzzy EDAS Method 459
h i h i
Where ll ð xÞ; lu ð xÞ and #l ð xÞ; #u ð xÞ 2 ½0; 1 respectively, represent membership
eA eA eA eA
degree and non-membership degree of the element x 2 X to A e which must satisfy the
q q
lu ð xÞ þ #u ð xÞ 1; 8q 1. Then the refusal degree of Interval-Valued q-
eA eA
Rung Orthopair Fuzzy Sets is shown as follow:
h i
l u
pe ð xÞ; pe ð xÞ
A A
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
q q ffi rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
q q ffi ð2Þ
q l l q u u
¼ 1 l ð xÞ þ # ð xÞ ; 1 l ð xÞ þ # ð xÞ
eA eA eA eA
h i h i
e ¼ ll ; lu ; #l ; #u
Definition 2 [5]. Let A be a IVq-ROFS, then Score function
eA eA eA eA
e
S A and Accuracy function H A e are calculated as follow:
h q q q q i
1 þ ll # l þ 1 þ lu # u
e ¼ eA eA eA eA
S A ð3Þ
4
q q q q
ll þ #l þ lu þ # u
e ¼ eA
H A
eA eA eA
ð4Þ
2
So the comparison rulls are shown as follow:

e1 [ S A
1. If S A e 2 , then A e 2;
e1 [ A

e1 ¼ S A
2. If S A e 2 and H A e1 [ H A e 2 , then A e 2;
e1 [ A

e1 ¼ S A
3. If S A e 2 and H A e1 ¼ H A e 2 , then A e 2;
e1 ¼ A

Definition 3 [5]. Let A e1 ¼ ll ; lu ; #l ; #u e2 ¼
, A ll ; lu ; #l ;
eA 1 eA 1 eA 1 eA 1 eA 2 eA 2 eA 2
h i h i
#u e ¼ ll ; lu ; #l ; #u
and A be three IVq-ROFSs. So some basic operation
eA 2 eA eA eA eA
rulls are defined as follow:
8 " sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi q q q qffi# 9
q q q qffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
>
> q q >
>
>
> l l þ l l l l l l ; lu þ lu lu lu ;>
>
< eA 1 eA 2 eA 1 eA 2 eA 1 eA 2 eA 1 eA 2 =
e1 A
A e2 ¼
> >
>
> >
>
>
: #l #l ; #u #u >
;
eA 1 eA 2 eA 1 eA 2
ð5Þ
8 9
>
> >
>
>
> ll ll ; lu lu ; >
>
< eA 1 eA 2 eA 1 eA 2 =
e e
A1 A2 ¼ " s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
q q q q q q q q s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi #
>
> q >
>
>
> #l þ #l #l #l ;
q
#u þ #u #u #u >
>
: eA 1 eA 2 eA 1 eA 2 eA 1 eA 2 eA 1 eA 2 ;
ð6Þ
(" rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi# )
q q k q q k k k
e¼
kA 1 1 ll ; 1 1 lu ; #e l u
#e ð7Þ
eA eA A A
( " ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi#)

k k k r q q k q q k
e ¼
A l
le u
le ; 1 1 #l ; 1 1 #u ð8Þ
A A eA eA

ej ¼
Definition 4 [5]. Let A ll ; lu ; #l ; #u for j ¼ 1; 2; . . .; n be a list of IVq-
eA j eA j eA j eA j
P
ROFSs and wj ¼ ðw1 ; w2 ; . . .; wn ÞT where wj 2 ½0; 1 and nj¼1 wj ¼ 1, then Interval-
Valued q-Rung Orthopair Fuzzy Weighted Arithmatic Mean (q-RIVOFWAM) and
Interval-Valued q-Rung Orthopair Fuzzy Weighted Geometric Mean (q-RIVOFWGM)
aggregation operators are defined as follow:
P
IVqROFWAM A e 1; A e 2 ; . . .; A
e n ¼ n wj A ej
j¼1
8 " sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
q wj sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
q wj # 9
>
> q Qn q Qn >
>
>
> 1 j¼1 1 l l ; 1 j¼1 1 lu ;>
>
< eA j eA j = ð9Þ
>
> >
>
>
> Qn l wj
Qn u wj >
>
: j¼1 e # ; j¼1 e # ;
Aj Aj
Q wj
IVqROFWGM A e 1; A e 2 ; . . .; A
en ¼ n ej
A
8 j¼1 9
> Q n l w j Q n u wj >
>
> l ; l ; >
>
< j¼1 e
Aj
j¼1 e
Aj = ð10Þ
¼ " s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
q wj s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
q wj #
>
> q Q q Q >
>
>
: 1 nj¼1 1 #l ; 1 nj¼1 1 #u >
;
eA j eA j
3 Interval Valued q-Rung Orthopair Fuzzy Weighted Power

Mean Aggregation Operator
Weighted Power Mean (WPM) aggregation operator is one of the integration operators
which is developed for IVq-ROFSs in this section and it’s properties are investigated.
Definition 5 [1]. Let x ¼ ðx1 ; x2 ; . . .; xn Þ be a collection of numbers with weight vector
w ¼ ðw1 ; w2 ; . . .; wn ÞT . The r-th power mean of these numbers is calculated as follow:
Xn 1
r r
WPMwr ðx1 ; x2 ; . . .; xn Þ ¼ i¼1
w i ð x i Þ ; 8r 2 R ð11Þ

Theorem 1. Let Ai ¼ ll ; lu ; #l ; #u for i ¼ ð1; 2; . . .; nÞ be a collection of
eA i eA i eA i eA i
P
IVq-ROFSs with vector w ¼ ðw1 ; w2 ; . . .; wn ÞT , where wi 2 ½0; 1 and ni¼1 wi ¼ 1. So
Interval Valued q-Rung Orthopair Fuzzy Weighted power Mean (IVq-ROFWPM)
aggregation operator is formulated as follow, which is still be a IVq-ROFSs and
satisfied the Idempotency, Monotonicity and Boundedness properties:

IvqROFWPMrw A e 1; A e 2 ; . . .; A en ¼
82 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
rq wi ffi 3 2 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi 39
> Qn q r wi 1r >
>
> rq l q Qn >
>
>
> 6 1 i¼1 1 l ; 7 6 1 1 i¼1 1 1 #l ;7 >
<6 eA i 7 6 e
A 7>=
6 7 6 i 7
6 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7 ; 6 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7
>
> 6 rq wi 7 6 q r wi 1r 7>
>4
> pq Q 5 4 q Qn 5>>
>
>
: 1 ni¼1 1 lu 1 1 1 1 # u >
;
eA i i¼1 eA i
ð12Þ
4 The Extended EDAS Method for IVq-ROFSs
Evaluation based on Distance from Average Solution (EDAS) method which was used
Average Solution (AS) to compare candidate alternatives method is an efficient method
to handle MAGDM problems [10]. In this section, an extended version of the EDAS
method is proposed to deal with multi-criteria group decision-making problems in
interval valued q-rung orthopair fuzzy environment. It means that, in this study, the
decision-makers express their opinions for alternatives with respect to each attribute by
IVq-ROF terms. Let A ¼ fA1 ; A2 ; . . .; Am g be set of alternative, C ¼ fC1 ; C2 ; . . .; Cn g
be a set of attributes with weight vector x ¼ ðx1 ; x2 ; . . .; xn ÞT and D ¼
T
fD1 ; D2 ; . . .; Dk g be a set of decision-makers with weight vector u ¼ Pðnnu1 ; u2 ; . . .; uk Þ
where xh 2 ½0; 1ðh ¼ 1; 2; . . .; nÞ, ui 2 ½0; 1ði ¼ 1; 2; . . .; kÞ, h¼1 xh ¼ 1 and
Pk
u
i¼1 i ¼ 1. Suppose that the attribute C h with respect to alternative Al assessed by
decision makerh D i with
i h Interval Valued
i q-Rung Orthopair Fuzzy Number, is expressed
l u l u
by xihl ¼ lihl ; lihl ; #lihl ; #lihl for all h ¼ 1; 2; . . .; n; l ¼ 1; 2; . . .; m and
i ¼ 1; 2; . . .; k. So the extended EDAS for IVq-ROFSs based on traditional EDAS
method is presented as follow steps:
Step 1. Choose the best Attributes to measure the performance of alternatives.
Step 2. Building the evaluation matrix for each decision-maker.
Step 3. Aggregate comments of decision-maker in one table with IVq-ROFWPM
aggregation operator in Eq. 12.
Step 4. Determine the Average Solution (AS) of all alternatives under each attri-
butes the same as given in Eqs. (13–14):
82 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1ffi 3 " 9
< q Qm l
q
m Qm l
m1 #=
l¼1 1 lhl
e h ¼ 1 m ex hl ¼ 4 q1ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; l¼1 #hl ;
AS ffi5
m l¼1 : q Qmm u
q
m1 ; Qm u
m1 ; ð13Þ
1 l¼1 1 lhl l¼1 #hl
Step 5. Calculate the Positive distance from Average Solution (PDAS) and the
Negative distance from Average Solution (NDAS) matrixes according to either
benefit or not_benefit attribute as given by Eqs. (14–15):

PDASB ¼ NDASN B
e h
¼ max 0; Sðex hl Þ S AS e h
=S AS ð14Þ

NDASB ¼ PDASN B
e h Sðex hl Þ =S AS
¼ max 0; S AS e h ð15Þ

e h and Sðex hl Þ be the score function which determined by Eq. 3.
Where S AS
Step 6. Determine the weighted sum of PDAS (SP) and NDAS (SN) for all alter-
natives as given by Eqs. (16–17):
Xn
SPl ¼ h¼1
xh :PDAShl ð16Þ
Xn
SNl ¼ h¼1
xh :NDAShl ð17Þ
Step 7. Normalize the weighted sum vector of positive and negative distances, as
follow:
NSPl ¼ SPl =ðmaxfSP1 ; SP2 ; . . .; SPm gÞ ð18Þ
NSNl ¼ 1 ðSNl =maxfSN1 ; SN2 ; . . .; SNm gÞ ð19Þ
Step 8. Derive the Integrative Appraisal Score (IAS) for all alternative according to
Eq. 20 as follow:
IASl ¼ ðNSPl þ NSNl Þ=2 ð20Þ
Where IASl 2 ½0; 1

Step 9. Sort the alternatives according to decreasing values of the Integrative
Appraisal Score (AS). The highest one is the best alternative.
5 A Case Study of Supplier Selection

One company supplies raw materials from four major suppliers A1 ; A2; ; ::; A4 . The
company used the method which is presented in this study to rank its suppliers and
select the strategic supplier for long-term business relationships. For this purpose, 8
attributes ðC1 ; C2 ; . . .; C8 Þ were defined such Cost, Quality, Distance, Delivery Time,
Flexibility, Service, Innovation and Green Product with weight vector

ð0:1; 0:2; 0:05; 0:13; 0:08; 0:17; 0:12; 0:15ÞT , respectively. The chief executive officer
of the company formed a group of three experts ðDM1 ; DM2 ; DM3 Þ from the com-
pany’s employees with weight vector ð0:5; 0:3; 0:2ÞT to express their views regardless
of the interdependence of the attributes. The results of using the proposed EDAS
method for this supplier selection is summarized in Tables 1 and 2 with q ¼ 4 and
r ¼ 3.
Table 1. Results of IVq-ROFWPM aggregation operator and Average Solution (AS)

C1 C2 C3 C4 C5 C6 C7 C8

A1 ½0:67; 0:77; ½0:19; 0:4; ½0:47; 0:67; ½0:57; 0:67; ½0:57; 0:87; ½0:57; 0:67; ½0:47; 0:67; ½0:3; 0:4;
½0:28; 0:39 ½0:55; 0:8 ½0:24; 0:45 ½0:55; 0:74 ½0:24; 0:49 ½0:39; 0:49 ½0:35; 0:45 ½0:55; 0:68

A2 ½0:76; 0:86; ½0:73; 0:82; ½0:57; 0:67; ½0:65; 0:8; ½0:53; 0:63; ½0:61; 0:7; ½0:53; 0:62; ½0:52; 0:61;
½0:18; 0:28 ½0:22; 0:34 ½0:38; 0:48 ½0:3; 0:43 ½0:51; 0:61 ½0:45; 0:55 ½0:35; 0:45 ½0:45; 0:67

A3 ½0:47; 0:57; ½0:34; 0:45; ½0:57; 0:76; ½0:55; 0:8; ½0:36; 0:64; ½0:76; 0:86; ½0:54; 0:71; ½0:57; 0:7;
½0:59; 0:71 ½0:47; 0:6 ½0:31; 0:5 ½0:28; 0:39 ½0:36; 0:5 ½0:29; 0:39 ½0:55; 0:67 ½0:3; 0:47

A4 ½0:63; 0:72; ½0:36; 0:49; ½0:44; 0:53; ½0:58; 0:84; ½0:57; 0:8; ½0:67; 0:77; ½0:59; 0:74; ½0:57; 0:69;
½0:37; 0:49 ½0:55; 0:68 ½0:34; 0:52 ½0:28; 0:45 ½0:3; 0:45 ½0:34; 0:44 ½0:24; 0:35 ½0:33; 0:46

AS ½0:66; 0:76; ½0:54; 0:64; ½0:52; 0:67; ½0:59; 0:79; ½0:53; 0:77; ½0:67; 0:77; ½0:54; 0:69; ½0:52; 0:63;
½0:32; 0:44 ½0:42; 0:58 ½0:32; 0:49 ½0:34; 0:49 ½0:34; 0:51 ½0:36; 0:46 ½0:36; 0:47 ½0:4; 0:56
Table 2. SP, SN, NSP, NSN and IAS value of each alternative
SP SN NSP NSN IAS
A1 0.038 0.104 0.585 0 0.293
A2 0.057 0.061 0.890 0.412 0.651
A3 0.064 0.046 1 0.555 0.777
A4 0.033 0.039 0.521 0.622 0.572
According to Table 3, A3 is selected as the best alternative. To illustrate the validity

of the proposed model, the same problem is solved by IVq-ROFWAM and IVq-
ROFWGM methods and results are shown in Table 3. Results of comparing the pro-
posed model with two aggregation operators are a little bit different in ranking of
alternatives but the best alternative is same.
Table 3. Order of alternatives by different methods

Ranking
IVq-ROFWAM A3 [ A4 [ A2 [ A1
IVq-ROFWGM A3 ¼ A4 [ A1 [ A2
IVq-ROF-EDAS A3 [ A2 [ A4 [ A1
6 Discussion and Conclusion
Interval Valued q-Rung Orthopair Fuzzy Sets provide a new way to accept information
and make decisions in uncertain and complex environments. Firstly, this paper
reviewed the definition of IVq-ROFSs and its’ score function, accuracy function and
operational laws. Next, the manuscript introduced a new aggregation operator based on
weighted power mean to combine the IVq-ROFSs. Furthermore, This paper presented
the IVq-ROF-EDAS based on the traditional EDAS method and developed computing
steps of a novel MAGDM method. Proposed method is a more accurate, effective and
confident method for considering the conflicting attributes. Finally, an illustrative
example for supplier selection has been given to represent the validity and applicability
of the proposed model and some comparisons between new model, IVq-ROFWAM
and IVq-ROFWAM operators are also conducted to illustrate advantages of the new
method. The linguistic form of IVq-ROF-EDAS and other extension models can be
applied to other MAGDM problems and many other uncertain and fuzzy environments.
References
1. Du, W.S.: Weighted power means of q-rung orthopair fuzzy information and their
applications in multiattribute decision making. Int. J. Intell. Syst. 34(11), 2835–2862 (2019)
2. Joshi, B.P., Singh, A., Bhatt, P.K., Vaisla, K.S.: Interval valued q-rung orthopair fuzzy sets
and their properties. J. Intell. Fuzzy Syst. 35(5), 5225–5230 (2018)
3. Farrokhizadeh, E. Seyfi Shishavan, S.A. Donyatalab, Y. Kutlu Gündoğdu, F. and Kahraman,
C.: Spherical fuzzy bonferroni mean aggregation operators and their applications to multiple-
attribute decision making. in decision making with spherical fuzzy sets-theory and
applications. In: Kutlu Gündoğdu, F., Kahraman, C. (eds.) Springer (2020). https://doi.org/
10.1007/978-3-030-45461-6_5
4. Ju, Y., Gonzalez, E.D.R.S., Wang, A.: Some interval valued q-Rung orthopair weighted
averaging operators and their applications to multiple attribute decision making. Int. J. Intell.
Syst. 34(10), 2584–2606 (2019)
5. Wang, J., Gao, H., Wei, G.: Methods for multiple-attribute group decision making with q-
Rung interval-valued orthopair fuzzy information and their applications to the selection of
green suppliers. SS symmetry (2019)
6. Gao, H., Ran, L., Wei, G., Wei, C.: VIKOR method for MAGDM based on Q-Rung interval-
valued orthopair fuzzy information and its application to supplier selection of medical
consumption products. Environ. Res. Public Heal. Artic. 17(2), 525–539 (2020)
7. Xu, Y., Shang, X., Wang, J., Zhao, H., Zhang, R., Bai, K.: Some interval-valued q-Rung
dual hesitant fuzzy muirhead mean operators with their application to multi-attribute
decision-making. IEEE Access 7, 54724–54745 (2019)
8. Ghorabaee, M.K., Zavadskas, E.K., Olfat, L., Turskis, Z.: Multi-criteria inventory
classification using a new method of evaluation based on distance from average solution
(EDAS). Informatica 26(3), 435–451 (2015)
9. Kahraman, C., Keshavarz Ghorabaee, M., Zavadskas, E.K., Cevik Onar, S., Yazdani, M.,
Oztaysi, B.: Intuitionistic fuzzy EDAS method: an application to solid waste disposal site
selection. J. Environ. Eng. Landsc. Manag. 25(1), 1–12 (2017)
10. Ghorabaee, M.K., Amiri, M., Zavadskas, E.K., Turskis, Z., Antucheviciene, J.: A new multi-
criteria model based on interval type-2 fuzzy sets and EDAS method for supplier evaluation
and order allocation with environmental considerations. Comput. Ind. Eng. 112, 156–174
(2017)
11. Peng, X., Liu, C.: Algorithms for neutrosophic soft decision making based on EDAS, new
similarity measure and level soft set. J. Intell. Fuzzy Syst. 32(1), 955–968 (2017)
12. Ilieva, G.: Group decision analysis algorithms with EDAS for interval fuzzy sets. Cybern.
Inf. Technol. 18, 51–64 (2018)
13. Feng, X., Wei, C., Liu, Q.: EDAS method for extended hesitant fuzzy linguistic multi-
criteria decision making. Int. J. Fuzzy Syst. 20(8), 2470–2483 (2018)
14. Li, Z., Wei, G., Wang, R., Wu, J., Wei, C., Wei, Y.: EDAS method for multiple attribute
group decision making under q-rung orthopair fuzzy environment. Technol. Econ. Dev.
Econ. 26(1), 86–102 (2020)
15. Beliakov, G., Bustince Sola, H., Calvo, Sánchez T.: A Practical Guide to Averaging
Functions. Springer, Heidelberg (2016)
16. Tan, C., Chen, X.: Generalized archimedean intuitionistic fuzzy averaging aggregation
operators and their application to multicriteria decision-making. Int. J. Inf. Technol. Decis.
Mak. 15(2), 311–352 (2016)
17. Wei, G., Lu, M.: Pythagorean fuzzy power aggregation operators in multiple attribute
Hamacher Aggregation Operators Based
on Interval-Valued q-Rung Orthopair Fuzzy
Sets and Their Applications to Decision
Making Problems
Yaser Donyatalab1 , Elmira Farrokhizadeh1 ,

Seyed Amin Seyfi Shishavan1(&) , and Seyyed Hadi Seifi2
1
[email protected]
2
Industrial and Systems Engineering Department, Mississippi State University,
Starkville, MS 39762, USA
[email protected]
Abstract. As a generalization of Intuitionistic Fuzzy Sets (IFSs), q-Rung

Orthopair Fuzzy Sets (q-ROFSs) provide a larger space for decision makers and
can express their uncertain information with more flexibly. This article deals
with the Multiple Attribute Group Decision Making (MAGDM) problem based
on Interval Valued q-Rung Orthopair Fuzzy Sets. First, we introduce some new
operational rules of Interval Valued q-Rung Orthopair Fuzzy Sets (IVq-ROFs),
then based on these operations, some Interval Valued q-Rung Orthopair Fuzzy
Hamacher Aggregation Operators (IVq-ROFHAO) such as Interval Valued q-
Rung Orthopair Fuzzy Sets Hamacher Weighted Arithmetic Mean ðIVq
HWAMHc Þ are developed for IVq-ROFSs. Finally, an illustrative numerical
example is presented to test the effectiveness of the proposed methods in
MAGDM problems.
Keywords: IVq-ROFSs MAGDM Hamacher t-norm Hamacher

t-conorm Hamacher aggregation operators IVq-ROFHAO IVq HWAMHc
1 Introduction
Due to the uncertainty and variability nature of events in the real-world, decision
information includes alternative performance measures that are shown in the form of
ambiguous variables rather than precise numerical ones [1]. The fuzzy set theory has an
important impact on the progress of decision-making problem in an uncertain envi-
ronment [2]. Various researchers have developed several extensions of ordinary fuzzy
sets and the extended fuzzy sets have utilized these extensions in the solution of multi-
criteria decision-making problems [3]. Type-2 Fuzzy Sets [4], Intuitionistic Fuzzy Sets
[5], Neutrosophic Fuzzy Sets [6], Hesitant Fuzzy Sets [7], Pythagorean Fuzzy Sets [8],
Picture Fuzzy Sets [9] and Spherical Fuzzy Sets [10] are the most popular extensions of
Hamacher Aggregation Operators 467
ordinary fuzzy sets. One of the novel extensions of ordinary fuzzy sets is q-Rung
Orthopair Fuzzy Sets (q-ROFSs), which was developed by [11] is generalized form of
IFS and PFS. q-ROFSs can fully express decision-makers’ comments and describe the
decision information precisely by parameters which can flexibly adjust the scope of
information expression [12]. In many practical decision processes, due to insufficiency
in available information, it is difficult for DMs to exactly express their opinions by a
crisp number, but they can be represented by an interval number within ½0; 1. Then
Interval Valued q-Rung Ortho Pair Fuzzy Sets (IVq-ROFSs) are introduced in [13].
In this paper, inspired by the capabilities of the IVq-ROFSs, we extend the
Hamacher t-conorm and t-norm to IVq-ROFSs and propose some novel aggregation
operators for IVq-ROFSs, which include Interval Valued q-Rung Orthopair Fuzzy Sets
Hamacher Weighted Arithmetic Mean ðIVq HWAMHk Þ and Interval Valued q-Rung
Orthopair Fuzzy Sets Hamacher Weighted Geometric Mean ðIVq HWGMHk Þ.
In order to do so, the paper will be set out as follows: In Sect. 2, we introduce the
concept and operations of Interval Valued q-Rung Orthopair Fuzzy Sets (IVq-ROFSs).
In Sect. 3, some hamacher operations for IVq-ROFS are developed based on Hamacher
t-norm and s-norm. In Sect. 4, we develop and propose new Hamacher aggregation
operators. Section 5, gives an illustrative example to show the feasibility and practi-
cability of the developed approach, and provides a comparative analysis to discuss the
effect on the results when the parameters k and C taking different values also include
other related group decision-making methods. Finally, the main conclusions of this
paper are summarized in Sect. 6.
2 Preliminaries and Basic Concepts
2.1 Interval Valued q-Rung Orthopair Fuzzy Sets (IVq-ROFSs)
Definition 1. Let X be a nonempty fixed set, then Interval Valued q-Rung Orthopair
Fuzzy Sets (IVq-RDHFS) A~ on X is defined as:
L
~¼ x; lA ð xÞ; lU
A ð xÞ; ½#A ð xÞ; #A ð xÞ jx 2 X
L U
A
L
Where lLA ð xÞ; lU
A ð xÞ and #A ð xÞ; #A ð xÞ denote the membership degree and non-
U
membership degree
of x 2 X respectively,
L which satisfies
the following condition: for
q q
every x 2 X: lLA ð xÞ; lU
A ð x Þ 2 ½ 0; 1 , #A ð x Þ; #U
A ð x Þ 2 ½ 0; 1 and lU
A ð xÞ þ #A ð xÞ 1,
U
ðq 1Þ, and,
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
U
q ffi
q q
½pLA ð xÞ; pU
A ð xÞ ¼ ½ 1 ðlU A ð xÞÞ #A ð xÞ ;
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð1Þ
q
1 ðlLA ð xÞÞq ð#LA ð xÞÞq
Denotes the hesitancy degree of x 2 X.

468 Y. Donyatalab et al.
For convenience, we call ½lLA ð xÞ; lU ð xÞ; ½#LA ð xÞ; #U

A ð xÞ an IVq-ROFN, which is
L U L U A
denoted by ~a ¼ la ; la ; #a ; #a .

Definition 2. Let A ~ ¼ hðlL ; lU Þ; ð#L ; #U Þi, A ~ 1 ¼ lL ; lU ; #L ; #U ~2 ¼
and A
L U
L U
1 1 1 1
l2 ; l2 ; #2 ; #2 be three IVq-ROFNs, and k be a positive real number, then the
operational laws are defined as follows [13]:

1

1
~1 A
A ~ 2 ¼½ lL q þ ðlL Þq ðlL Þq lL q q ; lU q þ ðlU Þq ðlU Þq lU q q ;
1
L L U U
2 1 2 1 2 1 2 ð2Þ
#1 #2 ; #1 #2

1
~1 A
A ~ 2 ¼ lL lL ; lU lU ; ½ð #L q þ #L q ð#L Þq #L q Þq ;
1 2 1 2 1 2 1 2

q
q
q ð3Þ
ð #L1 þ #L2 ð#L1 Þq #L2 Þ1=q
1q

L q k
U
q
k 1q h L
k U
k i
~¼
kA 1 1l Þ ; 1 1 l ; # ; # ; k[0 ð4Þ

L
k U
k
1=q
1=q
~k ¼
A L q k
l ; l ;½ 1 1 ð# Þ U q k
; 1 1 ð# Þ ; k [ 0 ð5Þ
Furthermore, a score function and an accuracy function of a q-ROFN are defined as

follows:
~ ¼ ðlL ; lU Þ; ð#L ; #U Þ be an IVq-ROFN, then the score function

~ and the accuracy function H A

S A ~ of A ~ are defined as follows [13]:

1

~ ¼ 1 þ lL q #L q þ 1 þ lU q #U q ; S A
S A ~ 2 ½0; 1 ð6Þ
A A A A
4

1 L
q U
q L
q U
q
~ ¼
H A lA þ lA þ #A þ #A ; H A ~ 2 ½0; 1 ð7Þ
2
Based on the score and the accuracy function, a comparison method for IVq-
ROFNs is defined as follows:
e 1 and A
Definition 4. Let A e 2 be two IVq-ROFNs, then

e1 [ S A
1. If S A e 2 , then A
e1 [ Ae 2;

e1 ¼ S A
2. If S A e 2 , then

(a) If H Ae1 [ H Ae 2 , then A
e1 [ A e 2;

(b) If H Ae1 ¼ H Ae 2 , then A
e1 ¼ A e 2:
3 Hamacher Operations of Interval Valued q-Rung

Orthopair Fuzzy Sets
In this section, some Hamacher operations, i.e., Hamacher product and Hamacher sum
of two IVq-ROFNs h1 and h2 , c > 0 are presented as follows:
Dh i h iE
Definition 5. Let A ~ q ðxi Þ ¼ lL~ ðxi Þ; lU ðxi Þ ; #L~ ðxi Þ; #U ðxi Þ ; 8i ¼ 1; 2; . . .n and
~q ~q
Dh i h Aq A
iE Aq A
e q ðxi Þ ¼ lL~ ðxi Þ; lU~ ðxi Þ ; #L~ ðxi Þ; #U~ ðxi Þ ; 8i ¼ 1; 2; . . .n be two Interval
B B q B q B q B q
Valued q-Rung Ortho Pair Fuzzy Numbers (IVq-ROFNs) which is defined on universe
discourse X and xi 2 X, which for more computation convenient, be represented as
follows:

e q ðxi Þ ¼ a ~ ðxi Þ; b ~ ðxi Þ ; c ~ ðxi Þ; d ~ ðxi Þ ; 8i ¼ 1; 2; . . .n and B
A ~ q ðxi Þ ¼ ½aB~ ðxi Þ;
A A A A q
bB~ ðxi Þ; ½cB~ ðxi Þ; dB~ ðxi Þiq ; 8i ¼ 1; 2; . . .n and q 1, then Hamacher Addition, Multi-
plication, Scalar Multiplication and Scalar Power Operations will be as follows:
2 !1q 3
aq~ ðxi Þ þ aq~ ðxi Þaq~ ðxi Þaq~ ðxi Þð1kÞaq~ ðxi Þaq~ ðxi Þ
6 A B A B A B
;7
6 1ð1kÞaq ðxi Þaq~ ðxi Þ 7
6 eA B 7
6 7;
6 q 1 7
4 q q q q q
b ~ ðxi Þ þ b ~ ðxi Þb ~ ðxi Þb ~ ðxi Þð1kÞb ~ ðxi Þb ~ ðxi Þ q 5
A B A B A B
1ð1kÞbq~ ðxi Þbq~ ðxi Þ
A B
e q ðxi Þ Hk B
8i : A e q ð xi Þ ¼ h i ð8Þ
2 !1q 3
cq ðxi Þcq ðxi Þ
6
~
A
eB
;7
6 k þ ð1kÞ cq ðxi Þ þ cq ðxi Þcq ðxi Þcq ðxi Þ 7
6 ~ ~ ~ ~ 7
6 7
A B A B
6 1 7
4 q q
d ðx Þd ðx Þ q 5
q A~ i qB~ i q q
k þ ð1kÞ d ~ ðxi Þ þ d ~ ðxi Þd ~ ðxi Þd ~ ðxi Þ

A B A B
2 1q 3
aq~ ðxi Þaq~ ðxi Þ
6 A B
;7
6 k þ ð1kÞ aq~ ðxi Þ þ aq~ ðxi Þaq~ ðxi Þaq~ ðxi Þ 7
6 A B A B
7
6 !1q 7;
6 bq~ ðxi Þbq ðxi Þ 7
4
A
eB
5
k þ ð1kÞ bq~ ðxi Þ þ bq~ ðxi Þbq~ ðxi Þbq~ ðxi Þ
A B A B
e q ð xi Þ H B
8i : A e q ðxi Þ ¼ h i ð9Þ
k
2 1q 3
cq~ ðxi Þ þ cq~ ðxi Þcq~ ðxi Þcq~ ðxi Þð1kÞcq~ ðxi Þcq~ ðxi Þ
6 A B AB
1ð1kÞcq~ ðxi Þcq~ ðxi Þ
;7 A B
6 7
6 7
A B
6 q 1 7
4 d~ ðxi Þ þ d~q ðxi Þd q~ ðxi Þdq~ ðxi Þð1kÞd~q ðxi Þd~q ðxi Þ q 5
A B A B A B
1ð1kÞd ~q ðxi Þd ~q ðxi Þ
A B
2
!1q 3
a a
1ð1kÞaq~ ðxi Þ 1aq~ ðxi Þ
6
A A
;7
6 a
7
1ð1kÞaq~ ðxi Þ ð1kÞ 1aq~ ðxi Þ
a
6 7
6 7;
A A
; 6 ! 1 7
6
a
a q 7
4 1ð1kÞbA~ ð
xi Þ 1b
q
A~ qi
a 5
q
ðx Þ
a
q
1ð1kÞb ~ ðxi Þ ð1kÞ 1b ~ ðxi Þ
A A
8i : a Hk e q ð xi Þ ¼ h
A i ð10Þ
2 q
a !1q 3
k c ~ ðxi Þ
6
A a
a ; 7
6 1ð1kÞcq~ ðxi Þ ð1kÞ cq~ ðxi Þ 7
6 7
6 7
A A
6 q
a !1q 7
6 7
4
k d ~ ð xi Þ

A
5
a a
1ð1kÞd q~ ðxi Þ ð1kÞ d q~ ðxi Þ
A A
2 !1q 3

a
6 k aq~ ðxi Þ
6
A
a
a ; 77
6 1ð1kÞaq~ ðxi Þ ð1kÞ aq~ ðxi Þ 7
6 A A
7
6 ;
6 q
a !1q 7
7
6 7
4
k b ~ ðxi Þ

A a q
a 5
q
1ð1kÞb ~ ðxi Þ ð1kÞ b ~ ðxi Þ
A A
ð11Þ
e q ð xi Þ H a ¼ h
8i : A i
k
2 !1q 3
a
a
6 1ð1kÞcq~ ðxi Þ 1cq~ ðxi Þ

6
A
a
A a
;7 7
6 1ð1kÞcq~ ðxi Þ ð1kÞ 1cq~ ðxi Þ 7
6 A A
7
6 ! 7
6

1 7
6 a a q 7
4 A~ qi
a 5
q q
1 ð 1k Þd ð x Þ 1d ð x Þ
q
A~ i

a
1ð1kÞd ~ ðxi Þ ð1kÞ 1d ~ ðxi Þ
A A
Where k 2 ð0; 1Þ and a [ 0 is a crisp number.
4 Interval Valued q-Rung Orthopair Fuzzy Sets Hamacher

Aggregation Operators
In this section, we present some IVq-ROFS Hamacher aggregation operators based on

the operational laws that expressed in Sect. (3).
4.1 Interval Valued q-Rung Orthopair Fuzzy Sets Hamacher Weighted

Arithmetic Mean ðIVqHWAM H c Þ
Definition 6. Let A e q ðxi Þ; 8i ¼ 1; 2; . . .n be a collection of IVq-ROFNs are be repre-

e q ðxi Þ ¼ a ~ ðxi Þ; b ~ ðxi Þ ; c ~ ðxi Þ; d ~ ðxi Þ ; 8i ¼ 1; 2; . . .n is defined on
sented as: A A A A A q
universe discourse X and xi 2 X, and aA~ ðxi Þ and bA~ ðxi Þ are lower bound and upper
bound for membership degree respectively, cA~ ðxi Þ and dA~ ðxi Þ are lower bound and
upper bound for non-membership degree respectively, with corresponding aggregation
Pn
weight vector wi ¼ ðw1 ; w2 ; . . .; wn Þ where wi 2 ½0; 1 and wi ¼ 1 and having the
i¼1
power q [ 0, then IVq HWAMHk will be defined as follows:
0 1
n n

B C
IVq HWAMHk A~ q ðxi Þ ¼ Hk Bwi Hk A
~ q ðxi Þ C ¼
@ A
i¼1 i¼1
2 3
Qn q

wi Qn q
wi !1q
6 ð Þa ð Þ ð Þ
A~ qi
wi ; 7
1 1k x 1a x
6 Qn i¼1 A~ i

wi Q
i¼1
n 7
6 1 ð1k Þa q
~
ðx i Þ ð 1k Þ 1a ~ ðxi Þ 7
6 i¼1 A i¼1 A
7
6 ;
6 Qn
Q
!1q 7
7
6 wi n w i 7
4
q
A q
wi 5
q
1ð1kÞb ~ ðxi Þ 1b ~ ðxi Þ
Qn i¼1 q
A
wi Q
i¼1
n ð12Þ
i¼1
1ð1kÞb ~ ðxi Þ ð1kÞ i¼1
1b ~ ðxi Þ
A A
2 !1q 3
Qn q
w i
6 k c ðx Þ
6 Qn
wi A~ i Qn
i¼1
wi ; 7 7
6 q
1ð1kÞc ~ ðxi Þ ð1kÞ cq~ ðxi Þ 7
6 i¼1 A i¼1 A
7
6 ! 7
6 Qn q
1 7
6 wi q 7
4 Qn
k

wi A~ i Qn q
wi 5
i¼1
d ð x Þ
q
i¼1
1ð1kÞd ~ ðxi Þ ð1kÞ i¼1
d ~ ðxi Þ
A A
Where Hamacherity coefficient k 2 ð0; 1Þ.
4.2 Interval Valued q-Rung Orthopair Fuzzy Sets Hamacher Weighted

Geometric Mean ðIVq HWGM H c Þ
Definition 7. Let A e q ðxi Þ; 8i ¼ 1; 2; . . .n be a collection of IVq-ROFNs are be repre-

e
sented as: A q ðxi Þ ¼ aA~ ðxi Þ; bA~ ðxi Þ ; cA~ ðxi Þ; dA~ ðxi Þ q ; 8i ¼ 1; 2; . . .n is defined on
universe discourse X and xi 2 X, and aA~ ðxi Þ and bA~ ðxi Þ are lower bound and upper
bound for membership degree respectively, ce ðxi Þ and de ðxi Þ are lower bound and
A A
upper bound for non-membership degree respectively, with corresponding aggregation

P
n
weight vector wi ¼ ðw1 ; w2 ; . . .; wn Þ where wi 2 ½0; 1 and wi ¼ 1 and having the
i¼1
power q [ 0, then IVq HWGMHk will be defined as follows:
0 1
n n

IVq HWGMHk A~ q ðxi Þ ¼ Hk B

@A~ q ðxi Þ ^Hk wi C
A¼
i¼1 i¼1
2 !1q 3
Qn q
w i
k a ~ ðxi Þ
6 Qn i¼1
A
Qn q
w i ; 7
6 1ð1kÞ 1aq~ ðxi Þ
wi
ð1kÞ a ~ ð xi Þ 7
6 7
6 7;
i¼1 A i¼1 A
6 Qn q
w i ! 7
6 q 7
1
4 k a ðx Þ 5 ð13Þ
Qn q i¼1
Awi Qn q
wi
~ i
i¼1
1ð1kÞ 1b ~ ðxi Þ ð1kÞ i¼1
b ~ ð xi Þ
A A
2
wi Qn q
wi !1q 3
Qn
1ð1kÞcq~ ðxi Þ 1c ~ ðxi Þ
6 Qn i¼1 A
Q
i¼1 A q
wi ; 7
6 q
wi
1ð1kÞc ~ ðxi Þ ð1kÞ
n
1c ~ ðxi Þ 7
6 7
6 7
i¼1 A i¼1 A
6 Qn
Q
!1 7
6 wi wi q 7
4 A q
wi 5
q n q
1ð1kÞd ~ ðxi Þ 1d ~ ðxi Þ
Qn i¼1 q
A
wi Q
i¼1
n
i¼1
1ð1kÞd ~ ðxi Þ ð1kÞ i¼1
1d ~ ðxi Þ
A A
Where Hamacherity coefficient k 2 ð0; 1Þ.
5 An Illustrative Example
In this section, a numerical example which that adapted from Gao et al. [14] is pre-
sented to show feasibility and advantages of the proposed new method. There exist a
panel with five possible suppliers, and three experts by considering four criteria are
comment on these suppliers based on IVq-ROFSs to select the best supplier. Experts
are going to comment for alternatives based on four criteria as: environment
improvement quality, supplier’s transportation n convenience, green o image and envi-
ronmental competencies. So we have A e¼ A e i j8i ¼ 1; 2; . . .; 5 set of possible alter-
n o
natives and Ce¼ C e j j8j ¼ 1; 2; . . .; 4 set of criteria are defined based on IVq-ROFSs

with crisp weight vector w ¼ f0:27; 0:37; 0:16; 0:2g, and E e¼ E e k j8k ¼ 1; 2; 3 is the
set of experts who comments on each alternative based on criteria and defined based on
IVq-ROFSs, with crisp weight vector h ¼ f0:35; 0:25; 0:4g. The proposed method will
be used to rank the alternatives.
Step 1. Construct alternative-criteria decision matrix based on IVq ROFNs
ðq ¼ 3Þ for every decision maker tables.
Step 2. Aggregate the decision matrices for each decision makers based on pro-
posed (IVq HWAMHk ) aggregation operator.
Step 3. Aggregate the achieved decision matrix from step 2 for criteria based on
proposed (IVq HWAMHk ) aggregation operator.
Step 4. Use the comparison function which is defined in Definition (4) to compare
the alternatives.
Step 5. Rank the alternatives. Results are shown in Table 1.
Table 1. Results of the proposed aggregation operators

IVq HWAMHk ðq ¼ 3; k ¼ 1Þ Score Ranking
h h
lLIVqHWAMH ; #LIVqHWAMH ;
k k
IVqHWAMH
lU IVqHWAMH
#U
k k
e 1 0.5066 0.6212 0.3078 0.4693 0.5593 4

A
e 2 0.5963 0.7224 0.2314 0.3456 0.6338 1
A
e 3 0.4103 0.5496 0.3741 0.5266 0.5092 5
A
e 4 0.5150 0.6551 0.3289 0.5097 0.5624 2
A
e 5 0.4293 0.6858 0.2697 0.5105 0.5622 3
A
Based on IVq HWAMHk with ðq ¼ 3; k ¼ 1Þ it is aggregated to the final decision

matrix and then the score function calculated for every alternative to get ranking. It is
mentioned in Table 1 that when ðk ¼ 1Þ the IVq HWAMHk will turn into IVq
WAMH1 aggregation operator. Therefore, the resulted final aggregated decision matrix
and rank of that are belong to IVq WAMH1 . The illustrative example based on pro-
posed MAGDM method with different k amounts for IVq HWAMHk is solved and
results are shown in Tables 2. Table 2 related to third degree ðq ¼ 3Þ of IVq-ROFNs,
and with different values for q results for rank of alternatives will change.
Table 2. Result of IVq HWAMHk with different ðq; kÞ values

IVq HWAMHk
q¼3 q¼4 q¼5
k¼1 k¼2 k¼3 k¼4 k¼1 k¼2 k¼3 k¼4 k¼1 k¼2 k¼3 k¼4
e1
A 4 3 3 3 4 2 2 3 4 3 2 2
e
A2 1 1 1 1 1 1 1 1 1 1 1 1
e3 5
A 5 5 5 5 5 5 5 5 5 5 5
e4 2
A 4 4 4 2 4 4 4 2 4 4 4
e5 3
A 2 2 2 3 3 3 2 3 2 3 3
6 Conclusion
This study aims to propose a new technique to evaluate alternatives for a group
decision making based on Hamacher aggregation operators. In this paper, we proposed
some basic operations of IV-qROFSs based on Hamacher t-norms and t-conorms and
then introduce two aggregation operators. An illustrative example is applied using the
proposed aggregators in IVq-ROFSs and the improvements in ranking quality are
clearly shown. Future studies will allow more detailed comparison of the proposed
aggregation techniques with other aggregation techniques and presenting better deci-
sions in a complex environment.
References
1. Jin, Y., Wu, H., Merigó, J.M., Peng, B.: Generalized Hamacher aggregation operators for
intuitionistic uncertain linguistic sets: multiple attribute group decision making methods.
Information 10(6), 1–22 (2019)
2. Donyatalab, Y. Farrokhizadeh, E. Seyed Garmroodi, S. D. and Seyfi Shishavan, S. A. :
Harmonic Mean Aggregation Operators in Spherical Fuzzy Environment and Their Group
Decision Making Applications. J. Mult.-Valued Log. Soft Comput 33(6), 565–592 (2019)
6. Smarandache, F., Neutrosophy: neutrosophic probability, set, and logic: analytic synthesis &
synthetic analysis. ISBN 1-879585-76-6. Contents : Preface by Charles, T. Le : 3. (2000)
8. Yager, R.R.: Pythagorean fuzzy subsets. In: Joint IFSA World Congress NAFIPS Annual
Meeting, vol. 2, no. x, pp. 57–61 (2013)
9. Cường, B.C.: Picture fuzzy sets. J. Comput. Sci. Cybern. 30(4), 409–420 (2015)
10. Gündoǧdu, F.K., Kahraman, C.: Spherical fuzzy sets and spherical fuzzy TOPSIS method.
11. Yager, R.R.: Generalized Orthopair Fuzzy Sets. IEEE Trans. Fuzzy Syst. 25(5), 1222–1230
(2017)
C.: Spherical Fuzzy Bonferroni mean aggregation operators and their applications to
multiple-attribute decision making. In: Kutlu Gündoğdu, F., Kahraman, C. (eds.) Decision
Making with Spherical Fuzzy Sets-Theory and Applications. Springer (2020). https://doi.
org/10.1007/978-3-030-45461-6_5
13. Joshi, B.P., Singh, A., Bhatt, P.K., Vaisla, K.S.: Interval valued q-rung orthopair fuzzy sets
and their properties. J. Intell. Fuzzy Syst. 35(5), 5225–5230 (2018)
14. Gao, H., Ran, L., Wei, G., Wei, C., Wu, J.: VIKOR method for MAGDM based on Q-Rung
interval-valued orthopair fuzzy information and its application to supplier selection of
medical consumption products. Int. J. Environ. Res. Public Health 17(2), 525 (2020)
Similarity Measures of q-Rung Orthopair
Fuzzy Sets Based on Square Root Cosine
Similarity Function
Yaser Donyatalab1(&), Elmira Farrokhizadeh2,

and Seyed Amin Seyfi Shishavan3
1
Industrial Engineering Department, University of Moghadas Ardabili,
Ardabil, Iran
[email protected]
2
Industrial Engineering Department, Urmia University of Technology,
Urmia, Iran
[email protected]
3
[email protected]
Abstract. Similarity measures as an important branch of Multiple Attribute

Decision Making (MADM) problems have been played important roll to
determine the degree of similarity between two objects. Currently, cosine sim-
ilarity based on euclidean distance is one of the most widely used similarity
measures in fuzzy environment. q-Rung Orthopair Fuzzy Sets (q-ROFSs) is one
of the most favorable extensions of fuzzy sets to declare the uncertainty by as
membership and non-membership degrees. In this paper, a new similarity
measure is extended based on square root cosine similarity for q-Rung Orthopair
Fuzzy Sets in two state of with and without hesitancy degree. Four different
forms of square root cosine similarity measures are developed in this paper for
q-Rung Orthopair Fuzzy Sets (q-ROFSs). In order to demonstrate the applica-
bility of the proposed methods, an illustrative example is presented.
Keywords: q-ROFSs MADM Square Root Similarity measure Cosine

similarity
1 Introduction
To determine the similarity degree between two objects in Multiple Attribute Decision-
Making (MADM) problems, similarity measures have been regarded as very beneficial
tools. The applications of similarity measures have been investigated in a variety of
fields, for instance, pattern recognition, scheme selection, machine learning, image
processing, and decision-making [1]. Many similarity measures have been developed in
the literature. Chen and Tan [2], proposed the application of similarity measures for
measuring the degree of similarity between ambiguous sets. Szmidt and Kacprzyk [3],
proposed some efficient similarity measures based on the Hamming distance and the
Euclidean distance between IFSs. Hwang and Yoon [4], proposed a new similarity
measure based on the TOPSIS method to avoid the strong similarity between two IFSs
when their distance is quite little. Also, the similarity measures for hesitant fuzzy sets
(HFSs) extended by [5]. They also proposed the variety of similarity measures for
HFSs. Zeng et al. [6], proposed a variety of distance measures for Pythagorean fuzzy
sets, which considered the five parameters (membership degree, non-membership
degree, hesitancy degree, strength of commitment about membership, and direction of
commitment) of Pythagorean fuzzy sets. Wei [7], developed cosine similarity mea-
sures, such as, weighted cosine similarity measure, set-theoretic similarity measure,
weighted set-theoretic cosine similarity measure, grey similarity measure and weighted
grey similarity measure between picture fuzzy sets. In the recent years, Liu et al. [13]
and Wang et al., considered some cosine similarity measures and distance measures
between q-rung orthopair fuzzy sets (q-ROFSs) considering the function of member-
ship degree, non-membership degree, and indeterminacy membership degree, based on
the traditional cosine similarity.
Cosine Similarity is considered as one of the prevalent and frequently used simi-
larity measures among the similarity measures, which is defined in vector space [8].
The cosine similarity can be obtained directly from Euclidean distance; however,
Euclidean distance is generally not a desirable metric for high-dimensional data mining
applications and probabilities. In order to overcome the shortcomings of Euclidean
distance, we propose Square Root Cosine Similarity Function based on Hellinger
distance for the q-rung Orthopair Fuzzy Sets. Using fuzzy logic [9], can be an
appropriate approach to deal with uncertainty nature of this type of problems. In order
to better expressing the uncertain information, the q-rung Orthopair fuzzy Sets (q-
ROFSs) are useful tools.
The rest of the presented paper is listed in the following. In Sect. 2, the fundamental
notions and concepts of IFSs, PFSs, q-ROFs, cosine similarity measure and square root
cosine function are shortly retrospect, which will be employed in the analysis in the
next section. In Sect. 3, we propose the new similarity measures based on square root
cosine function for q-ROFs. In Sect. 4, a numerical example is given to show the
effectiveness of the proposed similarity measure. The paper is concluded in Sect. 5.
2 Preliminaries and Basic Concepts
In this section, we review some basic knowledge about IFS, PFS and q-ROFSs, and
their related properties. Throughout the paper, let X ¼ fx1 ; x2 ; . . .; xn g be a finite set,
discrete and nonempty.
2.1 Intuitionistic Fuzzy Set (IFS)
Definition 1 [10]. Let X ¼ fx1 ; x2 ; . . .; xn g be a fixed set; then the IFS I on X is defined
as:
~I ¼ fhxi ; lI ðxÞ; #I ðxÞijxi 2 X g ð1Þ

Similarity Measures of q-Rung Orthopair Fuzzy Sets 477
where lI ð xÞ : X 2 ½0; 1 and #I ð xÞ : X 2 ½0; 1 represent the membership and non-

membership degrees of xi 2 X, respectively. For each xi 2 X, the below condition must
be satisfied: 0 lI þ #I 1;
Also, the hesitancy degree of IFS can be defined as pI ðxÞ ¼ 1 lI #I , where
0 pI ðxÞ 1.
2.2 Pythagorean Fuzzy Set (PFS)
Definition 2 [11]. Let X ¼ fx1 ; x2 ; . . .; xn g be a fixed set; then the PFS P on X is

defined as:
~ ¼ fhxi ; lP ðxÞ; #P ðxÞijxi 2 X g

P ð2Þ
where lP ð xÞ : X 2 ½0; 1 and #P ð xÞ : X 2 ½0; 1 represent the membership and non-

membership degrees of xi 2 X, respectively. For each xi 2 X, the below condition must
be satisfied: 0 l2P þ #2P 1;
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Also, the hesitancy degree of PFS can be defined as pP ðxÞ ¼ 1 l2P #2P , where
0 pP ðxÞ 1.
2.3 q-Rung Orthopair Fuzzy Sets (Q-ROFSs)

As mentioned in the Sect. 1, the q-ROFSs, which is general form of Intuitionistic fuzzy
sets and Pythagorean fuzzy sets, proposed by [12] in order to express more potential
evaluation of uncertain information.
In this section, some definition and fundamental concepts of q-ROFS are given.
~ on X is defined by:
Definition 3 [12]. Let X be a non-empty and finite set, a q-ROFS Q

~¼
Q xi ; lQ ðxÞ; #Q ðxÞ jxi 2 X ð3Þ
where lQ ðxÞ : X 2 ½0; 1 and #Q ðxÞ : X 2 ½0; 1 represent the membership degree and
non-membership degree respectively. The lQ ðxÞ and #Q ðxÞ have to satisfy the below
condition: 0 lQ ð xÞq þ #Q ðxÞq 1, ðq 1Þ;
The indeterminacy degree is defined as:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pQ ð xÞ ¼ q
lQ ð xÞq þ #Q ð xÞq lQ ð xÞq #Q ð xÞq :
Definition 4 [12]. Let Q ~ ¼ ðl; #Þ, Q~ 1 ¼ ðl1 ; #1 Þ and Q

~ 2 ¼ ðl2 ; #2 Þ be three
q-ROFNs, and k be a positive real number, then:
~1 Q
Q ~2 ¼ q
lq1 þ lq2 lq1 lq2 ; #1 #2 ð4Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~1 Q
Q ~2 ¼ l1 l2 ; #q1 þ #q2 #q1 #q2
q
ð5Þ
~¼
kQ
q
1 ð1 lq Þk ; #k ð6Þ
~k ¼
Q lk ;
q
1 ð1 #q Þk ð7Þ
~ c ¼ ð#; lÞ
Q ð8Þ
~1 [ Q
Q ~ 2 ¼ hmaxfl1 ; l2 g; minf#1 ; #2 gi ð9Þ
~1 \ Q
Q ~ 2 ¼ hminfl1 ; l2 g; maxf#1 ; #2 gi ð10Þ
~1 Q
Q ~ 2 if and only if lQ lQ and #Q1 #Q2 ð11Þ
1 2
2.4 Cosine Similarity Function

Cosine similarity is a measure of similarity between two non-zero vectors of an inner
product space that measures the cosine of the angle between them. In this section, we
first define the cosine similarity measure between two positive vectors.
Definition 5. Let A ¼ ða1 ; a2 ; . . .; an Þ and B ¼ ðb1 ; b2 ; . . .; bn Þ be the two positive
vectors. Then the cosine similarity measure between two vectors is defined as follows:
Pn
ai b i
CosðA; BÞ ¼ Pn i¼1 P ð12Þ
ð i¼1 ai Þð ni¼1 b2i Þ
2
The Cosine similarity measure satisfies the following properties:

(1) 0 CosðA; BÞ 1;
(2) CosðA; BÞ ¼ CosðB; AÞ;
(3) CosðA; BÞ ¼ 1, if and only if A ¼ B, (two vectors with the same orientation have a
cosine similarity of 1).
2.5 Square Root Cosine Similarity and Hellinger Distance

The Square Root Cosine Similarity is based on Hellinger distance, is a new similarity
measurement which proposed by [13].
Pn pffiffiffiffiffiffiffiffi
ai bi
SqrtCosðA; BÞ ¼ pffiffiffiPn i¼12 pffiffiffiPn 2 ð13Þ
4 i¼1 ai 4 i¼1 bi
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xm pffiffiffiffi pffiffiffiffi 2ffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xm pffiffiffiffiffiffiffiffiffi
H ða; bÞ ¼ i¼1
ð a i b i Þ ¼ 2 2 i¼1
ai bi ð14Þ
3 Some Similarity Measures Based on the Square Root

Cosine Similarity Function for q-Rung Orthopair Fuzzy
Sets
The cosine function is monotone decreasing function, thus, by considering the distance
measure between any two alternatives, the bigger the distance values is the smaller the
calculating result by cosine function is the lower similarity measures. Therefore, to
select the best alternatives in decision-making problems, the cosine similarity measure
utilized to obtain the similarity degree between each alternative and the ideal alter-
native. In this section, some square root cosine similarity measures based on the
Hellinger distance will be provide.
Definition 6. Let A ~ q ¼ fxi hlA ðxi Þ; #A ðxi Þijxi 2 X g; 8i ¼ 1; 2; . . . n and B
~ q ¼ fxi hlB
ðxi Þ; #B ðxi Þijxi 2 Xg; 8i ¼ 1; 2; . . . n be two q-ROFSs, and lðxi Þ and #ðxi Þ are mem-
bership and non-membership degrees respectively and q 1, then the Square Root
Cosine Similarity measure of A ~ q and B ~ q is as follows:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

1 Xn lqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ
~ ~ pffiffiffi ð15Þ
qROFS Aq ; Bq
SqrtCS1H ¼ i¼1 pffiffiffi
n 4 l2q 2q
4 l2q 2q
A ð xi Þ þ # A ð xi Þ B ðxi Þ þ #B ðxi Þ
Theorem 1. Let A ~ q ¼ fxi hlA ðxi Þ; #A ðxi Þijxi 2 X g and B

~ q ¼ fxi hlB ðxi Þ; #B ðxi Þijxi 2
Xg be any two q-ROFSs, then the Square Root Cosine similarity measure

~ ~
qROFS Aq ; Bq satisfies the following properties:
SqrtCS1H

(1) 0 SqrtCS1H ~ ~
qROFS Aq ; Bq 1

~ ~ ~ ~
qROFS Aq ; Bq ¼ SqrtCSqROFS Bq ; Aq
(2) SqrtCS1H 1H

~ ~ ~ ~
qROFS Aq ; Bq ¼ 1 if Aq ¼ Bq ! lA ðxi Þ ¼ lB ðxi Þ and #A ðxi Þ#B ðxi Þ
(3) SqrtCS1H
Proof

~q; B
~ q as two q-ROFSs. Then SqrtCS1H ~ ~
1. Let us assume A qROFS Aq ; Bq according to
Eq. (12) is defined as follow:

~ ~ pffiffiffi
qROFS Aq ; Bq
n 4 l2q 2q
4 l2q 2q
A ð xi Þ þ # A ð xi Þ B ð xi Þ þ # B ð xi Þ
~ ~
qROFS Aq ; Bq 0. So, according to inequality a þ b
We know that SqrtCS1H 2 2
2ab:

l2q
A ð x i Þ þ # 2q
A ð x i Þ þ l 2q
B ð x i Þ þ # 2q
B ð x i Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 l2q 2q
A ðxi ÞlB ðxi Þ þ #2q 2q
A ðxi Þ#B ðxi Þ
As a result, 0 SqrtCS1H ~ ~
qROFS Aq ; Bq 1.

~q; B
~ q as two Q-ROFSs. Then SqrtCS1H ~ ~
2. Let us assume A qROFS Aq ; Bq according to
Eq. (12) is defined as follow:

~ ~ pffiffiffi
qROFS Aq ; Bq
n 4 l2q 2q
4 l2q 2q
A ðxi Þ þ #A ðxi Þ B ðxi Þ þ #B ðxi Þ

¼ SqrtCS1H ~ ~
qROFS Bq ; Aq
~q ¼ B
3. We know that if A ~ q then, lA ðxi Þ ¼ lB ðxi Þ and #A ðxi Þ ¼ #B ðxi Þ.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 Xn lqA ðxi ÞlqA ðxi Þ þ #qA ðxi Þ#qA ðxi Þ 1 Xn l2qA ðxi Þ þ #A ðxi Þ
2q
1 Xn
p ffiffi
ffi p ffiffiffi ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ¼ ð1Þ
n i¼1 2q 2q
4 lA ðxi Þ þ #A ðxi Þ 4 lA ðxi Þ þ #A ðxi Þ2q 2q n i¼1 n i¼1
l2qA ðxi Þ þ #A ðxi Þ
2q
1
¼ n¼1
n
Definition 7. Let A ~ q ¼ fxi hlA ðxi Þ; #A ðxi Þ; pA ðxi Þijxi 2 X g and B

~ q ¼ fxi hlB ðxi Þ; #B ðxi Þ;
pB ðxi Þijxi 2 Xg be two q-ROFSs, and lðxi Þ, #ðxi Þ and pðxi Þ are membership, non-
membership and hesitancy degrees respectively and q 1, then the Square Root Cosine
Similarity measure of A ~ q and B~ q , considering hesitancy degree is as follows:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

1 Xn lqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ þ pqA ðxi ÞpqB ðxi Þ
SqrtCS2H
qROFS
~ ~
Aq ; Bq ¼ i¼1 p ffiffi
ffi pffiffiffi ð16Þ
n 4 l2q 2q 2q
4 l2q 2q 2q
A ðxi Þ þ #A ðxi Þ þ pA ðxi Þ B ðxi Þ þ #B ðxi Þ þ pB ðxi Þ
Definition 8. Let A ~ q ¼ fxi hlA ðxi Þ; #A ðxi Þijxi 2 X g and B

~ q ¼ fxi hlB ðxi Þ; #B ðxi Þijxi 2
Xg be two q-ROFSs, and lðxi Þ and #ðxi Þ are membership and non-membership degrees
respectively and q 1, with weight vector wi ¼ ðw1 ; w2 ; . . .; wn Þ which wi 2 ½0; 1 and
Pn
~ ~
i¼1 wi ¼ 1. Then, the Weighted Square Root Cosine Similarity measure of Aq and Bq
is as follows:

Xn lqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ
~ ~ pffiffiffi
qROFS Aq ; Bq ¼
WSqrtCS1H i¼1 i pffiffiffi
w
4 l2q A ð x i Þ þ # 2q
A ð x i Þ 4 l 2q
B ð x i Þ þ # 2q
B ð x i Þ
ð17Þ
Remark 1. It could be deduced that if weight vector of xi become an equal amount like

that wi ¼ 1n ; 1n ; . . .; 1n , the WSqrtCSH ~ ~

qROFS Aq ; Bq will be reduced to

H
SqrtCSqROFS A ~q; B~q .
Definition 9. Let ~ q ¼ fxi hlA ðxi Þ; #A ðxi Þ; pA ðxi Þijxi 2 X g

A and B~q ¼
fxi hlB ðxi Þ; #B ðxi Þ; pB ðxi Þijxi 2 X g be two q-ROFSs, and lðxi Þ, #ðxi Þ and pðxi Þ are
membership, non-membership and hesitancy degrees respectively Pn and q 1, with
weight vector wi ¼ ðw1 ; w2 ; . . .; wn Þ which wi 2 ½0; 1 and i¼1 wi ¼ 1. Then, the
Weighted Square Root Cosine Similarity measure of A ~ q and B ~ q , considering hesitancy
degree is as follows:
~ ~
WSqrtCS2H qROFS Aq ; Bq
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn lqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ þ pqA ðxi ÞpqB ðxi Þ ð18Þ
¼ pffiffiffi
i¼1 i pffiffiffi
w
2q 2q 2q 2q 2q 2q
4 lA ðxi Þ þ #A ðxi Þ þ pA ðxi Þ 4 lB ðxi Þ þ #B ðxi Þ þ pB ðxi Þ
4 Numerical Example
In this section to illustrate validity and effectiveness of our proposed similarity mea-
sures we shall give an application of them in pattern recognition field based on q-
ROFSs. The numerical example is adopted from article by [1]. Consider a pattern
recognition problem about construction material classification. Assume there are five
known construction materials based on five criteria from feature space which are
depicted by q-ROFSs which presented in Table 1.
Table 1. Evaluation matrix of construction materials.

~1
X ~2
X ~3
X ~4
X ~4
X
~1
A [0.5, 0.8] [0.6, 0.4] [0.8, 0.3] [0.6, 0.9] [0.1, 0.4]
~ 2 [0.6, 0.7] [0.7, 0.3] [0.6, 0.2]
A [0.8, 0.6] [0.3, 0.5]
~ 3 [0.3, 0.4] [0.7, 0.5] [0.9, 0.3]
A [0.4, 0.8] [0.2, 0.3]
~ 4 [0.5, 0.3] [0.4, 0.4] [0.6, 0.2]
A [0.4, 0.7] [0.2, 0.6]
~ 5 [0.4, 0.7] [0.6, 0.5] [0.5, 0.4]
A [0.5, 0.3] [0.4, 0.2]
Let’s assume an unknown material which are going to be used in a construction

project, it is needed to evaluate that which classes of construction material, it is belong
to. Therefore, referred unknown materials are depicted based on feature space (eval-
uated for all five features) and presented as q-ROFSs as below:
A 2 q ROFSs: A
¼ fX1 h0:7; 0:6i; X2 h0:8; 0:2i; X3 h0:4; 0:3i; X4 h0:7; 0:8i; X5 h0:4; 0:2ig
To determine that which class material A belong to, we apply the proposed square
root cosine similarity measures and results are presented in Table 2.
Table 2. Results of similarity degree using different proposed similarity measures.

SqrtCS1H
qROFS
Rank SqrtCS2H
qROFS
Rank WSqrtCS1H
qROFS
Rank WSqrtCS2H
qROFS
Rank
~ 1 ; AÞ 0.832
ðA 3 0.964 2 0.775 4 0.966 3
~ 2 ; A) 0.881
ðA 2 0.984 1 0.847 2 0.988 1
~ 3 , A) 0.884
ðA 1 0.947 4 0.856 1 0.945 5
~ 4 , A) 0.828
ðA 4 0.962 3 0.772 5 0.968 2
~ 5 , A) 0.780
ðA 5 0.943 5 0.803 3 0.960 4
In above table to calculate the WSqrtCS1H 2H

qROFS and WSqrtCSqROFS , the corre-
sponding weight vector of criteria assumed as wi ¼ ð0:15; 0:2; 0:25; 0:1; 0:3Þ and it is
supposed q ¼ 3 for q-ROFSs.
Based on proposed square root cosine similarity measures, classification of con-
struction material with considering the uncertainty based on q-ROFSs calculated and
results are shown in Table 2. It concluded that material A could be classified in the
construction material class A2 because from SqrtCS2H 2H
qROFS and WSqrtCSqROFS the
similarity of A to A2 stands in first place.
5 Conclusion
q-ROFSs as mentioned previous is a general form of Intuitionistic fuzzy sets and

Pythagorean fuzzy sets, proposed by [12] in order to express more potential evaluation
of uncertain information. In this work we shall to propose some novel square root
cosine similarity measures for q-ROFSs, and also some properties of them are dis-
cussed. Then validity and effectiveness of the proposed similarity measures through an
application in pattern recognition problem illustrated. It is also presented the numerical
example’s results for pattern recognition problem about construction material
classification.
References
1. Wang, P., Wang, J., Wei, G., Wei, C.: Similarity measures of q-rung orthopair fuzzy sets
based on cosine function and their applications. Mathematics 7(4), 1–29 (2019)
2. Chen, S.M., Tan, J.M.: Handling multicriteria fuzzy decision-making problems based on
vague set theory. Fuzzy Sets Syst. 67(2), 163–172 (1994)
3. Szmidt, E., Kacprzyk, J.: Entropy for intuitionistic fuzzy sets. Fuzzy Sets Syst. 118(3), 467–
477 (2001)
4. Hwang, C.L., Yoon, K.: Multiple Attribute Decision Making, Methods and Applications A
State-of-the-Art Survey. Springer, Heidelberg (1981)
5. Xu, Z., Xia, M.: Distance and similarity measures for hesitant fuzzy sets. Inf. Sci. 181(11),
2128–2138 (2011)
6. Zeng, W., Li, D., Yin, Q.: Distance and similarity measures of Pythagorean fuzzy sets and
their applications to multiple criteria group decision making. Int. J. Intell. Syst. 33(11),
2236–2254 (2018)
7. Wei, G.W.: Some similarity measures for picture fuzzy sets and their applications. Iran.
J. Fuzzy Syst 15(1), 77–89 (2018)
8. Sohangir, S., Wang, D.: Document understanding using improved sqrt-cosine similarity. In:
Proceedings - IEEE 11th International Conference on Semantic Computing, pp. 278–279.
IEEE (2017)
(2017)
13. Zhu, S., Liu, L., Wang, Y.: Information retrieval using Hellinger distance and sqrt-cos
similarity. In: Proceedings of 2012 7th International Conference on Computer Science and
Education, pp. 925–929. ICCSE (2012)
The Dice (Sorensen) Similarity Measures
for Optimal Selection with q-Rung Orthopair
Fuzzy Information
Elmira Farrokhizadeh1, Seyed Amin Seyfi Shishavan2,

Yaser Donyatalab3(&), and Sohrab Abdollahzadeh1
1
Urmia University of Technology, Urmia, Iran
[email protected]
2
3
University of Moghadas Ardabili, Ardabil, Iran
[email protected]
Abstract. The q‐Rung Orthopair Fuzzy set (q‐ROFS) which was developed by
Yager and characterized by two functions expressing the degree of membership
and non-membership, is more capable than similar fuzzy sets such as Pytha-
gorean and Intutionistic fuzzy sets to deal uncertainty in real life. This paper
develops an approach for Multi Attribute Decision-Making problems (MADM)
based on the Sorensen similarity or Dice similarity measure introduced by Dice
in 1945 and Sorensen in 1948. We will present Dice similarity for q-Rung
Orthopair Fuzzy Sets and other types of Dice similarities such as weighted,
generalized and weighted generalized Dice similarity measures form in two state
of with hesitancy and without hesitancy degree. Finally, a practical example has
been considered to illustrate the application of the proposed method.
Keywords: q‐Rung Orthopair Fuzzy Set Similarity measure Sorensen

similarity Dice similarity MADM
1 Introduction
Multi-Attribute Decision Making (MADM) is an important and interesting research topic

of modern decision science, having received more and more attention during the last
several years part. It has been extensively applied to various areas such as society,
economics, management, military, and engineering technology [1–3]. The aim of the
decision-making is to choose the best-performing alternative among the possible ones in
order to satisfy a certain objective represented by various attributes [4]. MADM is a
process that can give the ranking results for the finite alternatives according to the attribute
values [5]. The decision environment may be deterministic or uncertain. In the real world
and specially, when the data are based on humans’ perceptions rather than precise and
adequate numerical observations, the fuzzy set theory can be preferred in order to express
uncertainty [6]. The q‐Rung Orthopair Fuzzy Sets (q-ROFSs) developed by Yager [7], is
generalized form of IFS [8] and PFS [9]. The q-ROFSs, is an efficient tool to describe
vagueness of the multi-attribute decision‐making problems.
The Dice (Sorensen) Similarity Measures for Optimal Selection 485
Similarity measure is a key tool in constructing MADM methods in many areas

such as medical diagnosis, pattern recognition, clustering analysis, physical anthro-
pology, numerical taxonomy, ecology, information retrieval, psychology, citation
analysis, and automatic classification, decision making, machine learning in uncertainty
environment [10].
To the best of our knowledge, these Dice similarity measures do not deal directly
with the similarity measures for q‐Rung Orthopair Fuzzy Sets (q-ROFSs).
The reminder of the paper is organized as follows: In Sect. 2, some basic and
mathematical concepts of q-ROFSs and Dice (Sorensen) similarity measure are briefly
reviewed. In Sect. 3, different Dice (Sorensen) similarity measures between q-ROFSs
are proposed. In Sect. 4, we give the application of the proposed Dice (Sorensen)
similarity measures for q-ROFSs on optimal selection. Finally, the conclusions are
given in Sect. 5.
2 Mathematical Preliminaries
2.1 q‐Rung Orthopair Fuzzy Sets (q-ROFSs)
Based on Intuitionistic fuzzy sets and Pythagorean fuzzy sets, Yager [7], proposed a
more general form of these fuzzy sets, (q-ROFSs) to express more potential evaluation
information of decision-maker, and developed the operations of q-ROFSs. In the
q-ROFSs, the membership grades are pairs of values, from the unit interval, one
indicating the degree of support for membership in the fuzzy set and the other support
against membership ðpower ¼ qÞ [7].
In this section, we give the definition of q-ROFS and summarize some fundamental
concepts.
~ defined
Definition 1 (Yager 2017). Let X be a non-empty and finite set, a q-ROFS Q
on X is defined by:

~¼
Q x; lQ ð xÞ; #Q ð xÞ jx 2 X ð1Þ
Where lQ ð xÞ and #Q ð xÞ represent the membership degree and non-membership

degree respectively. The lQ ð xÞ and #Q ð xÞ have to satisfy the below condition:
lQ ð xÞ 2 ½0; 1; #Q ð xÞ 2 ½0; 1 and 0 lQ ð xÞq þ #Q ð xÞq 1; ðq 1Þ:
The indeterminacy degree is defined as

pQ ð xÞ ¼ q
lQ ð xÞq þ #Q ð xÞq lQ ð xÞq #Q ð xÞq :

In this paper, for convenience, the lQ ð xÞ; #Q ð xÞ is called a (q-ROFS) denoted by

~ ¼ lQ ; #Q :
Q
~ ¼ ðl; #Þ, Q
Definition 2 (Yager 2017). Let Q ~ 1 ¼ ðl1 ; #1 Þ and Q
~ 2 ¼ ðl2 ; #2 Þ be three
q-ROFNs, and k be a positive real number, then:
~1 [ Q
Q ~ 2 ¼ hmaxfl1 ; l2 g; minf#1 ; #2 gi ð2Þ
~1 \ Q
Q ~ 2 ¼ hminfl1 ; l2 g; maxf#1 ; #2 gi ð3Þ
~1 Q
Q ~ 2 ¼ q lq þ lq lq lq ; #1 #2 ð4Þ
1 2 1 2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~1 Q
Q ~2 ¼ l1 l2 ;
q
#q1 þ #q2 #q1 #q2 ð5Þ
~¼
kQ
q
1 ð1 l q Þk ; # k ð6Þ
~k ¼
Q l ; 1 ð1 #q Þk
kq
ð7Þ
~ c ¼ ð#; lÞ
Q ð8Þ
Definition 3 (Yager 2017). Let Q ~ ¼ ðl; #Þ be a q-ROFN, then the score function of Q
~

is defined as S Q~ ¼ l # , and the accuracy of Q
q q ~ is defined as H Q ~ ¼ l þ# .
q q
~ ~
For any two qROFNs, Q1 ¼ ðl1 ; #1 Þ and Q2 ¼ ðl2 ; #2 Þ:

1. If S Q~ [S Q ~ , then Q ~1 [ Q~ 2;
1 2
2. If S Q~1 ¼ S Q~ , then
2
~
I. If H Q1 [ H Q ~ , then Q ~1 [ Q~ 2;
2
~1 ¼ H Q
II. If H Q ~ 2 , then Q
~1 ¼ Q~ 2.
2.2 Dice (Sorensen) Similarity Measure

Dice (Sorensen) similarity measure proposed by [11], is a similarity measure between
two vectors, which is introduced in the following.
Let X ¼ ðx1 ; x2 ; . . .; xn Þ and Y ¼ ðy1 ; y2 ; . . .; yn Þ be the two vectors of length n
where all the coordinates are positive real numbers. Then the Dice similarity measure
between two vectors is defined as follows:
Pn
2X:Y 2 xi yi
DðX; Y Þ ¼ ¼ Pn i¼1 P
n ð9Þ
k X k22 þ kY k22 i¼1 xi þ
2 2
i¼1 yi
P
where X:Y ¼ ni¼1 xi yi is the inner product of the vectors X and Y and kX k2 and kY k2
are the Euclidean norms of X and Y (also called the L2 norms).
The Dice similarity measure takes value in the interval ½0; 1. However, it is
undefined if xi ¼ yi ¼ 0; ði ¼ 1; 2; . . .; nÞ. In this case, let the Dice measure value be
zero when xi ¼ yi ¼ 0; ði ¼ 1; 2; . . .; nÞ.
The Dice similarity measure satisfies the following properties [11]:
P1. 0 DðX; Y Þ 1;
P2. DðX; Y Þ ¼ DðY; X Þ;
P3. DðX; Y Þ ¼ 1, if and only if X ¼ Y, i.e., xi ¼ yi for i ¼ 1; 2; . . .; n:
3 Expected Dice (Sorensen) Similarity Measures for q‐Rung

Orthopair Fuzzy Sets (q-ROFSs)
In this section, we propose some Dice similarity measure, weighted Dice similarity
measure, generalized Dice similarity measure and weighted generalized Dice similarity
measure in two state of with hesitancy and without hesitancy degree between q-ROFSs
based on the concept of the Dice similarity measure.
It is necessary to mention that because of the space constraints, we just prove the
Dice (Sorensen) similarity Measure for q-ROFSs without considering hesitancy degree.
3.1 Dice (Sorensen) Similarity Measure for q-Rung Orthopair Fuzzy Sets
Without Hesitancy Degree
Definition 4. Let A ~ q ¼ hlA ðxi Þ; #A ðxi Þjxi 2 X i and B

~ q ¼ hlB ðxi Þ; #B ðxi Þjxi 2 X i, i ¼
1; 2; . . .; n; be two groups of q-ROFSs which is defined on universe discourse X. The
Dice similarity measure without considering hesitancy degree between q-ROFSs A ~ and
~
B is proposed as follows:

Xn 2 lqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ
DS1qROFS A ~q ¼ 1
~q; B ð10Þ
n i¼1
l2q ðxi Þ þ #2q ðxi Þ þ l2q ðxi Þ þ #2q ðxi Þ
A A B B
This Dice similarity measure between q-ROFSs A ~ q and B ~ q also satisfies the fol-
lowing properties:

(1) 0 DS1qROFS A ~q; B
~q 1

(2) DS1qROFS A ~q; B
~ q ¼ DS1QROFS B ~q
~q; A

(3) DS1qROFS A ~q; B ~q ¼ B
~ q ¼ 1 if A ~ q ! lA ðxi Þ ¼ lB ðxi Þ and #A ðxi Þ ¼ #B ðxi Þ
Proof of the properties:

(1) Let us assume A ~ and B
~ as two q-ROFSs that have n terms, i ¼ 1; 2; . . .n. Then

~ ~
DSqROFS Aq ; Bq according to Eq. (10) is defined as follow:
1

1 Xn 2 lqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ
DS1qROFS ~ ~
Aq ; Bq ¼
n i¼1
A A B B

It is obvious that SS1qROFS A ~q; B
~ q 0 and according to inequality
a2 þ b2 2ab, we will have:

l2q
A ð x i Þ þ #2q
A ð x i Þ þ l 2q
B ð x i Þ þ # 2q
B ð x i Þ 2 lqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ

So it is proved that 0 DS1qROFS A~q; B
~ q 1.
(2) Let us assume A ~q; B
q
1
DSqROFS A ~q; B
~ q according to Eq. (10) is defined as follow:

DS1qROFS A ~q ¼ 1
~q; B ¼ DS1qROFS B ~q
~q; A
n i¼1 2q 2q 2q
l ðxi Þ þ # ðxi Þ þ l ðxi Þ þ # ðxi Þ 2q
A A B B
(3) Let us assume A ~q; B

q
~ ~
DSqROFS Aq ; Bq according to Eq. (10) is defined as follow:
1

1 Xn 2 lqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ
DS1qROFS ~ ~
Aq ; Bq ¼
n i¼1
A A B B
~ ¼B
We know that if A ~ then lA ðxi Þ ¼ lB ðxi Þ and #A ðxi Þ ¼ #B ðxi Þ and in this
q q
1 ~q; B
case DSqROFS A ~ q will be:
P 2ðlqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi ÞÞ P 2ðlqA ðxi ÞlqA ðxi Þ þ #qA ðxi Þ#qA ðxi ÞÞ
DS1qROFS A~q; B
~ q ¼ 1 ni¼1 ¼ 1n ni¼1 l2q ðx Þ þ
n ðl2qA ðxi Þ þ #2q
A
ð x i Þ Þ þ ð l 2q
B ð xi Þ þ # 2q
B ð x i Þ Þ ð A i #2q
A
ðxi ÞÞ þ ðl2q A
ðxi Þ þ #2q A
ðxi ÞÞ
P n 2 ðl 2q
ð x Þ þ # 2q
ð x Þ Þ P n
¼ n i¼1 2 l2q ðx Þ þ #2q ðx Þ ¼ n i¼1 1 ¼ n n ¼ 1
1 A i A i 1 1
ðA i A i Þ
3.2 Dice (Sorensen) Similarity Measure for q-Rung Orthopair Fuzzy

Sets Considering Hesitancy Degree
Definition 5. Let A ~ q ¼ hlA ðxi Þ; #A ðxi Þ; pA ðxi Þjxi 2 X i and B ~ q ¼ hlB ðxi Þ; #B ðxi Þ;
pB ðxi Þjxi 2 Xi, i ¼ 1; 2; . . .; n; be two groups of q-ROFSs. The Dice similarity measure
considering hesitancy degree between q-ROFSs A ~ and B~ is proposed as follows:

Xn 2 lqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ þ pqA ðxi ÞpqB ðxi Þ
DS2qROFS A ~q ¼ 1
~q; B
n i¼1
l2q ðxi Þ þ #2q ðxi Þ þ p2q ðxi Þ þ l2q ðxi Þ þ #2q ðxi Þ þ p2q ðxi Þ
A A A B B B
ð11Þ
This Dice similarity measure between q-ROFSs A ~ q and B

~ q also satisfies the fol-
lowing properties:

(1) 0 DS2qROFS A ~q; B~q 1

(2) DS2qROFS A ~q; B
~ q ¼ DS2qROFS B ~q
~q; A

DS2qROFS A ~q; B~ q ¼ 1 if A~q ¼ B~ q ! lA ðxi Þ
(3)
¼ lB ðxi Þ; #A ðxi Þ ¼ #B ðxi Þ and pA ðxi Þ ¼ pB ðxi Þ
3.3 Weighted Dice (Sorensen) Similarity Measure for q-Rung Orthopair

Fuzzy Sets Without Considering Hesitancy Degree

1; 2; . . .; n; be two
P groups of q-ROFSs, si ¼ ðs1 ; s2 ; . . .; sn Þ be the weight vector, which
si 2 ½0; 1 and ni¼1 si ¼ 1. The weighted Dice similarity measure without considering
hesitancy degree between q-ROFSs A ~ and B ~ is proposed as follows:

WDS1qROFS ~q; B
A ~q ¼ s ð12Þ
i¼1 i
l2q 2q 2q
A ðxi Þ þ #A ðxi Þ þ lB ðxi Þ þ #B ðxi Þ
2q
3.4 Weighted Dice (Sorensen) Similarity Measure for q-Rung Orthopair

Fuzzy Sets Considering Hesitancy Degree
Definition 7. Let A ~ q ¼ hlA ðxi Þ; #A ðxi Þ; pA ðxi Þjxi 2 X i and B ~ q ¼ hlB ðxi Þ; #B ðxi Þ;
pB ðxi Þjxi 2 Xi, i ¼ 1; 2; . . .; n; be twoPgroups of q-ROFSs, si ¼ ðs1 ; s2 ; . . .; sn Þ be the
weight vector, which si 2 ½0; 1 and ni¼1 si ¼ 1. The weighted Dice similarity mea-
sure considering hesitancy degree between q-ROFSs A ~ and B
~ is proposed as follows:

Xn 2 lqA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ þ pqA ðxi ÞpqB ðxi Þ
WDS2qROFS ~q; B
A ~q ¼ s
i¼1 i
l2q 2q 2q 2q 2q
A ðxi Þ þ #A ðxi Þ þ pA ðxi Þ þ lB ðxi Þ þ #B ðxi Þ þ pB ðxi Þ
2q
ð13Þ
3.5 Generalized Dice (Sorensen) Similarity Measure for q-Rung

Orthopair Fuzzy Sets Without Considering Hesitancy Degree
Definition 8. Let A ~ q ¼ hlA ðxi Þ; #A ðxi Þjxi 2 X i and B~ q ¼ hlB ðxi Þ; #B ðxi Þjxi 2 X i, i ¼
1; 2; . . .; n; be two groups of q-ROFSs, the generalized Dice similarity measure without
considering hesitancy degree between q-ROFSs A ~ and B ~ is proposed as follows:
q
Xn lA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ
GDS1qROFS A ~q ¼ 1
~q; B
n i¼1
C l2q ðxi Þ þ #2q ðxi Þ þ ð1 CÞ l2q ðxi Þ þ #2q ðxi Þ
A A B B
ð14Þ
Where C is a positive value like that 0 C 1.
3.6 Generalized Dice (Sorensen) Similarity Measure for q-Rung

Orthopair Fuzzy Sets Considering Hesitancy Degree
Definition 9. Let A ~ q ¼ hlA ðxi Þ; #A ðxi Þ; pA ðxi Þjxi 2 X i and B~ q ¼ hlB ðxi Þ; #B ðxi Þ;
pB ðxi Þjxi 2 Xi, i ¼ 1; 2; . . .; n; be two groups of q-ROFSs, the generalized Dice simi-
larity measure considering hesitancy degree between q-ROFSs A ~ and B~ is proposed as
follows:
q
Xn lA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ þ pqA ðxi ÞpqB ðxi Þ
GSS2QROFS A ~q ¼ 1
~q; B
n i¼1
C l2q ðxi Þ þ #2q ðxi Þ þ p2q ðxi Þ þ ð1 CÞ l2q ðxi Þ þ #2q ðxi Þ þ p2q ðxi Þ
A A A B B B
ð15Þ
3.7 Weighted Generalized Dice (Sorensen) Similarity Measure

for q-Rung Orthopair Fuzzy Sets Without Considering
Hesitancy Degree

1; 2; . . .; n; be two
P groups of q-ROFSs, si ¼ ðs1 ; s2 ; . . .; sn Þ be the weight vector, which
si 2 ½0; 1 and ni¼1 si ¼ 1. The weighted Generalized Dice similarity measure without
considering hesitancy degree between q-ROFSs A ~ and B~ is proposed as follows:

WGDS1qROFS ~ ~q ¼
Aq ; B s
i¼1 i
C l2q 2q 2q
A ðxi Þ þ #A ðxi Þ þ ð1 CÞ lB ðxi Þ þ #B ðxi Þ
2q
ð16Þ
3.8 Weighted Generalized Dice (Sorensen) Similarity Measure

for q-Rung Orthopair Fuzzy Sets Considering Hesitancy Degree
Definition 11. Let A ~ q ¼ hlA ðxi Þ; #A ðxi Þ; pA ðxi Þjxi 2 X i and B

~ q ¼ hlB ðxi Þ; #B ðxi Þ;
pB ðxi Þjxi 2 Xi, i ¼ 1; 2; . . .; n; be two groups
Pn of q-ROFSs, si ¼ ðs1 ; s2 ; . . .; sn Þ be the
weight vector, which si 2 ½0; 1 and i¼1 s i ¼ 1. The weighted Generalized Dice
similarity measure considering hesitancy degree between q-ROFSs A ~ and B ~ is proposed
as follows:
q
Xn lA ðxi ÞlqB ðxi Þ þ #qA ðxi Þ#qB ðxi Þ þ pqA ðxi ÞpqB ðxi Þ
WGSS2QROFS A~q; B
~q ¼ s
i¼1 i
C l2q 2q 2q 2q 2q
A ðxi Þ þ #A ðxi Þ þ pA ðxi Þ þ ð1 CÞ lB ðxi Þ þ #B ðxi Þ þ pB ðxi Þ
2q
ð17Þ
4 Optimal Selection Using the Proposed Dice Similarity

Measures
In this section, to validate the proposed similarity measure, we provide a numerical

example about selecting the best Enterprise resource planning (ERP) system. There are
four possible systems (alternatives) provided by different companies. These alternatives
are fA1; A2; A3; A4g. These alternatives are evaluated by an expert using q-ROFS
number so that q ¼ 3, under four criteria, technology (C1), strategic adaptability (C2),
supplier’s ability (C3), and supplier’s reputation (C4). The aim is select the best
alternative. The individual decision matrix provided by decision maker is given in
Table 1.
Table 1. Decision matrix.

E1 ~1
C ~2
C ~3
C ~4
C
~1
A (0.53, 0.33) (0.89, 0.08) (0.42, 0.35) (0.08, 0.89)
~ 2 (0.73, 0.12) (0.13, 0.64) (0.03, 0.82) (0.73, 0.15)
A
~ 3 (0.91, 0.03) (0.07, 0.79) (0.04, 0.85) (0.68, 0.26)
A
~ 4 (0.85, 0.09) (0.74, 0.16) (0.02, 0.89) (0.08, 0.84)
A
The similarity degree of each alternative with reference index (RI) is calculated
using Eq. (10). The results are shown in Table 2. We consider the reference index as:
RI ¼ fð0:9; 0:1Þ; ð0:91; 0:1Þ; ð0:91; 0:1Þ; ð0:91; 0:1Þg.
Table 2. Final results and ranking.

Alternatives Similarity degree Ranking
A1 0.39 4
A2 0.41 3
A3 0.43 2
A4 0.45 1
Then, the rank of the four alternatives is obtained as A4 [ A3 [ A2 [ A1.

Therefore, the optimal alternative is A4.
5 Conclusion
The main contribution of this paper is that a novel similarity measure is proposed for
q-ROFS in order to use in MCDM problems. We propose six different form of Dice
similarity measure considering weighted and generalized form. To validate the pro-
posed similarity measure, we apply it to an Enterprise resource planning (ERP) system
selection problem and get the best alternative. For the further research, the Interval
Valued of q-ROFS can be considered.
References
1. Xing, Y., Zhang, R., Wang, J., Bai, K., Xue, J.: A new multi-criteria group decision-making
approach based on q-rung orthopair fuzzy interaction Hamy mean operators. Neural Comput.
Appl. (8) (2019)
2. Wang, R., Li, Y.: A novel approach for green supplier selection under a q-rung orthopair
fuzzy environment. Symmetry (Basel) 10(12), 1–28 (2018)
3. Li, L., Zhang, R., Wang, J., Shang, X., Bai, K.: A novel approach to multi-attribute group
decision-making with q-rung picture linguistic information. Symmetry 10(5) (2018)
4. Donyatalab, Y., Farrokhizadeh, E., Seyed Garmroodi, S.D., Seyfi Shishavan, S.A.:
Harmonic mean aggregation operators in spherical fuzzy environment and their group
decision making applications. J. Mult. Valued Log. Soft Comput. 33(6), 565–592 (2019)
C.: Spherical fuzzy Bonferroni mean aggregation operators and their applications to
multiple-attribute decision making. In: Kutlu Gündoğdu, F., Kahraman, C. (eds.) Decision
Making with Spherical Fuzzy Sets-Theory and Applications. Springer (2020)
6. Kutlu Gündoğdu, F., Kahraman, C.: A novel spherical fuzzy QFD method and its application
to the linear delta robot technology development. Eng. Appl. Artif. Intell. 87, 103348 (2020)
(2017)
10. Uluçay, V., Deli, I., Şahin, M.: Similarity measures of bipolar neutrosophic sets and their
application to multiple criteria decision making. Neural Comput. Appl. 29(30), 739–748
(2018)
11. Dice, L.R.: Measures of the amount of ecologic association between species. Ecology 26(3),
297–302 (1945)
Technology Selection of Indoor Location
Systems Using Interval Valued Type-2
Intuitionistic Fuzzy WASPAS
Basar Oztaysi(&), Sezi Cevik Onar, and Cengiz Kahraman

Istanbul, Turkey
{oztaysib,cevikse,kahramanc}@itu.edu.tr
Abstract. As mobile devices are, highly adopted, indoor location tracking

systems started to emerge. Companies can use these indoor location-tracking
systems to better understand their customers and get an insight into their
intentions and needs. In this paper, different indoor location tracking systems are
evaluated. Since there are various evaluation perspectives, the problem is
defined as a multi-criteria decision-making (MCDM) problem, and a decision
model containing ten criteria, and five alternatives are formed. In this study, we
use Interval Valued Type-2 Intuitionistic Fuzzy WASPAS method to solve the
problem.
Keywords: Indoor location systems WASPAS Type-2 Intuitionistic Fuzzy
1 Introduction
The most important issue for a company to maintain its existence is to understand its
customers by communicating better with them [1]. Tracing visitors both on the web and
indoor locations enable the integration of online and offline channels [2]. Behavioral
analytics can be used to understand user needs and provide the best solutions to these
needs. Behavioral analytics produces information on the places visited by customers,
the time they spent in these places, their next visit, and even how many people looked
at the store window but did not come in. There are many alternative technologies to
collect data in physical stores such as RFID (Radio-frequency identification), WiFi
(wireless fidelity) and Bluetooth, infrared (IR), ZigBee, and computer. Since each
method has its own advantages and disadvantages, there is not a consensus [3] about
one single technology.
Since there are various alternatives and different decision perspectives, selection of
indoor location tracking technology is a multi-criteria decision-making (MCDM)
problem where there are multiple alternatives and criteria. In the literature fuzzy sets
have been widely used in MCDM problems in different fields such as health systems
[4], engineering problems [5], quality management [6], public transportation [7] firms
selection [8], disaster response [9], urban transformation [10]. In the literature,
WASPAS method has been used in various studies [11] and also extended by using
various extensions of fuzzy sets. In this paper, we use Interval Valued Type-2
Technology Selection of Indoor Location Systems 495
Intuitionistic Fuzzy WASPAS method for indoor location tracking technology selec-
tion problems.
The rest of this paper is organized as follows. Section 2 presents a literature review
on location tracking systems. Section 3 presents the methodology In Sect. 4, the
decision model is explained by summarizing the technology alternatives and the
evaluation criteria. In Sect. 5, the application with numerical calculation is given.
Finally, the concluding remarks are given in Sect. 6.
2 Current Studies
Various technologies exist that can be used to track indoor locations. The most tra-
ditional ways of understanding how people behave in a location is the questionnaire,
interviews with the people, and observations. In a study, Kirchberg and Tröndle [12]
focus on visitor experiences by utilizing the interview method. In contrast with tradi-
tional methods, using technology and automatically collecting data has become more
popular. In some of the recent studies, researchers use novel technologies to gather
data. For example, Fernandez-Llatas [13] analyzes hospital processes by using event
logs. In another study, Yuanfang et al. [14] use WiFi technology to detect the location
of customers in a closed area. Another similar study is conducted in a museum to
navigate customer [15]. ZigBee is another technology that can be used for collecting
location data. Chan [16] use this technology to determine the places of the objects.
Video cameras can also be used to determine the location of visitors [2]. Wu et al. [17]
use video recording to get an insight about customers and customer paths in a store.
The indoor location data has been in the scope of various studies. In one branch of
studies, the problem is handled as a technology selection problem [18, 19]. Another
branch of research focuses on using the data collected by the location tracking systems.
Dogan et al. [20] focus on segmenting visitor paths in a shopping mall. Lee et al. [21]
propose a random forest algorithm for recognizing indoor location recognition. Dogan
et al. [22] analyze gender visit behaviors by using process mining approach. Li et al.
[23] use deep convolutional autoencoder to define indoor trajectory data similarity.
Dogan and Oztaysi [24] propose a gender prediction model based on indoor customer
paths. Dogan et al. [25] propose a fuzzy clustering method for indoor customer paths.
3 Interval Valued Type-2 Intuitionistic Fuzzy WASPAS
In this study, Interval Valued Type-2 Intuitionistic Fuzzy WASPAS method proposed
by Ilbahar and Kahraman [26] is used. The calculation steps of the methodology is
given below:
Step 1: Decision-makers’ evaluations are collected in linguistic form and converted
to IVT2IF by using the scales given in Table 1.
Table 1. An IVT2IF scale used by decision-maker.

Linguistic term µL µU ѵL ѵU
Very Very Good (VVG) 0.75 0.90 0.03 0.18
Very Good (VG) 0.66 0.81 0.12 0.27
Good (G) 0.57 0.72 0.21 0.36
Medium Good (MG) 0.48 0.63 0.30 0.45
Fair (F) 0.39 0.54 0.39 0.54
Medium Bad (MB) 0.30 0.45 0.48 0.63
Bad (B) 0.21 0.36 0.57 0.72
Very Bad (VB) 0.12 0.27 0.66 0.81
Very Very Bad (VVB) 0.03 0.18 0.75 0.90
Step 2. The normalization process is applied by using Eqs. (2) and (3). In order to
this, maxi pij is calculated by defuzzifying the values by using Eq. (1).
qp
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffiffiffiffiffiffiffiffiffiffiffiffiffipffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffi
lL þ lU þ 1 v2L þ 1 v2u lL lU 1 v2L 1 v2u
p¼ þ ð1Þ
4 4
~xij
~rij ¼ ð2Þ
maxi pij
maxi pij
~rij ¼ ð3Þ
~xij
Step 3. The weight of each criterion is obtained by linguistic terms and transformed
to the corresponding IVT2IFs (Table 2).
Table 2. Linguistic terms associated with importance evaluation.

Linguistic term µL µU ѵL ѵU
Very Important (VI) 0.70 0.90 0.06 0.26
Important (I) 0.54 0.74 0.22 0.2
Medium (M) 0.38 0.58 0.38 0.58
Unimportant (U) 0.22 0.42 0.54 0.74
Very unimportant 0.06 0.26 0.70 0.90
Step 4. The values in normalized decision matrix and Eq. (4) is used to obtain
IVT2IF weighted sum values.
Step 5. The IVT2IF weighted product values are obtained by using Eq. (5).
Xn
~ ði 1Þ ¼
Q ~r ~j
w ð4Þ
j¼1 ij
Yn ~
~ ði 2Þ ¼
Q
w
~r j ð5Þ
j¼1 ij
Step 6. Total relative importance of the alternatives are calculated by integration

IVT2IF weighted sum values and IVT2IF weighted product values.
~ ði 1Þ þ ð1 kÞQ
~ i ¼ kQ
Q ~ ði 2Þ ; k ½0; 1 ð6Þ
Step 7. The results showing the relative importance are defuzzified using Eq. (1).
The alternative with the highest defuzzified value is selected as the best alternative.
4 Decision Model
For a company, the selection of the appropriate technology for indoor customer
tracking is an important issue. Since it contains various alternatives and conflicting
criteria, the problem can be modeled as a multicriteria decision making (MCDM)
problem. In this section, first, the alternative technologies are introduced then the
evaluation criteria are summarized.
As mentioned in the literature review section, various traditional and non-
traditional methods are being used to understand visitors’ actions and thoughts. As a
brief summary of the existing novel technologies: RFID technology uses radio fre-
quencies to communicate between the reader and the tags. Another technology is WiFi.
It is based on the signal communication between the WiFi point and mobile devices.
Bluetooth is another technology for indoor location tracking. The technology is based
on Bluetooth signals transferring between devices. The signals are sent, involve a MAC
address, which can be used to identify the device. One of the most important locations
tracking system is video cameras. The recording is processed with image processing
techniques so that the objects or individuals can be tracked.
After a literature review and expert interviews, ten evaluation criteria are deter-
mined. The evaluation criteria are listed as follows: the first criterion is Simplicity,
which shows the operational effort to collect data. The second criterion is Time, which
represents the time required to collect and analyze data. Total Cost is the criterion that
shows the total cost required to collect the data. Maintenance refers to the efforts for
maintaining the system. There are also some criteria for the content of data. For
example, the criterion, Volume, shows the volume of the data collected by using the
technology. Variety shows the details of the data collected. Another criterion about the
content of the data is Accuracy. This criterion shows the quality of the data collected by
the system. Introspection represents the ability of the data to give insight into visitors’
intentions and goals. Bias refers to the random error intrinsic to the collected data. The
final criterion is Ease of Analysis, which shows the amount of preprocessing efforts
needed to make the analysis and reach results.
5 Application
A company wants to develop a system to enhance the interaction with its customers. As
a result of an extended literature review and expert evaluations, five technology
alternatives are selected. In order t to select the most appropriate technology, ten
criteria are identified.
Table 3. Expert evaluations on the alternatives.

Criteria type AL1 AL2 AL3 AL4 AL5
C1 Cost VG F B B G
C2 Cost VVG G MB B VG
C3 Benefit MG B MG G B
C4 Benefit VG B MB B VG
C5 Benefit VB G F F VG
C6 Benefit VB VG VG G F
C7 Benefit VVG VG VG VG VG
C8 Benefit G VG MG VG F
C9 Benefit VVB VB B B VB
C10 Benefit F VG VG VG B
The experts are asked to evaluate the alternatives by using linguistic variables given
in Table 1. The resulting decision matrix for alternative evaluations is given in Table 3.
By applying the steps of the above-given methodology, the evaluations are trans-
formed into Interval Valued Type-2 Intuitionistic Fuzzy numbers (Table 4).
Table 4. Fuzzy representation of expert evaluations.

AL1 AL2 AL3 AL4 AL5
µL µu mL mU µL µu mL mU µL µu mL mU µL µu mL mU µL µu mL mU
C1 0.66 0.81 0.12 0.27 0.39 0.54 0.39 0.54 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72 0.57 0.72 0.21 0.36
C2 0.75 0.9 0.03 0.18 0.57 0.72 0.21 0.36 0.3 0.45 0.48 0.63 0.21 0.36 0.57 0.72 0.66 0.81 0.12 0.27
C3 0.48 0.63 0.3 0.45 0.21 0.36 0.57 0.72 0.48 0.63 0.3 0.45 0.57 0.72 0.21 0.36 0.21 0.36 0.57 0.72
C4 0.66 0.81 0.12 0.27 0.21 0.36 0.57 0.72 0.3 0.45 0.48 0.63 0.21 0.36 0.57 0.72 0.66 0.81 0.12 0.27
C5 0.12 0.27 0.66 0.81 0.57 0.72 0.21 0.36 0.39 0.54 0.39 0.54 0.39 0.54 0.39 0.54 0.66 0.81 0.12 0.27
C6 0.12 0.27 0.66 0.81 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27 0.57 0.72 0.21 0.36 0.39 0.54 0.39 0.54
C7 0.75 0.9 0.03 0.18 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27
C8 0.57 0.72 0.21 0.36 0.66 0.81 0.12 0.27 0.48 0.63 0.3 0.45 0.66 0.81 0.12 0.27 0.39 0.54 0.39 0.54
C9 0.03 0.18 0.75 0.9 0.12 0.27 0.66 0.81 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72 0.12 0.27 0.66 0.81
C10 0.39 0.54 0.39 0.54 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27 0.21 0.36 0.57 0.72
By using Eq. 2 and Eq. 3, the values are normalized. Table 5 shows the normalized
decision matrix.
Table 5. Normalized decision matrix.

AL1 AL2 AL3 AL4 AL5
C1 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72
C2 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72
C3 0.48 0.63 0.3 0.45 0.21 0.36 0.57 0.72 0.48 0.63 0.3 0.45 0.57 0.72 0.21 0.36 0.21 0.36 0.57 0.72
C4 0.66 0.81 0.12 0.27 0.21 0.36 0.57 0.72 0.3 0.45 0.48 0.63 0.21 0.36 0.57 0.72 0.66 0.81 0.12 0.27
C5 0.12 0.27 0.66 0.81 0.57 0.72 0.21 0.36 0.39 0.54 0.39 0.54 0.39 0.54 0.39 0.54 0.66 0.81 0.12 0.27
C6 0.12 0.27 0.66 0.81 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27 0.57 0.72 0.21 0.36 0.39 0.54 0.39 0.54
C7 0.75 0.9 0.03 0.18 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27
C8 0.57 0.72 0.21 0.36 0.66 0.81 0.12 0.27 0.48 0.63 0.3 0.45 0.66 0.81 0.12 0.27 0.39 0.54 0.39 0.54
C9 0.03 0.18 0.75 0.9 0.12 0.27 0.66 0.81 0.21 0.36 0.57 0.72 0.21 0.36 0.57 0.72 0.12 0.27 0.66 0.81
C10 0.39 0.54 0.39 0.54 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27 0.66 0.81 0.12 0.27 0.21 0.36 0.57 0.72
The experts are also asked to express the importance of each criterion. The lin-
guistic expressions they used for the evaluation and the associated IVT2IF values are
given in Table 6.
Table 6. Expert evaluations on the importance of the criteria.

Linguistic eval. µL µU ѵL ѵU
C1 VI 0.7 0.9 0.06 0.26
C2 VI 0.7 0.9 0.06 0.26
C3 VI 0.7 0.9 0.06 0.26
C4 I 0.54 0.74 0.22 0.42
C5 I 0.54 0.74 0.22 0.42
C6 M 0.38 0.58 0.38 0.58
C7 M 0.38 0.58 0.38 0.58
C8 U 0.22 0.42 0.54 0.74
C9 U 0.22 0.42 0.54 0.74
C10 U 0.22 0.42 0.54 0.74
By multiplying the values in normalized decision matrix and importance values

weighted sum decision matrix is formed by using Eq. 4. By using Eq. 5 weighted
product decision matrix is formed. Table 7 shows the weighted sum decision matrix
and WPM values.
WSM and WPM values are aggregated by using p = 0.5 to find the interval-valued
type-2 intuitionistic fuzzy WASPAS values. After defuzzifying these values defuzzified
results are obtained. The results are shown in Table 8. According to the results Alter-
native 4 is the best alternative and it is followed by Alternative 5 and Alternative 1.
Table 7. Weighted product decision matrix and WPM values.

AL1 AL2 AL3 AL4 AL5
C1 0.29 0.44 0.52 0.67 0.29 0.44 0.52 0.67 0.29 0.44 0.52 0.67 0.29 0.44 0.52 0.67 0.29 0.44 0.52 0.67
C2 0.29 0.44 0.52 0.67 0.29 0.44 0.52 0.67 0.29 0.44 0.52 0.67 0.29 0.44 0.52 0.67 0.29 0.44 0.52 0.67
C3 0.55 0.69 0.27 0.41 0.29 0.44 0.52 0.67 0.55 0.69 0.27 0.41 0.64 0.77 0.19 0.32 0.29 0.44 0.52 0.67
C4 0.76 0.87 0.1 0.22 0.36 0.51 0.48 0.62 0.45 0.59 0.4 0.53 0.36 0.51 0.48 0.62 0.76 0.87 0.1 0.22
C5 0.25 0.42 0.56 0.71 0.69 0.81 0.17 0.29 0.54 0.67 0.32 0.45 0.54 0.67 0.32 0.45 0.76 0.87 0.1 0.22
C6 0.34 0.51 0.5 0.65 0.81 0.9 0.09 0.2 0.81 0.9 0.09 0.2 0.75 0.84 0.15 0.26 0.62 0.73 0.28 0.4
C7 0.86 0.95 0.02 0.13 0.81 0.9 0.09 0.2 0.81 0.9 0.09 0.2 0.81 0.9 0.09 0.2 0.81 0.9 0.09 0.2
C8 0.81 0.88 0.13 0.22 0.86 0.92 0.07 0.17 0.76 0.84 0.19 0.28 0.86 0.92 0.07 0.17 0.7 0.79 0.24 0.35
C9 0.27 0.53 0.52 0.68 0.45 0.61 0.44 0.57 0.56 0.68 0.37 0.49 0.56 0.68 0.37 0.49 0.45 0.61 0.44 0.57
C10 0.7 0.79 0.24 0.35 0.86 0.92 0.07 0.17 0.86 0.92 0.07 0.17 0.86 0.92 0.07 0.17 0.56 0.68 0.37 0.49
Table 8. WASPAS values and defuzzified results.

µL µu mL mU Results Ranking
AL1 0.432 0.741 0.036 0.191 0.621 3
AL2 0.401 0.726 0.033 0.184 0.603 5
AL3 0.426 0.745 0.027 0.159 0.62 4
AL4 0.441 0.76 0.024 0.154 0.633 1
AL5 0.442 0.752 0.027 0.169 0.63 2
6 Conclusion
Multicriteria decision-making methods have been widely used in the field of tech-
nology selection. WASPAS is a relatively novel MCDM method that has been used in
the literature. The method has also been extended to its fuzzy versions [27, 28]. In this
paper, interval-valued IFS2 fuzzy WASPAS method is used for a technology selection
problem. For further studies, the other fuzzy extensions [29] such as fuzzy information
axiom [30], Pythagorean fuzzy sets [31], hesitant fuzzy sets [32] may be used, and the
results can be compared with the results of this paper.
References
1. Celikkan, U., Somun, G., Kutuk, U., Gamzeli, I., Cinar, E.D., Atici, I.: Capturing
supermarket shopper behavior using SmartBasket. In: Digital Information Processing and
Communications, pp. 44–53. Springer, Heidelberg (2011)
2. Merad, D., Aziz, K.E., Iguernaissi, R., Fertil, B., Drap, P.: Tracking multiple persons under
partial and global occlusions: application to customers’ behavior analysis. Pattern Recogn.
Lett. 81, 11–20 (2016)
3. Yucel, H., Edizkan, R., Ozkir, T., Yazici, A.: Development of indoor positioning system
with ultrasonic and infrared signals. In: International Symposium on Innovations in
Intelligent Systems and Applications. Davos, Switzerland (2012)
4. Otay, İ., Oztaysi, B., Onar, S.C., Kahraman, C.: Multi-expert performance evaluation of
Knowl. Based Syst. 133, 90–106 (2017)
5. Kahraman, C., Gülbay, M., Kabak, Ö.: Applications of fuzzy sets in industrial engineering: a
topical classification. Fuzzy Appl. Ind. Eng. 201, 1–55 (2006)
6. Kaya, İ., Kahraman, C.: Development of fuzzy process accuracy index for decision making
problems. Inf. Sci. 180(6), 861–872 (2010)
7. Kaya, I., Öztayşi, B., Kahraman, C.: A two-phased fuzzy multicriteria selection among
public transportation investments for policy-making and risk governance. Int. J. Uncertainty,
Fuzziness Knowl. Based Syst. 20, 31–48 (2012)
8. Estrella, F.J., Cevik Onar, S., Rodríguez, R.M., Oztaysi, B., Martínez, L.: Selecting firms in
University technoparks: a hesitant linguistic fuzzy TOPSIS model for heterogeneous
contexts. J. Intell. Fuzzy Syst. 33(2), 1155–1172 (2017)
9. Öztaysi, B., Behret, H., Kabak, Ö., Sarı, I.U., Kahraman, C.: Fuzzy inference systems for
disaster response. In: Decision Aid Models for Disaster Man. and Emergencies, pp. 75–94
(2013)
10. Oztaysi, B., Cevik Onar, S., Kahraman, C.: Fuzzy multicriteria prioritization of Urban
transformation projects for Istanbul. J. Int. Fuzzy Syst. 30(4), 2459–2474 (2016)
11. Zavadskas, E.K., Turskis, Z., Antucheviciene, J., Zakarevicius, A.: Optimization of weighted
aggregated sum product assessment. Elektronika ir Elektrotechnika Electron. Electr. Eng.
122(6), 3–6 (2012)
12. Kirchberg, V., Tröndle, M.: Experiencing exhibitions: a review of studies on visitor
experiences in museums curator. Museum J. 55(4), 435–452 (2012)
13. Fernandez-Llatas, C., Lizondo, A., Monton, E., Benedi, J.M., Traver, V.: Process mining
methodology for health process tracking using real-time indoor location systems. Sensors 15
(12), 29821–29840 (2015)
14. Yuanfeng, D., Dongkai, Y., Huilin, Y., Chundi, X.: Flexible indoor localization and tracking
system based on mobile phone. J. Netw. Comput. Appl. 69, 107–116 (2016)
15. Yoshimura, Y., Sobolevsky, S., Ratti, C., Girardin, F., Carrascal, J.P., Blat, J., Sinatra, R.:
An analysis of visitors’ behavior in The Louvre Museum: a study using Bluetooth data.
Environ. Plan. B Plan. Des. 41, 1113–1131 (2014)
16. Huang, C.N., Chan, C.T.: ZigBee-based indoor location system by k-nearest neighbor
algorithm with weighted RSSI. Procedia Comput. Sci. 5, 58–65 (2011)
17. Wu, Y.K., Wang, H.C., Chang, L.C., Chou, S.C.: Customer’s flow analysis in physical retail
store. Procedia Manuf. 3, 3506–3513 (2015)
18. Dogan, O., Öztaysi, B.: In-store behavioral analytics technology selection using fuzzy
decision making. J. Enterpr. Inf. Manag. 31(4), 612–630 (2018)
19. Budak, A., Kaya, İ., Karaşan, A., Erdoğan, M.: Real-time location systems selection by
using a fuzzy MCDM approach: an application in humanitarian relief logistics. Appl. Soft
Comput. 92, 106322 (2020)
20. Dogan, O., Bayo-Monton, J.L., Fernandez-Llatas, C., Oztaysi, B.: Analyzing of gender
behaviors from paths using process mining: a shopping mall application. Sensors (Basel) 19
(3), 557 (2019)
21. Lee, S., Kim, J., Moon, N.: Random forest and WiFi fingerprint-based indoor location
recognition system using smart watch. Hum. Centric Comput. Inf. Sci. 9, 6 (2019)
22. Dogan, O., Gurcan, O.F., Oztaysi, B., Gokdere, U.: Analysis of frequent visitor patterns in a
shopping mall. In: Industrial Engineering in The Big Data Era, pp. 217–227 (2019)
23. Li, L., Li, X., Yang, Y., Dong, J.: Indoor tracking trajectory data similarity analysis with a
deep convolutional autoencoder. Sustain. Cities Soc. 45, 588–595 (2019)
24. Dogan, O., Oztaysi, B.: Genders prediction from indoor customer paths by Levenshtein-
based fuzzy kNN. Expert Syst. Appl. 136, 42–49 (2019)
25. Dogan, O., Oztaysi, B., Fernandez-Llatas, C.: Segmentation of indoor customer paths using
intuitionistic fuzzy clustering: Process mining visualization. J. Intell. Fuzzy Syst. 38(1), 675–
684 (2020)
26. Ilbahar, E., Kahraman, C.: Retail store performance measurement using a novel interval-
valued Pythagorean fuzzy WASPAS method. J. Intell. Fuzzy Syst. 35(2), 1–12 (2018)
27. Ren, P., Xu, Z., Gou, X.: Pythagorean fuzzy TODIM approach to multi-criteria decision
making. Appl. Soft Comput. 42, 246–259 (2016)
28. Garg, H.: A novel accuracy function under interval-valued Pythagorean fuzzy environment
for solving multicriteria decision making problem. J. Intell. Fuzzy Syst. 31(1), 529–540
(2016)
29. Kahraman, C., Öztayşi, B., Çevik Onar, S.: A comprehensive literature review of 50 years of
fuzzy set theory. Int. J. Comput. Int. Syst. 9(1), 3–24 (2016)
30. Kahraman, C., Onar, S.C., Cebi, S., Oztaysi, B.: Extension of information axiom from
ordinary to intuitionistic fuzzy sets: an application to search algorithm selection. Comput.
Ind. Eng. 105, 348–361 (2017)
31. Onar, S.C., Oztaysi, B., Kahraman, C.: Multicriteria evaluation of cloud service providers
using pythagorean fuzzy TOPSIS. J. Mult. Valued Log. Soft Comput. 30, 263–283 (2018)
32. Kahraman, C., Onar, S.Ç., Öztayşi, B.: B2C marketplace prioritization using hesitant fuzzy
linguistic AHP. Int. J. Fuzzy Syst. 20(7), 2202–2215 (2018)
Evaluating the Influencing Factors
on Adoption of Self-driving Vehicles by Using
Interval-Valued Pythagorean Fuzzy AHP
Gozde Bakioglu(&) and Ali Osman Atahan
Department of Transportation Engineering, Faculty of Civil Engineering,

Istanbul Technical University, 34469 Maslak, Istanbul, Turkey
{bakioglugo,atahana}@itu.edu.tr
Abstract. As the use of self-driving vehicles become substantially possible,

attention should be turned to focus on people’s adoption of those vehicles.
Factors affecting the opinions and perception towards the behavioral intention to
use the self-driving vehicles have positive and negative impacts on individual’s
adaptation. One of the main characteristics of that of factors is its uncertainties
and vagueness in many aspects. Evaluation of those factors requires a multi-
criteria decision-making approach to better address the uncertainties associated
with decision-makers’ judgments. The main purpose of this study is to assess the
weight coefficients of factors affecting adoption of self-driving vehicles through
using interval-valued Pythagorean fuzzy AHP. The research’s contribution lies
in identifying and weighting the sub-criteria of factors pertaining to acceptance
of self-driving vehicle under Pythagorean fuzzy environment, in which the study
takes into consideration psychological, safety, driving-related and external
factors. The identification of criteria in accordance with their relative weight
coefficients would help policy-makers and those who develop and manufacture
self-driving vehicles.
Keywords: Self-driving vehicles Adoption behavior Pythagorean fuzzy

sets AHP
1 Introduction
Self-driving (also referred to as driverless, automated) vehicles have drawn substantial

attention in recent years. The introduction of self-driving vehicles will cause not only
changes in the overall transportation cost, but social, cultural and environmental
changes in roads and people’s life. The automated technology with different levels has
been infused into the market; the fully autonomous vehicle still needs development,
however. The Victoria Transport Policy Institute reported that commercially avail-
ability of level 4–5 vehicles will be widely used by the 2040–2050s [1].
Automated vehicles provide a unique solution to many problems in transportation.
The arrival of this technology will bring about a wide range of benefits. The potential
benefits include time efficiency, enhanced mobility for those who are unable to
drive, lower environmental impact, improved energy efficiency, reduction in crash,
efficient use of resources, and enhanced safety for pedestrians and cyclists [2–4].
504 G. Bakioglu and A. O. Atahan
The technology further offers better riding comfort with smoother acceleration and jerk.
The shared system of driverless cars will also enable users to only pay for their actual
use of the vehicles, thus potentially reducing overall cost in transportation system [5].
There are some potential negative outcomes associated with self-driving vehicle,
however. The enhanced overall road usage and traffic congestion, and driving-related
job losses and workplace changes can be listed as disadvantages of this technology.
Risk-taking behavior due to perceived safe condition of the technology among the auto
passengers and other road users is another issue which may increase the chance of
crashes occurring as well as increase their severity [6, 7].
Individuals are often hesitant to accept the new technology. Since driverless
vehicles have not been commercialized yet, adoption of them and intention to use these
vehicles are needed to be examined. According to Bansal et al. [8] concluded that, more
than three-quarters of participants in the US were slightly interested in adopting
autonomous vehicle (AV), while Pettigrew et al.’s [9] study stated that 40% of the
respondents in Australia are neutral to use AVs. Additionally, younger people and
those who living in urban areas such as metropolitan areas were more willing to
embrace these emergent technologies. Men were also found to be more likely to
anticipate pleasure and not anxiety when using self-driving vehicles [10].
Factors affecting the opinions and perception towards the behavioral intention to
use the self-driving vehicles have positive and negative impacts on individual’s
adaptation. Merat et al. [11] indicated the importance of knowing those factors that
influence the attitudes about the adoption of new technologies. Reliability and trust of
self-driving vehicles is found to be crucial psychological factor when it comes to
acceptability of automation technologies in automobile. Casley et al. [12] emphasized
that safety, legal issues and cost influence the embracement of driverless vehicles.
Ward et al. [13] stated that perceptions of risk and benefit, knowledge, and trust were
pertaining to the intention to use automated technology.
One of the main characteristics of those influencing factor is its uncertainties and
vagueness in many aspects. Fuzzy multi-criteria decision-making approach has been
developed to handle these uncertainties. Dogan et al. [14] employed interval-valued
intuitionistic fuzzy AHP and TOPSIS method to select a corridor for operating AVs.
Prakash and Barua [15] integrated AHP and TOPSIS methods to identify and rank the
solutions of reverse logistics adoption to overcome its barriers. Çolak and Kaya [16]
combined analytic hierarchy process (AHP) based on interval type-2 fuzzy sets and
hesitant fuzzy TOPSIS to prioritize the renewable energy alternatives.
Numerous studies have analyzed the attitudes towards self-driving vehicles, and
influencing factors to use those cars. The identified knowledge gap of the pre-literature
review is that identifying and weighting the sub-criteria of factors pertaining to
acceptance of self-driving vehicle does not exist. This paper aims to fill this knowledge
gap. The main purpose of this study is to assess the weight coefficients of factors
affecting adoption of self-driving vehicles through using interval-valued Pythagorean
fuzzy AHP.
The remaining part of the paper is structured as follows. Section 2 provides pre-
liminaries of Pythagorean fuzzy sets. Section 3 summarizes the methodology of this
study. Section 4 applies proposed method to influencing factor of AVs adoption
problem. Finally, conclusion and future extension is given in the last section.
Evaluating the Influencing Factors on Adoption of Self-driving Vehicles 505
2 Pythagorean Fuzzy Sets: Preliminaries

2.1 Pythagorean Fuzzy Sets (PFSs)
Pythagorean fuzzy set (PFS) introduced by Yager [17] can be defined by both mem-
bership and nonmembership degrees where the square sum of membership degree and
nonmembership degrees should be equal or less than 1. Some basic definition of the
PFS is given in Definition 1 as follows:
Definition 1: Let X be a universe of discourse, x is the element of X. A single-valued
~ in X is defined as [18]:
Pythagorean fuzzy set P
~ ¼ fhx; PðlP ð xÞ; vP ð xÞÞijx 2 X g

P ð1Þ
where lP ðxÞ [0, 1] are the degree of membership and vP ð xÞ [0, 1] shows the non-
membership degree of x to P ~ respectively, and, for every x X, it satisfies that:
0 ðlP ð xÞÞ2 þ ðvP ð xÞÞ2 1: ð2Þ
~ the degree of hesitancy of x to P

For any PFS P, ~ is given by Eq. (3):
p P ð xÞ ¼ 1 ðlP ð xÞÞ2 ðvP ð xÞÞ2 ð3Þ

Definition 2: Let P~ 1 = P lP ; vP1 and P~ 2 = P lP ; vP2 be two Pythagorean fuzzy
1 2
numbers, and k > 0, then the operations on these two Pythagorean fuzzy numbers are
defined as follows [18]:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2 2 2
~1 P
P ~2 ¼ P lP1 þ lP2 lP1 lP2 ; vP1 ; vP2 ð4Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~1 P
P ~ 2 ¼ P lP lP ; ðvP1 Þ2 þ ðvP2 Þ2 ðvP1 Þ2 ðvP2 Þ2 ð5Þ
1 2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
~1 ¼
kP 1 ð1 lP1 Þk ; ðvP1 Þk ; k [ 0; ð6Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
k
~1 ¼
P ðlP1 Þk ; 1 ð1 ðvP1 Þ2 Þk ; k [ 0: ð7Þ
2.2 Interval-Valued PFSs

Interval-valued PFSs (IV PFSs) are introduced by Zhang [19]. Some basic definitions
of the IV-PFSs is given as follows:
L U
L U
~ 1 = lL1 ; lU
Definition 3: Let P 1 ; v1 ; v1 and P~ 2 = lL2 ; lU
2 ; v2 ; v2 be two
IVPFLNs, and k > 0, then the operations of these two IVPFNs are given as follows:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

~1 P
P ~ 2 ¼ ð½ lL1 2 þ lL2 2 lL1 2 lL2 2 ;
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð8Þ
2 U 2 U 2 U 2
lU 1 þ l2 l1 l2 ; ½vL1 ; vL2 ; vU 1 ; v2 Þ
U
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2 2 2
~ ~
P1 P2 ¼ ð½l1 ; l2 ; l; l2 ; ½ vL1 þ vL2 vL1 vL2 ;
L L U
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð9Þ
2 U 2 U 2 U 2
vU1 þ v2 v1 v2 Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi h i
L 2 k U 2 k
~1 ¼
kP 1 ð1 l1 Þ ; 1 ð1 l1 Þ ; ðv1 Þ ; ðv1 Þ L k U k
ð10Þ
h i qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
k L 2 k qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
U 2 kffi
~1 ¼
P ðlL1 Þk ; ðlU
1Þ
k
; 1 ð1 v1 Þ ; 1 ð1 v1 Þ ð11Þ
L U
~ i = lL1 ; lU
Definition 4 [20]: Let P 1 ; v1 ; v1 i ¼ ð1; 2; . . .; nÞ be a collection of IV-
P
PFNs and wi = (w1, w2, …,wn) is the weight vector of P
T ~ i where n wi = 1. Then,
i¼1
Pythagorean fuzzy ordered weighted averaging (PFOWA) operator of dimension n is a
follows:
Yn w Yn w
PFWA ðP~1; P~ 2 ; . . .:P
~ n Þ ¼ ð½ð1 1 l2L Þ1=2 ð1 1 l2U Þ1=2 ;
Yn Y n
i¼1 i¼1
½ð i¼1 ðvL Þw Þð i¼1 ðvU Þw ÞÞ
ð12Þ
L U
Definition 5: Let P~ i = lL1 ; lU

1 ; v1 ; v1 be an IV-PFN and pL , pU are the hesitancy
~
degree of the lower and upper points of Pi then the defuzzified value of this IV PFN is
obtained by Eq. (13):
ð13Þ
3 Methodology: Pythagorean Fuzzy AHP
In this section, the methodology used in weighting the sub-criteria of factors pertaining
to acceptance of self-driving vehicle is given in detail. The procedural steps of PAHP
are given in the following:
Step 1: Structure the compromised pairwise comparison matrix R = (rik )mxm
according to experts’ linguistic evaluations based on Table 1.
Table 1. Weighting scale for the IVPF-AHP method

Linguistic terms Interval-valued
Pythagorean fuzzy
numbers
lL lU mL mU
Certainly low importance (CLI) 0 0 0.9 1
Very low importance (VLI) 0.10 0.20 0.80 0.90
Low importance (LI) 0.2 0.35 0.65 0.8
Below average importance (BAI) 0.35 0.45 0.55 0.65
Average importance (AI) 0.45 0.55 0.45 0.55
Above average importance (AAI) 0.55 0.65 0.35 0.45
High importance (HI) 0.65 0.8 0.2 0.35
Very high importance (VHI) 0.8 0.9 0.1 0.2
Certainly high importance (CHI) 0.9 1 0 0
Equal importance (EI) 0.5 0.5 0.5 0.5
Step 2: Calculate the difference matrices D = (dik )mxm between lower and upper
values of membership and non-membership functions using Eqs. (12) and (13):
dikL ¼ l2ikL v2ikU ð12Þ
dikU ¼ l2ikU v2ikL ð13Þ
Step 3: Compute the interval multiplicative matrix S = ðsik Þmxm using Eqs. (14)
and (15):
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sikL ¼ 1000dL ð14Þ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sikU ¼ 1000dU ð15Þ
Step 4: Calculate the determinacy value s ¼ ðsik Þmxm using Eq. (16):
sik ¼ 1 ðl2ikU l2ikL Þ ðv2ikU v2ikL Þ ð16Þ

Step 5: Find the matrix of weights T = ðtik Þmxm by multiplying determinacy value
with S = ðsik Þmxm using Eq. (17):
s þ sikU
ikL
tik ¼ sik ð17Þ
2
Step 6: Compute the normalized priority weight xi using Eq. (18).
Pm
j¼1 xij
x i ¼ Pm Pm ð18Þ
i¼1 j¼1 xij
4 Application
In this study, a MCDM methodology based on PF-AHP is proposed for weighting the
sub-criteria of factors associated with adoption of self-driving vehicle. The proposed
methodology consists of two basic stages: (1) Definition of criteria and sub-criteria in
which a literature review and experts interviews are conducted (2) Employing the PF-
AHP methodology to obtain main and sub-criteria weights.
The criteria for evaluation of influencing factors for intention to use driverless
vehicles include 4 main criteria together with 12 sub-criteria. The decision hierarchy is
presented in Fig. 1.
Fig. 1. Hierarchy for influencing factor of self-driving vehicle usage
The pairwise comparison matrix of main criteria (a 4 4 matrix) and four com-
parison matrices (4 4, 4 4, 2 2 and 2 2 matrices given in Table 2) for sub-
criteria were formed by the expert team. The consistency check is implemented to
ensure the experts’ judgments in the pairwise comparison matrix are sensible or not.
According to Saaty’s classical consistency ratio, the ratios are 0.00632, 0.00612,
0.00624 for main, psychological and safety factors, respectively. They are acceptable
because they are less than 0.1.
Table 2. Pairwise comparison matrix of main and sub-criteria
C1 C2 C3 C4 C21 C22 C23 C24

C1 EI LI BAI HI C21 EI AAI VHI VHI
C2 HI EI HI VHI C22 BAI EI VHI VHI
C3 AAI LI EI AAI C23 VLI VLI EI AI
C4 LI VLI BAI EI C24 VLI VLI AI EI
C11 C12 C13 C14 C31 C32

C11 EI BAI BAI LI C31 EI AAI
C12 AAI EI BAI BAI C32 BAI EI
C13 AAI AAI EI AI C41 C42
C14 HI AAI AI EI C41 EI AI
C42 AI EI
Table 3. Weights of main and sub-criteria.

Main criteria Sub-criteria Main criteria Sub-criteria Criteria
weight weight weight
Psychological 0.189
Factor (C1) Level of awareness 0.111 0.021
(C11)
Compatibility (C12) 0.193 0.037
Subjective Norms 0.286 0.054
(C13)
Perceived Benefits 0.409 0.077
(C14)
Safety Factor (C2) 0.599
Safety concern (C21) 0.469 0.281
Trust of stranger 0.440 0.263
(C22)
Trust of new 0.045 0.027
technology (C23)
Reliability (C24) 0.045 0.027
Driving-related 0.156
Factor (C3) Perceived ease of use 0.661 0.103
(C31)
Perceived usefulness 0.339 0.053
(C32)
External Factor 0.056
(C4) Environmental 0.5 0.028
concern (C41)
Extra expenses (C42) 0.5 0.028
Table 3 shows the priority weights of main and sub-criteria according to PF-AHP.
According to PF-AHP results, Safety factor is determined as the most important main
influencing factor with degree of 0.599. Psychological factor, driving-related factor and
external factor are ordered behind, respectively. For sub-criteria, safety concern, trust of
stranger, trust of new technology and reliability are found as the important four sub-
criteria, respectively.
5 Conclusion
Since self-driving vehicles are expected to be commercialized within a few years,

acceptability, attitudes and drivers’ intentions toward those vehicles gain importance.
Influencing factors for intention to use driverless vehicles have positive and negative
impact on owning them. In this study, influencing factors and sub-criteria are identified
through literature review and experts interviews. Then, Pythagorean fuzzy AHP
methodology is used in order to evaluate these factors in which the most influential
factor is revealed. As a result of computations, safety concern is found to be the most
effective factor that is needed to take into consideration before releasing this new
technology. Within the safety factor, safety concern, trust of stranger, trust of new
technology and reliability are determined as the important four sub-criteria, respec-
tively. The result of this study will be valuable to policy-makers and those who develop
and manufacture self-driving vehicles. In order to enhance the embracement of auto-
mated vehicles, government and the automotive industry pertaining to autonomous
vehicles should ensure the safety of these vehicles. The AV companies should also
comply with some safety standards. There are numerous possible research directions
for the future as a follow up of this study. This approach makes use of Pythagorean
fuzzy set. A promising research topic could be the use of regular fuzzy logic and
comparison of its results with a sensitivity analysis.
References
1. Litman, T.: Autonomous Vehicle Implementation Predictions, Implications for Transport
Planning. Victoria Transport Policy Institute (2018)
2. KPMG: Self-driving cars: The next revolution, New York (2012)
3. Fagnant, D.J., Kockelman, K.: Preparing a nation for autonomous vehicles: opportunities,
barriers and policy recommendations. Transp. Res. Policy Pract. 77, 167–181 (2015)
4. Wadud, Z., MacKenzie, D., Leiby, P.: Help or hindrance? The travel, energy and carbon
impacts of highly automated vehicles. Transp. Res. Part A Policy Pract. 86, 1–18 (2016)
5. Alessandrini, S., Delle Monache, L., Sperati, S., Cervone, G.: An analog ensemble for short-
term probabilistic solar power forecast. Appl. Energy 157, 95–110 (2015)
6. Adnan, N., Nordin, S.M., Rahman, I., Vasant, P.M., Noor, A.: A comprehensive review on
theoretical framework-based electric vehicle consumer adoption research. Int. J. Energy Res.
41, 317–335 (2016)
7. Gokasar, I., Bakioglu, G.: Modeling the effects of real time traffic information on travel
behavior: a case study of Istanbul Technical University. Transp. Res. Part F Traffic Psychol.
Behav. 58, 881–892 (2018)
8. Bansal, P., Kockelman, K.M.: Forecasting Americans’ long-term adoption of connected and
autonomous vehicle technologies. In: Paper Presented at the 95th Annual Meeting of the
Transportation Research Board (2016)
9. Pettigrew, S., Talati, Z., Norman, R.: The health benefits of autonomous vehicles: public
awareness and receptivity in Australia. Aust. N. Z. J. Public Health 42, 480–483 (2018)
10. Ipsos MORI: Ipsos MORI loyalty automotive survey. Technical report (2014)
11. Merat, N., Jamson, A.H., Lai, F.C., Carsten, O.: Highly automated driving, secondary task
performance, and driver state. Hum. Factors 54(5), 762–771 (2012)
12. Casley, S.V., Jardim, A.S., Quartulli, A.M.: A study of public acceptance of autonomous
cars, interactive qualifying project. Worcester Polytechnic Institute (2013)
13. Ward, C., Raue, M., Lee, C., D’Ambrosio, L., Coughlin, J.F.: Acceptance of automated
driving across generations: the role of risk and benefit perception, knowledge, and trust. In:
19th International Conference on Human-Computer Interaction, Vancouver, Canada (2017)
14. Dogan, O., Deveci, M., Canıtez, F., & Kahraman, C.: A corridor selection for locating
autonomous vehicles using an interval-valued intuitionistic fuzzy AHP and TOPSIS method.
Soft Comput. 1–17 (2019)
15. Prakash, C., Barua, M.K.: Integration of AHP-TOPSIS method for prioritizing the solutions
of reverse logistics adoption to overcome its barriers under fuzzy environment. J. Manuf.
Syst. 37, 599–615 (2015)
16. Çolak, M., Kaya, İ.: Prioritization of renewable energy alternatives by using an integrated
fuzzy MCDM model: a real case application for Turkey. Renew. Sustain. Energy Rev. 80,
840–853 (2017)
17. Yager, R.R.: Pythagorean fuzzy subsets. In: Joint IFSA World Congress and NAFIPS
Annual Meeting (IFSA/NAFIPS), Edmonton, AB, Canada, 24–28 June, pp. 57–61 (2013)
Pythagorean fuzzy sets. Int. J. Intell. Syst. 29(12), 1061–1078 (2014)
19. Zhang, X.: Multicriteria pythagorean fuzzy decision analysis: a hierarchical QUALIFLEX
approach with the closeness index-based ranking methods. Inf. Sci. 330, 104–124 (2016)
20. Garg, H.: A new generalized pythagorean fuzzy information aggregation using Einstein
operations and its application to decision making. Int. J. Intell. Syst. 33(12), 886–920 (2016)
Evaluation of Maintenance Strategies Using
Pythagorean Fuzzy TOPSIS Method
Tolga Gedikli , Beyzanur Cayir Ervural(&) ,

and Durmus Tayyar Sen
Konya Food and Agriculture University, Konya, Turkey

{tolga.gedikli,beyzanur.ervural,
tayyar.sen}@gidatarim.edu.tr
Abstract. Maintenance strategy selection (MSS) is a key issue that is followed

worldwide and contributes significantly to the performance and productivity of
manufacturing companies. Proper maintenance planning increases the reliability,
efficiency, and profitability of enterprises. In today’s conditions where compe-
tition is getting more severe, maintenance planning activities have an important
share in operating costs. Therefore, it is important to correctly identify and
implement appropriate maintenance management planning to survive. In this
study, it is aimed to select the most appropriate maintenance strategy (MS) uti-
lizing the Pythagorean fuzzy TOPSIS (PF-TOPSIS) method for a food company
in Turkey. Five main criteria (safety, cost, reliability, feasibility, and added-
value), twenty sub-criteria, and six alternative maintenance strategies (corrective
maintenance (CM), time-based preventive maintenance (TBPM), opportunistic
maintenance (OM), condition-based maintenance (CBM), predictive mainte-
nance (PdM), and reliability-centered maintenance (RCM)) were identified in
order to choose the most appropriate MS. The best MS has been determined
after the implementation of the PF-TOPSIS methodology, and alternatives are
ranked according to PF-TOPSIS results.
Keywords: Maintenance Strategy Selection Multi-Criteria Decision Making

Pythagorean fuzzy TOPSIS
1 Introduction
The global competitive environment puts pressure on businesses to constantly reduce

their production costs. In order to cut production costs, enterprises need to focus on
maintenance management planning, which is an important part of total operating costs
[1, 2]. This serious burden on total production costs shows the importance of the
appropriate Maintenance Strategy Selection (MSS) problem. The appropriate MS to be
used in the enterprises not only reduces the possibility of equipment failure but also
improves the working conditions of the equipment and ensures higher product quality.
As the MSS problem consists of sequential complex processes, it is considered as a
Multi-Criteria Decision Making (MCDM) problem [1, 2]. Since the process contains
some uncertainties such as ambiguity of decision-maker (DM) opinions and hesitation
about criteria and alternatives, it is necessary to evaluate MCDM methods within the
Evaluation of Maintenance Strategies Using Pythagorean Fuzzy TOPSIS Method 513
framework of fuzzy set theory [3]. There are many fuzzy MCDM methods in the
literature. However, the Pythagorean fuzzy TOPSIS method was performed to better
reflect the views of DMs in the study.
In the literature, many studies have been conducted in recent years on the MSS
problem. The most frequently used alternative maintenance strategies are Time-Based
Preventive Maintenance (TBPM) [1, 4], Corrective Maintenance (CM) [2], Condition-
Based Maintenance (CBM) [5], Predictive Maintenance (PdM) [2], Total Productive
Maintenance (TPM) [6], Opportunistic Maintenance (OM) [1], Reliability-Centered
Maintenance (RCM) [6], respectively. When the studies are examined, it is seen that
the MSS is applied in many different working areas such as the manufacturing industry
[7], energy sector [1, 2], transportation [8], automotive [9], textile [10] sectors. The
most commonly used MCDM methods are AHP [1, 2], TOPSIS [10], ANP [11], SAW
[12], VIKOR [13], ELECTRE [4], DEMATEL [6]. It is seen that different maintenance
strategies are selected as the most appropriate MS [1, 6, 13] when the results of the
studies are evaluated in detail.
In this study, five main criteria (safety, cost, reliability, applicability, and value-
added) and five alternative maintenance strategies (CM, TBPM, OM, CBM, PdM, and
RCM) were defined to determine the most appropriate maintenance alternative for a
food company in Turkey. The main contribution of the study is the first use of the PF-
TOPSIS method in evaluating of appropriate MS in a fuzzy environment.
The rest of this study proceeds as follows. Alternative maintenance strategies are
explained in Sect. 2. Background information about research methodology is given in
Sect. 3. Application of the PF-TOPSIS methodology is shown in Sect. 4. Finally,
Sect. 5 presents a brief conclusion, and some future directions.
2 Alternative Maintenance Strategies
This section provides an overview of the MS evaluated in the study.

Corrective Maintenance: The critical feature of CM is that operations are performed
only when a machine fails. In this strategy, no maintenance is performed until dete-
rioration occurs [1]. The CM strategy can be used where the failure of any
machinery/equipment will not affect the production system [2].
Time-Based Preventive Maintenance: TBPM includes planned and coordinated
inspections, adjustments, repairs, and renovation work of machinery and equipment
within a specific program, without the condition of failure. One of the main objectives
of this MS is to detect any situation that could cause the machine to malfunction before
the malfunction occurs [10]. However, the use of this MS does not mean that the
machine/equipment will not fail during this time.
Opportunistic Maintenance: OM is the maintenance of a machine/equipment in the
system when it fails, and in the meanwhile, it is the maintenance of
machinery/equipment maintenance time of which is approaching or which is worn out
[1, 12].
514 T. Gedikli et al.
Condition-Based Maintenance: The CBM strategy uses a series of measurement and

data acquisition systems to monitor machine performance in real-time, and then a
maintenance decision is made [1]. Continuous monitoring of operating conditions
ensures that maintenance decisions are made based on historical data and the current
state of the equipment. However, insufficient and inaccurate information reduces the
effectiveness of CBM.
Predictive Maintenance: PdM is an MS that can predict performance degradation and
machine downtime by analyzing the data of monitored parameters [2]. In contrast to the
CBM strategy, the data obtained in PdM are analyzed to find a possible downward
trend on the equipment/machines. This can be predicted when the controlled parameter
values reach or exceed the threshold values [1]. The right prediction of downtime
prevents situations that can lead to high costs and security hazards [10].
Reliability-Centered Maintenance: RCM defined a systematic approach to optimize
TBPM and PdM programs to increase equipment efficiency (uptime, performance, and
quality) while aiming to minimize maintenance costs. The RCM strategy aims to
maintain system function, rather than to make the equipment ideal [14]. RCM is very
suitable for types of failures where economic losses are high. On the other hand, the
disadvantage of RCM is the high cost of investment and its complexity.
3 Pythagorean Fuzzy TOPSIS
The procedural steps of PF-TOPSIS algorithm are provided in the following [15, 16]:
Step 1: The importance of evaluation criteria and the weights of alternatives are
combined with the Pythagorean fuzzy weighted average (PFWA) operator [17].
PFWAðP1 ;s ; . . .; Pn Þ ¼ w1 P1 w2 P2 . . . wn Pn
P2ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Qn wj Q n w
¼ Pð 1 1 ðuPj Þ2 ; vPj j Þ ð1Þ
j¼1 j¼1
where wj shows the significance of Pj , wj 0ðj ¼ 1; 2; . . .; nÞ and Rnj¼1 wj ¼ 1.

Step 2: For an MCDM problem with PF numbers, we construct the decision matrix
R ¼ Cj ðxi Þ mxn where the elements Cj ðxi Þði ¼ 1; . . .; m; j ¼ 1; . . .; nÞ are the assess-
ments of the alternative xi 2 X with respect to the criterion Cj 2 C. Each Cj ðxi Þ ¼

P uij ; vij element is PF numbers. According to the Cj criteria of the alternative xi , the
degree of membership is defined as uij and the degree of non-membership is defined
as vij .
Step 3: PF positive ideal solution (PF-PIS) determined by using Eq. (2) and PF
negative ideal solution (PF-NIS) determined by using Eq. (3).

x ¼ Cj; min

i s C
j ðxi Þ jj ¼ 1; 2; . . .; n
ð2Þ
¼ C 1 ; P u
1 ; v1 ; C2 ; P u 2 ; v2

; . . . ; Cn ; P u
n ; vn

x ¼ Cj; min

i s C
j ðxi Þ jj ¼ 1; 2; . . .; n
ð3Þ
¼ C 1 ; P u
1 ; v1 ; C2 ; P u 2 ; v2

; . . . ; Cn ; P u
n ; vn
2 2
where s Cj ðxi Þ ¼ uij vij is defined.
Step 4: The distance between the alternative xi and the PF-PIS x þ can be calculated by
using Eq. (4).
Pn
D ð xi ; x þ Þ ¼ wj d Cj ðxi Þ; Cj ðx þ Þ
j¼1
P
ð4Þ
n
2 2 2 2 2
¼ 12 wj uij ðu þ Þ þ vij ðv þ Þ þ pij ðp þ Þ ;
2
j¼1
i ¼ 1; 2; . . .; n
Step 5: The distance between the alternative xi and the PF-NIS x can be calculated by
using Eq. (5).
Pn
D ð xi ; x Þ ¼ wj d Cj ðxi Þ; Cj ðx Þ
j¼1
P
ð5Þ
n
2 2 2 2
¼ 2 wj uij ðu Þ þ vij ðv Þ2 þ pij ðp Þ2 ;
1
j¼1
i ¼ 1; 2; . . .; n
Step 6: The revised closeness of each alternative can be calculated by using Eq. (6).
Dðxi ; x Þ D ð xi ; x þ Þ
fð xi Þ ¼ ð6Þ
Dmaks ðxi ; x Þ Dmaks ðxi ; x þ Þ
The alternative with the largest fðxi Þ is selected as the best alternative.
4 Application of Choosing the Most Appropriate

Maintenance Strategy
In this study, the application was conducted on one of the largest food manufacturing
companies in Turkey. The proposed MSS study consists of the following steps shown
in Fig. 1.
The characteristics of the company where the study is carried out are as follows.
• It has a flow type of production system.
• Some machines can cause loss of production.
• Production should be done in a way to minimize downtime.
• Currently, CM and TBPM strategies are implemented in the company.
The application was carried out with seven DMs, consisting of three maintenance
managers (0.5), three maintenance engineers (0.33), and a foreman (0.17). Then, face-
to-face interviews were conducted with DMs.
Step 1: Determination of the Step 2: Construct the decision Step 3: Application

problem matrix
Determine the PF-PIS and PF-NIS
Determination of DMs, criteria
Obtain the decision matrix of DM
and alternatives
Calculate the weight distance
between the alternative
Conducting questionnaire with Calculate the group aggregated
focus expert group decision matrix
Calculate the relative closeness of
MS, and rank MSS
Fig. 1. The flowchart of the application
4.1 Identify of Criteria and Alternatives Maintenance Strategies

Evaluation criteria should be determined to realize the most suitable MSS problem for
the company. These criteria were selected by considering the similar studies in the
literature and the opinions of the decision-making team at the company to be applied.
The defined evaluation criteria of problem are explained below.
Safety (C1): The safety criterion represents conditions for avoiding errors, accidents,
or other undesirable situations [1, 2].
• Personnel Safety (C11): It contains personal damages [2].
• Facility Safety (C12): It contains machine and facility damages [8].
• Environment Safety (C13): It contains environmental damages [2].
Cost (C2): Cost refers to any costs necessary to implement an MS. Different main-
tenance strategies result in different costs [8].
• Hardware Cost (C21): It covers all hardware costs required for the MS [2].
• Software Cost (C22): It covers all software costs required to implement the selected
MS [2].
• Personnel Training Cost (C23): It covers all the training to be given to mainte-
nance personnel to implement the selected MS.
• Maintenance Cost (C24): It includes factors such as consultancy services, spare
parts stocks, material costs, which are necessary for maintenance.
Reliability (C3): Reliability is the ability of equipment to perform the functions

expected of it without failure.
• Frequency of Failure (C31): It refers to the average time between failures [1].
• Repair Time (C32): It refers to the average repair time [1].
• Machine İmportance for the Process (C33): It states how heavy the machine is to
keep production going. More complex machines have higher maintenance costs
(material and labor) and longer repair time [1].
• Technique Reliability (C34): Expresses the technical and technological infras-
tructure necessary for the implementation of the MS [2, 18].
Feasibility (C4): The MS chosen should be feasible.

• Acceptance by Personnel/Manager (C41): Managers and employees prefer
maintenance strategies that are easy to understand and implement [2].
• Available Monetary Resource (C42): It refers to the financial resources required to
implement the selected MS [8].
• Maintainability (C43): It refers to the ease of maintenance operations.
• Access Difficulty (C44): It covers the difficulty of maintenance intervention as
machinery/equipment are placed in restricted areas or located in high places [1].
Added-Value (C5): Value-added refers to the benefits of maintenance activities [2].

• Production Loss (C51): It means the production stop due to the malfunction of the
machine/equipment [2].
• Spare Parts Inventory (C52): Refers to parts used to repair defective parts [18].
• Fault Identification (C53): Fault identification refers to the ability of the mainte-
nance team to predict where and when the fault may occur [2].
• Production Quality (C54): Refers to machine/equipment failures affecting the
quality of the product produced at the facility.
• Enhanced Competitiveness (C55): Providing less costly, high quality, high cus-
tomer satisfaction, and producing the right product at the right time in the devel-
oping world conditions increases the competitiveness of a business.
As a result of the evaluation of the study, the main criterion weights were deter-
mined equally, and the weights of main criteria were distributed equally to the sub-
criteria. (C11 = 0.067, C12 = 0.067, C13 = 0.067, C21 = 0.05, C22 = 0.05, C23 = 0.05,
C24 = 0.05, C31 = 0.05, C32 = 0.05, C33 = 0.05, C34 = 0.05, C41 = 0.05, C42 = 0.05,
C43 = 0.05, C44 = 0.05, C51 = 0.05, C52 = 0.05, C53 = 0.05, C54 = 0.05, C55 = 0.05).
In this study, considering the general situation of the company, CM, TBPM, OM,
CBM, PdM, and RCM are considered as alternative maintenance strategies. The final
decision model is developed with evaluation criteria and alternative maintenance
strategies after the problem is identified.
4.2 Application of Pythagorean Fuzzy TOPSIS Method

This section shows the application of the PF-TOPSIS method to the problem described
in Sect. 3.
The decision maker’s opinions are aggregated with the PFWA aggregation operator
and are shown as in Table 1. Also, Table 1 shows the degree of hesitation, member-
ship, and non-membership of alternatives according to each criteria.
The PF-PIS and PF-NIS in Table 2 are calculated using Eq. (2) and (3).
The distances from the PF-PIS and PF-NIS are computed using Eq. (4) and (5) and
are shown in Table 3. Again, the relative closeness of the alternatives is calculated
using Eq. (6) and is shown in Table 3. According to obtained results, RCM is the most
appropriate strategy for the food company. For other alternatives, the order of pref-
erence is PdM, TBPM, CBM, OM, and CM, respectively.
518
Table 1. Aggregated PF decision matrix

Criteria CM TBPM OM CBM PdM RCM
l ѵ p l ѵ p l ѵ p l ѵ p l ѵ p l ѵ p
C11 0.16 0.87 0.47 0.52 0.47 0.71 0.21 0.84 0.50 0.54 0.46 0.71 0.71 0.25 0.66 0.85 0.10 0.52
T. Gedikli et al.
C12 0.16 0.87 0.46 0.48 0.42 0.77 0.20 0.81 0.56 0.52 0.50 0.69 0.66 0.29 0.69 0.85 0.13 0.52
C13 0.15 0.84 0.52 0.50 0.46 0.73 0.18 0.83 0.52 0.52 0.44 0.73 0.67 0.27 0.69 0.82 0.13 0.55
C21 0.83 0.09 0.55 0.54 0.44 0.72 0.89 0.09 0.44 0.47 0.49 0.74 0.47 0.57 0.67 0.30 0.70 0.65
C22 0.84 0.09 0.54 0.54 0.45 0.71 0.84 0.07 0.54 0.55 0.42 0.72 0.38 0.67 0.64 0.48 0.64 0.59
C23 0.79 0.15 0.60 0.64 0.32 0.70 0.81 0.16 0.56 0.60 0.36 0.71 0.41 0.61 0.67 0.49 0.48 0.73
C24 0.24 0.81 0.53 0.50 0.49 0.72 0.30 0.78 0.55 0.52 0.42 0.74 0.68 0.24 0.69 0.85 0.17 0.51
C31 0.18 0.82 0.54 0.52 0.50 0.70 0.27 0.80 0.53 0.49 0.47 0.73 0.70 0.23 0.68 0.84 0.15 0.53
C32 0.23 0.78 0.58 0.62 0.35 0.70 0.24 0.75 0.62 0.56 0.40 0.73 0.79 0.20 0.58 0.87 0.12 0.47
C33 0.22 0.81 0.54 0.54 0.49 0.69 0.22 0.79 0.57 0.56 0.45 0.70 0.71 0.25 0.66 0.86 0.10 0.49
C34 0.48 0.49 0.73 0.67 0.36 0.64 0.43 0.57 0.70 0.58 0.38 0.72 0.67 0.27 0.70 0.85 0.14 0.50
C41 0.85 0.14 0.50 0.72 0.23 0.66 0.80 0.17 0.57 0.59 0.39 0.70 0.52 0.50 0.69 0.24 0.77 0.59
C42 0.84 0.16 0.51 0.49 0.49 0.72 0.75 0.21 0.62 0.50 0.51 0.70 0.30 0.66 0.69 0.15 0.82 0.55
C43 0.89 0.10 0.45 0.66 0.32 0.68 0.75 0.22 0.62 0.54 0.41 0.73 0.33 0.65 0.68 0.25 0.76 0.59
C44 0.14 0.80 0.58 0.55 0.44 0.71 0.20 0.80 0.56 0.40 0.58 0.71 0.67 0.29 0.68 0.82 0.20 0.53
C51 0.17 0.80 0.57 0.57 0.41 0.71 0.21 0.78 0.59 0.46 0.54 0.70 0.63 0.33 0.70 0.71 0.29 0.64
C52 0.07 0.85 0.53 0.63 0.36 0.69 0.22 0.80 0.55 0.55 0.50 0.67 0.64 0.35 0.69 0.80 0.23 0.56
C53 0.22 0.81 0.55 0.61 0.36 0.71 0.19 0.82 0.54 0.63 0.37 0.68 0.86 0.10 0.50 0.86 0.10 0.50
C54 0.18 0.81 0.56 0.55 0.41 0.72 0.23 0.81 0.54 0.49 0.43 0.75 0.66 0.30 0.69 0.80 0.15 0.59
C55 0.17 0.83 0.53 0.56 0.36 0.75 0.21 0.79 0.57 0.54 0.44 0.71 0.62 0.25 0.75 0.81 0.18 0.55
Table 2. PF-PIS and PF-NIS

Criteria Positive ideal Negative ideal
solution solution
l ѵ p l ѵ p
C11 0.85 0.10 0.52 0.16 0.87 0.47
C12 0.85 0.13 0.52 0.16 0.88 0.46
C13 0.82 0.13 0.55 0.15 0.84 0.52
C21 0.89 0.09 0.44 0.30 0.70 0.65
C22 0.84 0.07 0.54 0.38 0.67 0.64
C23 0.81 0.16 0.56 0.41 0.61 0.67
C24 0.85 0.17 0.51 0.24 0.81 0.53
C31 0.84 0.15 0.53 0.18 0.82 0.54
C32 0.87 0.12 0.47 0.23 0.78 0.58
C33 0.86 0.10 0.50 0.22 0.82 0.54
C34 0.85 0.14 0.50 0.43 0.57 0.70
C41 0.85 0.14 0.50 0.24 0.77 0.59
C42 0.84 0.17 0.51 0.15 0.82 0.55
C43 0.89 0.10 0.45 0.25 0.77 0.59
C44 0.82 0.20 0.53 0.14 0.81 0.58
C51 0.71 0.29 0.64 0.18 0.80 0.57
C52 0.80 0.23 0.56 0.07 0.85 0.53
C53 0.86 0.10 0.50 0.19 0.82 0.54
C54 0.80 0.16 0.59 0.18 0.81 0.56
C55 0.81 0.18 0.55 0.17 0.83 0.53
5 Conclusion
Correct maintenance management activities are one of the competitive tools for
companies due to the inevitable operating costs. Suitable maintenance management
Table 3. Distance to PF-PIS and PF-NIS and the relative closeness of the alternatives
Alternatives D D f Rank
CM 0.48 0.19 −2.20 6
TBPM 0.38 0.45 −1.12 3
OM 0.47 0.21 −2.15 5
CBM 0.42 0.42 −1.42 4
PdM 0.32 0.46 −0.77 2
RCM 0.18 0.48 0.00 1
practices offer organizations opportunities not only for emergencies, but also to achieve
long-term strategic and operational goals. The best MS increases the availability and
reliability levels of facility equipment while reducing unnecessary investments for
maintenance [2]. The ideal MSS can vary depending on different working fields, the
variety of production systems, the presence of various criteria, and the opinion of the
DMs. Therefore, the most proper MS decision considers, various criteria and alterna-
tives with a multidimensional perspective for each company’s qualification.
In this study, the most proper MS policy is selected under uncertainty for a food
manufacturing company with a flow-type production system, operating in Turkey.
According to Pythagorean fuzzy TOPSIS results, the most suitable MS alternative is
obtained as RCM for the company. Other alternative maintenance strategies are PdM,
TBPM, CBM, OM, and CM, respectively. The obtained results are parallel to the
literature, so the method gives consistent and successful results.
For future research, PF-TOPSIS method can be employed in new decision-making
processes. The most appropriate MSS problem can be applied to different production
systems using various criteria and alternatives.
References
1. Bevilacqua, M., Braglia, M.: The analytic hierarchy process applied to maintenance strategy
selection. Reliab. Eng. Syst. Saf. 70(1), 71–83 (2000)
2. Wang, L., Chu, J., Wu, J.: Selection of optimum maintenance strategies based on a fuzzy
analytic hierarchy process. Int. J. Prod. Econ. 107(1), 151–163 (2007)
3. Cayir Ervural, B., Zaim, S., Demirel, O.F., Aydin, Z., Delen, D.: An ANP and fuzzy
TOPSIS-based SWOT analysis for Turkey’s energy planning. Renew. Sustain. Energy Rev.
82, 1538–1550 (2018)
4. Vahdani, B., Hadipour, H.: Extension of the ELECTRE method based on interval-valued
fuzzy sets. Soft. Comput. 15(3), 569–579 (2011)
5. Pariazar, M., Shahrabi, J., Zaeri, M.S., Parhizi, S.: A combined approach for maintenance
strategy selection. J. Appl. Sci. 8(23), 4321–4329 (2008)
6. Aghaee, M., Fazli, S.: An improved MCDM method for maintenance approach selection: a
case study of auto industry. Manag. Sci. Lett. 2, 137–146 (2012)
7. Seiti, H., Tagipour, R., Hafezalkotob, A., Asgari, F.: Maintenance strategy selection with
risky evaluations using RAHP. J. Multi-Criteria Decis. Anal. 24(5–6), 257–274 (2017)
8. Emovon, I., Norman, R.A., Murphy, A.J.: Hybrid MCDM based methodology for selecting
the optimum maintenance strategy for ship machinery systems. J. Intell. Manuf. 29(3), 519–
531 (2018)
9. Nezami, F.G., Yildirim, M.B.: A sustainability approach for selecting maintenance strategy.
Int. J. Sustain. Eng. 6(4), 332–343 (2013)
10. Shyjith, K., Ilangkumaran, M., Kumanan, S.: Multi-criteria decision-making approach to
evaluate optimum maintenance strategy in textile industry. J. Qual. Maint. Eng. 14(4), 375–
386 (2008)
11. Pourjavad, E., Shirouyehzad, H., Shahin, A.: Selecting maintenance strategy in mining
industry by analytic network process and TOPSIS. Int. J. Ind. Syst. Eng. 15(2), 171–192
(2013)
12. Gedikli, T., Cayir Ervural, B.: Selection of optimum maintenance strategy using multi-
criteria decision making approaches. In: Calısır, F., Korhan, O. (eds.) Industrial Engineering
in the Digital Disruption Era, pp. 156–170. Springer, Cham (2020)
13. Vahdani, B., Hadipour, H., Sadaghiani, J.S., Amiri, M.: Extension of VIKOR method based
on interval-valued fuzzy sets. Int. J. Adv. Manuf. Technol. 47(9–12), 1231–1239 (2010)
14. Sharma, R.K., Kumar, D., Kumar, P.: FLM to select suitable maintenance strategy in process
industries using MISO model. J. Qual. Maint. Eng. 11(4), 359–374 (2005)
pythagorean fuzzy sets. Int. J. Intell. Syst .29, 1061–1078 (2014)
16. Onar, S.C., Oztaysı, B., Kahraman, C.: Multicriteria evaluation of cloud service providers
using pythagorean fuzzy TOPSIS. J. Mult. Log. Soft Comput. 30(2–3), 263–283 (2018)
17. Zhang, X.: A novel approach based on similarity measure for pythagorean fuzzy multiple
criteria group decision making. Int. J. Intell. Syst. 31(6), 593–611 (2016)
18. Kirubakaran, B., Ilangkumaran, M.: Selection of optimum maintenance strategy based on
FAHP integrated with GRA–TOPSIS. Ann. Oper. Res. 245(1–2), 285–313 (2016)
Multicriteria Decision Making-
Applications
Process Robot Automation Selection
with MADM in Airline Cargo Sector
Gul Durak1 and A. Cagri Tolga2(&)

1
Logistics and Financial Management Department,
Galatasaray University, Ortakoy, 34349 Istanbul, Turkey
[email protected]
2
Ortakoy, 34349 Istanbul, Turkey
[email protected]
Abstract. In an environment of rapid growth sectors, it is inevitable for

companies to benefit from advanced technology. The airline cargo sector is one
of these sectors. Robotic Process Automation (RPA) is one of the latest
developments of our time, and it is a shape of service process automation.
Companies that prefer Robotic Process Automation should initially deposit
money on this technology, but will eventually turn black; so, it will be useful.
The robot interprets each process and ensures that the repetitive tasks are per-
formed correctly, its program follows the instructions defined by the developer.
However, the presence of many Robotic Process Automation companies makes
it difficult for the airline to make a choice. Robotic Process Automation selection
for international, large-scale companies; considering how much they have their
business processes and business partners; it represents a universe with many
criteria and expectations. Therefore, it is thinkable to use Multi-Attribute
Decision Making (MADM) methods as one of the biggest decision support
methods. In this paper; a new MADM; The Weighted Euclidean Distance-Based
Approach (WEDBA) method was carried out for making the decision. One of
the biggest advantages of using the WEDBA method is it brings a new outlook
to the MADM issue and takes into account both quantitative and qualitative
attributes. The method is applied to an RPA process in an airline cargo sector
and challenging results are acquired.
Keywords: Robot selection Multi attribute decision making Robotic

process automation WEDBA Airline cargo
1 Introduction
Robotic process automation (RPA) is a relatively new technological approach in the

industry that working with a programmed robot. Following the definition of the
software-based task which can be methodically reproduced; RPA results could be
utilized in order to program a robot that can perform a series of software interactions to
complete the process [1].
According to Asquit et al. [1]; the efficiency and process speed of companies that
choose Robotic Process Automation are increasing day by day. The achievement that
526 G. Durak and A. Cagri Tolga
companies have obtained is can be thought the return of investment. Zarco et al. [2]
says that; nowadays, it is possible to do many things thanks to robots, as known. The
great deal of information usable in such smart organizations ought to be common-
sensical processed for the technician and for machine builders [3].
RPA is the most suitable approach for procedural styled jobs that are logical and
repeated. One of the most important gains of execution of RPA software is that it only
makes interaction by means of the display level of software; that which humans could
see. The display level implements information conversion, compaction, and encoding
among applications [1].
Uipath defines RPA robots as it is using the user interface to catch information and
manipulate applications as the same anybody does [4].
Over the next 20 years, the experts predict that world air cargo traffic will grow
more than double although it has a smaller ratio among all cargo types. This is because
of decreasing oil prices and new changing trends about timing needs [5].
An Airline company is thinking about using RPA technology nowadays. This
company must have an agreement with an RPA company. In this environment where
there is competition and many RPA companies; it is difficult to choose among many
process robot companies. For this reason, this project aims to prefer the most desirable
process robot company for the air cargo sector by implementing multi attribute decision
making methods. There are many decision-making methods in the literature, but the
chosen methodology has been chosen for the use of three types of weights, considering
both qualitative and quantitative features, and being a relatively new study. In addition,
this method has not been studied in the literature for the air cargo sector before.
A multi-attribute decision-making method (MADM) is a policy shaping how the
data are developed to determine the most successful choice amid the potential options
[3]. The weighted Euclidean distance-based approach is one of these methods.
In the literature; it exists many versatile decision-making applications for the
solution of decision-making problems in the industrial platform [6]. In this study; one
of the relatively new Multi-Attribute Decision Making Methods; Weighted Euclidean
Distance-Based Approach was applied to select the most suitable Robotic Process
Automation company.
Process robot automation has gained much interest in banking sector while they are
digitalizing all their processes. However, this digitalization trend is newly seen in air
cargo sector which can be seen as an intelligent system. This study aims to choose the
best RPA company among ten concurring softwares with an MADM technique called
WEDBA which integrates both objective and subjective weights.
WEDBA; the method used in this study; briefly; the highest and lowest probability
depends on the weight of the options in the likely state. In this method, the most
feasible situation is defined as the ideal point (i.e., optimum point) and the least feasible
situation is defined as the anti-ideal point (i.e. non-optimum point). Ideal and anti-ideal
points are determined as the most appropriate and non-optimal values within a range of
property ratios [7]. It is possible that one option has the finest values for all attributes or
worst values for all attributes. Hence, in this study, the ideal and anti-ideal points are
also taken into as feasible solutions and are used as a reference to which other options
are quantitatively measured against. What the decision problem is to reach a feasible
solution which as close as to the ideal point and at the same time far away from anti-
Process Robot Automation Selection with MADM in Airline Cargo Sector 527
ideal point. This approach allows the decision-maker to make decisions based on either
the objective weights of importance of the attributes or his/her subjective preferences or
considering both the objective weights and the personal preferences. This approach
allows the decision maker to make decisions both according to their subjective pref-
erences and the weight of the qualifications. In this method, integrated attribute weights
are also mentioned for ranking that contained the subjective and objective weights of
attributes [8]. If the decision maker is not very clear about the significance of the
qualities relative to each other, she uses objective weights, and if she is sure about the
importance of the qualities, the subjective weights are taken into account. Integrated
weights are calculated with subjective and objective weights. Hence, here is the
decision problem aims to find the most favorite solution, which is as close as possible
to the ideal point. In this approach, three types of attribute weights are calculated: 1-
objective weights, 2-subjective weights and 3-integrated weights.
This study is arranged as: in the next part, attributes will be defined. And in third
section it is mentioned Robotic Process Automation. After that WEDBA steps will be
expressed, and in the sixth paragraph real case application will be showed.
2 Attributes
In the decision-making, thinking only about one dimension or at most three dimensions
together is a very poor way for the quality of the process. However, thinking in many
dimensions is out of the scope of humankind without the usage of any method. As
explained before, the WEDBA method aims to help the poly-dimensional thinking
procedure. And in the decision-making process of the RPA selection in airline cargo,
the attributes should be determined at first. A team in the company was assigned to this
process and they’ve decided the attributes that affect the decision with the help of our
recommendations. Here are attributes when used in robot selection in this study:
• Speed: This attribute refers to the average time the robot spends to complete a
process based on minutes.
• Cost: It expresses how much the choice of robotic firm will cost the company.
• Security: The connection can reveal large and sensitive information, and if unau-
thorized access to this information is made, system security and data privacy can be
greatly affected. That is the reason why security is one of the attributes in the
selection. Security criteria has also three sub-criteria. These are given below:
– Recording and reporting of all logs

– Recording and reporting of all data
– Work and manage on the cloud
• Modularity: With the high modularity; robots can increase the work zones of theirs.
Modularity is used in many departments to increase service quality, information
accuracy, and to reduce risk. Today, modular robots are not widely used in the air
cargo industry. In this application, modularity also means how many processes can
robot make. There are 26 process that robots can make in the company. Modularity
attribute has two sub–criteria.
– Ability to make independent identification of coordination (product is inactive when

coordination changes)
– Roboform infrastructure support
• Flexibility: It should be specified the flexibility use of (mobile) robots in various

applications was not named as a reason for selection, however, first experiences are
obtained in stationary pilot applications. It means easily adapt to change.
– DB - SQL connection capability

– Attended and unattended interoperability
– Ability to make independent identification of coordination (product is inactive when
coordination changes)
• Life Cycle: It shows how long the robot will use. And this attribute is defined with
hours.
• Power Consumption: It refers to the total amount of energy that the robotic structure
needs.
• Labor Hour: It means how much workforce does robots can create as equal to the
capacity of employees.
3 Robotic Process Automation
RPA is the processes performed by humans by robots. According to the Uipath defi-
nition, RPA is a technology that uses a computer software or robot to complete tasks,
doing it by imitating human actions. Robotic Process Automation is the future way of
improvement, which will change outsourcing [4].
Some service automation practitioners have automated more than 35% of their
processes with this technology. In business processes, RPA most ordinarily used for
configuring the software to do the work previously done by human, such as transferring
information from multiple input sources like email and spreadsheets to systems of
record such as Enterprise Resource Planning (ERP) and Customer Relationship
Management (CRM) systems.
RPA is a tool that includes process elimination, process improvement and other
business processes. RPA is one tool along with process elimination, process
improvement, and other business process tools. Although it cannot be said that RPA is
the only solution for everything, it is a great tool for processes. RPA focuses on change
human by automation done in an ‘‘outside-in’’ way.
After defining a software-based job that can be methodically replicated; RPA

outputs can be used to program a robot that performs the same software interactions
required to complete the task. Therefore, RPA is best suitable for procedural manner
tasks that are consistent and repetitive. Hence this is not limited by screen coordinates,
which supplies for a more ‘smart’ and dynamic interaction with a graphical user
interface that could connect different interface elements (see Fig. 1).
Traditional Robotic Process Automation Works that can

process (RPA) only be done by
automation candidates people
Many cases follow the

same structured pro-
cess, making automa-
tion economically fea-
sible
occasion
frequency There is repetitive
(number of similar Infrequent /exceptional
cases in a given period)
work, but not fre- cases that need to be han-
quent enough to only
dled in an ad-hoc manner
automation
Various types of
occasions
(sorted by frequency)
Fig. 1. Locating RPA.
4 WEDBA Steps
The novel MADM technique called WEDBA which integrates both subjective and
objective weights is utilized in this study. The application procedure of the WEDBA
methods by steps is given as follows:
Step 1 To find the Decision matrix: In this stage, the decision matrix is created, in
which all the characteristics of the problem addressed are expressed in a numerical
way [9].
0 1
y11 ... y1j . . . y1m
B... ... ... ... ... C
B C
D¼B
B yi1 ... yij . . . yjm C
C ð1Þ
@... ... ... ... ... A
yn1 ... ynj . . . ynm
Step 2 To standardization of attribute data: To make the decision-making vetire

easier, standardization is applied in this step. Standard values of qualifications are
sometimes defined as standard scores [10].
0 1
Z11 ... Z1j ... Z1n
B ... ... ... ... ... C
B C
D ¼ B
B Zi1 ... Zij ... Zin C
C ð2Þ
@ ... ... ... ... ... A
Zn1 ... Znj ... Zmn
xij lj
Zij ¼ ð3Þ
rj
yij
Xij ¼ ; if jth attribute is benefical ð4Þ
maxj ðyij Þ
mini ðyij Þ
Xij ¼ ; if jth attribute is non benefical ð5Þ
yij
1 Xm
l¼ i¼1 ij
X ð6Þ
m
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2
i¼1 ðXij lj Þ
rj ¼ ð7Þ
m
Step 3 To find the Ideal and Anti-Ideal points: The definition of ideal points are the
most craved points. On the other hand, non-ideal points are clusters with no undesirable
or least desired properties. The ideal points are displayed as ‘a*’ and anti-ideal points,
displayed as ‘b*’ are found from standardized decision matrix [10].
a* = { a*j} and b* = { b*j} where j = {1,2, ….,n}
Step 4 To find the attribute weights: Objective weights are utilized when the
decision maker is unsure of the relative importance of the attributes; on the other hand,
if we are sure of relative importance, subjective weights are used. Finally; integrated
weights are calculated using both objective and subjective weights [10].
For Objective Weights
Pm
ð pij ln pij Þ
Ej ¼ i¼1
; ðj ¼ 1; 2; . . .; nÞ ð8Þ
ln m
yij
Pij ¼ Pm ð9Þ
k¼1 ykj
dj ¼ 1 Ej ; ð1 j nÞ ð10Þ
dj
wj ¼ Pn ð11Þ
j¼1 dj
Integrated Weights
woj wsj
wIj ¼ Pn ð12Þ
k¼1 wk wk
o s
Step 5 To find the weighted Euclidean distance (WED), index score and ranking.
The WEDBA method used focuses on how close any alternative is to the ideal
solution or how far it is from the non-ideal solution. Calculations ensure that the chosen
alternative is closest to the ideal solution and farthest to the non-ideal solution. Defi-
nition of Euclidean distance is the shortest interval between two points. The overall
performance index score of an alternative is calculated by the Euclidean distance to
ideal and anti-ideal solutions. This distance is interrelated with the attributes’ weights
and should be incorporated in the distance measurement. Because all alternatives are
compared with ideal and non-ideal solutions, not among themselves. Therefore,
weighted Euclidean distances are taken into in the selected method [10].
5 Real Case Application
In this application, an air cargo company has to select software robot company. And
there are 10 alternative companies and 8 Attributes. Attributes are given Table 1.
Alternatives cannot be presented due to secrecy of the project. As a result of WEDBA
method; final Index Scores and ranks are given in Table 2.
Table 1. Attributes
Attributes
C1 Speed
C2 Cost
C3 Security
C4 Modularity
C5 Flexibility
C6 Life cycle
C7 Power consumption
C8 Labor hour
For subjective weights of attribute; the ranking of alternatives are 1-6- 2- 10–9- 4-8-
7- 3- 5 and for the objective and integrated weights rankings are 6–9 – 8 – 5 – 4 – 3 –
1–10–2 – 7 and 1–6 – 10–4 – 2–9 – 8 – 3 – 7 – 5.
Table 2. Index scores and rankings

Alternatives Objective Rank Subjective Rank Integrated Rank
weights weights weights
1 0.7158 1 0.2992 7 0.6699 1
2 0.5983 3 0.2809 9 0.4971 5
3 0.3141 9 0.3984 6 0.3009 8
4 0.5470 6 0.4053 5 0.5557 4
5 0.2493 10 0.4235 4 0.1926 10
6 0.6724 2 0.6194 1 0.6553 2
7 0.3811 8 0.5085 3 0.2181 9
8 0.4725 7 0.2327 10 0.4128 7
9 0.5613 5 0.5961 2 0.4809 6
10 0.5679 4 0.2927 8 0.5562 3
6 Conclusion
An approach amid the MADM; Weighted Euclidean distance-based is utilized in this

study to deal with the process robot selection problem of the airline cargo sector. Air
cargo industry has a critical importance in transportation. Both the objective and
subjective weights of importance of the attributes are utilized to calculate the integrated
weights. After all calculations; it is concluded that the Alternative 1 (Firm 1) is the
main alternative for integrated and subjective weights. It can be said that the results are
consistent. The uniqueness of this methodology is that it offers a general path that could
be applied to various choice problems.
The proposed method is logical and largely mathematical, so everything can be
explained with numbers.
For future studies: The attributes offered in this study is general for the RPA
selection. However, the researchers can adjust the attributes to their special areas. In
addition, other MADM techniques can be applied to this procedure to find out the
robustness of the result. And the results could be compared.
References
1. Asquit, A., Horsman, G.: Let the robots do it! – taking a look at robotic process automation
and its potential application in digital forensics. Forensic Sci. Int. Rep. 1, 100007 (2019)
2. Zarco, L, Siegert, J., Bauernhansl, T.: Software model requirements applied to a cyber –
physical modular robot in a production environment. In: Science Direct: Conference,
pp. 352–357 (2019)
3. Van der Aalst, W.M.P., Bichler, M., Heinzl, A.: Robotic process automation. Bus. Inf. Syst.
Eng. 60(4), 269–272 (2018)
4. Uipath.com. https://www.uipath.com/rpa/robotic-process-automation. Accessed 10 Mar
2020
5. Tolga, A.C., Durak, G.: Evaluating innovation projects in air cargo sector with fuzzy
COPRAS. In: Advances in Intelligent Systems and Computing, vol. 1029, pp. 702–710.
6. Jain, V., Ajmera, P.: Application of MADM methods as MOORA and WEDBA for ranking
of FMS flexibility. Int. J. Data Netw. Sci. 3, 119–136 (2019)
7. Ilbahar, E., Cebi, S., Kahraman, C.: A state-of-the-art review on multi-attribute renewable
energy decision Making. Energy State Rev. 25, 18–33 (2019)
8. Rao, R.V., Singh, D.: Weighted Euclidean distance-based approach as a multiple attribute
decision making method for plant or facility layout design selection. Int. J. Ind. Eng.
Comput. 3, 365–382 (2012)
9. Attri, R., Grover, S.: Application of preference selection index method for decision making
over the design stage of production system life cycle. J. King Saud Univ. Eng. Sci. 27(2),
207–216 (2015)
10. Rao, R.V.: Decision Making in the Manufacturing Environment Using Graph Theory and
Fuzzy Multi Attribute Decision Making Methods, vol. 2, pp. 159–191 (2007)
Smart System Evaluation in Vertical Farming
via Fuzzy WEDBA Method
Murat Basar1 and A. Cagri Tolga2(&)

1
Graduate School of Science and Engineering, Galatasaray University,
[email protected]
2
[email protected]
Abstract. By 2050, The World is expected to be populated by 9.1 billion

people and global food demand is estimated to rise approximately 50%. Due to
global warming agricultural output is expected to fall by 10% by 2050. Since
arable land is limited, new ways to harvest more food from the unit area such as
vertical farming are gaining popularity. Fourth industrial revolution and the AI
are expected to contribute to these efficiency efforts. Innovators use even vessel
containers to set up smart farms. In this paper, we asked experts to evaluate
urban smart farming alternatives. Weighted Euclidean Distance Based
Approximation (WEDBA) method is employed to assist the decision-making
process.
Keywords: Smart farming WEDBA method Urban agriculture
1 Introduction
Urban population is still growing worldwide. According to UN 2018 Revision of

World Urbanization Prospects, urban population outnumbered rural population by
2007. Feeding this ever-growing population is an important problem waiting for
alternative solutions. Urban agriculture is an area gaining attention and investment
among alternative solutions.
This study offers the evaluation of three vertical farm alternatives called Basic, IoT
and Automated. All three alternatives are based on hydroponic agriculture and use an
area of 200 m2. We utilized MCDM methods to assess the overall aspects of alter-
natives. We are extremely excited to see the outcomes of the study because not only the
alternatives but also experts’ enthusiasm for technology investment for agriculture will
be evaluated through this study. This study is, to our knowledge, the first one to ever
evaluate technology selection in agriculture via MCDM methods. We also integrated
fuzzy attribute weights to WEDBA method.
A literature review on hydroponics is written in the next section. Decided criteria
are introduced in the third section. A brief methodology of WEDBA is delivered in the
fourth section. A real case study is presented in the fifth section. The last section is the
conclusion.
Smart System Evaluation in Vertical Farming via Fuzzy WEDBA Method 535
2 Literature Review
There are numerous works related to hydroponics. Main articles to our concern are
those focused on process optimization; automation; hydroponic greenhouses, vertical
farm, and plant factory designs; proposed IoT applications for hydroponics and AI
(including machine learning, Big Data, and Artificial Neural Networks) applications
regarding hydroponic systems.
There are some papers, in the area of our concern, on different growing techniques.
Kaewmard and Saiyod [1], study traditional farms and greenhouses; Shamshiri et al.
[2], work on hydroponic greenhouses; Kodmany [3] uses vertical farm technique and
Graamans et al. [4] use plant factories to study the subjects.
Lina et al. [5], for optimization, examine water values (pH, EC and temperature
values), light values, nutrition values, room temperature and humidity values.
There are some studies dealing with a specific part of IoT implementation of
hydroponic system. Kaewmard and Saiyod [1] propose water management; Domingues
et al. [6] present pH and nutrient values; Kim et al. [7] study a computer controlled
macronutrient system; Graamans et al. [4] examine energy consumption of hydro-
ponics; Gwynn-Jones et al. [8] study economics of hydroponics.
De Silvas [9] and Palande et al. [10] propose a total automation system integrated
with IoT applications, on model scale works, thanks to sensors added to the farm,
farmers enjoy the capability of monitoring real time values of the farm such as tem-
perature, humidity and nutrition values. These systems not only monitor the farm, but
also can take necessary actions if needed. Researchers, generally, suggest similar
systems. These systems consist of hardware to collect and store the data and software to
process, evaluate and share the results. Researchers use different sensors to collect data
for water temperature, ph values, EC values, nutrient values, humidity, light density.
Collected data is processed via a microcomputer and the system has a board (such as
Ardunio) to connect sensors, microcomputer and gain internet and cloud access.
A more complex and promising area of study is integrating AI applications such as
machine learning, deep learning, and big data to smart farming procedures. As
Kamilaris and Prenafeta-Boldú [11] stated this area of study is relatively new since
there are no papers published prior to 2014.
There are some researches on applying machine learning algorithms to track and
control hydroponic plant growth, Alipio et al. [12] applies Bayesian network; Mehra
et al. [13] applies Deep Neural Networks to hydroponic systems.
3 Evaluation Criteria
Hydroponics and vertical farming are popular subjects gaining ever growing interest of
researchers but still there are not any MCDM study on the subject. We discussed with
farmers and scholars to detect evaluation criteria. For this study following are the
determined criteria for the evaluation process: (C1) Venture Capital Attractiveness:
Smart farming is a new field for entrepreneurs. Sustainable farming options take grants
from governments and multinational organizations such as EU. (C2) Effective Man-
ufacturing Processes: Product demand may vary according to season; it is important
536 M. Basar and A. C. Tolga
to adjust production capacity accordingly. Keeping high product quality is also an

asset. (C3) Workforce Requirement: There is constant manpower requirement to
continue production. Since the proposed systems are more complicated than conven-
tional food producing methods, educated workforce has the utmost importance. This
situation increases the workforce costs and training time. (C4) Security: There are two
aspects of security regarding the alternatives. First aspect is physical measurements and
the other one is cyber security. (C5) Space Requirement: Vertical farming enables
farmers to harvest maximum crop quantity from unit land. Since urban land is an
expansive asset, Farmers have to decide for the usage of available area. Technological
upgrades increase vertical space usage, but extra equipment also occupy floor. (C6)
R&D Capabilities: Each alternative’s advantages for developing new crop species,
ability to grow special species and increasing production, shorten crop cycles and
product quality. (C7) Expansion Opportunities: Once the production begins, investor
might increase production volume according to market demands. Alternatives’ initial
investment cost and set up time are important aspects regarding this criterion. (C8)
Investment and Maintenance Costs: Indoor farming is still a relatively young idea.
Equipment and urban land are still very expensive. Proposed alternatives require an
initial investment and the amount differs according to technology upgrades. Once the
systems are prepared there will be constant depreciation and maintenance costs.
4 WEDBA
The weighted Euclidean distance-based approach is based on the weighted distance of

alternatives from the most and the least desired situations, respectively. In this method, the
ideal point resembles the most desired situation and anti-ideal point resembles the least
desired situation. The index score of alternatives shows the effectiveness of alternatives
according to comparison of numerical comparison. Higher index score indicates close-
ness to ideal point. There are three types of attribute weights for WEDBA method:
1-objective weights, 2-subjective weights and 3-integrated weights. We will employ
integrated weights for this study. Utilization of WEDBA method according to Rao and
Singh [14]’s study is described briefly below:
Step 1: Establish the average decision matrix. Decision matrix is the collection of
each alternative’s attribute data. In a decision matrix, with m alternatives and n attri-
butes, one may express ith alternative’s performance value as dij of alternative ‘i’ for
attribute ‘j’. We will employ Ghorabee et al. [15]’s method to construct decision matrix
to include several decision-makers to process. Decision matrix construction process is
as follows:
D ¼ ½e
d ij n m ð1Þ
e 1X z
e
d ij ¼ dp ð2Þ
z i¼1 ij
where ed ijp is performance value of alternative ‘i’ for attribute ‘j’ given by pth
decision maker. The linguistic importance terms for preference ratings of alternatives
can be selected from Table 1 by decision makers. To continue standardization step, one
must defuzzificate fuzzy values attributed to alternatives by decision-makers. We
preferred Zeng et al. [16]’s study for defuzzification, following formula is employed:
x1 þ 2x2 þ 2x3 þ x4
x ¼ ð3Þ
6
Table 1. Preference ratings for alternatives.

Verbal expression Fuzzy number
Very Poor (VP) (0, 0, 0, 1)
Poor (P) (0, 1, 2, 3)
Medium Poor (MP) (2, 3, 4, 5)
Medium/Fair (M) (4, 5, 5, 6)
Medium Good (MG) (5, 6, 7, 8)
Good (G) (7, 8, 9, 10)
Very Good (VG) (9, 10, 10, 10)
Step 2: Standardization. This step allows us to calculate standard scores. Standard

scores are standardized form of attribute data. Standard scores have a variance of 1 and
mean of 0. Standardized decision matrix, D’, is as follows:
D0 ¼ ½Zij n m ð4Þ
xij lj
zij ¼ ð5Þ
rj
dij
xij ¼ ; for beneficial attributes; ð6Þ
maxj dij

maxj dij
xij ¼ ; for non-beneficial attributes; ð7Þ
dij
1 Xm
lj ¼ i¼1 ij
x ð8Þ
m
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm 2
i¼1 xij lij
rj ¼ ð9Þ
m
Step 3: Find ideal and anti-ideal points. Ideal point is the most wanted attribute
value and shown as ‘a*’ and least wanted attribute value is anti-ideal point and shown
as ‘b*’. Both ideal and anti-ideal points are picked from standard matrix.
n o n o
a ¼ aj and b ¼ bj and j ¼ f1; 2; 3; . . .; ng ð10Þ
Step 4: Attribute weights. Decision maker may prefer to use objective, subjective
or integrated weights which is a combination of both objective and subjective weights.
Objective attribute weights. Objective weights are calculated according to following
steps:
dij
rij ¼ Pm ð11Þ
i¼1 dij
Pm
i¼1 rij ln rij
Ej ¼ ð12Þ
ln m
lij ¼ 1 Ej ð13Þ
lij
wj ¼ Pm ð14Þ
i¼1 lij
Subjective attribute weights. For this study we preferred to use Ghorabee et al. [15]’s
method to calculate subjective attribute weights matrix:

ej 1m
W¼ w ð15Þ
1X z
ej ¼
w ej
w ð16Þ
z i¼1
where we pj is the criterion Cj’s weight given by pth decision-maker. The linguistic
importance terms for criteria can be selected from Table 2 by decision makers. To
continue to calculate weighted Euclidian distance, one must defuzzificate fuzzy values
attributed to criteria by decision-makers. We will, again use Eq. (3) to find defuzzifi-
cated weights.
Integrated attribute weights. woj is objective weights and wsj is subjective weights of
attribute ‘j’. Integrated weight wij of attribute ‘j’ is calculated as follows:
Table 2. Importance terms for criteria.

Verbal expression Fuzzy number
Very Poor (VP) (0, 0, 0, 1)
Poor (P) (0, 1, 2, 3)
Medium Poor (MP) (2, 3, 4, 5)
Medium/Fair (M) (4, 5, 5, 6)
Medium Good (MG) (5, 6, 7, 8)
Good (G) (7, 8, 9, 10)
Very Good (VG) (9, 10, 10, 10)
woj wsj
wij ¼ P ð17Þ
m
i¼1 w o
j w s
j
Step 5: WEDBA method depends on Euclidian distance which is the shortest

distance to rank the alternatives. Best option among the alternatives is the one closest to
ideal point. WEDiþ denotes Weighted Euclidian Distance (WED) between alternative
‘i’ and ideal point a* and WED i denotes WED between alternative ‘i’ and anti-ideal
point b*.
" #1=2
X
n
2
WEDiþ ¼ wj : Zij ai ð18Þ
j¼1
" #1=2
X
n
2
WED
i ¼ wj : Zij bi ð19Þ
j¼1
WED
Index Scorei ¼ i
ð20Þ
WEDiþ þ WED
i
Index score of an alternative resembles its closeness to ideal solution. Highest index
score shows the best option between alternatives.
5 Application
In a previous study [17], we asked five decision-makers to evaluate three hydroponics

alternatives. First alternative was to set up a smart farm in a vessel container, second
alternative was to build a modular smart farm which was flexible and could be scaled
according to available area, last alternative was to convert an industrial zone warehouse
to a smart farm. Decision makers selected the last alternative as the best alternative
among all.
This time we asked three experts (a farmer, an entrepreneur, and a data scientist) to
evaluate three alternatives. All three alternatives are hydroponic smart farms build in an
industrial zone warehouse. LEDs for illumination, HVACs, vertical farming slots,
refrigerators, harvest equipment and an area of 200 m2 are common features for all
alternatives. First alternative is a basic system. Farmer runs the system as in traditional
farms. pH value, EC value, humidity, water temperature, facility temperature and crop
growth have to be supervised by the farmer manually. Second alternative is an IoT
integrated model. This alternative collects data via sensors to monitor vital values of the
farm. Data collection occurs through sensors and the farmer intervenes to system
accordingly. System is based upon Arduino modules and farmer may monitor system
values even from distance. Third alternative not only enjoys IoT capabilities but also
has automation and AI technology integrated. Data collected thorough sensors are
processed and required actions, such as adding nutrient solution to water when EC
value decrease, are taken automatically. Crop growth is also monitored closely, and
system learns to produce best plant via machine learning algorithms.
System costs are calculated for each alternative separately. Both smart farming and
IoT materials are easily accessible due to subject’s ever-growing popularity. Since there
are not enough space, detailed cost calculations are not shared here but a brief cost data
for unit square meter of alternatives is as showed in Table 3. Cost data is added to
solution as a non-beneficial attribute. According to given steps above, survey and cost
data are evaluated and obtained outcomes are offered in Table 4. Automation alter-
native is found as the best option among alternatives thanks to advanced R&D capa-
bilities and minimum need for workforce requirements. IoT alternative performed as
the second option due to average investment cost and easy expansion opportunities.
Even tough Basic alternative requires lowest investment costs, it performed as the third
option.
Table 3. Expected cost data of alternatives. Table 4. Index scores and overall ranking.
Alternative Cost (TL) Alternative Index score Ranking
Basic (1227,1371,1516,1660) Basic 0.083 3
IoT (1240,1386,1532,1678) IoT 0.677 2
Automation (1278,1429,1579,1730) Automation 0.916 1
6 Conclusion
Global warming, growing urban population and regional conflicts are driving people to
find feeding alternatives. Indoor farming is a logical method to keep urban population
nourished and a promising alternative for countries of harsh environment. After years
of mechanized agriculture, precision agriculture might become the new norm.
This study proposed alternatives for precision agriculture and to our knowledge
there are not earlier MCDM work on the very subject. Automation alternative out-
performed other alternatives according to WEDBA method since decision makers
attached importance to R&D capabilities and workforce requirements.
For future studies, alternatives might be modified. Another MCDM method might
be used to evaluate same subject. Option valuation of alternatives is also an important
study option for the future.
References
1. Kaewmard, N., Saiyod, S.: Sensor data collection and irrigation control on vegetable crop
using smart phone and wireless sensor networks for smart farm. In: 2014 IEEE Conference
on Wireless Sensors, ICWiSE, Subang, pp. 106–112 (2014)
2. Shamshiri, R.R., Kalantari, F., Ting, K.C., Thorp, K.R., Hameed, I.A., Weltzien, C., et al.:
Advances in greenhouse automation and controlled environment agriculture: a transition to
plant factories and urban agriculture. Int. J. Agric. Biol. Eng. 11(1), 1–22 (2018)
3. Kodmany, K.: The vertical farm: a review of developments and implications for the vertical
city. Buildings 8, 24 (2018)
4. Graamans, L., et al.: Plant factories: crop transpiration and energy balance. Agric. Syst. 153,
138–147 (2017)
5. Lina, K.H., Huangb, M.Y., et al.: The effects of red, blue, and white light-emitting diodes on
the growth, development, and edible quality of hydroponically grown lettuce (Lactuca sativa
L. var. capitata), Scientia Horticulturae, 150, 86–91 (2013)
6. Domingues, D.S., Hideaki, W.T., Carlos, A., Camara, P., Suzana, L.: Automated system
developed to control pH and concentration of nutrient solution evaluated in hydroponic
lettuce production. Comput. Electron. Agric. 84, 53–61 (2012)
7. Kim, H.J., And, W.K., Roh, M.Y., Kang, C.I., Park, J.M., Sudduth, K.A.: Automated
sensing of hydroponic macronutrients using a computer-controlled system with an array of
ion-selective electrodes. Comput. Electron. Agric. 93, 46–54 (2013)
8. Gwynn-Jones, D., et al.: Can the optimisation of pop-up agriculture in remote communities
help feed the world? Glob. Food Secur. 18, 35–43 (2018)
9. De Silva, P.C.P., Ipanera P.C.A.: An industry 4.0 based architecture for distributed soil-less
food production systems. In: Proceedings of the 1st Manufacturing & Industrial Engineering
Symposium (2016)
10. Palande, V., Zaheer, A., George, K.: Fully Automated hydroponic system for indoor plant
growth. Proc. Comput. Sci. 129, 482–488 (2018)
11. Kamilaris, A., Prenafeta-Boldú, F.X.: A review of the use of convolutional neural networks
in agriculture. J. Agri. Sci. 156, 312–322 (2018)
12. Alipio, M.I., Dela Cruz, A.E.M., Doria, J.D.A. Fruto, R.M.S.: A smart hydroponics farming
system using exact inference in Bayesian network. In: IEEE 6th Global Conference on
Consumer Electronics (2017)
13. Mehra, M., Saxena, S., Sankaranarayanan, S., Jackson, T.R., Veeramanikandan, M.: IoT
based hydroponics system using deep neural networks. Comput. Electron. Agric. 155, 473–
486 (2018)
14. Rao, R.V., Singh, D.: Weighted Euclidean distance-based approach as a multiple attribute
decision making method for plant or facility layout design selection. Int. J. Ind. Eng.
Comput. 3, 365–382 (2012)
15. Ghorabaee, M.K., Zavadskas, E.K., Amiri, M., Turskis, Z.: Extended EDAS Method for
Fuzzy Multi-criteria Decision-making: An Application to Supplier Selection. Int. J. Comput.
Commun. Control 11, 358–371 (2016)
16. Zeng, J., An, M., Smith, N.J.: Application of a fuzzy based decision-making methodology to
construction project risk assessment. Int. J. Project Manage. 25, 589–600 (2007)
17. Tolga, A.C., Gamsiz, B., Basar, M.: Evaluation of Hydroponic System in Vertical Farming
via Fuzzy EDAS Method. In: Kahraman, C., Cebi, S., Cevik Onar, S., Oztaysi, B., Tolga, A.
C., Sari, I.U. (eds.) INFUS 2019. AISC, vol. 1029, pp. 745–752. Springer, Cham (2020).
Blockchain Software Selection for a Maritime
Organization with MCDM Method
Melis Kaska1 and A. Cagri Tolga2(&)

1
Logistics and Financial Management Department, Galatasaray University,
[email protected]
2
[email protected]
Abstract. Technological development has begun with the Mainframe in 1970’s

and the blockchain appears in near 2010s. The blockchain is known as a dis-
tributed ledger system that keeps records of the transactions between various
parties. These records are unalterable and immutable. Maritime industry is one
of the most preferable method of transport due to pricing and it plays a huge role
in the global supply chain, however, due to the customs operations and the
slowness of the ship, it lags behind in terms of time compared to its competitors.
This situation also questions its efficiency. With the help of the blockchain,
maritime industry can become smarter, more efficient and trustworthy. This
study aims to help a maritime organization to choose Blockchain software with
an MCDM Method such as technique for order of preference by similarity to
ideal solution (TOPSIS). All the selection criteria are taken from the literature
and evaluated based on the opinions of the experts of the industry. A real case
study is applied in a maritime enterprise and the results are explicated in the last
section.
Keywords: Blockchain Intelligent contracts Maritime industry TOPSIS

Supply chain Smart solutions
1 Introduction
Blockchain is well known technology which entered our lives with a Blockchain
concept. It is almost used in every area of our lives like banking, communication,
transportation. Maritime industry is one of the most preferable and well-known actors
of the method of transport. More than 90% of the world trade is carried by the sea.
Approximately 4 trillion dollars’ worth of goods and materials are moving around the
world according to Shirani [1].
Blockchain is a distributed ledger system. It keeps a record of transaction between
various parts in a permanent way. A consensus mechanism secures the validity of the data
with the help of peer-to-peer system. Every transaction uses cryptographic hash. All the
transactions have a time stamp and digital signature. Key characteristics of a blockchain
technology can be listed as follows, decentralization, transparency, immutability,
544 M. Kaska and A. C. Tolga
privacy, anonymity and security. With the help of the blockchain intermediaries can be
eliminated, transaction costs can be lowered, data privacy can be maintained, customer
satisfaction can be augmented.
Maritime industry constitutes a huge part of the global trade. In the current shipping
supply chain, almost all the operational procedures are run manually and contain too
many third parties. Many of the processes are paper based. Paper-based actions due to
delays and huge operational costs. The intervention of the third parties may cause theft
and fraud. To avoid all these negativities, there should be transparency on the trans-
actions. This study aims to find the ideal smart solution that fits the needs the of the
selected company. With that solution company can benefit from the opportunities of
blockchain and turn these advantages into a milestone in their transformation through
the digitalization. The contribution of this study to the literature and practical life is to
help the decision-makers in the maritime sector to select the most suitable blockchain-
based smart solution to their company.
This paper is structured as follows: the following section contains the late studies
about this topic. The methodology is explained in the third section. At the fourth
section, the application of the selected is discussed. Then the conclusion and further
works are given.
2 Literature Review
Blockchain and maritime industry are addressed in many different ways in the litera-
ture. K. Czachorowski mentioned that blockchain could reduce the pollution and
provide a green supply chain [2]. Petersan et al. [3] suggest that companies should
try small-scaled blockchain technologies before making huge investments. Ivanic
and Jardas et al. [4] list the benefits of blockchain as consistency and transparency in
the network, minimization of the errors, protection against fraud and real time feed-
back. Cherniavska et al. [5] supports the transparency and traceability of the real time
information in the supply chain. Weernink claims that blockchain can improve the
efficiency in supply chains by digitalizing the port logistics. [6] H. Tang et al. [7] come
up with a research that shows that the most influential indicators are the recognition,
activity and technology in selecting the right technology. Sadouskaya [8] concludes
that the firms can build a trust for the blockchain over time to spread it all over their
supply chain. Every day, the word Blokchain is spreading all over the world. By the
time, firms will build a trust for the blockchain and its capabilities. According to the
researches conduct by Gausdal et al. [9], main barrier in accepting the blockchain is the
lack of knowledge. That absence of knowledge can be eliminated by being open
minded to the new possibilities and efficient solutions.
3 Methodology
Selected methodology for this paper is a MCDM method such as TOPSIS. The word
MCMD signifies the Multi Criteria Decision Making and TOPSIS means the Tech-
nique for Order Preference by Similarity to an Ideal Solution. The main advantage of
Blockchain Software Selection for a Maritime Organization with MCDM Method 545
this method is to measure a relative performance among the alternatives. Selected

method is MCDM. The alternatives can be evaluated and sorted in the appropriate
manner to help the decision makers to choose the best smart solution that fits to their
companies. Entropy is used to evaluate the weights of the decision maker ideas.
Selected criteria are from the literature. At the beginning, there were 58 criteria.
After several eliminations, the number of criteria is reduced to 20. Total number of
alternatives to be compared are 12. There are 10 decision makers. The alternatives are
listed as follows:
1. Tallysticks – Smart Contracts
2. Wave – Electronic Bill of Laden
3. Skuchain – Handling Cargo Paper Documentation
4. Hyperledger – Configuration of a consensus mechanism, open distributed ledger
5. Provenance – Transparency
6. Fluent – Low cost and secure payment
7. Chain of Things – Monitoring and Recording Data
8. Solas VGM – Collaboration between different parties
9. Tu Delft Blockchain Lab – Transaction Speed and cryptography
10. Global Shared Container Platform (GSCP) – Real time tracking
11. Blockfreight – Fast transaction with the elimination of third parties
12. BitNautic – Transparency and tracking
4 Application
As mentioned above there are 20 criteria. Table 1 shows the criteria. All 20 criteria are
evaluated based on the selected alternatives by the decision makers.
10 decision makers from the maritime industry evaluate all the criteria for each
alternative. Evaluation scale is from 1 to 5. 1 represents the “Very Low” and 5 rep-
resents “Very High”. Their evaluations are taken and the average score of the evalu-
ation computed.
4.1 Entropy Method and Results

As mentioned before, Entropy method is used to evaluate the criteria weights. Suppose
that there is decision matrix D with m alternatives and n criteria.
1 Step: Projected weight Pij is calculated as follows
Xij
Pij ¼ Pm 8i; j ð1Þ
i¼1 Xij
Table 1. Criteria’s explanation.

Criteria Explanation
C1 Operating without a trusted third party
C2 Fast and inexpensive transactions
C3 Security and privacy
C4 Storing transactions openly across many nodes
C5 Greater automation of business processes and less human impact
C6 Investment and Return on Investment (ROI)
C7 Efficiency and transparency on custom administration
C8 Efficiency on Import-Export procedures
C9 Carrier to terminal coordination and planning
C10 Supply chain visibility and information sharing
C11 Transparency and speed
C12 Recognition of the technology
C13 Legal Complexity
C14 Authenticity
C15 Long-term digital preservation
C16 Waste reduction
C17 Size of the organization
C18 IT readiness
C19 Competitive pressure
C20 Sector support
2 Step: Entropy value Eij is calculated as follows

Xm
Eij ¼ k i¼1
PijlnPij 8j ð2Þ
3 Step: k value, which is denoted as a positive constant is calculated with the fol-
lowing equation
1
k¼ ð3Þ
lnm
4 Step: Calculation of the diversification degree is computed as follows
dj ¼ 1 Ej; 8j ð4Þ
5 Step: Weight is computed with the following equation
dj
W j ¼ Pn ; 8j ð5Þ
j¼1 dj
With respect to the equations above, the entropy value, diversification degree and
the weights for each criterion can be seen in Table 2.
4.2 TOPSIS Method and Results

Hwang and Yoon [10] propose TOPSIS Method in 1981. The main assumption in this
method is the alternative that is chosen, should have the shortest distance from the
positive ideal solution and farthest distance from the negative ideal solution. In the
TOPSIS Method, the weights obtained from the Entropy Method are used to find the
ideal alternative.
Table 2. Entropy calculations.

Ej dj Wj
C1 0,994 0,006002 0,5
C2 0,997638 0,002362 0,2
C3 0,993745 0,006255 0,6
C4 0,992427 0,00757 0,7
C5 0,951684 0,048316 4,4
C6 0,993998 0,006002 0,5
C7 0,995628 0,004372 0,4
C8 0,99579 0,00421 0,4
C9 0,997519 0,002481 0,2
C10 0,995992 0,004008 0,4
C11 0,997892 0,002108 0,2
C12 0,990403 0,009597 0,9
C13 0,995628 0,004372 0,4
C14 0,988828 0,011172 1
C15 0,995718 0,004282 0,4
C16 0,98032 0,01968 1,8
C17 0,979964 0,020036 1,8
C18 0,99579 0,00421 0,4
C19 0,993998 0,006002 0,5
C20 0,994984 0,005016 0,5
First step is the calculation of the normalized decision matrix. Then weights
obtained from the entropy are multiplied by the normalized performance values. This
multiplication takes us to the weighted normalized decision matrix. After that we
calculate the ideal positive and negative values. After that step, Euclidean distance from
the ideal best and ideal worst solution are calculated. Performance scores are deter-
mined to rank the alternatives among each other.
Table 3 shows the final ranking of the alternatives. The bigger positive ideal
solution takes the first place. A12 has the biggest and A9 has the smallest ideal
solution. That is why, in the ranking A12 takes the first place and A9 takes the last
place. A12 represents the BitNautic and A9 represents the Tu Delft Blockchain smart
solutions. The calculations show that the best smart solution for the selected company
is the BitNautic. BitNautic eases transparency and tracking.
5 Conclusion
Blockchain can be implemented in maritime industry and the maritime companies in

various ways. The maritime sector become smarter, efficient and pioneer in the sector
with the application of the blockchain technology. Smart solutions can apply to the
companies to ease real time tracking, obtain maximum transparency, eliminate the third
parties, utilize the smart contracts create a consensus mechanism and many other ways.
Table 3. The final ranking of the alternatives.

Alternatives Performance score Rank
A1 0,331307 9
A2 0,631825 4
A3 0,38779 8
A4 0,22095 10
A5 0,205408 11
A6 0,66716 3
A7 0,427592 6
A8 0,409736 7
A9 0,184078 12
A10 0,752914 2
A11 0,621873 5
A12 0,863162 1
Entropy and TOPSIS methods are used to evaluate 20 criteria and 12 alternatives.
There were highly qualified 10 decision makers that evaluate each criterion. Entropy is
used to obtain the weights, which are used in TOPSIS to find the ideal alternative
among twenty others. After the computations with the help of TOPSIS method, the best
smart solution that fits the needs of the selected company is the BitNautic. That smart
solution ease tracking and provides transparency. With the help of that smart solution,
tracking can be done from anywhere and anytime. One of the main advantages of the
blockchain is the transparency. This transparency builds trust between the companies
and customers and like a chain, this trust brings customer satisfaction. Higher customer
satisfaction means loyal customers and more profit to the companies.
In the future research, combination of different smart solutions and their impact can
be analyzed.
References
1. Shirani, A.: Blockchain for global maritime logistics. Issues Inf. Syst. 19(3), 175–183 (2018)
2. Czachorowski, K., Solesvik, M., Kondratenko, Y.: The application of blockchain technology
in the maritime industry. In: Kharchenko, V., Kondratenko, Y., Kacprzyk, J. (eds.) Green IT
Engineering: Social Business and Industrial Applications. Studies in Systems, Decision and
Control, vol. 171. Springer, Cham (2019)
3. Petersan, M., Hackius, N., See, B.: Mapping the sea of opportunities: blockchain in supply
chain logistics. IT Inf. Technol. 60(5–6), 263–271 (2018)
4. Tijan, E., Aksentijevic, S., Ivanavic, K., Jardas, M.: Blockchain technology implementation
in logistics. Sustainabiliy 11, 1185 (2019)
5. Cherniavska, B.T., Cherniavskyi, B.: Blockchain Technology into the Logistics Supply
Chain Implementation Effectiveness. Priority Sectors for Small and Medium Enterprises as
Drivers of Economic Growth (2018)
6. Weernink, M.O., Engh, W., Francisconi, M., Thorborg, F.: The blockchain potential for port
logistics. White Paper-Blockhain (2017)
7. Tang, H., Shi, Y., Dong, P.: Public blockchain evaluation using entropy and TOPSIS. Expert
Syst. Appl. 117, 204–210 (2019)
8. Sadouskaya, K.: Adoption of Blockchain Technology in Supply Chain and Logistics. Master
thesis, XAMK (2017)
9. Gasudal, A.H., Czachorowski, K., Solesvik, M.: Applying blockchain technology: evidence
from norwegian companies. Sustainability 10, 1985 (2018)
10. Olson, D.L.: Comparison of weights in TOPSIS models. Math. Comput. Model. 40(7–8),
721–727 (2004)
Prioritization of Influence Factors
for Selecting E–Learning Systems
Ali Karasan1(B) and Melike Erdogan2

1
Department of Industrial Engineering, Yildiz Technical University,
Besiktas, 34347 Istanbul, Turkey
[email protected]
2
Department of Industrial Engineering, Duzce University,
Konuralp, 81620 Duzce, Turkey
[email protected]
Abstract. COVID-19 pandemic affects not only daily life activities but
also traditional education systems. Based on the current developments,
to stick by their academic calendars, most of the educational institu-
tions continue their classes via online channels. Since the selection of
the most appropriate e–learning platform depends on multi–criteria, the
evaluation of this selection process can be dealt with decision support
systems. In this study, cognitive mapping extended with intuitionistic
fuzzy sets is introduced for prioritizing the e–learning platform selection
factors under fuzzy environment based on the multi–expert judgments.
Based on the results, infrastructure and ease of use are determined as the
most effective factors. For further studies, a sensitivity analysis based on
the initial vector determination can be studied to check its effect on the
outputs.
Keywords: E–learning · Intuitionistic fuzzy sets · Cognitive maps ·

Decision making
1 Introduction
One of the most important sectors affected by the extraordinary situations expe-
rienced by countries is the education. In cases of war, epidemic disease, earth-
quake, etc., students become unable to receive education and if the effects of
these situations continue for a long time, both students and all elements of the
education system will be irreversibly damaged. Due to the COVID–19 pandemic
that has been happening all over the world recently, countries have stopped tak-
ing face–to–face training activities at all levels. The absence of a clear prediction
on the ending date of the pandemic also required action to be taken as soon as
possible on behalf of educational activities. One of the precautions that can be
taken in order to prevent educational activities from disrupting is to apply to
e–learning platforms.
Today, most of the educational institutions continue their classes via online
channels to stick by their academic calendars. E–learning is the using of internet
Prioritization of Influence Factors for Selecting E-Learning Systems 551
so as to learn specific information and content of students [1]. As the synonym of

e–learning, web–based learning, online learning, course–learning, virtual learn-
ing, and digital collaboration learning are used [1]. E–learning is an effective
learning process created by combining digitally presented content with learning
support services [2]. E–learning is a very useful tool to ensure the continuity of
teaching and learning in order to allow students to have distance education. It
not only provides continuity of education, but also promotes to develop at the
institutional and personal level, including both faculty and students [3]. Uni-
versities have also started to use e–learning platforms due to the extraordinary
situation such as COVID19 pandemic in the world nowadays, but each university
can have different expectations in terms of the technical and flexibility features
of these platforms. For this reason, choosing the e–learning platform that meets
the expectations best in such an environment is a process that many different
criteria should be considered.
The concept of e–learning is a concept that attracts a lot of attention not
only in practice but also in the literature. The selection of the most suitable
e–learning platform is a subject previously addressed by multi–criteria decision
making methods. One of the outstanding papers in this regard, paper of Alptekin
and Karsak which is proposed a hybrid model in the selection of e-learning appli-
cations by the universities in Turkey [2]. In the proposed approach, the QFD
method was used to allocate resources and coordinate functions according to
user expectations, and fuzzy regression was used to determine the functional
relationship between customer expectations and product features. Begicevic et
al. evaluated e-learning practices in 4 stages as intelligence, design, choice and
implementation and applied AHP method to determine the best alternative in
the selection stage [3]. Bhuasiri et al. addressed the determination of critical suc-
cess factors affecting the acceptance of e–learning systems [4]. Relative impor-
tance of these factors was calculated by AHP with responses from 76 users for
information and communications technology experts’ perspective and faculty
members perspective separately. Chao and Chen identified five main criteria
and sixteen sub-criteria for a successful electronic learning application and used
consistent fuzzy preference relations in AHP to determine the relative impor-
tance of these criteria relative to each other [5]. Bo and Peng identified the main
factors to measure the effectiveness of network education in China and calcu-
lated index weights using fuzzy AHP [6]. In addition to these papers, the study
of Zare et al. could be accepted as the main reference for researchers who would
study on decision making for e–learning with their detailed literature review [1].
However, unlike all these studies, a different perspective is provided for the first
time by using the intuitionistic fuzzy cognitive mapping method. Intuitionistic
fuzzy cognitive mapping is applied and it is determined which criteria should
be taken into consideration primarily for the universities that want to select an
e-learning platform for the distance education during the pandemic period.
The rest of the paper is organized as follows. Section 2 presents the applied
methodology. In Sect. 3, the application is carried out. The paper ends with
obtained results and future work suggestions.
552 A. Karasan and M. Erdogan
2 Methodology
Since the prioritization problem consists of multi evaluation criteria and many
alternatives together with human knowledge on the operation of the system, the
structure of the problem can be constructed by using expert judgments. The
cognitive maps were firstly introduced by Axelrod to represent interrelation-
ships and their corresponded binary values [7] based on the human evaluations.
After that, Kosko constructed fuzzy cognitive maps to increase its ability by
representing not only cardinal information but also vague judgments [8]. In the
constructed environment, construction methodologies rely on the exploitation of
experts’ experience on system’s model and behavior [9]. Therefore, construct-
ing a scale based on the linguistic terms is an efficient way of representing the
human knowledge. On the other hand, experts can be hesitant while express
their judgments. In these systems where there is no exact determinacy about
the evaluations, the most appropriate way to use the data is to formulate hesi-
tancy and integrate it to mathematical formulations. In our paper, Intuitionistic
fuzzy sets are applied to extend cognitive maps to use this advantage while deal
with the hesitant information [9].
Step 1. Assign the linguistic terms for membership functions based on the scale
given in Table 1.
Table 1. Linguistic terms and their corresponded membership functions
Linguistic term Corresponded triangular membership function

Neg Very Strong –VS (−1, −1, −0.7)
Neg Strong –S (−1, −0.7, −0.5)
Neg Medium –M (−0.7, −0.5, −0.3)
Neg Weak –V (−0.5, −0.3, −0.1)
Neg Very Weak –VW (−0.2, −0.1, 0)
No Relation N (−0.1, 0, 0.1)
Pos Very Strong VW (0, 0.1, 0.2)
Pos Strong V (0.1, 0.3, 0.5)
Pos Medium M (0.3, 0.5, 0.7)
Pos Weak S (0.5, 0.7, 1)
Pos Very Weak VS (0.7, 1, 1)
Step 2. Determine the hesitancy of the membership functions based on the

scales given in Table 2.
Step 3. Convert linguistic terms to obtain triangular membership and hesitancy
functions.
Step 4. Construct the weight matrix.
Table 2. Linguistic terms and their corresponded hesitancy functions
Linguistic term Corresponded hesitancy function

Very Low VL (0, 0.1, 0.2)
Low L (0.1, 0.2, 0.35)
Medium M (0.25, 0.4, 0.6)
High H (0.5, 0.65, 0.8)
Very High VH (0.7, 0.9, 1)
Step 5. Compute the weights by using Eq. (1).

⎛ ⎞
N

⎜ μ π⎟
sk+1
i = f ⎝ski + skj · wji − skj · wji ⎠ (1)
j=1
j=i
where f (x) = 1+e1−x . The iteration is run until all the weights are converged or
a certain number of iterations is reached.
Step 6. Obtain the aggregated weights with respect to decision makers.
3 Application
Through the determination of the most influenced criteria for the e–learning
systems, 14 criteria are determined. 3 decision makers evaluated the system and
constructed to interrelationships. List of the criteria is given in Table 3.
Table 3. List of the criteria
List of the criteria

C1-Ease of use [10] C8-Pricing [11]
C2-Ease of exchanging learning with the others [10] C9-Flexibility [11]
C3-Capability of controlling learning progress [10] C10-Reporting [12]
C4-Network infrastructure [3] C11- Access (time & place) [11]
C5-Availability of technical support staff [3] C12-Security & privacy [11]
C6-Exam management system [3] C13-Trialability [11]
C7-Video & audio streaming [3] C14-Interactivity level [11]
Evaluations of the decision maker 1 is given in Table 4.
Table 4. Decision maker 1 evaluations
P C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14

P N VL N VLN VLNVL N VLN VL N VL NVLN VLNVLN VL N VLNVLN VL NVL
C1 –S L N VLN VLNVL N VLN VL N VL NVL–VWL NVL–S VHN VLNVLS H NVL
C2 N VL N VLN VLNVL N VLN VL N VL NVLN VLNVLN VL N VLNVL–S H NVL
C3 N VL N VLN VLNVL N VLN VL N VL NVLN VLNVLN VL N VLNVL–M VHNVL
C4 M L N VLN VLNVL N VLN VL N VL NVLN VLNVLN VL N VLNVLN VL NVL
C5 V H N VLN VLNL VWVLN VL N VL NVLN VLNVLN VL N VLNVLN VL NVL
C6 –VS M N VLN VLNVHV VLN VL N VL NVLN VLNVLN VL N VLNVLN VL NVL
C7 –VWVHN VLN VLNVL VS VLN VL –VSVHNVLN VLNVL–V VHN VLNVLS VL NVL
C8 M L N VLN VLNVL N VLN VL N VL NVLN VLNVLV VL N VLNVL–S VL NVL
C9 –S M N VLN VLNVL N VLN VL N VL NVLN VLNVL–MVHN VLNVL–VWVHNVL
C10N M N VLN VLNVL N VLN VL N VL NVLN VLNVLN VL N VLNVLN VL NVL
C11S VHN VLN VLNVL N VLN VL N VL NVLN VLNVLN VL N VLNVLN VL NVL
C12N VH–V M N VLNVL N VL–SVHN VL NVLN VLNVLN VL VWVLNVLN VL NVL
C13–S L –VWVL–VWVLNVL N VLS L N VL NVLN VLNVLN VL N VLNVLN VL NVL
C14–S L –VWH N VLNVL N VLN VL N VL NVLN VLNVLN VL N VLNVLN VL NVL

P N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL
C1 –S M N VL N VL N VL N VL N VL N VL N VL –VW L N VL –S VH N VL N VL S VH N VL
C2 N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL –S L N VL
C3 N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL –M L N VL
C4 N M N VL N VL N VL N VL N VL N VL –M VL N VL N VL N VL N VL N VL N VL N VL
C5 V VL N VL N VL N L VW H N VL N VL N VL N VL N VL N VL N VL –VS VH N VL N VL
C6 –VW VH N VL N VL N VH V VL N VL N VL N VL N VL N VL N VL –VS H N VL N VL N VL
C7 N L N VL N VL N VL VS H N VL –VS M N VL N VL N VL –V VH N VL N VL M L –M H
C8 N VL N VL N VL M VL N VL N VL N VL N VL N VL N VL –VW VL N VL N VL N VL N VL
C9 –S VL N VL N VL N VL N VL N VL N VL N VL N VL N VL VW L N VL N VL N VL N VL
C10 N M V VL N VL N VL N VL N VL N VL N VL –M VL N VL N VL N VL N VL N VL N VL
C11 N VH VW VL N VL N VL N VL N VL N VL N VL N VL N VL VW VH N VL N VL N VL N VL
C12 M H N M N VL N VL N VL –S VH N VL N VL N VL N VL N VL VW VL N VL N VL N VL
C13 V H –V VL –VW VL N VL N VL S VH N VL N VL N VL N VL N VL N VL N VL N VL V M
C14 VW M –VS H N VL N VL N VL N H N VL N VL N VL N VL N VL N VL N VL N VL N VL


P N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL
C1 M L N VL N VL N VL N VL N VL N VL N VL –VW L N VL –S VH N VL N VL M H N VL
C2 N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL –M H N VL
C3 V VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL VS VL N VL –M VH N VL
C4 –V L N VL N VL N VL N VL N VL –V VL N VL N VL N VL N VL N VL N VL N VL N VL
C5 VS H N VL N VL N L VW VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL
C6 –M M N VL N VL N VH N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL
C7 –S VH N VL –S VL N VL VS VL N VL –VS VH N VL S VL N VL VW VH –V VL N VL S VL N VL
C8 N L N VL N VL N VL N VL N VL –S VL N VL N VL VW VL V VL N VL N VL –S VL N VL
C9 N M VS VL N VL N VL N VL N VL N VL N VL N VL N VL V VH N VL N VL S VH N VL
C10 –VS M N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL
C11 –V VH N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL S VL N VL –S VL
C12 V VH VS M N VL N VL N VL N VH N VL N VL N VL N VL N VL N VL N VL N VL N VL
C13 VW L V VL N VL N VL N VL N L N VL N VL N VL N VL N VL N VL N VL N VL N VL
C14 –V L M H N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL N VL
For the initial vector, it is determined as [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1].

After 10 iterations, weights are converged. Results of the application are given
in Table 7.
Table 7. Results of the application
Weight C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14

0.3 DM1 0.599 0.642 0.659 0.930 0.749 0.637 0.659 0.645 0.659 0.857 0.676 0.659 0.645 0.659
0.3 DM2 0.817 0.645 0.726 0.881 0.657 0.566 0.535 0.549 0.659 0.636 0.620 0.638 0.542 0.780
0.4 DM3 0.959 0.525 0.659 0.807 0.659 0.410 0.659 0.750 0.678 0.899 0.758 0.781 0.669 0.504
Aggregated 0.808 0.596 0.679 0.866 0.686 0.525 0.622 0.658 0.667 0.807 0.692 0.701 0.624 0.633
Rank 2 13 7 1 6 14 12 9 8 3 5 4 11 10
Based on the results, C4–Network infrastructure and C1–Ease of use criteria

are determined as first and second most important criteria for the selection of
the e–learning systems.
4 Conclusions
One of the most effective ways to prevent educational activities from being inter-
rupted in extraordinary situations is to apply for online learning processes. In
today’s conditions, while all countries in the world struggle with the epidemic
disease COVID–19, it has been necessary to apply to e–learning platforms for
continuity of educational activities. E–learning becomes an effective tool by com-
bining digital content delivery with learning services. However, at this point, we
face the problem of determining which e–learning platform on the market univer-
sities should apply. For this reason, it should be investigated which criteria should
be taken into consideration in choosing the most suitable e–learning platform

for educational institutions in selecting the best platform from the alternatives.
The most critical point in this process is to apply a priority analysis based on
the evaluation criteria for obtaining their rankings. For this aim, we proposed a
decision-making approach to obtain the order of importance for the determined
criteria to use for the problem of selecting best online education platforms during
this pandemic period. We have determined 14 criteria based on the literature
research and expert opinions to be used in the process and apply intuitionistic
fuzzy cognitive mapping method to determine the most important criterion. As
a result of the calculations, it is determined that the most important factor is
‘C4-Network infrastructure’ and the least important factor in the evaluation of
e-learning platforms is ‘C6–Exam management system’.
For further studies, the results of the application can be compared with the
Pythagorean fuzzy cognitive mapping method’s outputs. Also, different scenarios
based on the changes in the initial vector can be studied to check its effects.
References
1. Zare, M., Pahl, C., Rahnama, H., Nilashi, M., Mardani, A., Ibrahim, O., Ahmadi,
H.: Multi-criteria decision making approach in E-learning: a systematic review and
classification. Appl. Soft Comput. 45, 108–128 (2016)
2. Alptekin, S.E., Karsak, E.E.: An integrated decision framework for evaluating and
selecting e-learning products. Appl. Soft Comput. 11(3), 2990–2998 (2011)
3. Begicevic, N., Divjak, B., Hunjak, T.: Prioritization of e-learning forms: a multi-
criteria methodology. Cent. Eur. J. Oper. Res. 15(4), 405–419 (2007)
4. Bhuasiri, W., Xaymoungkhoun, O., Zo, H., Rho, J.J., Ciganek, A.P.: Critical suc-
cess factors for e-learning in developing countries: a comparative analysis between
ICT experts and faculty. Comput. Educ. 58(2), 843–855 (2012)
5. Chao, R.J., Chen, Y.H.: Evaluation of the criteria and effectiveness of distance
e-learning with consistent fuzzy preference relations. Expert. Syst. Appl. 36(7),
10657–10662 (2009)
6. Bo, L., Xuning, P., Bingquan, B.: Modeling of network education effectiveness
evaluation in fuzzy analytic hierarchy process. In: 2009 International Conference
on Networking and Digital Society, Guiyang, pp. 198–200. IEEE (2009)
7. Axelrod, R.: Schema theory: an information processing model of perception and
cognition. Am. Polit. Sci. Rev. 67(4), 1248–1266 (1973)
8. Kosko, B.: Fuzzy cognitive maps. Int. J. Man-Mach. Stud. 24(1), 65–75 (1986)
9. Groumpos, P.P.: Fuzzy cognitive maps: basic theories and their application to com-
plex systems. In: Groumpos, P.P. (ed.) Fuzzy Cognitive Maps, pp. 1–22. Springer,
Berlin (2010)
10. Shee, D.Y., Wang, Y.S.: Multi-criteria evaluation of the web-based e-learning sys-
tem: a methodology based on learner satisfaction and its applications. Comput.
Educ. 50(3), 894–905 (2008)
11. Zhang, L., Wen, H., Li, D., Fu, Z., Cui, S.: E-learning adoption intention and its
key influence factors based on innovation adoption theory. Math. Comput. Model.
51(11–12), 1428–1432 (2010)
12. Colace, F., De Santo, M., Pietrosanto, A.: Evaluation models for e-learning plat-
form: an AHP approach. In: 36th Annual Conference of Frontiers in Education,
San Diego, pp. 1–6. IEEE (2006)
Special Agriculture Production Selection Using
Interval Type-2 Fuzzy AHP
Sarper Alem(&)
Department of Industrial Engineering, Graduate School of Science

and Engineering, Galatasaray University, Istanbul, Turkey
[email protected]
Abstract. Food and agriculture production cannot satisfy global demand which
is increasing day by day because of the world population growth. To meet this
demand, new technologies could be used to improve the production efficiency.
In this study, four different agriculture type was investigated which are “tradi-
tional farming”, “artificial intelligence aided farming”, “vertical farming” and
“plant-based meat”. Taking the most appropriate decision is complicated and
complex because these decisions are affected by many factors which are not
straight all the time. In this study, “cost”, “ecology” and “customer” were
selected as three main criteria. Any dependency couldn’t find between these
criteria therefore Interval Type-2 Fuzzy Analytic Hierarchy Process is selected
to determine factors’ weights. After weights are found, Hesitant Fuzzy TOPSIS
is used to rank alternatives. Using two different a multicriteria decision-making
model is provided to consider complexity and vagueness of decision.
Keywords: Agriculture type selection Analytic hierarchy process Hesitant

fuzzy sets Interval Type-2 fuzzy sets Vertical farming
1 Introduction
The world’s population is expected to reach 9.1 billion people in 2050 therefore
human’s farming needs will be increased too. However, climate change also brings
inefficiency along and traditional methods can not satisfy demand. Artificial intelli-
gence aided farming, vertical farming or plant-based meat can be considered as some of
the most effective solutions for more efficient and sustainable farming.
Vertical farming is a type of urban farming which help to decrease logistic cost
because of production place may be close to big cities. Plants are produced under a
controlled area which is generally controlling with IoT technology. Thus, it has high
efficiency on production. It also consume 95% less water than traditional farming.
Artificial intelligence aided farming became popular in 2010’s. It uses drone and
IoT technologies to measure plants and seeds online status and feed all plant by their
unique needs. Due to plants are growing in land, it’s still organic method if any other
supplementer or agriculture medicine are not used.
Plant based meat is not an agriculture system but it uses plant to make “fake meat”.
Thus, it might be called as farming system. This system uses heme and leg hemoglobin
558 S. Alem
molecules getting from soy roots to give meat taste and the system also uses coconut fat
to make it fatty.
In the literature, some of studies on vertical farming are examined. There is a lot of
decision technique used to select alternatives and these techniques can be adapted to
selection of farming systems. Decision making approaches can also be used with fuzzy
number which is capable of calculation with vagueness. Saaty’s [1] analytic hierarchy
process method and the fuzzy set theory of Zadeh [2] combine to achieve effective
results. Keller and Godse [3] were applied fuzzy sets on mcdm method for farming
system decision. Qureshi [4] was also used fuzzy mcdm to select crop pattern. Teniwut
[5] was used fuzzy AHP to ranking suitable location for seaweed production. Wu and
Xinwang [6] Liu has proposed fuzzy ANP method with Interval Type 2 Fuzzy sets
evaluate enterprise technology innovation ability. Xing, Cheng, and Keyu [7] have
used hesitant fuzzy linguistic analytic network process to evaluate the performance of
urban environmental governance.
This study aims to select most suitable agriculture system alternative, according to
determined criteria. Interval type-2 fuzzy Analytic Hierarchy Process is used to cal-
culate weights of criteria and Hesitant fuzzy TOPSIS is used as an MCDM approach
[8–13,16]. Experts pointed three main criteria (cost, ecology and customer) with their
sub-criteria (Capital cost, Transportation cost, Labor cost, Production cost, VC
Attractiveness, Attainability, Greenhouse gas emission, Energy usage, Water usage,
Waste reduction, Customer’s demand, Product range, Production volume, Organic
food perception) and they also evaluate three alternatives. These criteria are selected by
experts experiences. Because of these alternatives are so fresh, there is no similar study
on this issue in the literature which provides originality to this study.
This paper is organized as follows. The following section introduces criteria for the
selection of the farming alternatives. Interval type-2 fuzzy Analytic Hierarchy Process
and Hesitant fuzzy TOPSIS methods are presented in Sect. 3. A real case study and
results are shown in Sect. 4. Conclusions are given in last section.
2 Methodology
In this section, already existed methodologies Interval Type-2 Fuzzy AHP and Hesitant
Fuzzy TOPSIS are explained.
2.1 Interval Type-2 Fuzzy AHP

Saaty [1] developed Analytical hierarchy process (AHP) as multi-criteria decision
method. It uses individual experts’ judgments and experiences to calculate weights and
to rank alternatives as using continuous paired comparisons. Additionally, IT2 FAHP
use trapezoidal fuzzy numbers in three dimensions in linguistic variable scale [14].
Step 1: Defining problem as setting goal and determining criteria and sub-criteria
which are affecting to select the most suitable alternative.
Special Agriculture Production Selection 559
Step 2: Constructing pairwise comparison matrices and collecting expert judgments

using trapezoidal interval type-2 fuzzy scales used in the study are given in Table 1
[14]. In this study, AHP is used only to calculate weights of criteria and sub-criteria.
Table 1. Definition and interval type 2 fuzzy scales of the linguistic variables
Linguistic variables Trapezoidal interval type-2 fuzzy scales
Equally important (E) (1, 1, 1, 1; 1, 1) (1, 1, 1, 1; 1, 1)
Weakly important (WI) (1, 2, 4, 5; 1, 1) (1.2, 2.2, 3.8, 4.8; 0.8, 0.8)
Strongly important (S) (3, 4, 6, 7; 1, 1) (3.2, 4.2, 5.8, 6.8; 0.8, 0.8)
Very strongly important (VS) (5, 6, 8, 9; 1, 1) (5.2, 6.2, 7.8, 8.8; 0.8, 0.8)
Absolutely important (AS) (7, 8, 9, 9; 1, 1) (7.2, 8.2, 8.8, 9; 0.8, 0.8)
~
~
As a result, a sample comparison matrix that integrates interval type-2 fuzzy sets (AÞ
is formed as given in the following;
2 ~~a12 ~1n 3
1 ... ~
a
6 1
~~ ¼ 6 ~~a12 1 ... ~
a2n 7
~ 7
A 4... ... ð1Þ
... ... 5
1
~~a1n
1
~~a2n ... 1
where

~~a ¼ ððaU ; aU ; aU ; aU ; H1 ðau Þ; H2 ðau ÞÞ ððaL ; aL ; aL ; aL ; H1 aL ; H2 aL Þ
11 12 13 i4 12 13 21 22 23 24 22 23
and
1 1 1 1 1 1 1 1
1=~~a ¼ ðð U ; U ; U ; U ; H1 ða12 Þ; H2 ða13 ÞÞ; ððaL ; aL ; aL ; aL ; H1 ða22 Þ; H2 ða23 ÞÞ
u u L L
a14 a13 a12 a11 24 23 22 21
Step 3: Consistency is checked fuzzy pairwise comparison matrices. After that, the
comparison matrix is defuzzified and it is examined for consistency.
Step 4: Aggregating the expert evaluations using geometric mean.
h i1=n
~~aij ¼ ~~a1 . . . ~
~
a n
ð2Þ
ij ij
where
qffiffiffiffiffi qffiffiffiffiffiffiffi qffiffiffiffiffiffiffi qffiffiffiffiffiffiffi qffiffiffiffiffiffiffi qffiffiffiffiffiffiffi qffiffiffiffiffiffiffi qffiffiffiffiffiffiffi qffiffiffiffiffiffiffi
n
~~aij ¼ n aU ; n aU ; n aU ; n aU ; H U aij ; H U aij ; n aL ; n aL ; n aL ; n aL ; H L aij ; H L aij
ij1 ij2 ij3 ij4 1 2 ij1 ij2 ij3 ij4 1 2
560 S. Alem
Step 5: Calculating the fuzzy weights for each criterion in a comparison matrix. The
geometric mean of each row (~~r i ) is calculated.
Then fuzzy weight of the ith criterion (~~pi ) is calculated as using following equation.

~~pi ¼ ~~r i ½~~r 1 . . . ~~r i . . . ~
~r n 1 ð3Þ
The fuzzy weight calculated as a result of a pairwise comparison matrix is called the
local weights, in order to find the global weights of each sub-criteria, the local eights
should be multiplied by the local weight of the upper level criteria.
Step 6: Fuzzy weights are defuzzified. The DTtrT method [14] is used for defuzzifi-
cation in this step.
ðuU lU Þ þ ðbU :m1U lU Þ þ ðaU :m2U lU Þ

þ lU þ ½ðuL lL Þ þ ðbL :m1L 4 lL Þ þ ðaL :m2L lL Þ
DTtrT ¼ 4
2
ð4Þ
2.2 Fuzzy Hesitant TOPSIS

Huang and Yoon developed the TOPSIS (technique for order preference by similarity
to an ideal solution) [15] as an MCDM which evaluates alternatives according to their
distances to the optimal solution.
Step 1: The positive and negative ideal solutions are determined.
A ¼ f h1 ; h2 ; . . .: h
ng ð5Þ
where
[m
hj ¼ i¼1 hij ¼ Uy1j 2h1j ;...::;ymj 2hmj ; max y1j ; . . .:ymj j ¼ 1; 2; . . .; n
A ¼ f h
1 ; h2 ; . . .: hn g;
where
\m
hj ¼ i¼1
hij ¼ \ y1j 2h1j ;...::;ymj 2hmj ; max y1j ; . . .:ymj j ¼ 1; 2; . . .; n
Step 2: Separation measures of each alternative from the ideal solution is calculated.
The distance of an alternative form positive ideal is calculated as follows:

Diþ ¼ Rnj¼1 wj hij hjþ ð6Þ
where wj represents the weight of the j th criterion determined by IT2FAHP.

And, the distance of an alternative form negative ideal is given as

D i ¼ Rj¼1 wj hij hj
n
ð7Þ
Following equation is used to calculate distance between two hesitant fuzzy numbers;
kh1 h1 k ¼ Rlj¼1
h1rð jÞ h2rð jÞ
ð8Þ
l
Step 3: Relative closeness to the ideal solution is calculated using the following
equation:
D
Ci ¼ i
þ ð9Þ
D
i þ Di
Step 4: The alternatives are ranked according relative closeness; the alternative with the
highest value is selected as the best alternative.
3 Application
Determination of the Criteria Weights

Weights of the decision criteria and sub-criteria are determined as using interval type-2
AHP and criteria are determined by using experts’ experiences. Experts’ linguistic
evaluations formed first pairwise comparison matrices. Then consistency check is
completed for all matrices and all matrices found to be consistent. Experts’ linguistic
evaluations for main criteria are shown in Table 2.
Table 2. Expert Evaluations of main criteria with respect to the goal

C1 C2 C3
E1 E2 E3 E4 E5 E6 E1 E2 E3 E4 E5 E6 E1 E2 E3 E4 E5 E6
C1 E E E E E E AS AS VS VS S WI VS VS S S AS S
1
C2 AS 1 1 1 1 1 E E E E E E S WI S S WI E
AS VS VS S WI
C3 1 1 1 1 1 1 1 1 1 1 1 E E E E E E E
VS VS S S AS S S WI S S WI
Pairwise comparisons between C1 and C2 are shown in following. For the calcu-
lation of the aggregated comparison value of six experts, first the linguistic evaluations
(1/AS, 1/AS, 1/VS, 1/VS, 1/S, 1/WI) are transformed to related interval type-2 fuzzy
sets. (0.11, 0.11, 0.13, 0.14; 1, 1), (0.11, 0.11, 0.12, 0.14; 0.8, 0.8); (0.11, 0.11, 0.13,
0.14; 1, 1) (0.11, 0.11, 0.12, 0.14; 0.8, 0.8); (0.11, 0.13, 0.17, 0.2; 1, 1), (0.11, 0.13,
0.16, 0.19; 0.8, 0.8); (0.11, 0.13, 0.17, 0.2; 1, 1), (0.11, 0.13, 0.16, 0.19; 0.8, 0.8);
(0.14, 0.17, 0.25, 0.33; 1, 1), (0.15, 0.17, 0.24, 0.31; 0.8, 0.8); (0.2, 0.25, 0.5, 1; 1, 1),
(0.21, 0.26, 0.45, 0.83; 0.8, 0.8)
562 S. Alem
Then aggregation is completed.
~~a12 ¼½ð0:11; 0:11; 0:13; 0:14; 1; 1Þ; ð0:11; 0:11; 0:12; 0:14; 0:8; 0:8Þ
ð0:11; 0:11; 0:13; 0:14; 1; 1Þð0:11; 0:11; 0:12; 0:14; 0:8; 0:8Þ
ð0:11; 0:13; 0:17; 0:2; 1; 1Þ; ð0:11; 0:13; 0:16; 0:19; 0:8; 0:8Þ
ð0:11; 0:13; 0:17; 0:2; 1; 1Þ; ð0:11; 0:13; 0:16; 0:19; 0:8; 0:8Þ
ð0:14; 0:17; 0:25; 0:33; 1; 1Þ; ð0:15; 0:17; 0:24; 0:31; 0:8; 0:8Þ
ð0:2; 0:25; 0:5; 1; 1; 1Þ; ð0:21; 0:26; 0:45; 0:83; 0:8; 0:8Þ1=4
¼ ð0:13; 0:14; 0:2; 0:25; 1; 1Þð0:13; 0:14; 0:18; 0:24; 0:8; 0:8Þ
The aggregation of expert’s evaluations is shown in Table 3.
Table 3. Aggregated expert evaluations of criteria with respect to goal

C1 C2 C3
C1 (1, 1, 1, 1; 1, 1), (3.93, 5.14, 7.07, 7.83; 1, 1), (4.1, 5.14, 7.07, 7.94; 1, 1),
(1, 1, 1, 1; 1, 1) (4.19, 5.39, 6.86, 7.68; 0.8, 0.8) (4.31, 5.41, 6.86, 7.76; 0.8, 0.8)
C2 (0.13, 0.14, 0.2, 0.25; 1, 1), (1, 1, 1, 1; 1, 1), (1.73, 2.52, 3.89, 4.52; 1, 1),
(0.13, 0.14, 0.18, 0.24; 0.8, 0.8) (1, 1, 1, 1; 1, 1) (1.9, 2.7, 3.76, 4.4; 0.8, 0.8)
C3 (0.12, 0.14, 0.2, 0.24; 1, 1), (0.22, 0.26, 0.4, 0.57; 1, 1), (1, 1, 1, 1; 1, 1),
(0.13, 0.14, 0.19, 0.23; 0.8, 0.8) (0.23, 0.26, 0.38, 0.52; 0.8, 0.8) (1, 1, 1, 1; 1, 1)
Geometric means of reach row are calculated as using Table 3. First row is cal-
culated as follow (Table 4):
Table 4. Geometric means of each row

Fuzzy geometric means
~~r1 (2.53, 2.98, 3.68, 3.96; 1, 1), (2.62, 3.08, 3.61, 3.91; 0.8, 0.8)
~~r2 (0.61, 0.71, 0.92, 1.04; 1, 1), (0.63, 0.72, 0.88, 1.02; 0.8, 0.8)
~~r3 (0.3, 0.33, 0.43, 0.52; 1, 1), (0.31, 0.33, 0.42, 0.49; 0.8, 0.8)
~~r1 ¼ ½ð1; 1; 1; 1; 1; 1Þ; ð1; 1; 1; 1; 1; 1Þ

ð3:93; 5:14; 7:07; 7:83; 1; 1Þ; ð4:19; 5:39; 6:86; 7:68; 0:8; 0:8Þ
ð4:1; 5:14; 7:07; 7:94; 1; 1Þ; ð4:31; 5:41; 6:86; 7:76; 0:8; 0:8Þ1=3
¼ ð2:53; 2:98; 3:68; 3:96; 1; 1Þð2:62; 3:08; 3:61; 3:91; 0:8; 0:8Þ
Next the weight of each criteria is determined. For the C1 the operations are given
in the following:
~~r1 ¼ ð2:53; 2:98; 3:68; 3:96; 1; 1Þð2:62; 3:08; 3:61; 3:91; 0:8; 0:8Þ
½ð2:53; 2:98; 3:68; 3:96; 1; 1Þð2:62; 3:08; 3:61; 3:91; 0:8; 0:8Þ
ð0:61; 0:71; 0:92; 1:04; 1; 1Þð0:63; 0:72; 0:88; 1:02; 0:8; 0:8Þ
ð0:3; 0:33; 0:43; 0:52; 1; 1Þð0:31; 0:33; 0:42; 0:49; 0:8; 0:8Þ1
¼ ð2:53; 2:98; 3:68; 3:96; 1; 1Þð2:62; 3:08; 3:61; 3:91; 0:8; 0:8Þ
½ð3:44; 4:02; 5:03; 5:52Þ; ð3:56; 4:13; 4:9; 5:42Þ1
¼ ð2:53; 2:98; 3:68; 3:96; 1; 1Þð2:62; 3:08; 3:61; 3:91; 0:8; 0:8Þ
ð0:18; 0:2; 0:25; 0:29Þ; ð0:18; 0:2; 0:24; 0:28Þ
¼ ð0:46; 0:59; 0:92; 1:15; 1; 1Þð0:48; 0:63; 0:87; 1:1; 0:8; 0:8Þ
And then weights are defuzzified as using DTtrT method (Table 5):
Table 5. Fuzzy and Normalized weights of the criteria

Criteria Fuzzy weights Crisp Normalized crisp
weights weights
C1 (0.46, 0.59, 0.92, 1.15; 1, 1) (0.48, 0.63, 0.87, 0,74 0,73
1.1; 0.8, 0.8)
C2 (0.11, 0.14, 0.23, 0.3; 1, 1) (0.12, 0.15, 0.21, 0,18 0,18
0.29; 0.8, 0.8)
C3 (0.05, 0.07, 0.11, 0.15; 1, 1) (0.06, 0.07, 0.1, 0,09 0,09
0.14; 0.8, 0.8)

1 ð1; 15 0; 46Þ þ ð1 0; 59 0; 46Þ þ ð1 0; 92 0; 46Þ
w1 ¼ þ 0; 46
2 4

ð1; 1 0; 48Þ þ ð0; 8 0; 63 0; 48Þ þ ð0; 8 0; 87 0; 48Þ
þ þ 0; 48 ¼ 0; 74
4
All sub-criteria are calculated as using same steps on above and result are shown as
below (Table 6);
564 S. Alem
Table 6. Weights of sub-criteria

Criteria Defuzzified Normalized Criteria Defuzzified Normalized
weights weights weights weights
C11 0,28 0,2 C23 0,13 0,02
C12 0,3 0,22 C24 0,4 0,07
C13 0,06 0,04 C25 0,09 0,02
C14 0,14 0,1 C31 0,61 0,05
C15 0,23 0,17 C32 0,14 0,02
C21 0,16 0,03 C33 0,13 0,01
C22 0,21 0,04 C34 0,12 0,01
Production Volume and Organic Food Perception criteria are determined as the
lowest importance (0,01) and Transportation Cost criteria is determined as the highest
importance (0,22) criteria. These importance weights are used for Hesitant Fuzzy
TOPSIS in following section.
Experts’ judgments are collected as hesitant fuzzy numbers then they are aggre-
gated as below (Table 7);
Table 7. Hesitant fuzzy decision matrix

C11 C12 C13 C14 C15 C21 C22
TD {0.14, 0.24, 0.35} {0.24, 0.42, 0.61} {0.2, 0.3, 0.45} {0.22, 0.36, 0.55} {0.18, 0.3, 0.42} {0.19, 0.31, 0.44} {0.21, 0.32, 0.47}
AIF {0.2, 0.3, 0.4} {0.18, 0.28, 0.41} {0.31, 0.55, 0.75} {0.25, 0.39, 0.51} {0.33, 0.62, 0.82} {0.23, 0.38, 0.54} {0.31, 0.55, 0.75}
VF {0.54, 0.72, 0.9} {0.25, 0.39, 0.53} {0.33, 0.62, 0.82} {0.22, 0.4, 0.56} {0.34, 0.51, 0.71} {0.36, 0.58, 0.78} {0.26, 0.48, 0.66}
PBM {0.58, 0.73, 0.87} {0.2, 0.34, 0.48} {0.49, 0.7, 0.87} {0.21, 0.32, 0.44} {0.28, 0.43, 0.58} {0.31, 0.5, 0.69} {0.31, 0.5, 0.71}
C23 C24 C25 C31 C32 C33 C34
TD {0.22, 0.36, 0.53} {0.13, 0.23, 0.33} {0.13, 0.23, 0.33} {0.46, 0.68, 0.87} {0.43, 0.61, 0.8} {0.39, 0.59, 0.78} {0.58, 0.73, 0.88}
AIF {0.24, 0.42, 0.61} {0.25, 0.39, 0.6} {0.21, 0.34, 0.52} {0.39, 0.59, 0.77} {0.42, 0.66, 0.85} {0.39, 0.65, 0.85} {0.46, 0.62, 0.8}
VF {0.2, 0.35, 0.52} {0.68, 0.77, 0.9} {0.54, 0.71, 0.88} {0.27, 0.44, 0.64} {0.23, 0.35, 0.48} {0.24, 0.42, 0.59} {0.23, 0.38, 0.52}
PBM {0.27, 0.4, 0.56} {0.54, 0.71, 0.88} {0.5, 0.7, 0.88} {0.11, 0.21, 0.31} {0.16, 0.28, 0.43} {0.22, 0.36, 0.51} {0.11, 0.21, 0.33}
Next separation measure of each alternative from ideal solution is calculated using
Eq. 6 and Eq. 7. Finally, the relative closeness values are calculated using Eq. 9 for
each alternative (Table 8).
Table 8. Separation values and relative closeness of each alternative

Alternatives D+ D− Ci Rank
TF 0,42 0,18 0,31 4
AIF 0,35 0,25 0,41 3
VF 0,24 0,36 0,60 1
PBM 0,29 0,31 0,52 2
4 Conclusion and Further Research
This study aims at helping entrepreneurs to select most suitable alternative such as;
traditional farming, artificial aided farming, vertical farming and plant-based meat.
These alternatives are so fresh and there is no exact information or market research
about them which cause the vagueness on different decision makers’ idea. Criteria’s
weights are calculated by interval type-2 fuzzy AHP after that alternatives are evaluated
by using hesitant fuzzy TOPSIS.
In this study, Vertical farming is found to be the best alternative which has cost
advantage especially on sub-criteria “transportation cost”. Cost has a weight of 0.73
and it has huge effect on selecting most suitable alternative. It is known that entre-
preneurs are generally focused on cost therefore result are actualized as expected.
For further research, AHP can be extended to work with 3D fuzzy sets. Also,
MCDM methods could be changed from TOPSIS to SMART or ELECTRE or VIKOR.
The obtained results from this new integration can be compared by our results.
References
1. Saaty, T.L.: Fundamentals of the analytic hierarchy process. Mati. Modell. 161–176 (1987)
2. Zadeh, L.A.: Fuzzy sets. Inf. Control. 8, 338–353 (1964)
3. Keller’, J.M.: Multiple Criteria decision making (MCDM) using fuzzy logic: an innovative
approach to sustainable agriculture. In: NAFIPS, pp. 503–508 (1995)
4. Qureshi, M.R.N.: Decision support model to select crop pattern for sustainable agricultural
practices using fuzzy MCDM. Environ. Dev. Sustain. 20, 641–659 (2018)
5. Djatna, T.: GIS-based multi-criteria decision-making model for site selection of seaweed
farming information centre: a lesson from small Islands. Indonesia. Decis. Sci. Lett. 8,
137–150 (2019)
6. Wu, T.: An interval type-2 fuzzy ANP approach to evaluate enterprise technological
innovation ability. Kybernetes 45, 1486–1500 (2016)
7. Gao, S.: An evaluation of environmental governance in urban china based on a hesitant
fuzzy linguistic analytic network process. MDPI (2018)
8. Mendel, J.M., John, R.I.: Type-2 fuzzy sets. Trans. Fuzzy Syst. 10, 808–821 (2002)
9. Buckley, J.J.: Fuzzy hierarchical analysis. Fuzzy Sets Syst. 17, 233–247 (1985)
10. Chen, S.M., Lee, L.W.: Fuzzy multiple attributes group decision-making based on the
interval type-2 topsis method. Expert Syst. Appl. 37, 2790–2798 (2010)
11. Torra, V.: Hesitant fuzzy sets. Int. J. Intell. Syst. 25, 529–539 (2010)
566 S. Alem
12. Xia, M.M., Xu, Z.S.: Hesitant fuzzy information aggregation in decision making. Int.
J. Approx. Reason. 52, 395–407 (2011)
13. Senturk, S., Erginel, N., Binici, Y.: Interval type-2 fuzzy analytic network process for
modelling a third-party logistics (3PL) company. J. Mult. Valued Log. Soft Comput. 28,
311–333 (2016)
14. Kahraman, C., Sari, I.U., Turanoglu, E.: Fuzzy analytic hierarchy process with type-2 fuzzy
sets. In: Proceedings of the 19th International Flins Conference (2012)
15. Huang, C., Yoon, K.: Multiple Attribute Decision Making Methods and Applications.
16. Kahraman, C., Cevik Onar, S., Oztaysi, B.: Strategic decision selection using hesitant fuzzy
topsis and interval type-2 fuzzy AHP: a case study. Int. J. Comput. Intell. Syst. 7, 1002–1021
(2014)
Investigating Enablers to Improve
Transparency in Sustainable Food Supply
Chain Using F-BWM
Anish Kumar1 , Sachin Kumar Mangla2 , Pradeep Kumar1 ,

and Yasanur Kayikci3(&)
1
Department of Mechanical and Industrial Engineering,
Indian Institute of Technology, Roorkee 247667, Uttrakhand (UK), India
2
Knowledge Management and Business Decision Making,
Plymouth Business School, University of Plymouth, Plymouth PL48AA, UK
3
Turkish-German University, 34820 Istanbul, Turkey
[email protected]
Abstract. Food Supply Chains (FSC) are complex and dynamic in behavior
and prone to increasing risks of unsustainability. Consumers increasingly
demand food quality, safety, and sustainability, which are fast becoming issues
of great importance in FSC. Lack of real-time information sharing and con-
nectivity among stakeholders make these issues tougher to mitigate. Supply
chain transparency (SCT) is thus an essential attribute to manage these supply
chain complexities and enhance the sustainability of FSC. The paper identifies
and analyses key enablers for SCT in FSC. Several technical, as well as
sustainability-related enablers, contribute to the implementation of SCT. The
identified enablers are analyzed using Fuzzy-best worst methodology (F-BWM),
which determine the most critical factors using the decision maker’s opinion.
Extending BWM with fuzzy logic incorporates the vagueness of human-
behaviour into decision making approach. The results of this research provides
decision makers with the priority of enablers to the decision maker. Enhancing
these enablers in will help improve the transparency for better management of
FSC. The article expands upon the practical as well as theoretical implications of
SCT on sustainability in FSC. It addresses the requirement of including sus-
tainability in the decision-making process. The results demonstrate the effec-
tiveness of the F-BWM for the decision making process. The study is conducted
by considering downstream supply chain activities in Indian context. It is one of
the first studies that analyzes SCT enablers using F-BWM method in Indian
context. The study contributes towards improving the environmental, econom-
ical, and social sustainability of FSC.
Keywords: Supply chain transparency Fuzzy-best worst method

Sustainable food supply chain
568 A. Kumar et al.
1 Introduction
Food supply chain (FSC) in the 21st century is long and complex, with large number of
operations, covering large distances. A multitude of complex challenges such as-
product perishability, quality issues, poor demand management, poor traceability, and
supply chain (SC) visibility, logistics and cold chain performance etc., that make them
tough to manage [1]. All these issues in FSCs lead to a huge amount of wastage and
losses in the FSCs. Almost one-third of the total food produced is wasted/lost globally
[2]. Apart from these challenges, food quality, safety, and security are the essential
factors for a sustainable food supply chain (SFSC) [3, 4]. Various food scares such as
the melamine milk adulteration scandal in China, horsemeat scandal in Ireland, avian
influenza outbreak have added to the concerns of customers regarding food quality and
safety [5]. Managing these challenges call for transparency, monitoring, and trace-
ability, in FSC.
While most food companies express their willingness to improve transparency and
traceability for an SFSC, the shift in the Indian food industry has been rather slow. One
major reason for this could be that the motivators and enablers for improving trans-
parency in an SFSC are not fully understood. This serves as a research gap for us to
investigate the enablers for transparency in SFSC using advanced multi-criteria deci-
sion making (MCDM) tools.
A varied range of MCDM tools can be found in literature such as AHP [6], ANP
[7], VIKOR [8], TOPSIS [9] etc. Among these AHP is the most popular one, with ANP
it’s general form, which are both based on pair-wise comparison matrix formulation.
However, AHP has also been criticized for inconsistency in decision making and the
requirement of a large number of input comparisons. Owing to these issues, Rezaei
proposes the Best-Worst Method (BWM) for ranking factors, that essentially gave
better results than AHP [10]. The BWM uses reference comparisons with respect to the
best and worst criterion and essentially remedies the inconsistency issue. Further, to
integrate the uncertainty and the vagueness in the decision-makers perspective, we use
the fuzzy-BWM to analyze the enablers to improve transparency in SFSC. We proceed
with the following research objectives-
• Identify the enablers for improving transparency in sustainable food supply chain.
• Analyze the identified preference order of enablers using Fuzzy-Best Worst Method
multi-criteria decision making technique
The article is organized in 5 sections where after the introduction in the first section,
the following section presents a brief review of the literature related to the enablers of
transparency in FSCs. Next, Sect. 3 presents the fuzzy-BWM based research
methodology adopted for this research. Section 4 presents the results the research and
their discussion, with Sect. 5 concluding the article.
Investigating Enablers to Improve Transparency 569
2 Literature Review
Transparency is an essential characteristic of efficiency in SFSC. It may be defined as

the timely sharing of accurate, fact-based, reliable, relevant, and readable information,
of appropriate quality and quantity, among all the stakeholders [11]. It is essentially an
SC performance metric, a lack of which can lead to problems related to the business,
environment, and social sustainability [12]. Ringsberg conducting a review of the
literature on various perspectives on traceability in FSCs, found transparency and
interoperability as an important perspective for risk management in FSCs [13]. Wog-
num et al. studying the use of information systems for sustainability in FSCs, identified
that most of the systems were focussed on single actors and did not consider the whole
FSC [11]. Bastian and Zentes studied the antecedents and consequences of SC trans-
parency of sustainability in FSC, identifying significant effects on social, ecological,
operational, and relational dimensions of sustainability [14]. The majority of the studies
on transparency considered traceability as the focal concept affecting SFSC [12, 15].
A lot of support can be found in the literature for the technology based solutions such
as ICT, IoT, and blockchain [5, 16].
2.1 Enablers for Transparency

Addressing consumer concerns are central to sustainable development in FSC [17].
Gaining trust and confidence from the customer is one of the essential factors in the
food industry [18]. Consumers show a great interest in accessing product information
such as point of origin, date of manufacturing, expiry date, etc. [5]. A 2016 study finds
that 94% of respondents found transparency an important characteristic of food man-
ufacturers, with 84% finding in-depth information of the product of great value [19].
By enabling traceability and with a seamless flow if information, the governments
play a crucial role of avoiding the unforeseeable disasters [20]. Regulations through
legislations, drive implementation of transparency and traceability in FSC [3]. Wu et al.
identify a lack of regulatory supervision as a contributing factor of the 2008 melamine
scandal [21]. Health hazards make transparency and traceability the cornerstone of food
safety. Further, it protects FSC from various malpractices such as, use of hazardous
chemicals and additives, food frauds, processing malpractices, etc. Adulteration or
“food counterfeit” of food products or supply of inferior quality by suppliers is a
common complaint in FSC [22]. Counterfeits are generally cases of conscious fraud,
and are coupled with genuine scientific knowledge [23]. This makes the guilty espe-
cially tough to pinpoint without a strong transparency-based system. Thus, the detec-
tion of counterfeits is a major enabler for supply chain transparency.
Transparency enables auditing of sustainability practices of all the stakeholders in
the SC [24]. It ensures that operations are free from negative consequences such as
child labor, waste mismanagement, poor factory conditions, etc. [12]. In such a system,
governments and NGOs can monitor all the parties involved in the SC in real.
Therefore, preventing any social as well as environmental malpractice. From an
operational perspective, transparency and traceability help reap many benefits such as
quick and easy recalls, swift pin-pointing of the actor-at-fault in case of a product
performance issue, improved cooperation among organizations [11]. Significant market
570 A. Kumar et al.
competition in the globalized world makes it necessary to invest in transparency and

coordination based technologies [3].
Thus, a SC transparency-based system with information sharing about the sus-
tainability of participants across multiple tiers of SC is essential. Based on an extant
literature review, as shown in Table 1, eight enablers of SC transparency are identified,
which are next analysed using Fuzzy-BWM.
Table 1. Enablers for improving transparency in SFSC

S No. Enablers Source
1. Enhance operational efficiency [3, 11]
2. Mitigate food hazard and enable safety [22, 25]
3. Fulfil regulatory requirements [20, 21]
4. Promote sustainability in Supplier practices [12, 24]
5. Expanding e-retail business [2]
6. Quick and easy recall [3]
7. Product counterfeit prevention [2, 23]
8. Gain customer trust and confidence [5, 17]
The BWM, as proposed by Rezaei [10], provides a priority vector based on two
reference comparison vectors. The extension of BWM using fuzzy set theory, previ-
ously used in literature [26, 27], has been found give preference with better
consistency.
A triangular fuzzy number (TFN) A on the set R, with its membership function
lA ð xÞ : R ! ½0; 1, is defined as [6]
8 xa 9
< ba ; x 2 ½a; b =
lA ð xÞ ¼ bc
xc
; x 2 ½b; c ð1Þ
: ;
0; otherwise
Where, a b c, a is the lower, b the modal, and c the upper boundary of the TFN.
We use the Graded-Mean-Integration representation (GMIR), R(A j) to rank the TFN,
and de-fuzzify the solutions.
ai þ 4bi þ ci
j ¼
R A ð2Þ
6
The steps for the Fuzzy-BWM are as follows:
Step 1: Identification of decision factors. On the basis of literature, finalize a set of
factors, F = {f1, f2….fn}.
Step 2: Identification of best and worst factors based on discussion with experts.
Step 3: Provide the fuzzy preference comparisons of “Best over other” and “others
over Worst”.
Using the linguistic comparisons, shown in Table 2, conduct pairwise comparisons
to get the “Best over others” Vector, labelled XB, and “Others over Worst” vector
labelled XW.
XB ¼ ðxB1 ; xB2 ; . . .. . .. . .:xBn Þ; and XW ¼ ðx1W ; x2W ; . . .. . .. . .:xnW Þ
Table 2. Linguistic terms for reference comparison

Linguistic terms Fuzzy numbers
Absolutely more significant (7/2, 4, 9/2)
Strongly more significant (5/2, 3, 7/2)
Fairly more significant (3/2, 2, 5/2)
Weakly more significant (2/3, 1, 3/2)
Equally significant (1, 1, 1)
Step 4: Compute the fuzzy weights of the factors.

The fuzzy weights of the factors (Wj) are computed using the BWM programming
formulation proposed by [26]. The optimal weights for the factors are such that
W B
WB
¼ xBj and W j ¼ xjW . To achieve this, we try to minimize the gaps W xBj and

Wj
W Wj
Wj
W xjW , by solving the constrained optimization problem shown in Eq. (3).
W
min n
such that

WB Wj
x n; x
jW n
W Bj W
j W ð3Þ
Xn
R(Wj Þ ¼ 1
j¼1
aW
j 0; aW
j bW
j cW
j ; j ¼ 1; 2;. . .; n
where WB , Wj , WW , are fuzzy numbers. The fuzzy weights of the factor fj is given by
Wj = ðaW
j ; bj ; cj Þ, and n = (k, k, k)
W W
Step 5: Check the consistency ratio of the solution.

572 A. Kumar et al.
Table 3. Consistency index for fuzzy BWM

Linguistic Abs. more Strongly more Fairly more Weakly more Equally
terms Sig. Sig. Sig. Sig. Sig.
aBW (7/2, 4, 9/2) (5/2, 3, 7/2) (3/2, 2, 5/2) (2/3, 1, 3/2) (1, 1, 1)
Cons. index 8.04 6.69 5.29 3.80 3
k
Consistency ratio = Consistency Index, with consistency index given in Table 3.
which are the solutions of the Eq. (4).
n2 ð1 + 2xBW Þn + ðx2BW xBW Þ = 0; ð4Þ
Which can be transformed to Eq. (5)
n2 ð1 þ 2cBW Þn þ ðc2BW cBW Þ ¼ 0 ð5Þ
The eight identified enablers were discussed with the industry experts. Based on the
discussion, the following reference comparison inputs were provided.
XB ¼ ð7=2; 4; 9=2Þ; ð0:67; 1; 1:5Þ; ð1; 1; 1Þ; ð2=3; 1; 3=2Þ; ð3=2; 2; 5=2Þ; ð3=2; 2; 5=2Þ;
ð2=3; 1; 3=2Þ; ð2=3; 1; 3=2Þ; and XW ¼ ð1; 1; 1Þ; ð5=2; 3; 7=2Þ; ð7=2; 4; 9=2Þ; ð3=2; 2; 5=2Þ;
ð2=3; 1; 3=2Þ; ð3=2; 2; 5=2Þ; ð3=2; 2; 5=2Þ; ð2=3; 1; 3=2Þ:
The formulation in Eq. 1 was coded in LINGO 18.0 and solved, we get an
n = 0.7912, with the preference as shown in Table 4.
Table 4. Preference ranking of the enablers to transparency in SFSC

Enabler Fuzzy weight Crisp wight Preference
Enhance operational efficiency (.064, .064, .076) 0.066 8th
Mitigate food hazard and enable safety (.143, .173, .173) 0.168 2th
Fulfil regulatory requirements (.206, .206, .237) 0.211 1st
Promote sustainability in supplier practices (.114, .124, .140) 0.125 3nd
Expanding e-retail business (.083, .085, .096) 0.086 7th
Quick and easy recall (.086, .098, .116) 0.099 6th
Product counterfeit prevention (.114, .124, .140) 0.125 4rd
Gain customer trust and confidence (.106, .115, .143) 0.118 5th
Consistency Ratio = n / CI = 0:7912=8:04 = 0:098, is very good. Thus, regu-

latory requirements as the most significant enabler of improving transparency in FSCs,
followed by food hazards, supplier malpractices, and counterfeit prevention, and cus-
tomer trust. Operations focused enablers such as quick and easy recall, expanding
e-retail, and operational efficiency are the last, which is in concurrence with the pre-
vious literature [2, 14]. We infer that raising issues related to food safety and security
among customers can motivate businesses to be more transparent regarding sustain-
ability in their FSCs.
5 Conclusion
The concept of sustainability is fast gaining acceptance in FSCs across the globe as
well as in India. From this perspective, it is imperative that we businesses have a clear
view of the key factors that are of importance in SFSC. The current paper serves this
purpose by identifying key enablers from a social, environmental as well as economical
perspective, to improve transparency in SFSC. The identified enablers are analyzed
using the fuzzy-BWM, which still rather new and has been found to be of great use in
these decision-making problems.
The research in this direction can be further extended by validating the results
empirically while considering more factors and wider industrial perspective. Further,
simulation and game theory related formulations can also be used to test the results of
this research. From the methodological perspective, the results can also be compared
with the results of other methodologies such as VIKOR, ELECTRE, etc.
References
1. Kumar, A., Mangla, S.K., Kumar, P., Karamperidis, S.: Challenges in perishable food supply
chains for sustainability management: a developing economy perspective (2020). https://doi.
org/10.1002/bse.2470
2. Confederation of Indian Industry: Food Traceability in India A study report by CII FACE &
GS1 India (2017)
3. Shankar, R., Gupta, R., Pathak, D.K.: Modeling critical success factors of traceability for
food logistics system. Transp. Res. 119, 205–222 (2018). https://doi.org/10.1016/j.tre.2018.
03.006
4. Wang, J., Yue, H.: Food safety pre-warning system based on data mining for a sustainable
food supply chain. Food Control 73, 223–229 (2017). https://doi.org/10.1016/j.foodcont.
2016.09.048
5. Astill, J., Dara, R.A., Campbell, M., Farber, J.M., Fraser, E.D.G., Sharif, S., Yada, R.Y.:
Transparency in food supply chains: a review of enabling technology solutions. Trends Food
Sci. Technol. 91, 240–247 (2019). https://doi.org/10.1016/j.tifs.2019.07.024
6. Kumar, S., Kumar, P., Kumar, M.: Resources, conservation and recycling risk analysis in
green supply chain using fuzzy AHP approach: a case study. Resour. Conserv. Recycl. 104,
375–390 (2015)
574 A. Kumar et al.
7. Saaty, T.L., Vargas, L.G.: Decision making with the analytic network process economic,
political, social and technological applications with benefits, opportunities, costs and risks
(2013). https://doi.org/10.1007/978-1-4614-7279-7
8. Opricovic, S., Tzeng, G.H.: Compromise solution by MCDM methods: a comparative
analysis of VIKOR and TOPSIS. Eur. J. Oper. Res. 156, 445–455 (2004). https://doi.org/10.
1016/S0377-2217(03)00020-1
9. Tyagi, M., Kumar, P., Kumar, D.: Assessment of CSR based supply chain performance
system using an integrated fuzzy AHP-TOPSIS approach. Int. J. Logist. Res. Appl. 21, 378–
406 (2018). https://doi.org/10.1080/13675567.2017.1422707
10. Rezaei, J.: Best-worst multi-criteria decision-making method. Omega (United Kingdom). 53,
49–57 (2015). https://doi.org/10.1016/j.omega.2014.11.009
11. Wognum, P.M., Bremmers, H., Trienekens, J.H., Van Der Vorst, J.G.A.J., Bloemhof, J.M.:
Systems for sustainability and transparency of food supply chains - current status and
challenges. Adv. Eng. Inf. 25, 65–76 (2011). https://doi.org/10.1016/j.aei.2010.06.001
12. Bai, C., Sarkis, J.: A supply chain transparency and sustainability technology appraisal
model for blockchain technology. Int. J. Prod. Res. 1–21 (2020). https://doi.org/10.1080/
00207543.2019.1708989
13. Ringsberg, H.: Perspectives on food traceability: a systematic literature review. Supply
Chain Manag. Int. J. 19, 558–576 (2014). https://doi.org/10.1108/SCM-01-2014-0026
14. Bastian, J., Zentes, J.: Supply chain transparency as a key prerequisite for sustainable agri-
food supply chain management (2013). http://dx.doi.org/10.1080/09593969.2013.834836.
https://doi.org/10.1080/09593969.2013.834836
15. Bosona, T., Gebresenbet, G.: Food traceability as an integral part of logistics management in
food and agricultural supply chain. Food Control 33, 32–48 (2013). https://doi.org/10.1016/
j.foodcont.2013.02.004
16. Zhao, G., Liu, S., Lopez, C., Lu, H., Elgueta, S., Chen, H., Boshkoska, B.M.: Blockchain
technology in agri-food value chain management: a synthesis of applications, challenges and
future research directions. Comput. Ind. 109, 83–99 (2019). https://doi.org/10.1016/j.
compind.2019.04.002
17. Cagliano, R., Worley, C.G., Caniato, F.F.A.: The challenge of sustainable innovation in agri-
food supply chains. Organ. Sustain. Eff. 5, 1–30 (2016). https://doi.org/10.1108/S2045-
060520160000005009
18. Bilyea, T., Mcinnes, D.: With natural capital and trust, Canada can become an agrifood
powerhouse (2016). https://www.theglobeandmail.com/report-on-business/rob-commentary/
with-capital-and-trust-canada-can-become-an-agrifood-powerhouse/article30989002/
19. Insight, L.: How consumer demand for transparency is shaping the food industry (2016)
20. Faisal, M.N.: A study of inhibitors to transparency in red meat supply chains in Gulf
cooperation council (GCC) countries. Bus. Process Manag. J. 21, 1299–1318 (2015). https://
doi.org/10.1108/BPMJ-11-2014-0107
21. Wu, X., Lu, Y., Xu, H., Lv, M., Hu, D., He, Z., Liu, L., Wang, Z., Feng, Y.: Challenges to
improve the safety of dairy products in China (2018). https://doi.org/10.1016/j.tifs.2018.03.019
22. Gupta, R., Shankar, R.: Ranking of collusive behaviour in Indian agro-supply chain using
interval 2-tuple linguistic TOPSIS method. J. Model. Manag. 11, 949–966 (2016). https://
doi.org/10.1108/JM2-03-2015-0006
23. Creydt, M., Fischer, M.: Blockchain and more - algorithm driven food traceability. Food
Control 105, 45–51 (2019). https://doi.org/10.1016/j.foodcont.2019.05.019
24. Mani, V., Gunasekaran, A.: Four forces of supply chain social sustainability adoption in
emerging economies. Int. J. Prod. Econ. 199, 150–161 (2018). https://doi.org/10.1016/j.ijpe.
2018.02.015
25. Kaur, H.: Modelling internet of things driven sustainable food security system. Bench-
marking (2019). https://doi.org/10.1108/BIJ-12-2018-0431
26. Guo, S., Zhao, H.: Fuzzy best-worst multi-criteria decision-making method and its
applications. Knowl.-Based Syst. 121, 23–31 (2017). https://doi.org/10.1016/j.knosys.2017.
01.010
27. Karimi, H., Sadeghi-Dastaki, M., Javan, M.: A fully fuzzy best–worst multi attribute
decision making method with triangular fuzzy number: a case study of maintenance
assessment in the hospitals. Appl. Soft Comput. J. 86 (2020). https://doi.org/10.1016/j.asoc.
2019.105882
Resilience Measurement System: A Fuzzy
Approach
Merve Kadan(&), Gökhan Özkan, and Mehmet Hilmi Özdemir
STM Savunma Teknolojileri ve Mühendislik A.S., Ankara, Turkey

[email protected]
Abstract. The world retains its viability by the help of complex relations. Due
to these complex relations, any disruption in the basic factors is felt in various
dimensions at different levels of relevant systems. Governments, private sector
actors, and international institutions are examples of organizations with complex
relations. Owing to the complexity of organizations, they are required to be
resilient regarding the sustainability of basic needs such as water and food
resources, health services, communication, and transportation infrastructures in
the face of shocks such as natural disasters, cyber-attacks, wars, and drought. To
recover possible shocks, all stakeholders in exemplified organizations must
continuously check their capacities and take proactive measures in the required
areas that would contribute to the resilience. In this study, a prioritization and
weight determination mechanism are developed to be used in a decision support
system evaluating resilience capacity in the face of possible shocks. Within the
scope of the study, fuzzy methods were preferred to reflect the resilience
problem more realistically instead of deciding if an organization is resilient or
not by one-nil approach. Therefore, Fuzzy Analytical Hierarchy Process was
used to determine the weights of possible shocks and to analyze the organiza-
tion’s proximity to the optimum resilience level.
In further studies, it is planned to use shock priority values in a system
dynamics decision support model to be developed in the context of creating
what-if scenarios and calculating total resilience value.
Keywords: Fuzzy Analytical Hierarchy Process Resilience measurement

systems Intelligent decision support systems
1 Introduction
Governments, private sector actors and international institutions are examples of systems
that have complex relations [1]. Due to that, any deterioration in basic factors (i.e. baseline
parameters) in exemplified organizations is felt at various levels and with various mag-
nitudes. Therefore, these organizations are required to be resilient regarding the sus-
tainability of basic factors such as communication, transportation etc. [2]. There are many
definitions of resilience in the literature. The United Nations defines resilience as the
ability of a system, community or society exposed to hazards to resist, absorb and recover
from the effects of a hazard in a timely and efficient manner, including the preservation
and restoration of its essential basic structures and functions [3]. Manyena et al. defined
resilience as “a fuzzy concept which involves epistemological and operational
Resilience Measurement System: A Fuzzy Approach 577
challenges” [4]. In this study, considering the uncertainty of resilience, fuzzy methods are
preferred to reflect resilience more realistically [5]. This study aims to develop a priori-
tization and weight determination mechanism that could be used in a decision support
system evaluating resilience of an organization. In future studies, it is planned to develop a
system dynamics decision support model which includes the priority values of baseline
parameters and shocks. Using the model created it will be possible to simulate what-if
scenarios and calculating total resilience value.
It is believed that this study will be beneficial for literature due to it focuses on a
specific area such as resilience, and combines that area with fuzzy methodologies.
In the following section of the study, methodology and implementation are
described. Finally the outcomes of the study are given in the last section.
2 Methodology
A generic study flow is given in Fig. 1. According to this flow, basic factors that will be
addressed as resilience baseline parameters from now on were identified as Continuity
of Government, Energy Supplies, Movement of People, Food and Water, Mass casu-
alties, Communication, and Transportation all which disclosed by NATO [6, 7].
Organizations should maintain sufficient resilience levels regarding these baseline
parameters upon shocks occurrences such as cyber-attacks, state of war, electricity
blackout, natural disasters, epidemics and migrations [8]. The priority values of
baseline parameters and weights of shocks were determined by using FAHP.
Fig. 1. The study flow graphics
2.1 Fuzzy Analytical Hierarchy Process (FAHP)

The Fuzzy Analytical Hierarchy Process (FAHP) is a frequently used method in multi
criteria decision making problems that take into account the preferences of decision
makers, evaluate qualitative and quantitative variables together and have a flexible
structure [8]. In this study extent analysis method proposed by Chang is used, which
the most common used FAHP application [9].
First step of FAHP is creating pairwise comparison matrices using the corre-
sponding triangular fuzzy numbers as seen in fuzzy transformation scale (Table 1) due
to digitalize the expert opinions.
578 M. Kadan et al.
Table 1. The fuzzy transformation scale [9].

Linguistic scale Triangular fuzzy scale Triangular fuzzy reciprocal scale
Equally preferred (1, 1, 1) (1, 1, 1)
Weakly preferred (2/3, 1, 3/2) (3/2, 1, 2/3)
Fairly strongly preferred (3/2, 2, 5/2) (2/5, 1/2, 3/3)
Very strongly preferred (5/2, 3, 7/2) (2/7, 1/3, 2/5)
Absolutely preferred (7/2, 4, 9/2) (2/9, 1/4, 2/7)
Then, fuzzy synthetic extent value (Si ) is calculated.

Xn 1
Si ¼ j¼1
Mkjj X hP Pn i; ð1Þ
m
i¼1 j¼1 Mijj

where Mij ¼ lij ; mij ; uij
The degree of possibility of M1 M2 is defined as:
8
>
> 1; if M2 M1
>
<
0; if L1 U2
VðM1 M2 Þ ¼ ð2Þ
>
> l 1 u2
>
: ; etc
ð m 2 u2 Þ ð m 1 l 1 Þ
Where d is the ordinate of the highest intersection point in Fig. 2:
Fig. 2. The interaction between M1 and M2 [9].
We can assume that:
d 0 ðAiÞ ¼ minVSi [ Sk ð3Þ

Then, the weight vector is given by:
W 0 ¼ ðd 0 ðA1 Þ; d 0 ðA2 Þ; . . .; d 0 ðAn ÞT ð4Þ
where Ai (i = 1, 2,…., n) are n elements.

As a final step the normalized weight vectors (priority values) are determined as W.
W ¼ ðdðA1 Þ; dðA2 Þ; . . .; dðAn ÞT ð5Þ
3 Implementation
FAHP was used for determining priority values of baseline parameters and weights of
shocks. In context of FAHP, the main attributes are determined as baseline parameters
and alternatives are determined as shocks. The pairwise comparison matrices for
baseline parameters and shocks are shown in Table 2 and Table 3.
Table 2. The pairwise comparison matrice for baseline parameters

Continuity of Energy Movement of Food and Mass Communication Transportation
Government Supplies People Water casualties
Continuity of (1, 1, 1) (7/2, 4, 9/2) (7/2, 4, 9/2) (1/2, 1, 3/2) (5/2, 3, 7/2) (7/2, 4, 9/2) (7/2, 4, 9/2)
Government
Energy Supplies (2/9, 1/4, 2/7) (1, 1, 1) (5/2, 3, 7/2) (1, 1, 1) (5/2, 3, 7/2) (3/2, 2, 5/2) (5/2, 3, 7/2)
Movement of (2/9, 1/4, 2/7) (2/7, 1/3, 2/5) (1, 1, 1) (2/9, 1/4, 2/7) (2/5, 1/2, 2/3) (2/5, 1/2, 2/3) (1, 1, 1)
People
Food and Water (2/3, 1, 2) (1, 1, 1) (7/2, 4, 9/2) (1, 1, 1) (2/7, 4, 7/2) (7/2, 4, 9/2) (5/2, 3, 7/2)
Mass casualties (2/7, 1/3, 2/5) (2/7, 1/3, 2/5) (3/2, 2, 5/2) (2/9, 1/4, 2/7) (1, 1, 1) (5/2, 3, 7/2) 15/2, 3, 7/2)
Communication (2/9, 1/4, 2/7) (2/5, 1/2, 2/3) (3/2, 2, 5/2) (2/9, 1/4, 2/7) (2/7, 1/3, 2/5) (1, 1, 1) 15/2, 3, 7/2)
Transportation (2/9, 1/4, 2/7) (2/7, 1/3, 2/5) (1, 1, 1) (2/7, 1/3, 2/5) (2/7, 1/3, 2/5) (2/7, 1/3, 2/5) (1, 1, 1)
Table 3. The pairwise comparison matrice for shocks

Electricity Cyber attack Natural Migrations State of war Epidemics
blackout disasters
Electricity (1, 1, 1) (3/2, 2, 5/2) (2/5, 1/2, 2/3) (5/2, 3, 7/2) (2/5, 1/2, 2/3) (2/5, 1/2, 2/3)
blackout
Cyber attack (2/5, 1/2, 2/3) (1, 1, 1) (3/2, 2, 5/2) (5/2, 3, 7/2) (2/5, 1/2, 2/3) (3/2, 2, 5/2)
Natural Disasters (3/2, 2, 5/2) (2/5, 1/2, 2/3) (1, 1, 1) (5/2, 3, 7/2) (2/9, 1/4, 2/7) (1, 1, 1)
Migrations (2/7, 1/3, 2/5) (2/7, 1/3, 2/5) (2/7, 1/3, 2/5) (1, 1, 1) (2/7, 1/3, 2/5) (2/7, 1/3, 2/5)
State of war (3/2, 2, 5/2) (3/2, 2, 5/2) (7/2, 4, 9/2) (5/2, 3, 7/2) (1, 1, 1) (1, 1, 1)
Epidemics (3/2, 2, 5/2) (2/5, 1/2, 2/3) (1, 1, 1) (5/2, 3, 7/2) (1, 1, 1) (1, 1, 1)
580 M. Kadan et al.
The priority values of baseline parameters and shocks are shown in Table 4 and
Table 5.
Table 4. Priority values of baseline parameters

Baseline parameters Defuzzified normalized weights
Continuity of government 0.42
Energy supplies 0.17
Movement of people 0.03
Food and water 0.17
Mass casualties 0.06
Communication 0.09
Transportation 0.03
Table 5. Priority values of shocks

Shocks Defuzzified normalized weights
Electricity blackout 0.14
Cyber attack 0.19
Natural disasters 0.13
Migrations 0.03
State of war 0.31
Epidemics 0.13
When we analyze the result for baseline parameters, Continuity of Government has
the highest priority while Energy Supplies and Food and Water come after and the
lowest priority value (0,03) belong to movement of people and transportation.
When we analyze the priority values of shocks, the State of War shock is seen as
having the most impact on the baseline parameters. Contrary to this, migrations may
have the least effect on resilience.
4 Conclusion
In this study, it is aimed to develop a prioritization and weight determination mecha-

nism that could be used in a decision support system that can contribute to the resi-
lience capacity of an organization in the face of possible shocks. Fuzzy Analytical
Hierarchy Process was used to prioritize the baseline parameters and calculate weights
of possible shocks.
In further studies, it is planned to develop a system dynamics decision support
model. The priority values of baseline parameters and shocks calculated in this study
will be embed in this model. Using the model created it will be possible to simulate
what-if scenarios and calculating total resilience value.
It is believed that this study will be unique due to it focuses on a specific area such
as resilience, and combining that specific area with fuzzy techniques. In this respect it
will be useful for literature.
References
1. Orsini, A.J., Prestre, P.L., Haas, P.: Complex systems and international governance. Int. Stud.
Rev. viz020 (2019). https://doi.org/10.1093/isr/viz020
2. Ran, J., MacGillivray, B., Gong, Y., Hales, T.: The application of frameworks for measuring
social vulnerability and resilience to geophysical hazards within developing countries.
A systematic review and narrative synthesis. Sci. Total Environ. 711, 134486 (2010)
3. United Nations: Vulnerability Resilience Country Profile (2015)
4. Bernard, M., Fortunate, M., O’Keefee, P.: Disaster resilience integrated framework for
transformation (DRIFT): a new approach to theorising and operationalising resilience. World
Dev. 123, 104587 (2019)
5. Ribeiro, P., Gonçalves, L.: Urban resilience: a conceptual framework. Sustain. Cities Soc. 50,
101625 (2019)
6. NATO Web Site. https://www.nato.int/cps/en/natohq/topics_132722.htm. Accessed 12 Apr
2020
7. Civil-Military Cooperation Centre of Excellence. https://www.cimic-coe.org/wp-content/
uploads/2019/07/Resilience-through-Civil-Preparedness.pdf. Accessed 21 Apr 2020
8. Hodicky, J., Özkan, G., Özdemir, H., Stodola, P., Drozd, J., Buck, W.: Dynamic modelling
for resilience measurement: NATO resilience decision support model. Appl. Sci. 10, 2639
(2020)
9. Biswas, T.K., Akash, S.M., Saha, S.: A fuzzy-AHP method for selection best apparel item to
start-up with new garment factory: a case study in Bangladesh. Int. J. Res. Ind. Eng. 7(1), 32–
50 (2018)
Road Selection for Autonomous Trucks
in Turkey with Fuzzy AHP
Zeynep Hasgul(&) and Can Aytore
Bogazici University, Istanbul, Turkey

{zeynep.hasgul,can.aytore}@boun.edu.tr
Abstract. A smart city is an urban area that uses data collected from citizens,
devices, and assets to manage resources efficiently. It can be argued that
autonomous cars will spread related to this concept. However, the implemen-
tation and adaptation of the autonomous cars is a complex process due to human
factors and traffic. Therefore, it is expected that autonomous vehicle (AV) will
find its place first in the heavy transportation as it is more cost efficient and
easier to adapt. In order to adapt intercity transportation routes to autonomous
trucks, an investment should be made on Information and Communications
Technology such as sensors and cameras connected to the network with Internet
of Things. In this study, the multi-criteria decision making (MCDM) problem
for road selection for autonomous trucks in Turkey is examined. Although self-
driving trucks are not yet introduced, this study aims to pioneer by examining
truck roads in between industrial areas and decide which road would be better fit
to start an autonomous truck investment. As there are many uncertainties in the
nature of this decision, the fuzzy Analytic Hierarchy Process (AHP) method will
be used for decision-making.
Keywords: Fuzzy analytic hierarchy process Multi-criteria decision making

Autonomous trucks Smart city
1 Introduction
Decision-making has an important place of human life in helping individuals and

organizations overcome problems and have benefit from opportunities [1]. Decision-
making is a study of identifying and choosing alternatives based on the values and
preferences of the decision maker [2]. In the nature of decision there are alternatives to
be considered, between them decision maker wants to choose the one with highest
probability of success.
MCDM is an interdisciplinary study that combines mathematics, management,
informatics, social sciences, and economy in order to give decision maker a chance to
evaluate [3]. The studies on MCDM started in 1970’s with Saaty’s suggestion of AHP
and it has been widely accepted in the management and science community. AHP is a
technique, which takes the decision problems into a hierarchical structure and bases the
criteria and decisions on the double comparison logic [4].
Fuzzy sets provide mathematical explanations for these uncertainties and defi-
ciencies, increasing their applicability in real life. The concept of fuzzy sets was
Road Selection for Autonomous Trucks in Turkey with Fuzzy AHP 583
introduced in 1965 by Lotfi Zadeh to deal with uncertainties of the real world. Fuzzy
logic proved its power through the successful applications and fuzzy MCDM methods
are also example to these applications [5].
To make a logical decision, the decision maker requires to evaluate data. However,
there are many cases where it is impossible to get relevant data, or it may not be
possible to measure essential parameters [6]. The traditional decision making methods
were not designed to deal with uncertainty in MCDM problems. In the binary com-
parison process, it is insufficient to address uncertainty and instability [3]. For this
reason, in the fuzzy AHP the comparisons are given in a range of values, unlike
classical AHP, where exact values are used. Thus, the uncertainty in the decision-
making process can be overcome more easily.
In this study cutting edge technologies are examined which consist of many
uncertainties and decision making is based on mostly predictions. Since a scale does
not exist, it is a fuzzy problem in its nature. To be able to tackle these fuzziness, the
literature reviews on smart cities and AVs are analyzed.
Cities have existed with people and cities are affected by this as technology
advances in people’s lives. Smart cities represent interconnected, sustainable, com-
fortable, attractive, safe, and community of average technology [7].
Smart transportation is an important part of smart city enterprise in terms of the
experience of people using smart technology. Smart transport systems improve safety,
security, efficiency, and environment friendliness of surface transport system by
combine intelligent transportation systems with other vehicular technologies, such as
AVs [8]. When these assumptions are realized in the future, AVs will not only revo-
lutionize transportation, but also it will change the urban forms as it is known [9].
The transition of vehicles into autonomous mechanism will be slow and gradually.
The Society of Automotive Engineers International has created a five level hierarchy to
help clarify the definition of AV. The higher levels on the scale represent higher
proportion of automation to manual control. On this scale, level 0 represents no
automation. In level 1, the driver still has the full control; however, automated system
also provides added control to aid the driver. In level 2 and level 3, even though the
vehicle can function autonomously the driver still has to monitor the roadway and the
driver is expected to be available to take control in cases that it is needed. In level 4, the
vehicle is designed to manage all driving and safety functions and monitor roadway
conditions for an entire trip [10]. Literature review showed that, AV technology is not
far away in the horizon. Therefore, implementing AV technology in Turkey is expected
to begin with automated trucks due to higher expected benefits. This paper is written
considering AVs with level 3 and higher.
Legislations for AVs vary in different countries. However, in the countries with a
large number of vehicle manufacturers such as UK, USA, Germany, Sweden, Japan
and France, the governments have allowed testing of AVs on public roads [11]. Fur-
thermore, around 30% of the trucks in the UK could be automatically driven by 2022
and up to 75% of the vehicles on road could be fully automated by 2040 [12].
Autonomous trucks are viewed as a feasible option due to their expected benefits since
584 Z. Hasgul and C. Aytore
heavy vehicles have a significant role in contributing to the key issues such as high
usage of infrastructure and bad emissions, and energy demand characteristics. Besides,
according to the survey on executives of major car manufacturing firms, in ten years or
even sooner it is expected that AVs with higher levels of automation will be affordable
and mass marketed [13].
This paper combines a common method used for decision making with future
technologies. It contributes to literature by demonstrating a possible application for
fuzzy AHP. This study aims to pioneer by examining truck roads in between different
regions of Turkey and decide which road would be beneficial to begin the investment.
In the next part, the methodology is explained. The alternatives and criteria are defined.
Later the results are interpreted. The sensitivity analysis is applied for the robustness of
the study. Finally, paper is concluded with comments and future studies.
2 Methodology
The Fuzzy AHP method is used to select a motorway road for implementation instead of
Classical AHP since there is no exemplary and realized project, and level of uncertainty
is high in addition to lack of defined parameters. In order to build a Fuzzy AHP model,
five criteria are defined based on the cumulative information that was gathered from the
research. The triangular fuzzy numbers are the most commonly used type in the
membership functions and are used in this paper to represent experts’ opinions. Also,
Saaty’s well-known 1 to 9 scale is used [4]. To solve for the Fuzzy AHP method a web-
based program by Palacký University Olomouc is used [14]. The mathematical back-
ground used in this software, is based on the paper by Holeček P. and Talašová, J. [15].
3 Road Selection Using Fuzzy AHP
3.1 Defining Criteria

To define criteria literature on road selection problems that use MCDM methods were
analyzed and summarized in Table 1.
MCDM has been used to select routes in very different fields such as construction
engineering, environmental engineering, automation control engineering and man-
agement. From these articles it is concluded that the decision criteria have similarities.
The selected criteria are listed below.
Cost and Maintenance (C1): Integration costs of AV technology, proper placement
of sensors, accurate identification of the lanes and regular maintenance are important
factors for the sustainability of the project in terms of economic and liquidity. This
criterion corresponds to the road length in kilometers of the motorways which were
completed and put into service at the end of 2018.
Table 1. Literature of road selection.

Research Title Criteria Alternatives
A corridor selection for locating Availability, financial aspects, 4 corridors in
autonomous vehicles using an operational capability, security Istanbul that are
interval-valued intuitionistic and safety, brand impact, segregated from the
fuzzy AHP and TOPSIS integration traffic
method [16]
Forest roads locating based on Technical features, traffic 3 regions in a forest
AHP and COPRAS-G Methods: volume, quality and safety,
an empirical study based on environmental, social, and
Iran [17] economical features
Assessing risk and uncertainty Technology, contracts and legal 4 highway case
inherent issues, resources, quality, studies in China
in Chinese highway projects cultural considerations
using AHP [18]
Creating a risk-based network Risk of carriage explosion and 3 alternative routes
for hazmat: Gasoline logistics in road accidents, and
Rayong, Thailand by route consequences of an incident
prioritizing with AHP [19]
Traffic (C2): The traffic density on motorways is an important factor in terms of

fluency of transportation. This criterion corresponds to annual average daily traffic.
Demand of Region (C3): In determining the transportation needs of the regions
reached by the highways, it is appropriate to use the average daily number of heavy
vehicles passing through these highways. This criterion corresponds to the annual
average daily heavy vehicle traffic.
Trade Potential (C4): The amount of transportation carried on the region, highway
and provincial roads reached by the specific motorway is an important factor in terms
of showing the trade intensity. This criterion corresponds to the annual transportation
amount in tons-km of the provinces in which the motorway has crossed.
Weather Condition (C5): The weather conditions of the area where the motorway is
located are an important factor for the sensors to properly detect autonomous trucks and
road lanes. This criterion corresponds to the annual average precipitation in mm.
The reviews of six experts with different backgrounds were consulted to evaluate
the criteria weights. Information on the experts are given in the Table 2.
Table 2. Background information of experts.

Experts Occupation Background Title Experience (Years)
#1 Mechanical engineer Private sector Factory owner 38
#2 Industrial engineer Academic field Assoc. prof. 32
#3 Civil engineer Academic field Prof. 25
#4, #5, #6 Industrial engineer Private sector Logistics manager 18, 15, and 15
3.2 Defining Alternatives

Among the motorways in Turkey, eleven alternatives are selected with the consider-
ation of the expert’s opinion and the roads business. The selected alternatives, and
hierarchical structure based on the explained criteria and alternatives are demonstrated
in Fig. 1 and Fig. 2, respectively.
Fig. 1. The selected alternatives on the map of Turkey.
Fig. 2. Hierarchical structure.
3.3 Deriving Criteria Weights

In this step, the decision makers compared relative importance of the criteria one by
one. The combined the judgments of the decision makers are demonstrated in Table 3.
Table 3. Combined criteria weights.

C1 1 135 1 1 12 3
C2 1/5 1/3 1 1 1/5 1/4 1/3 1/3 1/2 1 12 3
C3 1 345 1 1/3 1/2 1 57 9
C4 1 123 123 1 35 7
C5 1/3 1/2 1 1/3 1/2 1 1/9 1/7 1/5 1/7 1/5 1/3 1
Since comparison matrix of the criteria was obtained by combining each expert’s
binary comparisons, it is not guaranteed to satisfy consistency as a whole. As Saaty
recommended consistency ratio (CR) was found to be below 10 percent for every
criteria and alternative comparison matrix [4].
Triangular numbers for this comparison matrix are demonstrated in Fig. 3.
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Fig. 3. Fuzzy numbers of criteria.
3.4 Alternative Evaluation with Respect to Criteria

For the alternatives, the motorways in Turkey are considered since they are more standard
in terms of infrastructure, road conditions and regulations. When determining alterna-
tives, data based on the most recently traffic, transportation and weather information
published by the General Directorate of Highways [20] and Turkish State Meteorological
Service [21] were used. The fuzzy numbers for the alternatives are demonstrated in Fig. 4.
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Fig. 4. Fuzzy sets of alternatives respect to criteria.

This result shows that when AV technology applied in Turkey, it is logical to start
implementing on O4. This alternative is followed by O3, O22, O30, O53, O52, O20,
O51, O31, O21, and O32, respectively. For the robustness of the result, the sensitivity
analysis is applied.
3.5 Sensitivity Analysis

Through sensitivity analysis it is aimed to examine how a change in weight of each
criterion affects the final evaluation. The current weight of each criterion and its effect
on alternative choices of the weight change in that criterion are given in Fig. 5.
C1: Cost & Maintenance C2: Traffic C3: Demand of Region
C4: Trade Potential C5: Weather Condition
Fig. 5. Sensitivity analysis with respect to each criterion.
4 Conclusion and Further Studies
The decision problem for road selection for autonomous trucks has many uncertainties
because there is no application in Turkey yet. Therefore, the fuzzy AHP method was
suggested for this problem. According to literature review, the five criteria: cost and
maintenance, traffic, demand of region, trade potential, and weather condition were
determined. The problem was evaluated according to both qualitative and quantitative
criteria and expert opinions was used for comparing the criteria. To compare
alternatives, real world transportation data were used. This study can be repeated with
more recent data and upcoming alternatives.
As further study, international logistics networks that are prone to failing due to
human factors can be analyzed with the fuzzy AHP. During Covid19 pandemic,
international trucking and logistics industry were effected. This showed that AV can
play a major role in near future.
References
1. Evans, G.W.: Multiple Criteria Decision Analysis for Industrial Engineering: Methodology
and Applications. CRC Press, Boca Raton (2017)
2. Harris, R.: Introduction to Decision Making (2012). http://www.virtualsalt.com/crebook5.
htm
3. Paksoy, T., Pehlivan, N.Y., Özceylan, E.: Fuzzy Set Theory. Nobel, Ankara (2013)
4. Saaty, T.L.: How to make a decision: the analytic hierarchy process. Eur. J. Oper. Res. 48,
9–26 (1990)
6. Bandemer, H., Gottwald, S.: Fuzzy Sets, Fuzzy Logic Fuzzy Methods with Applications.
Wiley, West Sussex (1995)
7. Lazaroiu, G., Roscia, M.: Definition methodology for the smart cities model. Energy 47,
326–332 (2012)
8. Song, H., Srinivasan, R., Sookoor, T., Jeschke, S.: Smart Cities: Foundations, Principles, and
Applications. Wiley, Hoboken (2017)
9. Meyer, J., Becker, H., Bösch, P.M., Axhausen, K.W.: Autonomous vehicles: the next jump
in accessibilities? Res. Transp. Econ. 62, 80–91 (2017)
10. SAE International, Taxonomy and Definitions for Terms Related to Driving Automation
Systems for On-Road Motor Vehicles (2018). https://www.sae.org/standards/content/j3016_
201806/
11. Bishop, R., et al.: Automated Driving and Platooning Issues and Opportunities. In: Meeting
of the ATA Technology and Maintenance Council Future Truck Program, Virginia, USA,
American Trucking Associations, September 2015. http://orfe.princeton.edu
12. Wadud, Z., MacKenzie, D., Leiby, P.: Help or hindrance? The travel, energy and carbon
impactsof highly automated vehicles. Transp. Res. Part A 86, 1–18 (2016)
13. KPMG: Global Automotive Executive Survey 2019. https://automotive-institute.kpmg.de/
GAES2019/downloads/GAES2019PressConferenceENG_FINAL.PDF
14. Palacký University Olomouc, 2016. Fuzzy AHP. http://fuzzymcdm.upol.cz/FuzzyAHP/
15. Holeček, P., Talašová, J.: A free software tool implementing the fuzzy AHP method. In:
Proceedings of the 34th International Conference on Mathematical Methods in Economics,
Liberec, Czech Republic, pp. 266–271 (2016)
16. Dogan, O., Deveci, M., Canıtez, F., Kahraman, C.: A corridor selection for locating
autonomous vehicles using an interval-valued intuitionistic fuzzy AHP and TOPSIS method.
Soft. Comput. 24, 1–17 (2019). https://doi.org/10.1007/s00500-019-04421-5
17. Zolfani, S.H., Rezaeiniya, N., Kazimieras, Z.E., Turskis, Z.: Forest roads locating based on
AHP and COPRAS-G methods: an empirical study based on Iran. Ekonomie a Manage.
14(4), 6–21 (2011)
18. Zayed, T., Amer, M., Pan, J.: Assessing risk and uncertainty inherent in Chinese highway
projects using AHP. Int. J. Project Manage. 26, 408–419 (2008)
19. Sattayaprasert, W., Taneerananon, P., Hanaoka, S.: Creating a risk-based network for
HazMat case study: Gasoline logistics in Rayong, Thailand. Int. Assoc. Traffic Saf. Sci.
32(1), 74–87 (2008)
20. Turkish General Directorate of Highways, Traffic and Transportation Survey of Highways.
https://www.kgm.gov.tr/SiteCollectionDocuments/KGMdocuments/Istatistikler/Trafikve
UlasimBilgileri/18TrafikUlasimBilgileri.pdf
21. Turkish State Meteorological Service, Statistics on the seasonal normal of the provinces.
https://www.mgm.gov.tr/veridegerlendirme/il-ve-ilceler-istatistik.aspx?k=H
Multi-criteria Oil Station Location Evaluation
Using Spherical AHP&WASPAS: A Real-Life
Case Study
Irem Otay(&) and Serhat Atik
Faculty of Engineering and Natural Sciences, Department of Industrial

Engineering, Istanbul Bilgi University, Eski Silahtarağa Elektrik Santrali,
34060 Eyüpsultan, Istanbul, Turkey
Abstract. The world’s dependence on oil and gas has been increasing as global
economies and infrastructure continue to rely heavily on petroleum-based
products. Thus, there is a significant increase in the number of stations to meet
the growing needs of the world. In this study, we concentrate on the location
evaluation and selection problem of an oil station. Location selection of an oil &
gas station is a Multiple Criteria Decision Making (MCDM) problem com-
prising several quantitative & qualitative criteria such as environmental, eco-
nomic, and traffic related factors. In the literature, there are a number of studies
focusing on oil-gas station location selection problem using different MCDM
methods. The literature review has pointed out that traditional methods for
location selection are insufficient for dealing with the indefinite or uncertain
nature of linguistic assessment. Thus, considering uncertainties and subjectivity
in human judgments, many researchers have been solving the multicriteria
decision making problems employing MCDM methods based on the fuzzy sets.
There are many extensions of the fuzzy sets such as Type-2, Hesitant, Pytha-
gorean, and Neutrosophic fuzzy sets. In this study, a recently popular extension
namely Spherical fuzzy set is used for solving a real-life oil station location
evaluation and selection problem. In the study, Spherical fuzzy Analytical
Hierarchy Process (SF-AHP) and Spherical fuzzy Weighted Aggregated Sum
Product ASsessment (SF-WASPAS) methods are proposed for the solution of
the problem. Then, sensitivity analysis is performed.
Keywords: Location selection Spherical fuzzy sets Spherical fuzzy AHP

Spherical fuzzy WASPAS MCDM
1 Introduction
Location evaluation and selection topic has been studying since 1980s. There are several
reasons why it still preserves its importance and criticality. As studies have highlighted
that location selection decision has effects on reducing costs (such as production and
transportation costs), increasing profits and the return profitability; so that, it enables
companies to increase competitiveness in the market place [1, 2]. On the other hand,
considering the world’s dependency on oil and gas, high competency in energy sector,
592 I. Otay and S. Atik
increase in population, traffic, and environmental concerns, location selection decision

for gas/oil stations takes the attentions of researchers and practitionars.
In the literature, there are several valuable studies focusing on location evaluation
and selection problem for gas & oil stations. For instance, Mokhtarian [1] proposed a
new fuzzy weighted average method using left and right scores for multi-criteria
facility location selection problems and applied the proposed method to a real life
gas/oil station location problem. Semih and Seyhan [3] were also dealing with gas
station site selection problem. The researchers suggested using AHP method evaluating
both quantitative and qualitative criteria such as the total number of other stations in the
area, traffic, and social criteria. Njoku and Alagbe [4] developed a multi-dimensional
Geographic Information System (GIS) based method for petrol filling station problem
by analyzing a set of criteria, and applied the proposed method for evaluating 113
petrol filling stations in Oyo town. Niu et al. [5] proposed a novel ranking system
considering waiting time of refueling events by minimizing human mobility data for a
gas station location selection problem. Alavipoor et al. [6] developed an MCDM
method based on GIS for locating a gas power plant in Iran. The authors used AHP, the
fuzzy set theory and weighted linear combination by taking into account uncertainties
in real life. Li et al. [7] proposed an integrated multi-objective decision-making method
for the natural gas siting project related to public transportation. The proposed method
was composed of multi-objective optimization, fuzzy evaluations, and radar charting
by considering environmental and social concerns. The proposed method was applied
to the case of Wuhan in China. Fakhari et al. [8] also focused on the gas power plant
location-allocation problem. The authors evaluated 25 cities in Iran by implementing
fuzzy data envelopment analysis based on Z-numbers considering several criteria i.e.
the pollution rate, land cost, population, and electricity generation.
On the other hand, Spherical Fuzzy Set (SFS) introduced by Kahraman and Kutlu
Gündoğdu [9], is one of the recent extensions of the ordinary fuzzy sets. SFSs
including three dimensions which are membership, non-membership and hesitancy
degrees, aim to define uncertainties in a better way with the coverage of a larger
domain. In SFSs, the square sum of membership, non-membership and hesitancy
degrees is defined as less than “1”. In the literature, there are a few numbers of studies
proposing MCDM methods based on SFSs. For instance, Kutlu Gündoğdu and
Kahraman [10] extended ordinary fuzzy AHP using SFSs and applied the proposed
method to provide a solution to the renewable energy location selection problem. The
researchers also compared the results with the ones obtained from neutrosophic
AHP. Kutlu Gundoğdu and Kahraman [11] presented arithmetic, aggregation, and
defuzzification operations for SFSs, proposed Spherical fuzzy WASPAS method,
which is a combination of Weighted Sum and Weighted Product Models, and applied it
to a multi-criteria industrial robot selection decision. The researchers also constructed
comparative and sensitivity analyses. In addition to these, Spherical fuzzy aggregation
operators [12] and other Spherical fuzzy (SF) MCDM methods such as SF-VIKOR
[13] and SF-TOPSIS [14] have also been using for a variety of real-life MCDM
problems. However, according to the best knowledge of the authors, in multi-criteria
location selection literature there is no study conducting an integrated Spherical fuzzy
MCDM method employing fuzzy AHP and fuzzy WASPAS based on single-valued
Spherical fuzzy sets.
Multi-criteria Oil Station Location Evaluation 593
The remaining of the paper is organized as follows: Sect. 2 gives information on

Spherical fuzzy sets. Section 3 presents the proposed integrated Spherical fuzzy
method. Section 4 implements the proposed Spherical fuzzy MCDM method to an oil
station location selection problem while Sect. 5 concludes the paper and presents
suggestions for future studies.
2 Spherical Fuzzy Sets
Spherical fuzzy sets (SFSs) introduced by Kutlu Gündoğdu and Kahraman [10] was
defined as a generalization of Pythagorean and Neutrosophic fuzzy sets. The researchers
highlighted that SFSs provide a larger preference domain and cover uncertainties better
for decision making problems.
Definition 1. General representation of a single-valued Spherical fuzzy set of the
universe of a discourse set U is given in Eq. (1). The equation points out that the square
sum of the membership ðlA~ S ðuÞÞ, non-membership ðvA~ S ðuÞÞ and hesitancy ðpA~ S ðuÞÞ
degrees can not exceed “1”.
n o
~ S ¼ hu; l ~ ðuÞ; v ~ ðuÞ; p ~ ðuÞ ju 2 U where 0 l2~ ðuÞ þ v2~ ðuÞ þ p2~ ðuÞ 1 8u 2 U
A AS AS AS AS AS AS
and lA~ S : U ! ½0; 1; vA~ S : U ! ½0; 1; pA~ S : U ! ½0; 1

ð1Þ

~ p ¼ l ~ ðuÞ; v ~ ðuÞ; p ~ ðuÞ and B
Let A ~ p ¼ lB~ ðuÞ; vB~ ðuÞ; pB~ ðuÞ be SVSFSs.
Ap Ap Ap p p p
Then, some of the arithmetic operations are presented in Eqs. (2–5) [15]:
Addition
8 1=2 9
>
< l2A~ þ l2B~ l2A~ l2B~ ; v ~ v ~ ; >
=
Ap pB
~ ~
Ap Bp ¼
p p

p

p
ð2Þ
: 1 l2 p2 þ 1 l2 p2 p2 p2 1=2 >
> ;
~p
B ~p
A ~p
A ~p
B ~p B
A ~p
Multiplication
8 1=2 9
>
< lA~ p lB~ p ; v2A~ þ v2B~ v2A~ v2B~ ; >
=
~ ~
Ap Bp ¼
p p

p p
ð3Þ
: 1 v2 p2 þ 1 v2 p2 p2 p2 1=2 >
> ;
~p
B ~p
A ~p
A ~p
B ~p B
A ~p
~p
Multiplication by a scalar and Power of A
( )
k 1=2 k k 1=2
~p ¼
kA 1 1 l2A~ p ; vkA~ p ; 1 l2A~ p 1 lA~ p pA~ p
2 2
ðk [ 0Þ ð4Þ
( )
k 1=2 k k 1=2
~k
A ¼ lkA~ p ; 1 1 vA~ p
2
; 1 vA~ p 1 vA~ p pA~ p
2 2 2
ðk [ 0Þ
p
ð5Þ
Definition 2. Assume that A ~ S is a single-valued Spherical fuzzy number (SVSFN).

~
Score ðScðAS ÞÞ and accuracy ðAccðA ~ S ÞÞ functions for A
~ S ¼ ðl ~ ; v ~ ; p ~ Þ are given in
AS AS AS
Eq. (6) [15].

lA~ S 2vA~ S pA~ S þ 2
~SÞ ¼
ScðA ~ S Þ ¼ l2~ þ v2~ þ p2~
; AccðA ð6Þ
3 AS AS AS
3 Methodology
In this section, the steps of an integrated Spherical fuzzy AHP-WASPAS method using
single-valued Spherical fuzzy sets are summarized based on the studies of [9, 10, 14, 15].
Step 1. Initially, a multi-criteria decision making problem is defined explicitly and a

hierarchical structure of the problem is designed.
Step 2. A linguistic scale with corresponding Spherical fuzzy numbers is determined.

In the study, we employ the linguistic scale proposed by [10] presented in Table 1.
Table 1. Linguistic terms with corresponding SFNs [10]

Linguistic terms ðl; v; pÞ
Absolutely More Importance (AMI) (0.9, 0.1, 0)
Approximately Equal Importance (EI) (0.5, 0.4, 0.4)
Absolutely Low Importance (ALI) (0.1, 0.9, 0)
Evaluation of Criteria
Step 3. Expert/s are asked to make pairwise evaluations of criteria ðC ¼
fC1 ; C2 ; . . .. . .Cn gÞ using Table 1.
Step 4. Once the pairwise comparison matrices are obtained, then consistency of the
pairwise matrices are controlled by means of the classical consistency analysis based
on the score indices proposed by [10].
Step 5. Then, the judgments of the experts are aggregated and the Spherical fuzzy
weights of criteria are calculated by Spherical Weighted Arithmetic Mean Operator,
shortly SWAM operator, as in Eq. (7).
* Q
n Q
n
+
½1 ð1 l2A Þwi 1=2 ; vwA i ;
si si
i¼1 i¼1
SWAMw ðAS1 ; . . .; ASn Þ ¼ Q
n Q
n w ¼ 1=n
½ ð1 l2A Þwi ð1 l2A p2A Þwi 1=2
si si si
i¼1 i¼1
ð7Þ
Step 6. The Spherical weights of criteria are defuzzified using the equation for score
function and the weights are normalized by dividing each weight by the sum of the
criteria weights. Hence, the defuzzified weights of criteria ðw ¼ fw1 ; w2 ; . . .. . .wn g
where 0 wj 1Þ are obtained.
Evaluation of Alternatives
Step 7. In this step, alternatives are analyzed
by applying Spherical fuzzy WASPAS
method. A decision matrix ~xij ¼ Cj ð~xi Þ mn with Spherical evaluations of the alter-
natives ðX ¼ fx1 ; x2 ; . . .; xm gÞ for each set of criteria are collected from the expert/s by
means of Table 1, as illustrated in Eq. (8).
0 1
ðl11 ; v11 ; p11 Þ ðl12 ; v12 ; p12 Þ : : : ðl1n ; v1n ; p1n Þ
B ðl21 ; v21 ; p21 Þ ðl22 ; v22 ; p22 Þ : : : ðl2n ; v2n ; p2n Þ C
B C
~xij ¼ Cj ð~xi Þ mxn ¼ B
B C
C
@ A
ðlm1 ; vm1 ; pm1 Þ ðlm2 ; vm2 ; pm2 Þ : : : ðlmn ; vmn ; pmn Þ
ð8Þ
Step 8. In Spherical fuzzy WASPAS, firstly Weighted Sum Model (WSM) is applied.
In WSM model, the calculation of Q ~ ð1Þ
i is divided into two operations which are
multiplication (Eq. (3)) and addition operations (Eq. (2)). Then, the values of the
Weighted Product Model (WPM) method ðQ ~ ð2Þ
i Þ are calculated using Eq. (5) and
Eq. (3), respectively.
X
n Y
n
w
~ ð1Þ
Q i ¼
~ ð2Þ
~xij wj ; Q i ¼ ~xij j ð9Þ
j¼1 j¼1
Step 9. The values of Q~ i are obtained considering the threshold value (k). Finally, the
~ i values are defuzzified and the alternatives are ranked in descending order.
Q
Q ~ ð1Þ
~ i ¼ kQ ~ ð2Þ
i þ ð 1 kÞ Q i ð10Þ
4 Implementation
The proposed model is applied to solve a real life problem of one of the international
energy company operating in Turkey since 1992. The company being the country’s 5th
largest company with more than 500 stations, is looking for a new location for its oil
station project. Based on literature review and meetings with the management, several
quantitative and qualitative criteria which are Traffic (C1), Environment & safety (C2),
Market (C3), Economic factors (C4) and Land structure (C5) are determined.
According to the management, there are three alternative locations (A1, A2 and A3) for
constructing the station. After several meetings, a pairwise comparison matrix is col-
lected using the linguistic scale as presented in Table 2. Then, consistency of the
pairwise comparison matrix based on the score indices, is checked using the consis-
tency analysis. By following the steps of Spherical fuzzy AHP, the weights of criteria
are calculated as 0.15, 0.29, 0.20, 0.25 and 0.11, respectively. After calculating the
criteria weights, the decision matrix is collected from the expert as in Table 3.
Table 2. Pairwise comparison matrix

C1 EE VL ML L MH
C2 EE H MH VH
C3 EE ML H
C4 EE VH
C5 EE
After then, the alternatives are evaluated and ranked using Spherical fuzzy
WASPAS method. Herein, WSM and WPM methods are applied and the calculated
values of Q ~ ð1Þ ~ ð2Þ ~ ð1Þ
i and Qi are listed in Table 4. Table 4 also presents the values of kQi ,
ð1 kÞQ ~ ð2Þ ~ ~
i , and Qi . When the score values of Qi are obtained by means of Eq. (6), the
alternatives are ranked in descending order. The results indicate that A is the best
appropriate location with the normalized score value of “0.406”. It is followed by B
with “0.323” while the worst alternative is obtained as “C” (0.271).
Table 3. Decision matrix

Criteria Alternatives
A B C
C1 H MH ML
C2 VH MH H
C3 VH H MH
C4 VH H MH
C5 MH H ML
Table 4. The results of Spherical fuzzy WASPAS

~ ið1Þ
Q ~i
kQ
ð1Þ ~ ið2Þ
Q ~i
ð1 kÞQ
ð2Þ ~i
Q
A (0.77, 0.23, 0.14) (0.6, 0.48, 0.12) (0.76, 0.25, 0.16) (0.59, 0.5, 0.14) (0.77, 0.24, 0.15)
B (0.66, 0.34, 0.24) (0.5, 0.58, 0.2) (0.65, 0.35, 0.25) (0.49, 0.59, 0.21) (0.66, 0.34, 0.26)
C (0.6, 0.41, 0.27) (0.45, 0.64, 0.22) (0.57, 0.45, 0.28) (0.42, 0.67, 0.22) (0.58, 0.43, 0.3)
The sensitivity analysis is also performed to check the robustness of the proposed
approach. The developed model is solved for 10 different scenarios for each criterion.
Each time, the weight of a criterion is changed from “0.1” to “1.0” by setting the same
weights for the remaining criteria. The results demonstrate that there is no change in the
ranking of the alternatives for different weights of C1, C3 and C4. Figure 1 illustrates
sensitivity results based on the normalized Q values. Since there is almost no change in
the normalized Q values for different weights of C3 and C4, their graphics aren’t
presented in Fig. 1. The results also show that when the weight of C2 is greater than
“0.6”, the rank is A > C > B. However, for the weight of C5 greater than “0.5”, the
ranking of the alternatives is found as: B > A > C.
0.5
C1 0.5 C2
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.5
C5
0.4
0.3 A
0.2 B
0.1 C
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 1. The results of sensitivity analysis
5 Conclusion
In this study, an oil station location evaluation and selection decision is analyzed. The
authors proposed an integrated fuzzy AHP-WASPAS method by means of single-
valued spherical fuzzy sets. Because of its capability in defining uncertainties in a better
way, the authors prefer to employ Spherical fuzzy sets. The sensitivity analysis is also
conducted for checking the robustness of the study. The results highlight that the
proposed model provides robust solutions.
For future research, other integrated MCDM methods can be applied to compare
the results. On the other hand, for further studies the results can be also compared with
other recent extensions of the fuzzy sets including q-Rung Orthopair, Fermatean and/or
Hexagonal fuzzy sets.
References
1. Mokhtarian, M.N.: A new fuzzy weighted average (FWA) method based on left and right
scores: an application for determining a suitable location for a gas oil station. Comput. Math
Appl. 61(10), 3136–3145 (2011)
2. Cebi, F., Otay, I.: Multi-criteria and multi-stage facility location selection under interval
type-2 fuzzy environment: a case study for a cement factory. Int. J. Comput. Intell. Syst. 8
(2), 330–344 (2015)
3. Semih, T., Seyhan, S.: A multi-criteria factor evaluation model for gas station site selection.
J. Glob. Manag. Glob. Res. Agency 2(1), 12–21 (2011)
4. Njoku, C., Alagbe, A.O.: Site suitability assessment of petrol filling stations (PFSs) in Oyo
Town, Oyo State, Nigeria: a geographic information systems (GIS) approach. IOSR J. Env-
iron. Sci. Toxicol. Food Technol. (IOSR-JESTFT) 9(12), 08–19 (2015)
5. Niu, H., Liu, J., Fu, Y., Liu, Y., Lang, B.: Exploiting human mobility patterns for gas station
site selection. In: Lecture Notes in Computer Science, pp. 242–257 (2016)
6. Alavipoor, F.S., Karimi, S., Balist, J., Khakian, A.H.: A geographic information system for
gas power plant location using analytical hierarchy process and fuzzy logic. Glob.
J. Environ. Sci. Manag. 2(22), 197–207 (2016)
7. Li, S., Su, B., St-Pierre, D.L., Sui, P.-C., Zhang, G., Xiao, J.: Decision-making of
compressed natural gas station siting for public transportation: integration of multi-objective
optimization, fuzzy evaluating, and radar charting. Energy 140, 11–17 (2017)
8. Fakhari, F., Tavakkoli-Moghaddam, R. Tohidifard, M., Farid Ghaderi, S.: Location
optimization of gas power plants by a Z-number data envelopment analysis. In: World
Congress on Global Optimization WCGO 2019: Optimization of Complex Systems: Theory,
Models, Algorithms and Applications, pp. 926–936 (2019)
10. Kutlu Gündoğdu, F., Kahraman, C.: Extension of WASPAS with spherical fuzzy sets.
Informatica 30(2), 269–292 (2019)
11. Ashraf, S., Abdullah, S.: Spherical aggregation operators and their application in
multiattribute group decision-making. Int. J. Intell. Syst. 34(3), 493–523 (2019)
12. Kutlu Gündoğdu, F., Kahraman, C.: A novel VIKOR method using spherical fuzzy sets and
its application to warehouse site selection. J. Intell. Fuzzy Syst. 37(1), 1197–1211 (2019)
13. Kutlu Gündoğdu, F., Kahraman, C.: A novel fuzzy TOPSIS method using emerging interval-
14. Kutlu Gündoğdu, F., Kahraman, C.: Spherical fuzzy analytic hierarchy process (AHP) and
its application to industrial robot selection. In: Proceedings of International Conference on
Intelligent and Fuzzy Systems INFUS 2019, Istanbul, pp 988–996. (2019e)
15. Otay, I., Kahraman, C., Öztayşi, B., Onar, S.Ç.: A novel single-valued spherical fuzzy AHP-
WASPAS methodology. In: Fuzzy Logic and Intelligent Technologies in Nuclear Science
2020 (FLINS 2020) Conference and the 15th International Conference on Intelligent
Systems and Knowledge Engineering (ISKE2020), Cologne, Germany, 18–21 August 2020
Managing Cultural Built Heritage in Smart
Cities Using Fuzzy and Interval Multi-criteria
Decision Making
Mimica Milošević1(&), Dušan Milošević2, and Ana Stanojević3

1
Faculty of Business Economics and Entrepreneurship, University of Belgrade,
Belgrade, Serbia
[email protected]
2
Faculty of Electronic Engineering, University of Niš, Niš, Serbia
3
Faculty of Civil Engineering and Architecture, University of Niš, Niš, Serbia
Abstract. Including human, technological, organizational, and natural resour-

ces, contemporary cities represent an urban reflection of civilization develop-
ment. The globalization and competitiveness have shaped cities commercially,
politically, socially, and in terms of their design and growth. Looking for the
new urban patterns that could meet the global challenges, culture was recently
recognized, as an indispensable part of the sustainability, and the significant
pillar for grounding the Smart City initiatives. The management of cultural
heritage is a significant factor for recognization, definition, and affirmation of
cultural identity. Given that technology primarily enhances the urban infras-
tructure, the management of built heritage is significant within the Smart City
development paradigm. For this paper, in the decision-making process, indica-
tors related to the management of the cultural built heritage are divided into six
groups: governing, economic, social, environmental, technological, and archi-
tectural factors. The paper aims to identify priority indicators, as necessary in the
further development of heritage management, using the Fuzzy Analytic Hier-
archy Process (FAHP) and Interval Analytic Hierarchy Process (IAHP) meth-
ods. The isolation of dominant indicators should enable a more precise
definition of future activities and measures. The application of these methods
should address the extent to which differences in fuzzy and interval grading
approaches may affect the final ranking.
Keywords: Multi-Criteria Decision-Making Fuzzy Analytic Hierarchy

Process Interval Analytic Hierarchy Process Smart city Cultural built
heritage Managing
1 Introduction
Contemporary cities are under pressure caused by the increased urban population,
requirements and needs of society, political and economic changes, technological
innovations, cultural transformations, as well as global ecological problems and
challenges. All of this has been reflected in urban matrices, questioning the further
preservation of existing urban heritage and identity. The concept of Smart City
600 M. Milošević et al.
developed as a global paradigm and framework for promotion and integration of

sustainable innovative solutions, which encourage cities to be more resilient, has
become an indispensable segment of urban development strategies and actual social
issues [1–3]. Using advanced technologies, such as ICT, many urban sectors (mobility,
water waste management, energy management, health, education, etc.) were success-
fully enhanced in cities across the world.
Cultural heritage is a significant urban resource and part of the cultural dimension
of sustainability [4, 5]. Built heritage, as one type of cultural heritage, defines the urban
landscapes, reflects the history and shapes a recognizable urban matrix for the foun-
dation of future activities. As physical evidence of cultural development, built heritage
includes different types and forms of architectural facilities and spatial historic-cultural
areas, that are often under the protection regime as cultural assets or have recognized
historical, cultural and architectural values that classify them as representatives of
epochs, styles or contemporary design. Managing cultural heritage represents a mul-
tidimensional process, viewed from multiple perspectives, requires continued interac-
tion between different stakeholders. It is defined by appropriate groups of indicators
that are directly related to the postulates of current spatial sustainable development
strategies.
This study of cultural built heritage has conceived as a short review through three
temporal dimensions, past, present, and future across spaces in the world. The paper
deals with the managing of cultural built heritage in contemporary smart cities, using a
multi-criteria decision-making framework. The research aims to rank indicators,
applying the methods of the Analytical Hierarchical Process to determine priority ones
in the decision on the protection and preservation of a particular cultural property.
A group of selected experts opted for two ways of evaluating the indicators. The
approach in evaluating the indicators, given by fuzzy numbers and intervals, influences
the concluding ranking results.
The paper is divide into three sections. In the first section, the indicators that affect
the management of cultural built heritage in a smart city have defined. The second
section defines applied algorithms of Fuzzy and Interval AHP method, through
appropriate steps, while the third section gives obtained results and rankings. In con-
clusion, final remarks have presented comparing results from both methods.
2 Indicators for Managing Cultural Built Heritage
Based on the previous studies on smart and sustainable cities, as well as the preser-
vation and protection of cultural heritage, the indicators are divided into six groups:
governing factors (A), economic factors (B), social factors (C), environmental factors
(D), technological factors (E) and architectural factors (F) [6, 7]. Governing factors
refer to indicators that are related to all aspects of heritage management at different
hierarchical levels, including actors involved in this process, as well as legislative and
other planning and legal frameworks for implementing management procedures. They
obtain:
Managing Cultural Built Heritage in Smart Cities 601
• development of a strategic and legal framework for managing and protecting the
cultural built heritage and harmonization with international standards (A1);
• a collaboration between different stakeholders (institutions for the protection and
promotion of cultural heritage, non-government organizations, ministries, etc.) (A2);
• increasing transparency and accessibility of information, written and digitalized
material on cultural heritage to the public (A3);
• active participation of the citizens in heritage management (A4).
• Economic factors include indicators related to investing in the restoration and
management or financial analysis of investments and returns. In this study, they
obtain:
• the rate of economic income after the heritage restoration and activation (B1);
• investment costs on heritage restoration (studies, state analyzes, development of
reconstruction plan, realization) (B2);
• investment costs on heritage promotion (organization of public debates, exhibitions,
production of written and digitized catalogs, marketing, tourist offers) (B3).
Social factors concern the influence that society has on the perception of cultural
heritage and preferences regarding its preservation and protection, but also inver-
sely, the impact which heritage has on urban development. They obtain:
• jobs opportunities for the population through activation of cultural heritage (C1);
• creating urban identity and reflecting the historical and cultural development of the
local community (C2);
• connecting citizens with the importance and value of cultural heritage (C3).
Environmental factors are related to the impact of existing heritage on the envi-
ronment and its restoration and the potential for pollution degree. They obtain:
• degree of pollution and the amount of waste generated during the restoration (D1);
• degree of pollution that comes with the cultural built heritage and as such endangers
environment (military and industrial abandoned heritage) (D2);
• recycling of the cultural built heritage through the development of revitalization and
adaptive reuse plans, as well as sustainable regeneration strategies (D3).
Technological factors relate to the modern application of technologies in the search,
digitization, mapping, and promotion of cultural heritage. They obtain:
• mapping of cultural built heritage using GIS and creation of databases (E1);
• development of infrastructure that will enable the continuous interoperable process
of digitalization of cultural heritage and networking (E2);
• development, and implementation of applications, platforms and other multimedia
solutions for education, as well as to promote local heritage for tourists (E3);
• implementation of smart applications for cultural heritage diagnosis (E4).
Finally, architectural factors relate to the real possibilities that built heritage offer in
terms of restoration and recognized cultural and historical values as a reflection of
different architectural styles. They include:
• state of the heritage (structural integrity, degree of material conservation, etc.) (F1);
• the ability to use existing spatial capacities and infrastructure for the new purpose (F2);
• the lifespan of the building after restoration (F3);
• authenticity and architectural value of the cultural built heritage (F4);
• the scope of rehabilitation and restoration construction works (F5).
3 Fuzzy and Interval AHP Algorithm
The Fuzzy AHP method is an extension of the crisp AHP method, in which human
estimates have presented as fuzzy values. Many methods and applications of fuzzy
AHP have expressed by numerous researchers. The fuzzy analytic hierarchy process
(FAHP) method has used to determine the preference weightings of criteria by sub-
jective perception. The decision-maker’s uncertainty can also have solved by using the
Interval AHP (IAHP) method, which uses interval comparison matrices where crisp
numerical values have replaced by intervals ½a; b; a b where a represents the lower
bound, and b the upper.
3.1 Fuzzy AHP Algorithm

Step 1. The overall goal is identified and clearly defined.
Step 2. The criteria and sub-criteria that contribute to the overall goal are identified and
the hierarchical structure is formed.

Step 3. Using triangular fuzzy numbers we form a comparison matrix A e¼ e a ij n n for
a fuzzy comparison of alternatives by pairs, where e a ij is a fuzzy value that expresses
the relative importance of one alternative to another. For such defined fuzzy values
e
a ij ¼ ð1; 1; 1Þ; if i ¼ j and e a ij ¼ 1= e
a ij for i 6¼ j: The meaning of triangular fuzzy
numbers according to Saaty’s scale is: Equal importance ð1; 1; 1 þ dÞ Weak dominance
ð3 d; 3; 3 þ dÞ; Strong dominance ð5 d; 5; 5 þ dÞ; Demonstrated domination
ð7 d; 7; 7 þ dÞ; Absolute domination ð9 d; 9; 9Þ: To get the most consistent solu-
tion, we used a fuzzy distance d ¼ 2: For values 2, 4, 6 and 8 we used the fuzzy
distance d ¼ 1 [8].
Step 4. Using
the procedure of the crisp AHP method, based on the comparison matrix
A ¼ aij n n , we calculate the values CI ¼ ðkmax nÞ=ðn 1Þ and CR ¼ CI=RI;
where kmax is the maximum eigenvalue of a matrix A and RI is a random index
(consistency index of the matrix). The value CI ¼ ðkmax nÞ=ðn 1Þ indicates the
consistency of the comparison matrix A e [9].
Step 5. For matrix A, e the priority fuzzy weighting vectors Mi ¼ ðli ; mi ; ui Þ; i ¼
1; . . .; n; are evaluated using the eigenvalue method [10].
Step 6. For the obtained triangular fuzzy numbers Mi the total integral value is cal-
culated as follows [11]:
wi ¼ ITk ðMi Þ ¼ 0:5ðkui þ kmi þ ð1 kÞli Þ; k 2 ½0; 1; ð1Þ
where k represents an optimism index. It describes the decision maker’s attitude toward
risk: the smaller value of k indicates a higher degree of risk (a lower degree of
optimism). Values 0, 0.5, and 1 have used to represent the pessimistic, moderate, and
optimistic views of the decision-maker, respectively.
3.2 Interval AHP Algorithm

Step 1 and Step 2 are the same as in the method FAHP.
Step 3. We construct an interval pairwise comparison matrix [12]:
2 3
1 ½a12 ; b12 ... ½a1n ; b1n
6 ½a21 ; b21 1 ... ½a2n ; b2n 7
6 7
Aint ¼ 6 .. .. . 7 ð2Þ
4 . . 1 .. 5
½an1 ; bn1 ½an2 ; bn2 ... 1
In the matrix Aint for all i; j ¼ 1; 2; . . .; n inequalities aij bij ; aij 0; bij 0; aij ¼
1=bij and bij ¼ 1=aji are satisfied. So the matrix Aint is a reciprocal matrix.

Step 4. Based on the matrix Aint , we construct matrices P ¼ pij nn , Q ¼ qij nn and

R ¼ rij nn :
8 8
< bij ; i\ j < aij ; i\ j
pij ¼ 1; i ¼ j ; qij ¼ 1; i ¼ j ; rij ¼ paij q1a
ij ; i; j ¼ 1; 2; . . .; n; ð3Þ
: :
aij ; i [ j bij ; i [ j
where a is an arbitrary parameter for which double inequality applies 0 a 1:

Step 5. The introduction of non-interval P and Q matrices allows testing the consis-
tency of the interval matrix Aint by examining the consistency of matrices P and Q. If
matrices P and Q have an acceptable value CR 0:1; that is, if they are consistent, then
an interval matrix Aint will also be consistent.
Step 6. To calculate the interval weights of an interval matrix Aint , we will use the
method of convex combination : Suppose that w ðRÞ is a vector containing weights of a
Q 1=n
n
matrix R, where wi ðRÞ ¼ j ¼ 1 rij , for all i ¼ 1; 2; . . .; n and all 0 a 1. If
Qn Qn 1=n Q 1=n
n a 1a
i¼1 w i ¼ 1; then w i ð R Þ ¼ j¼1 r ij ¼ p q
j ¼ 1 ij ij ¼ wai ðPÞw1a
i ðQÞ;
where wðPÞ and wðQÞ are weighting vectors of matrices P and Q, respectively. For a
matrix Aint can be obtained, based on the weighting vector w of the matrix R, interval
weight w ðAint Þ, where
wi ðAint Þ ¼ ½minfwi ðRÞj0 a 1g; maxfwi ðRÞj0 a 1g

ð4Þ
¼ ½minfwi ðPÞ; wi ðQÞg; maxfwi ðPÞ; wi ðQÞg:
Step 7. We calculate the probability that one interval weight is bigger than
h the other
i
L U
[13]. Interval weight wi ¼ wi ; wi is bigger than the interval weight wj ¼ wLj ; wU
j if

P wi wj [ P wj wi
with probability

max 0; wU
i wLj max 0; wLi wU
j
pij ¼ P wi wj ¼ ; ð5Þ
ðwU L U L
i wi Þ þ ðwj wj Þ
for all i; j ¼ 1; 2; . . .; n; i 6¼ j:
In the case where the interval weights wi and wj are equal, then it is pij ¼ 0:5:
Specially pii ¼ 0:5; for all i =1,…,n. By calculating comparison probabilities for all

intervals, we obtain a probability matrix (preferences) Pp ¼ pij :
nn
Step 8. Based on the probability matrix obtained in this way, we rank the interval
weights using the row-column elimination method [14].
4 Results
Table 1 provides a fuzzy matrix of criteria comparison obtained by experts. Based on

the obtained value CR < 0.1, we conclude that the comparison matrix is consistent.
Table 1. Pairwise comparison matrix of criteria, CI = 0.0244927, CR = 0.0197522.

FAHP A F B E D C
A (1,1,1) (1,2,3) (1,3,5) (3,4,5) (3,5,7) (5,6,7)
F (1/3,1/2,1) (1,1,1) (1,2,3) (1,3,5) (3,4,5) (3,5,7)
B (1/5,1/3,1) (1/3,1/2,1) (1,1,1) (1,2,3) (1,3,5) (3,4,5)
E (1/5,1/4,1/3) (1/5,1/3,1) (1/3,1/2,1) (1,1,1) (1,2,3) (1,3,5)
D (1/7,1/5,1/3) (1/5,1/4,1/3) (1/5,1/3,1) (1/3,1/2,1) (1,1,1) (1,2,3)
C (1/7,1/6,1/5) (1/7,1/5,1/3) (1/5,1/4,1/3) (1/5,1/3,1) (1/3,1/2,1) (1,1,1)
As for the criteria, the experts provided comparison matrices for all six groups of
sub-criteria, which were not listed due to space savings. Table 2 gives the corre-
sponding interval comparison matrix for the criteria. Since CRp and CRq are less than
0.1, we note that the interval matrix is consistent.
Table 2. Interval pairwise comparison matrix of criteria, CIp = 0.0145849, CRp = 0.011762,
CIq = 0.0370986, CRq = 0.0299183.
IAHP A F B E D C
A [1] [1, 2] [3] [3, 4] [4] [5]
F [1/2,1] [1] [2, 3] [3] [3, 4] [4, 5]
B [1/3,1/3] [1/3,1/2] [1] [1, 2] [2] [3]
E [1/4,1/3] [1/3,1/3] [1/2,1] [1] [1, 2] [2, 3]
D [1/4,1/4] [1/4,1/3] [1/2,1/2] [1/2,1] [1] [2]
C [1/5,1/5] [1/5,1/4] [1/3,1/3] [1/3,1/2] [1/2,1/2] [1]
As with the fuzzy AHP method, the experts also gave interval matrices of com-
parison for the six sub-criteria groups. In Table 3, the indicators have ranked using the
FAHP and IAHP method. We denote by wc the weight of the criterion, and by the wsc
of the weight of the sub-criterion. By creating two comparison matrices, we wanted to
discover the extent to which different representations of comparison matrices and
further analysis may affect the final order of the indicators.
Table 3. Ranking of indicators by FAHP and IAHP.

FAHP k¼1 k = 0.5 k=0 IAHP
wc wsc wc∙wsc wc∙wsc wc∙wsc wL wU p*
A1 0.326 0.436 0.142 A1 0.144 A1 0.148 A1 5.255 6.588 1
B1 0.180 0.529 0.095 A2 0.096 A2 0.109 F4 3.862 4.640 0.9923
F4 0.250 0.374 0.093 F4 0.094 F4 0.096 A2 3.117 3.962 0.8534
A2 0.326 0.280 0.091 B1 0.093 B1 0.087 F1 2.911 3.498 1
A4 0.326 0.210 0.068 A4 0.067 F1 0.067 F2 2.276 2.529 0.8411
F1 0.251 0.265 0.066 F1 0.067 A4 0.064 E2 1.877 2.500 0.6598
B2 0.180 0.298 0.053 B2 0.054 B2 0.055 A4 1.800 2.367 1
E2 0.125 0.427 0.053 E2 0.052 E2 0.049 E1 1.630 1.742 1
F2 0.251 0.188 0.047 F2 0.045 F2 0.042 F3 1.354 1.413 0.8648
D3 0.074 0.529 0.039 D3 0.038 D3 0.035 B1 0.978 1.447 0.8594
E1 0.125 0.297 0.037 E1 0.035 B3 0.031 A3 1.015 1.073 1
B3 0.180 0.172 0.031 B3 0.031 E1 0.030 D3 0.832 1.009 0.6695
F3 0.251 0.109 0.027 F3 0.027 A3 0.027 B2 0.796 0.978 0.6842
A3 0.326 0.072 0.023 A3 0.024 F3 0.027 F5 0.847 0.860 1
C2 0.042 0.529 0.022 C2 0.022 D2 0.022 D2 0.555 0.832 0.5005
E3 0.125 0.177 0.022 D2 0.022 C2 0.022 C2 0.658 0.729 0.9937
D2 0.074 0.298 0.022 E3 0.021 E3 0.019 E3 0.580 0.667 1
F5 0.251 0.062 0.015 F5 0.015 F5 0.016 E4 0.443 0.541 0.7908
C3 0.042 0.298 0.012 C3 0.013 C3 0.013 B3 0.438 0.489 1
D1 0.074 0.172 0.012 D1 0.012 D1 0.012 C3 0.362 0.401 0.6886
E4 0.125 0.098 0.012 E4 0.012 E4 0.011 D1 0.305 0.416 1
C1 0.042 0.172 0.007 C1 0.007 C1 0.007 C1 0.199 0.220
The obtained results show, in this case, that the use of fuzzy numbers and intervals
does not prioritize the same indicators. We took an optimistic view. The results of
applying the FAHP method significantly over the IAHP method favor the economic
indicator of financial gains after the restoration and activation of cultural heritage as one
of the significant in the decision-making process which cultural asset is of importance
for restoration and protection. In the same way, the importance of investment costs in
the process of cultural heritage restoration by the FAHP method has emphasized.
5 Conclusion
Heritage management is a global challenge that seeks equal attention and importance in
society, which constantly strives for sustainable development principles and smart
cities. In this paper, we have explored the issue of managing built heritage, as a part of
cultural heritage, in urban areas that actively apply the principles of a smart city. It has
observed that the FAHP method emphasizes external factors, while the IAHP method
places the heritage itself first. The result of applying the FAHP method does not
emphasize technological factors, so although we are talking about smart cities, it seems
that technologies in themselves are a secondary tool when it comes to the issues of
cultural heritage management. From another side, the development of infrastructure
that will facilitate the digitization process and thus enable the integration of cultural
heritage at national as well as at the global level is at the top of the priority indicators
by the IAHP method. Energy rehabilitation seems to become a strategically necessary
and legally obligatory task. The future research aim is to identify the most optimal
measures to be implemented to increase energy efficiency, depending on the degree of
protection of the building obtained by the fuzzy approach.
Acknowledgments. This work has been supported by the Ministry of Education, Science and
Technological Development of the Republic of Serbia.
References
1. Li, C.X., Fong, P.S.W., Dai, S., Li, Y.: Towards sustainable smart cities: an empirical
comparative assessment and development pattern optimization in China. J. Cleaner Prod.
215, 730–743 (2019)
2. Camero, A., Alba, E.: Smart City and information technology: A review. Cities 93, 84–94
(2019)
3. Eremia, M., Toma, L., Sanduleac, M.: The smart city concept in the 21st century. Proc. Eng.
181, 12–19 (2017)
4. Soini, K., Dessein, J.: Culture-sustainability relation: towards a conceptual framework.
Sustainability 8, 167 (2016)
5. Rouhi, J.: Definition of cultural heritage properties and their values by the past. Asian J. Sci.
Technol. 8(12), 7109–7114 (2017)
6. Li, J., Krishnamurthy, S., Pereira-Roders, A., Wesemael, P.: Community participation in
cultural heritage management: a systematic literature review comparing Chinese and
international practices. Cities 96, 102476 (2020)
7. Guzman, P.C., Pereira-Roders, A.R., Colenbrander, B.J.F.: Measuring links between cultural
heritage management and sustainable urban development: an overview of global monitoring
tools. Cities, 60(A), 192–201 (2017)
8. Srdjević, B., Medeiros, Y.: Fuzzy AHP assessment of water management plans. Water
Resour. Manage 22, 877–894 (2008)
9. Milošević, M.R., Milošević, D.M., Stević, D.M., Stanojević, A.D.: Smart city: modeling key
indicators in Serbia using IT2FS. Sustainability 11(13), 3536 (2019)
10. Dimić, V., Milošević, M., Milošević, D., Stević, D.: Adjustable model of renewable energy
projects for sustainable development: a case study of the Nišava district in Serbia.
Sustainability 10(3), 775 (2018)
11. Kahraman, C., Süder, A., Kaya., I.: Fuzzy multicriteria evaluation of health research
investments. Technol. Econ. Dev. Econ. 20(2), 210–226 (2014)
12. Ghorbanzadeh, O., Moslem, S., Blaschke, T., Duleba, S.: Sustainable urban transport
planning considering different stakeholder groups by an interval-AHP decision support
model. Sustainability 11(1), 9 (2019)
13. Liu, F.: Acceptable consistency analysis of interval reciprocal comparison matrices. Fuzzy
Sets Syst. 160, 2686–2700 (2009)
14. Wang, Y.M., Yang, J.B., Xu., D.L.: A two-stage logarithmic goal programming method for
generating weights from interval comparison matrices. Fuzzy Sets Syst. 152(3), 475–498
(2005)
Criteria Weighting for Blockchain Software
Selection Using Fuzzy AHP
Ferhat Karayazi(&) and Ilke Bereketli
Galatasaray University, 34349 Istanbul, Turkey

Abstract. Supply chain has an intrinsically complex structure with its perma-
nently growing supplier relationships, ever-increasing milestones, complexified
inventory across multiple parties, and complicated customer orders. Through-out
the chain, this complexity aggravates to monitor information and material flows
peer-to-peer, to do business with paper-based manual business processes and to
aim having an optimum cost by using limited resources. Blockchain system
brings reliability, integrity, and transparency into supply chain with its dis-
tributed, secured and immutable data structure. Today’s technology touching
solutions stand at a high price, and blockchain software including a high tech-
nology in its background is one of them. For this reason, it is crucial for com-
panies to select the most appropriate blockchain software in one go. The aim of
this study is to propose a multi-criteria decision model to assist a global logistics
company on the blockchain software selection problem using Buckley’s Fuzzy
Analytical Hierarchy Process (Fuzzy AHP) in accompany with expert opinions.
Fuzzy Logic is used in this study since some of the criteria are defined with
linguistic variables.
Keywords: Blockchain Software selection Logistics Fuzzy AHP
1 Introduction
Blockchain system has been widely recognized with bitcoin in the finance world,
however, with the understanding of its ability to know simultaneously who is per-
forming what kind of activities or actions including the information of location and
time through supply chain, blockchain has expanded widely its utilization. The first
idea of digitalizing of peer-to-peer business processes are based on Haber and Ston-
ernetta’s cryptographically secured chain of blocks generating time-stamping digital
documents [1], and developed more by use of Merkle Tree enabling to save lots of
documents into one block [2]. Following Nick Szabo’s bit gold [3] and Nakamato’s
white paper [4], the first source code for Bitcoin was released that would be open
source software [5]. Since then, blockchain technology’s applicability beyond currency
and its new opportunities have been understood. For instance, tracking shipping con-
tainers [6], preventing overfishing [7], detecting data manipulation [8], using block-
chain as an assurance service in diamond and jewellery supply chains [9], and
recording an art piece’s history [10] are outstanding examples to these opportunities.
Criteria Weighting for Blockchain Software Selection Using Fuzzy AHP 609
Firms take advantages of bringing reliability, integrity, and full transparency into
their business processes with blockchain’s distributed, secured and immutable data
structure [11]. For these reasons, a global logistics company in Turkey has decided to
make activities through its supply chain traceable and trackable by using blockchain
system in its logistics processes. To provide a basis for this business’s selection
problem, criteria are determined and weighted with Buckley’s Fuzzy Analytical
Hierarchical Process (AHP) in accompany with expert opinions from the company.
Technological solutions cause remarkable expenses for businesses. A wrong
selection can create a risk of reaching the expected utility value from the technological
investment in case an adopted solution implemented into business process is incapable
of realizing expectations from it. Therefore, it is crucial for companies to select opti-
mum blockchain software among alternatives to create a better collaborative business
environment and bring concrete and abstract benefits to organization [12].
Blockchain is relatively a hot debate both in literature and business life. The
literature has been flourished ever-increasingly with many academic works related
blockchain systems, however, there is not a suchlike work in literature. This study
proposes a comprehensive framework of blockchain software selection by presenting
an objective hierarchy, specifying related attributes, and weighting criteria in order to
provide a detailed instruction for evaluation of the potential blockchain softwares.
Following Introduction Section, the organization of the study as it follows: In
Sect. 2, studies related selection in blockchain systems are reviewed. In Sect. 3, steps
of the proposed method is given, and criteria weights are calculated. In Sect. 4, weights
of criteria are examined in detail. In Sect. 5, we conclude by giving a short evaluation
of results, and a further research suggestion.
2 Literature Review
Yizhong et al. proposed a fair selection protocol consisting of the mining process and
confirmation of the new nodes in a permissionless blockchain [13]. Liu et al. studied
transcoder selection for blockchain-enabled video streaming platforms [14]. Wu et al.
used scheduling algorithm to realized node selection for Sapiens Chain, which refers to
a cybersecurity framework of protection of the privacy of the anonymous users, by
covering and rewarding nodes based on history included performance [15]. Feng et al.
came up with a novel mechanism called proof-of-negotiation evaluating the trust-
worthiness of miners that gives opportunity to select random-honest miners for block
creation [16]. Gorczyca and Decker developed an approach by using Directed Acyclic
Graphs for witness, which is the only determinant of order of events in a distributed
ledger, selection [17]. Farshidi et al. suggested a decision model for blockchain plat-
form selection [18]. Wang et al. studied game-theoretic analysis of mining pool
strategy selection in proof-of-work-based blockchain networks [19].
610 F. Karayazi and I. Bereketli
3 Methodology
The main purpose of the study is to propose a multi-criteria decision model for the
strategic decision of blockchain software selection. We decided to use AHP Method in
the study since it is widely used in literature for weighting of decision criteria. Because
some of our criteria were not convenient to determine with a crisp value, and usage of
fuzzy numbers would assist in tackling with the vagueness of decision makers’ verbal
judgements in evaluation of criteria, triangular fuzzy numbers are utilized in pairwise
comparison. Buckley’s Fuzzy AHP, which is one of the extended fuzzy AHP methods,
is used in this study.
3.1 Criteria Weighting with Buckley’s Fuzzy AHP

The entire process in the study as it follows [20, 21, and 22]:
Step 1: Form hierarchical structure.

Step 2: Make pairwise comparison using linguistics variables in Table 1.
Table 1. Linguistics variables for pairwise comparison.

Abbreviation Linguistics variable Triangular fuzzy number
VL Very Low (1, 1, 3)
L Low (1, 3, 5)
M Medium (3, 5, 7)
H High (5, 7, 9)
VH Very High (7, 9, 9)
Step 3: Construct pairwise comparison matrix.

2 3
~ak11 ~ak12 ... ~ak1n
6 ~ak21 ~ak12 ... ~ak1n 7
~ 6 7
A ¼6 .
k
.. .. .. 7 ; ð1Þ
4 .. . . . 5
~akn1 ~akn2 ... ~aknn
where ~akij is kth decision maker’s preference of ith criterion over j criterion.
Step 4: Construct aggregated decision matrix using weighted average method in case
there is more than one decision maker.
P
K
~akij
k¼1
~aij ¼ ð2Þ
K
~ is n n matrix.
Step 5: Update pairwise comparison matrix. A
2 3
~a11 ~a12 ... a~1n
6 ~a21 ~a22 ... ~a2n 7
~¼6
A 6 .. .. ..
7
.. 7 ð3Þ
4 . . . . 5
~an1 ~an2 ... ~ann
Step 6: Calculate the geometric mean of each parameter of ~

akij .
!1=n
Y
n
~ri ¼ ~aij ; where i; j ¼ 1; 2. . .; n ð4Þ
j¼1
Step 7: Find the vector summation of each ~ri . Then find the −1 power of them and
replace them with triangular fuzzy numbers to find the weight of each criterion.
~ i ¼ ~ri ð~r1 ~r2 . . .~rn Þ1 ;

w
ð5Þ
¼ ðlwi ; mwi ; uwi Þ
where i = 1, 2, …, n, and (l, m, v) refers to a triangular fuzzy number.

Step 8: Fuzzy weight of w ~ i is defuzzified with Chou and Chang’s Centre of Area
Method [23].
lwi þ mwi þ uwi

Mi ¼ ð6Þ
3
Step 9: Mi is a non-fuzzy number and requires to be normalized so that sum of weights

of criteria is equal to 1. An obtained normalized value denotes Ni.
3.2 Application of the Proposed Method

Step 1: As given in Table 2, criteria and sub-criteria are determined in hierarchical
structure.
Step 2–3: Each decision maker’s pairwise comparison for main criteria are as it is in
Table 3. Three experts who are IT Group Manager, R&D Group Manager and Business
Analyst respectively from the company accompanied to decision process. Their
opinions were obtained via pairwise comparisons were realized by them and then
aggregated with weighted average method.
Step 4–5: Aggregated decision matrix is calculated using (2) and (3) as in Table 4.
Step 6: Using (4), geometric mean of each main criterion is calculated as in Table 5.
~ is converted to n 1 matrix.
With this step, A
Step 7: With summation of each colunmn’s geometric mean value, we obtain vector
summation values. Then, reverse of vector summation is realized in Table 5. For
instance, reverse of vector summation is 1/(0.14) for 7.37. These reverse values are
sequenced as increasing order. Using (l, m, w) values (it is (0.08, 0.1, 0.14) in Table 5
weights of criteria are calculated, and demonstrated in Table 6.
Table 2. Hierarchical structure.

Main criteria Sub-criteria
C1: Cost C11: Purchase Cost
C12: Maintenance Cost
C13: Additional Cost
C2: Speed C21: Encryption Speed
C22: Speed of Transactions
C23: Latency in Results of Queries
C3: Privacy Concerns C31: Security
C32: Anonymity
C33: IP Tracking
C4: Logistics Issues C41: Work-in-harmony With Internet of Things
C42: Interoperability With Other Supply Chains
C43: Range of Services Provided
C44: Adaptability to Logistics Business Process
C5: Functionality –
C6: Developer Availability –
Table 3. Decision Makers(DM)’ pairwise comparisons for main criteria.

DM1 C1 C2 C3 C4 C5 C6 DM2 C1 C2 C3 C4 C5 C6 DM3 C1 C2 C3 C4 C5 C6
C1 1 M H L VH VH C1 1 L L M H VH C1 1 H M M H VH
C2 1 H M VH VH C2 1 M L H VH C2 1 L H VH VH
C3 1 H H VH C3 1 M L VH C3 1 M H H
C4 1 H VH C4 1 VH VH C4 1 H H
C5 1 M C5 1 H C5 1 VH
C6 1 C6 1 C6 1
Table 4. Aggregated matrix is constructed according to decision makers’ preferences.

C1 C2 C3 C4 C5 C6
C1 (1, 1, 1) (3, 5, 7) (3, 5, 7) (7/3, 13/3, 19/3) (17/3, 23/3, 9) (7, 9, 9)
C2 (1/7, 2/9, 1/2) (1, 1, 1) (11/3, 17/3, 23/3) (3, 7, 7) (19/3, 25/3, 9) (7, 9, 9)
C3 (1/7, 2/9, 1/2) (1/8, 2/7, 1/6) (1, 1, 1) (11/3, 17/3, 23/3) (11/3, 17/3, 23/3) (19/3, 25/3, 9)
C4 (1/8, 1/6, 2/7) (1/7, 2/9, 1/2) (1/8, 1/6, 2/7) (1, 1, 1) (17/3, 23/3, 9) (19/3, 25/3, 9)
C5 (1/8, 1/8, 1/6) (1/9, 1/8, 1/6) (1/7, 1/5, 1/2) (1/9, 1/8, 1/6) (1, 1, 1) (5, 7, 25/3)
C6 (1/9, 1/9, 1/7) (1/9, 1/9, 1/7) (1/9, 1/8, 1/6) (1/9, 1/8, 1/6) (1/8, 1/7, 2/9) (1, 1, 1)
Table 5. Geometric mean of each main criterion.

Geometric mean of criteria Fuzzy number of
geometric mean
(Step 6) ~r1 (3.07, 4.42, 5.41)
~r2 (2.05, 2.8, 3.61)
~r3 (1.09, 1.49, 2.07)
~r4 (0.68, 0.88, 1.23)
~r5 (0.32, 0.28, 0.52)
~r6 (0.16, 0.17, 0.02)
(Step 7) Vector Summation 7.37 10.1 13.1
Reverse (power of −1) 0.14 0.1 0.08
Increasing Order 0.08 0.1 0.14
Step 8–9: Mi values are computed using (6). Then normalized weights are calculates as
it is in Table 6. Sum of Nis equal to 1.
Table 6. Fuzzy weights of criteria are obtained using (l, m, w) values.

Geometric mean of criteria Fuzzy weights Mi N i
(Step 7) ~r1 (0.23, 0.44, 0.73) (Step 8) 0.47 0.42
~r2 (0.16, 0.28, 0.49) 0.31 0.28
~r3 (0.08, 0.15, 0.28) 0.17 0.15
~r4 (0.05, 0.09, 0.17) 0.1 0.09
~r5 (0.02, 0.04, 0.07) 0.04 0.04
~r6 (0.01, 0.02, 0.03) 0.02 0.02
4 Results
As it is shown in Table 6, Cost criterion is the dominant weight with 0.42, and Speed
follows it with 0.28. With Privacy Concerns having third biggest weight, these three
weights account for 0.85, which purports they will take a big role in blockchain
software selection problem. Logistics Issues has relatively bigger weight than Func-
tionality and Developer Availability, however, their effects will be less with 0.15
weight selection procedure. This result shows us that software cost, software speed, and
privacy concerns are main determinants of the selection.
Using same steps given in methodology, sub-criteria weights are calculated and
overall weights of them computed as it is in Table 7.
Table 7. Weights of criteria.

Main criteria Sub-criteria Weights Overall
C1: Cost C11: Purchase Cost 0.760 0.32
C12: Maintenance Cost 0.182 0.08
C13: Additional Cost 0.058 0.02
C2: Speed C21: Encryption Speed 0.656 0.18
C22: Speed of Transactions 0.274 0.08
C23: Latency in Results of Queries 0.070 0.02
C3: Privacy Concerns C31: Security 0.703 0.11
C32: Anonymity 0.242 0.04
C33: IP Tracking 0.055 0.01
C4: Logistics Issues C41: Work-in-harmony With IoT 0.608 0.06
C42: Interoperability With Other Supply 0.218 0.02
Chains
C43: Range of Services Provided 0.078 0.01
C44: Adaptability to Logistics Business 0.096 0.01
Process
C5: Functionality – 0.04
C6: Developer – 0.02
Availability
5 Conclusion
A real world’s strategic decision of blockchain software selection is introduced. Our

aim for this study is to provide a multi criteria decision-making model for blockchain
software selection problems. With this study, a basis for blockchain software selection
is provided in terms of specifying and weighting criteria. Results reveal that purchase
cost, encryption speed of the software, and security are highly important in evaluation
of alternative with the weight of 0.61 in total. With the 0.85 weight in overall, criteria
of Cost, Speed and Privacy Concerns are dominant criteria through selection process.
With the sum of 0.48 weight, Speed and Privacy Concern are slightly more significant
in the determination of optimum alternative compared to Cost criterion having 0.47
weight. The rest of the main criteria with 0.15 weight in total expresses that technical
concerns related to blockchain software remain at the forefront of Logistics Issues,
Functionality and Developer Availability. For further researches, we will apply our
model in other extensions of Fuzzy AHP methods in order to observe how these
methods can affect the criteria weights.
References
1. Haber, S., Stornetta, W.S.: How to time-stamp a digital document. J. Cryptol. 3, 99–111
(1991)
2. Tijan, E., Aksentijevic, S., Ivanic, K., Jardas, M.: Blockchain technology implementation in
logistics. Sustainability 11, 1185 (2019)
3. Szabo, N.: Bit Gold. https://nakamotoinstitute.org/bit-gold/. Accessed 03 Apr 2020
4. Nakamato, S.: Bitcoin: A Peer-to-Peer Electronic Cash System. https://bitcoin.org/bitcoin.
Accessed 03 Apr 2020
5. Swan, M.: Blockchain: Blueprint for a New Economy, 1st edn. O’Reilly Media, Sebastopol
(2015)
6. IT World Canada. https://www.itworldcanada.com/article/canadas-first-commercial. Acces-
sed 03 Apr 2020
7. Forbes. https://www.forbes.com/sites/keshiahannam/2016/09/30/this-emerging-tech-
company-has-put-asias-tuna-on-the-blockchain/. Accessed 03 Apr 2020
8. CoinDesk. https://www.coindesk.com/defense-giant-lockheed-martin-integrates-blockchain.
9. The TrustChain Initiative. https://www.trustchainjewelry.com/. Accessed 03 Apr 2020
10. VERISART. https://verisart.com/. Accessed 03 Apr 2020
11. Perboli, G., Musso, S., Rosano, M.: Blockchain in logistics and supply chain: a lean
approach for designing real-world use cases. IEEE Access 6, 18–28 (2018)
12. Meng, F., Wang, N., Xu, Y.: Triangular fuzzy neutrosophic preference relations and their
app in enterprise resource planning software selection. Cogn. Comput. 12, 261–295 (2019)
13. Yizhong, L., Jianwei, L., Zongyang, Z., Hui, Y.: A fair selection protocol for committee-
based permissionless blockchains. Comput. Secur. 91, 101718 (2020)
14. Liu, M., Yu, F.R., Teng, Y., Leung, V.C.M., Song, M.: Deep reinforcement learning (DRL)-
based transcoder selection for blockchain-enabled video streaming. In: 2018 IEEE
Globecom Workshops (GC Wkshps), Abu Dhabi, pp. 1–6 (2018)
15. Wu, Y., Wang, Z., Ruan, Q., Shi, J., Fang, B.: Node scheduling: a blockchain-based node
selection approach on sapiens chain. In: 2019 IEEE International Conference on Big Data
and Smart Computing (BigComp), Kyoto, Japan, pp. 1–7 (2019)
16. Feng, Y., Zhao, X., Chen, K., Zhao, F., Zhang, G.: Towards random-honest miners selection
and multi-blocks creation: proof-of-negotiation consensus mechanism in blockchain
networks. Future Gener. Comput. Syst. 105, 248–258 (2020)
17. Gorczyca, A., Decker, A.: Distributed ledger witness selection in bounded width directed
acyclic graphs. In: 2019 IEEE International Conference on Blockchain and Cryptocurrency
(ICBC), Seoul, Korea (South), pp. 124–127 (2019)
18. Farshidi, S., Jansen, S., España, S., Verkleij, J.: Decision support for blockchain platform
selection: three industry case studies. IEEE Trans. Eng. Manag. 1–20 (2020)
19. Wang, Y., Tang, C., Lin, F., Zheng, Z., Chen, Z.: Pool strategies selection in PoW-based
blockchain networks: game-theoretic analysis. IEEE Access 7, 1 (2019)
21. Buckley, J.J.: Fuzzy hierarchical analysis. Fuzzy Sets Syst. 17, 233–247 (1985)
22. Ayhan, M.B.: A fuzzy AHP approach for supplier selection problem: a case study in a
gearmotor company. Int. J. Managing Value Supply Chains 4, 11–23 (2013)
23. Chou, S.Y., Chang, Y.H.: A decision support system for supplier selection based on a
strategy-aligned fuzzy SMART approach. Expert Syst. Appl. 34(4), 2241–2253 (2008)
A Proposed Decision Making Methodology
to Select IT Suppliers for Software
Development Outsourcing in Banking Sector
Using EDAS Method
Murat Levent Demircan1(&) and Algı Acarbay2

1
Galatasaray University, Ortaköy, 34349 Istanbul, Turkey
[email protected]
2
Payment Gateway Department, Wirecard, Munich, Germany
Abstract. Payment is an important part of people’s daily lives, and although

both online and offline transactions made by credit cards increase every day with
the growing interest in e-commerce, 85–90% of worldwide payment transac-
tions are still made by cash. However, the banks, giant payment institutions like
Visa and MasterCard, and Payment Service Provider companies aim to increase
the number of card transactions and enhance their platforms by supporting a
wide range of channels with digital solutions, enhancing the security and pro-
viding value-added services. Having the user-friendly and secure payment
platforms, and easy and quick integration opportunities with the merchants are
quite critical for the banks; which shift them to outsource Virtual Point of Sale
platform and focus on their core business. In this article, the vendor selection
process of a bank is researched and Evaluation based on Distance to Average
Solution (EDAS) method by Interval Type-2 Fuzzy Sets (IT2FS) is used for
vendor selection.
Keywords: Multiple criteria decision making EDAS method Software

outsourcing
1 Introduction
Although about 85% to 90% of the transactions are still cashed today, the payment
world is becoming more digitalized every day. Visa and MasterCard are still the
biggest payment institutions, aiming to decrease the cash transactions and making
people prefer card transactions more both in virtual and physical environments. The
time’s importance, the companies’ improved delivery services, the easiness of refunds,
and the technological innovations push individuals to make online purchases more and
more. That makes a strong Virtual Payment Gateway infrastructure a major issue for
the banks. An online standard authorization transaction process is shown in Fig. 1.
3D (Domain) Secure process prevent fraud in online transactions. Three domains
are Acquirer Domain, Interoperability Domain, and Issuer Domain including Merchant
Plug-in (MPI), Directory Server (DS), and Access Control Service (ACS) additionally
A Proposed Decision Making Methodology to Select IT Suppliers 617
Fig. 1. Authorization process.
to the Merchant Website, Acquirer Bank and Issuer Bank. 3D Secure’s aim is to
authenticate the cardholder, and it is performed by a static password or a dynamic
password that is generated during the payment process. When the cardholder enters that
password in the authentication page, Issuer Bank confirms the transaction is made by
the real cardholder; and becomes liable for the transaction. Liability means being
responsible for the chargeback, in case of fraud.
Since the technology and our lives have changed dramatically since 2002, 3DSe-
cure v.1.0 is insufficient in a lot of aspects. EMVco declared 3DS Secure version 2.0 in
2016 as with the improvements (see Fig. 2).
Fig. 2. 3D secure 2.0 workflow [1].
Section 2 represents literature review. In Sect. 3, the criteria and sub-criteria for the
decision-making problem are defined and explained, and 4 alternatives described
briefly. In Sect. 4 the proposed model is explained, and the numerical results are
618 M. L. Demircan and A. Acarbay
provided in Sect. 5. Finally, in Sect. 6 conclusion is introduced by the evaluation of the

model and numerical results.
2 Literature Review
A comprehensive definition of IT outsourcing is an organization’s decision to contract

out part of or all software development activities to a third party suppliers who in return
is responsible for providing and managing software development activities and services
to its customer for monetary returns over a decided time period based on service level
agreements [2]. Third party suppliers consist of consulting firms, technology vendors,
and system integrators and contractors [3].
According to Gilley and Rasheed, in another perspective, outsourcing is “the
fundamental decision of rejecting the internalization of an activity” [4]. Lacity and
Willcocks [5] stated that two main reasons behind the preference of IT outsourcing in
the industry as focus on core business and the perception of Information Systems
function as a cost charge considering it is a non-core activity. Additionally to those
factors, IT outsourcing has the advantage of improving the product or service quality,
increasing the customer satisfaction [6]. Another benefit of T outsourcing is lower risk
of obsolescence, and companies manage it by renting the technology instead of pur-
chasing it. Moreover, companies choose IT outsourcing due to the advantage of saving
the staff costs; since it is easier to work with a person already specialized in that area,
instead of training its own employee [7].
In addition, by outsourcing most companies transferring the responsibility to their
vendors which might be quite attractive especially for unexperienced and small com-
panies. MCDM is mostly used in vendor/supplier selection in both public and private
organizations. Determining the attributes to interpret and compare the alternatives
based on is the substantial part of MCDM process. A comprehensive study on the
analysis of vendor selection was made by Dickson [8]. Ghodsypour and O’Brien [9]
proposed a supplier selection method integrating AHP and linear programming by
approaching this problem with three main criteria; cost, quality, and service. Their
research provided a systematic and objective approach for the supplier selection, and
also expressed in addition to the tangible factors, the intangible factors should also be a
part of the assessment. Wang et al. [10] analyzed nine factors including price, location,
flexible contract terms, cultural match, reputation, existing relationship, commitment to
quality, scope of resources, added capability with a mixed Analytic Hierarchy Process
(AHP) and Preference Ranking Organization Methods for Enrichment Evaluations
(PROMETHEE) method for supplier selection. Venkatraman [11] identified the con-
cept of a value center for Information Technologies Resources, including Cost Center,
Profit Center, Service Center, and Investment Center components. He pointed out that
managing IT resources accurately based on those values also enable IT outsourcing to
be successful. Buck-Lew [12] emphasized the difference between hybrid outsourcing
and pure outsourcing Information Systems functions. The EDAS method stands for
Evaluation Based on Distance Average Solution, and it is a MCDM method that has
been developed lately. It was originally suggested by Ghorabaee and his friends, with
an experimental study in which EDAS method is compared with the other MCDM
techniques to prove it is a valid method [13]. While evaluating the alternatives, the
EDAS method considers the average solution. EDAS method was extended by
Ghorabaee et al. [14] with fuzzy logic and linguistic terms. EDAS method was inte-
grated with the neutrosophic cluster method by Peng and Liu [15]. It was combined
with the single-valued neutrosophic numbers, in a study which investigates the
selection of a software development project for an internet company. Kahraman et al.
[16] improved an intuitive fuzzy EDAS method by integrating the EDAS method with
an interval-valued intuitionistic fuzzy numbers. This method has been used in the
evaluation of solid waste disposal sites. To solve the problem of solid waste disposal
site selection, three different forms of EDAS method was used; crisp EDAS, ordinary
fuzzy EDAS and interval-valued intuitionistic fuzzy EDAS methods.
3 Software Development Outsourcing Vendor Selection

in Banking Sector
This section is about the criteria and sub-criteria specification and explanation.
Deciding the criteria has always been a major issue in the vendor selection process.
Weber et al. [18] evaluated 74 articles and pointed out that “net price”, “quality, and
“delivery” criteria are stated as the most considerable factors. Dickson [8] also made a
research on the factors to be scored during the vendor selection process, and found out
the elements to be decided depends on the product/service to be purchased. Garg et al.
[19] stated the high-level criteria were stated as accountability, agility, cost, perfor-
mance, assurance, security/privacy, and usability for Cloud Service Outsourcing. In
this article, the main criteria are economic factors, quality factors, and the service.
Besides, the explanations of the factors are made by considering a bank’s Virtual POS
provider company selection, which is the basic research subject of this article.
Initial Cost (C1): Initial cost is the fixed cost that is paid initially for the use of the
service.
Development Cost (C2): Software development cost contains all the development
related costs, after the initial purchase.
Change Request Cost (C21): Change requests are the additional demands for cus-
tomizations that are not met with the standard product.
Maintenance Cost (C22): Maintenance includes fixing the bugs that arise in the
production environment.
Quality (C3): Quality is another considerable factor to be evaluated while deciding
the Service Provider. In this study, bug-free and performance are selected as quality
sub-criteria.
Bug-Free (C31): Having a bug free product and deploying packages for additional
development without bug is quite an important factor.
Performance (C32): Performance is an element which should be reckoned during
the evaluation process of a product/service.
Availability (C4): Availability is defined as a customer’s access to a product/service.
Virtual POS should remain available for 7/24. In case of unavailability, the mer-
chant either proceeds with the backup acquirer bank or loses giro and both con-
sequences harm the acquirer bank of the merchant.
Reliability (C5): Reliability is an indicator of how a product/service operates

without collapsing along a specified time and condition. For this reason, it can also
be described in the meantime the Service Provider faces the failure [8]. Further-
more, the company reputation affects the reliability factor based on past
experiences.
Service (C6): The last main criteria to be analyzed in this article is the service.
Transparency, on-time delivery, and monitoring is the sub-criteria for the service.
Transparency (C61): Transparency is about Service Provider to have the system
clear and apparent to its client. For example, a system can be defined as transparent
if the client should be able to export the reports for transactions/activities, access the
support tickets, and sight what is going on in real-time.
On-time delivery (C62): Time is valuable to companies as much as it is to indi-
viduals in their daily lives. Failing about the on-time deliveries can even result in
the termination of the contract between the companies which makes this criterion a
very significant one.
Monitoring (C63): Monitoring is related to the continuous control of the system’s
working, and making sure that the functionality is smooth and the performance is
good. The advantage of the monitoring is the ability to notice a problem before it
arises, or the quick intervention in case of urgency.
Security (C7): Security is another key concern for the banks since both the personal
and sensitive data can be exposed in case of a security bug, which can result in a
disaster.
Data Security (C71): The important data stored in the database is encrypted,
however in case the encryption algorithm is found out, the data can be decrypted by
the attackers. The warehouse security is also significant in data security.
Connection Security (C72): Transaction channels are between the merchant and PSP
and between PSP and the bank. These channels’ security and resistance to the
attacks are major for the sensitive information to not be uncovered; otherwise, there
might be enormous financial losses. The criteria that were stated and explained in
the previous section are applied to three different privately held software companies
named A1, A2 and A3. We received the proposals from those companies and scored
the financial factors based on them. The proposals also can’t be shared also due to
privacy regulations. The quality and service factors are scored based on the repu-
tation and past experiences of the companies. Out of the companies, A1 is a
company that is new in the market. A2 is the market leader which provides out-
sourcing service to 18 banks both in Turkey and Europe. A3 is the company 3rd in
this sector which can be referred to as a medium scale as shown in Table 1.
Table 1. Information about software development company alternatives.

Alternative Scale Age Market share
A1 Small Less than 1 year Less than 5%
A2 Large 20 years More than 75%
A3 Medium 5 years 15%
4 Methodology
Ghorabaee et al. [13] firstly introduced the EDAS method and Ghorabaee et al. [15]
extended the EDAS method by using IT2FS. The definitions which are presented above
are used for extending the EDAS method by using IT2FS.
Suppose that we have a set of n alternatives A ¼ fA1 ; A2 ; . . .; An g, a set of m criteria
C ¼ fC1 ; C2 ; . . .; Cm g, and k decision makers D ¼ fD1 ; D2 ; . . .; Dk g. The steps of
EDAS IT2FS method are presented as follows [17];
Step 1: The average decision matrix (X), is defined as follows:
1
X ¼ ~~xij nxm ; ~~xij ¼ kp¼1 ~
~xpij ð1Þ
k
and ~~xpij denotes the performance value of alternative Ai ð1 i nÞ with respect to cri-
terion Cj ð1 i mÞ assigned by the pth decision maker ð1 p kÞ.
Step 2: The matrix of criteria weights, is defined as follows:

~~ j
W¼ w ~~ j ¼ 1 k w
; w ~
~p ð2Þ
1xm k p¼1 j
~~ p denotes the weight of criterion Cj ð1 i mÞ assigned by the pth decision maker
and w j
ð1 p kÞ.
Step 3: Determine the matrix of average solutions, shown as follows:
h i
~~ j
AV ¼ M ~~ j ¼ 1 ni¼1 ~
; M ~xpij ð3Þ
1xm n
~~ represents the average solutions with respect to each criterion. Therefore, the
M j
dimension of the matrix is equal to the dimension of criteria weights matrix.
Step 4: If B is the set of beneficial criteria and N is the set of non-beneficial criteria.
Then the matrices of positive distance from average (PDA) and negative distance from
average (NDA) are calculated with regard to the type of criteria as follows:

PDA ¼ ~p~ij nxm ; NDA ¼ ~
~
nij nxm ð4Þ
8 ~ 8 ~
> z ~~xij M ~j
> z M ~ j ~~xij
< ~~ if j 2 B
> < ~~ if j 2 B
>
~p~ij ¼ S~ M j ~~nij ¼ S M~j ð5Þ
>
> z ~ j ~~xij
M >
> z ~~x M ~
: ~ if j 2 N : ij ~ j if j 2 N
~j
S M ~j
S M
where ~~pij and ~n~ij denote the positive and negative distance of performance value of ith
alternative from the average solution in terms of jth criterion, respectively.
Step 5: The weighted sum of positive and negative distances for all alternatives are

f
e i ¼ m
sp ~~ j ~~pij fe i ¼ m ~
~j ~
j¼1 w sn j¼1 w nij ð6Þ
Step 6: The normalized values of f

e i and f
sp e i for all alternatives are calculated as
sn
follows:
f
e
sp f
e
sn
f
f
np i ¼ i f
f
nn i ¼ 1 i ð7Þ
f
ei
maxi S sp maxi S fei
sn
Step 7: The appraisal score ~~hi for all alternatives is calculated as follows:

~~h ¼ 1 f f
n p f
f
n n ð8Þ
i i i
2
Step 8: The method proposed by Ghorabaee et al. [20] is used in this step for com-
puting the ranking value (RV) of trapezoidal IT2FS. The alternatives according to the
decreasing ranking values (RV) of appraisal scores are ranked and the alternative with
the highest appraisal score is accepted as the best option among other alternatives.
5 Numerical Results
Based on linguistic variables corresponding IT2FS shown in Table 2, performance

values for A1, A2 and A3 are evaluated as follows: C1(ML, VH, H), C2(M, H, H),
C3(M, VH, H), C4(MH, H, MH), C5(MH, H, MH), C6(M, VH, MH), C7(MH, VH, H),
respectively. Based on average solutions shown in Table 3, ranking values for A1, A2
and A3 are calculated as 0.6734, 0.6802, 0.6072, respectively.
Table 2. Linguistic variables corresponding IT2FS values.

Linguistic variable Interval type-2 fuzzy sets
Very Low (VL) (0.00 0.00 0.00 0.10; 1.00 1.00) (0.00 0.00 0.00 0.05; 0.90 0.90)
Low (L) (0.00 0.10 0.15 0.30; 1.00 1.00) (0.05 0.10 0.15 0.20; 0.90 0.90)
Medium Low (ML) (0.10 0.30 0.35 0.50; 1.00 1.00) (0.20 0.30 0.35 0.40; 0.90 0.90)
Medium (M) (0.30 0.50 0.55 0.70; 1.00 1.00) (0.40 0.50 0.55 0.60; 0.90 0.90)
Medium High (MH) (0.50 0.70 0.75 0.90; 1.00 1.00) (0.60 0.70 0.75 0.80; 0.90 0.90)
High (H) (0.70 0.85 0.90 1.00; 1.00 1.00) (0.80 0.85 0.90 0.95; 0.90 0.90)
Very High (VH) (0.90 1.00 1.00 1.00; 1.00 1.00) (0.95 1.00 1.00 1.00; 0.90 0.90)
Table 3. Matrix of average solutions.

Criteria Fuzzy value Crisp score
C1 (0.20 0.25 0.27 0.30 0.38 0.38, 0.23 0.25 0.27 0.28 0.34 0.34) 0.26
C2 (0.21 0.28 0.29 0.34 0.38 0.38, 0.25 0.28 0.29 0.31 0.34 0.34) 0.28
C3 (0.21 0.27 0.29 0.32 0.38 0.38, 0.25 0.27 0.29 0.30 0.34 0.34) 0.27
C4 (0.20 0.26 0.28 0.33 0.38 0.38, 0.23 0.26 0.28 0.30 0.34 0.34) 0.27
C5 (0.21 0.28 0.30 0.35 0.38 0.38, 0.25 0.28 0.30 0.32 0.34 0.34) 0.29
C6 (0.23 0.29 0.30 0.33 0.38 0.38, 0.26 0.29 0.30 0.31 0.34 0.34) 0.29
C7 (0.27 0.33 0.34 0.36 0.38 0.38, 0.30 0.33 0.34 0.35 0.34 0.34) 0.33
6 Conclusion
In this study, a decision-making problem for IT outsourcing is illustrated with the

payment service provider selection problem of a bank. Study represents a unique
contribution to IT outsourcing evaluation based on methodology approach and based
on the domain specified. Moreover, since the problem is about payment systems
domain, a deep knowledge of the domain was provided. Alternative companies are
evaluated based on initial cost, development cost, quality, reliability, availability,
service, and security criteria. Security, quality, and service criteria are the highest
weights respectively; whereas initial cost and development cost criteria have the lowest
weights. A1 is a recently founded company, A2 is a large company that has the furthest
market share, where A3 is a medium scale company that has a market share of around
15%. A1 offers relatively low prices while other criteria is about medium level. A2 has
high prices but provides very high quality and security. A3 also has high prices
although not as much as A2 does, and offers high quality and service. Despite the fact
that all the alternatives are appealing, A2 is selected as the best option. However, it
should be mentioned that the overall ranking scores are quite close to each other. The
reason A2 stands out is that its product has very high quality and security. This result
indicates that the new companies in the market should have a very good strategy and
offering good prices is not adequate unless the expected service is not provided in terms
of quality, security, and service.
References
1. EMVCo: EMV ® 3-D Secure Protocol and Core Functions Specifications (2016)
2. Kern, T., Willcocks, L.: Exploring information technology outsourcing relationships: theory
and practice. J. Strateg. Inf. Syst. 9(4), 321–350 (2000)
3. Palvia, P.C.: A dialectic view of information systems outsourcing: pros and cons. Inf.
Manag. 29, 265–275 (1995)
4. Gilley, K.M., Rasheed, A.: Making more by doing less: an analysis of outsourcing and its
effects on firm performance. J. Manag. 26(4), 763–790 (2000)
5. Kern, T., Lacity, M., Willcocks, L.: Application Service Provision. Prentice Hall,
Englewood Cliffs (2001)
6. Clark, T.D., Zmud, R.W., McCray, G.E.: The outsourcing of information services:
transforming the nature of business in the information industry. J. Inf. Technol. 10, 221–237
(1995)
7. Gonzalez, R., Gasco, J., Llopis, J.: Information systems outsourcing reasons and risks: an
empirical study. Int. J. Hum. Soc. Sci. 4(3), 181–192 (2009)
8. Dickson, G.W.: An analysis of vendor selection systems and decisions. J. Supply Chain
Manag. 2(1), 5–17 (2009)
9. Ghodsypour, S.H., O’Brien, C.: A decision support system for supplier selection using an
integrated analytic hierarchy process and linear programming. Int. J. Prod. Econ. 56–57,
199–212 (1998)
10. Wang, J.J., Gu, R., Diao, X.J.: Using a hybrid multi-criteria decision aid method for
outsourcing vendor selection. In: Proceedings of the 4th International Conference on
Wireless Communications, Networking and Mobile Computing, WiCOM 2008, Dalian, 12–
14 October, pp. 1-4 (2008)
11. MIT Sloan Management Review- Beyond Outsourcing: Managing IT Resources as a Value
Center, https://sloanreview.mit.edu/article/beyond-outsourcing-managing-it-resources-as-a-
value-center/. Accessed 27 Apr 2020
12. Buck-Lew, M.: To outsource or not? J. Int. J. Inf. Manag. 12(1), 3–20 (1992)
13. Ghorabaee, K.M., Zavadskas, E.K., Olfat, L., Turskis, Z.: Multi-criteria inventory
classification using a new method of evaluation based on distance from average solution
(EDAS). Informatica 26(3), 435–451 (2015)
14. Ghorabaee, K.M., Zavadskas, E.K., Amiri, M., Turskis, Z.: Extended EDAS method for
fuzzy multi criteria decision-making: an application to supplier selection. Int. J. Comput.
Commun. Control 11(3), 358–371 (2016)
15. Peng, X., Liu, C.: Algorithms for neutrosophic soft decision making based on EDAS and
new similarity measure. J. Intell. Fuzzy Syst. 32, 955–968 (2017)
16. Kahraman, C., Ghorabaee, K.M., Zavadskas, E.K., Cevik Onar, S., Yazdani, M., Oztaysi, B.:
Intuitionistic fuzzy EDAS method: an application to solid waste disposal site selection.
J. Env. Eng. Landscape Manag. 25(1), 1–12 (2107)
17. Ghorabaee, K.M., Amiri, M., Zavadskas, E.K., Turskis, Z.: Multi-criteria group decision-
making using an extended EDAS method with interval type-2 fuzzy sets. Ekonomika
Manag. 1, 48–68 (2017)
18. Weber, C.A., Current, J.R., Benton, W.C.: Vendor selection criteria and methods. Eur.
J. Oper. Res. 50(1), 2–18 (1991)
19. Mendel, J.M., John, R.I., Feilong, L.: Interval type-2 fuzzy logic systems made simple. IEEE
Trans. Fuzzy Syst. 14(6), 808–821 (2006)
20. Ghorabaee, K.M., Amiri, M., Salehi Sadaghiani, J., Hassani Goodarzi, G.: Multiple criteria
group decision-making for supplier selection based on COPRAS method with interval type-2
fuzzy sets. Int. J. Adv. Manuf. Technol. 75(5–8), 1115–1130 (2014)
Analysis of Traffic Accidents to Identify
Factors Affecting Injury Severity with Fuzzy
and Crisp Techniques
Tutku Tuncalı Yaman1(&) , Emrah Bilgiç2 ,

and M. Fevzi Esen3
1
Department of Management Information Systems, Beykent University,
Istanbul, Turkey
[email protected]
2
School of Social Sciences and Humanities,
Kayseri University, Kayseri, Turkey
[email protected]
3
Department of Management Information Systems,
University of Health Sciences, Istanbul, Turkey
[email protected]
Abstract. Injury severity in motor vehicle traffic accidents is determined by a

number of factors including driver, vehicle, and environment. Airbag deploy-
ment, vehicle speed, manner of collusion, atmospheric and light conditions,
degree of ejection of occupant’s body from the crash, the use of equipment or
other forces to remove occupants from the vehicle, model and type of vehicle
have been considered as important risk factors affecting accident severity as well
as driver-related conditions such as age, gender, seatbelt use, alcohol and drug
involvement. In this study, we aim to identify important variables that contribute
to injury severity in the traffic crashes. A contemporary dataset is obtained from
National Highway Traffic Safety Administration’s (NHTSA) Fatality Analysis
Reporting System (FARS). To identify accident severity groups, we performed
different clustering algorithms including fuzzy clustering. We then assessed the
important factors affecting injury severity by using classification and regression
trees (CRT). The results indicate that the most important factor in defining injury
severity is deployment of air-bag, followed by extrication, ejection occurrences,
travel speed and alcohol involvement.
Keywords: Traffic accidents Data mining Injury severity
1 Introduction
According to the Global Road Safety Report of 2018, there are almost 1.2 million
people killed in road traffic crashes each year, while the number of injured or disabled
who suffer life-altering injuries with long lasting effects can be as high as 50 million
worldwide [1]. Traffic injuries are a multidimensional problem that causes considerable
economic and social costs to society. To reduce the risk of road traffic injuries and
deaths, it is essential to identify the factors influencing exposure to risk. These factors
626 T. T. Yaman et al.
are considered as facilitating for the acquisition of the knowledge on the nature of
traffic crashes and analyzing the factors leading to injuries. This provides vital infor-
mation for preventive actions, which are taken by decision makers [2].
Injury severity in motor vehicle traffic accidents is determined by a number of
factors including drivers, vehicle, and environment. Airbag deployment, vehicle speed,
manner of collusion, atmospheric and light conditions, degree of ejection of occupant’s
body from the crash, the use of equipment or other forces to remove occupants from the
vehicle, model and type of vehicle have been considered as important risk factors
affecting accident severity as well as driver-related conditions such as age, gender,
seatbelt use, alcohol and drug involvement [3–8].
The main purpose of this study is to find out unapparent but important factors in
traffic crashes that result in severe injuries. This purpose can support the investigation
of preventive actions regarding disruptive results of traffic accidents and appropriate
design of traffic measures that address road and transport systems. By pursuing the
stated aim, both Fuzzy and CRISP techniques are performed and Classification and
Regression Tree (CRT) is found superior in definition of important factors.
The following sections of the paper describes the dataset along with preprocessing
steps and present the methodology and crucial results of used techniques, respectively.
In the closing section, discussion of the results is provided.
2 Data Set and Data Preprocessing
For the analysis, we used Person, Accident and Vehicle level datasets from FARS of
NHTSA. Accident dataset includes case (accident) basis variables such as location,
date, time, occasion, number of vehicles involved in the crash, type of intersection that
the crash is related, manner of collision, and route signing. Person dataset harbours the
information regarding motor vehicle occupants such as age, sex, person type
(driver/passenger in/not in transport), injury severity, ejection, extrication, and use of
alcohol or other related factors. Vehicle dataset contains number of occupants, vehicle
model and year, body, type, travel speed, fire occurrence in the vehicle during the
crash, vehicle trailing, hit-and-run and other related factors. To obtain reliable results
from the datasets, which constituted in different levels, it was decided to reduce data in
person level and each case was represented by only the driver information of the
vehicle(s). In other words, more than one driver information can be covered according
to the nature of the crash. After the data reduction task, pursued methodology started
with variable selection and data preprocessing. The variables were selected among the
entire dataset according to aforementioned literature.
Dataset constituted with many categorical variables with no inter-relational
numerical coding. In addition to that, there were missing values in each variable and
in some cases these were coded as 98 or 99. The coding handbook is also available in
FARS/CRSS Coding and Validation Manual [9]. For the sake of clarity of our results,
all cases with intentional or unintentional missing values were omitted. Then, some
categories that have low frequency or similar meanings were coded in the same cat-
egory. The final dataset included 23 variables belonging to 8,403 traffic crashes and
11,615 drivers in total.
Analysis of Traffic Accidents to Identify Factors Affecting Injury Severity 627
3 Methodology
3.1 Classification and Regression Tree (CRT) Analysis
According to Han, Kamber & Pei, classification and regression may need to be pre-
ceded by relevance analysis, which attempts to identify attributes that are significantly
relevant to the classification and regression process [10]. Such attributes are selected
for the classification and regression process. Other attributes, which are irrelevant, can
then be excluded from consideration. One of the classification techniques is decision
trees, which can be used in data reduction and variable screening by selecting a useful
subset of predictors from a large set of variables [11]. With the CRT algorithm, the
classification capability of factors is surveyed comparative with a split or cut-point. The
absolute best indicator is chosen as the beginning variable at the highest point of a
various leveled tree. Subjects with values not exactly the slice direct advance toward
one classification, while those with values more noteworthy than the cut-point move
into a second box of the various leveled tree. Cut-points are then surveyed in a stepwise
procedure for the rest of the indicators. A decision tree is produced until maximal
arrangement is accomplished, or further parting is made a decision to be not “cost-
effective” [12]. CRT relies on pruning to cut the branches that do not perform well.
Since the technique is growth two-way branches, it is preferred for the ease of
understanding of generated rules [13]. Detailed characteristics of the algorithm can be
found in [14].
In the nature of the algorithm, splitting criteria in classification procedure selected
by Gini index (see Eq. (1)), which measures the degree or probability of a particular
variable being wrongly classified when it is randomly chosen. The degree of Gini index
varies between 0 and 1, where 0 denotes that all elements belong to a certain class or if
there exists only one class, and 1 denotes that the elements are randomly distributed
across various classes. A Gini Index of 0.5 denotes equally distributed elements into
some classes [10].
Xn
Gini ¼ 1 i¼1
ð pi Þ 2 ð1Þ
3.2 Cluster Analysis

Cluster analysis is one of the unsupervised learning techniques, which divides the
dataset into meaningful groups by considering the similarities of the points with each
other. Relationships among the data points or objects are represented in a proximity
matrix. If the objects mentioned are data points, which placed in the d-dimensional
space, proximities are the distances between each data points (e.g. Euclidean distance).
If there is no distance between the point of pairs or if there is no similarity, then cluster
analysis cannot be performed. The reason for this issue is that similarity (proximity)
matrix is the only input to perform cluster analysis [15, 16].
Steps to follow while performing a clustering are generally: selection of variables,
distance function and clustering algorithm to use, validity of clustering and interpre-
tation of the results. All of these steps are interconnected and they feed on each other
and they appear as the determining factors of the clustering results. Methods used in
cluster analysis in general can be divided into two subgroups. These are hierarchical
and non-hierarchical methods.
Gower’s Distance Matrix
There is a distance measure to calculate the required distance matrix for mixed-type
data sets [17]. It is possible to perform cluster analysis with these kinds of dissimilarity
measures. Kaufman and Rousseeuw [18] defined how to handle mixes type of variables
for dissimilarity matrix as follows:
Object j
1 0
1 a b a+b
Object i
0 c d c+d
a+c b+d
Fig. 1. Dissimilarity matrix
Let data set contains p variables, which are in different types. The dissimilarity
between objects i and j is defined as;
Pp ðfÞ ðfÞ
f ¼1 dij dij
d ði; jÞ ¼ Pp ðfÞ
ð2Þ
f ¼1 dij
ðf Þ
where, dij is equal to 1 when both measurements x if and xjf for the fth variable are not
ðf Þ
missing and 0 otherwise. dij is contribution of the fth variable to the dissimilarity
ðf Þ
between i and j. If variable f is interval-scaled, then dij is given by;

ðfÞ
xif xjf
dij ¼ ð3Þ
Rf
where Rf is the range of variable f. If variable f is asymmetric binary scaled, then

Jaccard Coefficient (1908) is appropriate for calculating dissimilarity matrix;
bþc
d ði; jÞ ¼ ð4Þ
aþbþc
k-medoids Algorithm
The steps followed in the k-medoids algorithm can be summarized as follows: Instead
of minimizing distance squares as in k-means algorithm, k-medoids minimizes the
absolute distances between the observation values and the observation selected as
medoid.
The number of points set at the beginning elected as their representatives at initial
step. In assignment step, each data point is assigned to the cluster closest to it. In
distance calculation step, the sum of the intra-cluster distances are calculated, when the
smallest distance is found the algorithm stops. If a smaller distance is possible the
assignment step is repeated until the smallest distance is found.
Fuzzy Clustering
In a fuzzy clustering, each observation is “spread out” over the various clusters. Denote
by u(i,v) the membership of observation i to cluster v. The memberships are nonneg-
ative, and for a fixed observation i they sum to 1. FANNY is a fuzzy or soft clustering
algorithm, where each node in the graph is associated with a membership coefficient,
indicating degree of belongingness of each node to different clusters. The particular
method fanny fuzzy clustering stems from [18].
Two-Step Clustering Algorithm
The approach can be viewed as integration of k-means and hierarchical clustering. The
algorithm pre-clusters the cases into many small sub-clusters and then performs the
hierarchical clustering to group the cases into desired number of clusters using
Bayesian Information Criterion (BIC).
A technical introduction can be given as follows. The processed input variables
fXk ; k ¼ 1; 2. . .; K þ 1g are used to create a clustering model. The two-step clustering
algorithm consists of: (a) a pre-cluster step that pre-clusters cases into many sub-
clusters and (b) a cluster step that clusters the sub-clusters resulting from pre-cluster
step into the desired number of clusters. Given a case the closest cluster h is found. The
variable deviation index VDIk of variable Xk is defined as the contribution dk (h, s) of
the variable to its log-likelihood distance d(h, s). The corresponding norm value is Mhk ,
which is the cluster sample mean of Xk. The group deviation index GDI of a case is the
log-likelihood distance d(h, s), which is the sum of all the variable deviation indices
fVDIk ; k ¼ 1; . . .; K þ 1g.
4 Results and Discussions
In this study, over 8.000 accidents were analyzed. This section will provide results of
previously presented techniques respectively.
4.1 Gower’s Distance Matrix

To be able to identify different accident groups, which cause different injury severities,
cluster analysis was performed. The main idea behind performing cluster analysis is to
find similar accidents, which will help us to find out the featured variables within the
clusters. By doing so, one can understand the causes of severely injured accidents.
First of all, a distance measure for mixed-type of variables is calculated with
Gower’s formula (see Fig. 1). Then, clustering algorithms including fuzzy clustering
algorithm, hierarchical and non-hierarchical techniques were performed. However,
none of the algorithms could find a good clustering structure for the data set analyzed.
For mixed-type data (data set with different kinds of variables), one needs to
calculate a distance matrix, which is specifically designed. In this study, we use
Gower’s distance matrix as introduced in [17].
4.2 Pam Clustering Algorithm

We performed Pam Clustering Algorithm with pam function in R. As seen in the
Table 1 below, a good clustering structure could not be found with this algorithm.
Table 1. Average Silhouette Widths for each number of clusters

Number of clusters Average Silhouette Width
2 0.12
3 0.17
4 0.11
5 0.10
4.3 Fuzzy Clustering

Fanny algorithm is performed for fuzzy clustering with funny function in R. Any
clustering structure could not be found with this algorithm either. Then fuzzification
process is conducted and all variables converted to triangular fuzzy numbers. Cluster
analysis is performed to those fuzzy numbers as follows:
First, Two-step clustering algorithm in SPSS 23.0 is used and a good clustering
structure has been identified for three clusters. However, two-step algorithm has found
good clustering structures for four and five clusters too. Secondly, pam algorithm is
used to perform cluster analysis and the input for clustering is again fuzzy numbers
from fuzzification process. Five clusters seem the best solution with this algorithm.
Setting the cluster number to three and four yields good results too. However, these
results were misleading, since fuzzification process removes the variability in the data
set.
4.4 Classification and Regression Analysis (CRT)

CRT technique is mainly used for obtaining factor importance rather than prediction.
Implementation of the technique is performed in SPSS 23.0 with the following defi-
nitions. Target variable was set as Injury Severity and others introduced as independent
variables. In pre-pruning, three different parameters were set. First is maximum depth
of the tree, which was set as 7. With the purpose of obtaining a more detailed rule set,
which includes relatively important variables, this value was increased even though the
default value was 5. Minimum sample size of each parent node was 100 and minimum
sample size of child nodes was 50. These were the defaults of the criteria.
Validation of the model was obtained by splitting data to training (70%) and test
(30%) samples. Classification accuracy was evaluated by fit scores. According to the
results, “importance values” of independent variables according to injury severity were
calculated. Figures are shown in the following Table 2. If we investigated overall

classification results, overall fit of training sample would be 71.7% vs. test sample
would be 69.5%.
Table 2. Importance of factors affecting injuries

Independent variable Importance Normalized importance
Air bag deployed .089 100.0%
Extrication .085 95.5%
Ejection .078 87.5%
Travel speed .028 31.7%
Alcohol involvement .024 27.1%
Manner of collision .022 24.9%
Speed related factors .018 20.4%
Age .016 18.4%
Drug involvement .014 15.8%
Model year .011 12.6%
Land use .011 12.3%
Route signing .005 5.2%
Body type .004 4.9%
Functional system .004 4.3%
Fire occurrence .003 3.5%
Type of intersection .003 3.0%
Light condition .001 1.5%
Atmospheric conditions .001 1.4%
Sex .001 1.4%
Hit-and-Run .001 1.1%
Mis-use of restraint system/Helmet use .000 0.3%
Vehicle trailing .000 0.2%
School bus related 5.000E-5 0.1%
The procedure generated 15 nodes in total and a number of the important rule sets
in predicting major injuries were coded as 1, 7 and 15.
5 Conclusion Remarks
In this study, the authors performed classification and regression (CRT) and cluster
analysis with different clustering algorithms to identify important factors affecting the
traffic accident injuries. Clustering analysis was conducted to identify different accident
groups. However, any clustering structures could not be found with different clustering
algorithms including fuzzy clustering. Furthermore, the values of each variable were
converted to fuzzy numbers and then, clustering analysis was conducted again. This
time, the algorithms found perfect clustering structures due to the reduced variability in
the data. To identify the factors affecting injury severity occurred during the accidents,
CRT technique was performed. Deployment of air-bag was found the most important
factor in defining injury severity. This means that deployment of the airbag has a great
effect for preventing severe injuries. Extrication, ejection occurrence, travel speed and
alcohol involvement were other important factors affecting injury severity. Since these
factors have great potential to prevent fatal/severely injured accidents, the findings of
this study will guide car manufacturers, policy makers and insurance companies. The
sample size and time range of data can be expanded for the sake of more accurate
analysis. If more features are available, socioeconomic factors and incidence of traffic
accidents should be prioritized to identify characteristics, which are more related to
crashes. It should be noted that, the road safety authorities can implement preventive
measures towards certain factors such as geography, characteristics of the roads,
economic and social costs of accidents as well as the main causes of accidents as stated
in our results.
References
1. World Health Organization: Global Status Report on Road Safety. WHO Press, Switzerland
(2018)
2. World Health Organization: Risk Factors for Road Injuries. WHO Press, Switzerland (2018)
3. Vorko-Jović, A., Kern, J., Biloglav, Z.: Risk factors in urban road traffic accidents. J. Saf.
Res. 37(1), 93–98 (2006)
4. Elvik, R., Vaa, T.: The Handbook of Road Safety Measures, 1st edn. Emerald Group
Publishing Limited, Great Britain (2008)
5. Jägerbrand, A.K., Sjöbergh, J.: Effects of weather conditions, light conditions, and road
lighting on vehicle speed. SpringerPlus 5(1), 1–17 (2016)
6. Wåhlberg, A.: Driver Behaviour and Accident Research Methodology: Unresolved
Problems. CRC Press, Boca Raton (2017)
7. Al-Balbissi, A.H.: Role of gender in road accidents. Traffic Inj. Prev. 4(1), 64–73 (2003)
8. Regev, S., Rolison, J.J., Moutari, S.: Crash risk by driver age, gender, and time of day using
a new exposure methodology. J. Saf. Res. 66, 131–140 (2018)
9. FARS: Manuals and Documentation. https://www.nhtsa.gov/research-data/fatality-analysis-
reporting-system-fars. Accessed 20 Jan 2020
10. Han, J., Kamber, M., Pei, J.: Data Mining: Concepts and Techniques, 3rd edn. Morgan
Kaufman Publishers, San Francisco (2012)
11. IBM: Creating Decision Trees. https://www.ibm.com/support/knowledgecenter/. Accessed
23 Jan 2020
12. Knable, M.B., Barci, B.M., Bartko, J.J., Webster, M.J., Torrey, E.F.: Molecular abnormal-
ities in the major psychiatric illnesses: classification and Regression Tree (CRT) analysis of
post-mortem prefrontal markers. Mol. Psychiatry 7(4), 392–404 (2002)
13. De Ville, B., Neville, P.: Decision Trees for Analytics Using SAS Enterprise Miner. SAS,
New York (2020)
14. Breiman, L., Friedman, J.H., Olshen, R.A., Stone, J.C.: Classification and Regression Trees.
Taylor & Francis, New York (1984)
15. Jain, A.K., Dubes, R.C.: Algorithms for Clustering Data. Prentice Hall, New Jersey (1988)
16. Zaki, M.J., Meira, W.: Data Mining and Analysis: Fundamental Concepts and Algorithms.
Cambridge University Press, England (2014)
17. Gower, J.C.: A general coefficient of similarity and some of its properties. Biometrics 27(4),
857–871 (1971)
18. Kaufman, L., Rousseeuw, P.J.: Finding Groups in Data: An Introduction to Cluster Analysis.
Wiley, New York (1990)
Logistics Performance Index (LPI) Analysis
with Using AHP and Fuzzy AHP
Emir Şahin , İrem Pala , and Berrin Denizhan(&)
Sakarya University, Serdivan 54055, Turkey

[email protected]
Abstract. The Logistics Performance Index (LPI), which is published by the

World Bank at two-year intervals, evaluates and ranks the countries’ logistics
performance in six categories with a survey application. This study aims to
present an application in which companies within the country can individually
categorize according to LPI. The LP survey scores based on answers given in
each category, applying the sample business groups in Turkey have been cre-
ated. Unlike the LPI index is 2018 points considering the world and in Turkey,
were analyzed by the method of AHP and Fuzzy AHP decision-making. Based
on these analyses, the most important criteria were found. These criteria were
analyzed statistically and the differences between LPI rank and Fuzzy AHP and
AHP rankings were analyzed. These results provide a practical application that
shows which areas they are behind the LPI index and which criteria can improve
their performance by making individual analyses of the enterprises. Thus,
companies can also benefit from LPI analysis individually.
Keywords: LPI AHP Fuzzy AHP
1 Introduction
The Logistics Performance Index (LPI) is published by the World Bank. They are
published in the report entitled “Connecting for Competition: Foreign Trade Logistics
in the Global Economy”, which reveals the comparative situation of the world logistics
sector between countries. It measures the growth in the logistics industry, the quality,
and speed of doing business in countries; It is considered as an important assessment
tool in developing future strategies for decision makers, countries, and companies in
transportation policies. As of 2018, LPE ranks 160 countries according to their logistics
performance. The data is collected through two surveys, national and international. The
survey includes quantitative and qualitative assessments made by logistics profes-
sionals working in a country’s trade partner countries. The evaluations include the
indicators collected under six headings; Customs and border crossing efficiency, the
quality of the trade and transportation infrastructure, ease of shipment planning at
competitive prices, the adequacy and quality of logistics services, the ability to track
and track shipments, the frequency of shipment reaching the buyer at planned and
expected shipment times. While calculating the LPI index of the world bank, 1051
specialist logistics professions, which are carried out every two years, are scored
between 1 (worst) and 5 (best) of the countries in which they do business. Logistics
Logistics Performance Index (LPI) Analysis with Using AHP and Fuzzy AHP 635
Performance Index (LPE) and country rankings of the countries are created by taking
the average of these ratings. The data is compiled annually from businesses that do
multinational trade via an online survey. In 2018, 160 countries were asked to evaluate
their countries of logistics by covering their respondents and approximately 6000
country evaluations were received [15].
In line with the LPI published every two years, its surveys and reports, countries’
scores and evaluations in 6 categories can be seen. However, since the country’s
general scores are evaluated according to the arithmetic average method, it can be
difficult to make individual firm based comments. The score of any company in the
country may be below or above that country score. In order to see them in a healthier
on progressive way, an individual company based study was firstly conducted in this
research.
Considering the studies on the logistic performance index, the existence of the
studies in the last 9 years has been observed. It is seen that the general structure of these
studies is based on country analysis, the effect of LPI on countries and competition and
analysis of variables. Roy analyzed the relationship between Canada’s Logistics Per-
formance Index and business efficiency. Work efficiency positively affects the coun-
try’s logistics performance and supply chain management [10]. In the study conducted
by, it is aimed to reveal the effects of different types of organizational culture and
leadership styles in ensuring the organizational commitment of employees in enter-
prises in the logistics sector, which have been neglected in organizational behavior
research [1]. Factors adversely affecting logistics performance using LPR data deter-
mined in accordance with Turkey’s year and are highlighted in particular the effect of
infrastructure logistics performance of road transport [5]. Güngör and colleagues
analyzed and interpreted the relationship between economic size and logistics activities
of the countries by looking at the LPI data of 14 countries from the World Bank [6].
The intermediate variable role of logistic performance and global competition shown in
the impact of the perception of corruption on a gross domestic product by using LPI
data. There are also studies on LPI comparisons of many countries [2, 4, 7, 9, 14].
Turkish researchers also concentrated on Turkey’s LPI data were taken from the World
Bank.
These data have been analyzed according to years and the points to be developed
have been specified. In addition, rail transport has been evaluated in logistics activities
[6, 12, 15]. Within the context of the literature, we have reached the effects of LPI data
commonly has been studied on sectorial and country basis are at the forefront. The
techniques used are concentrated as surveys and statistical comparative analyzes. In
this study, the individual effect of LPI on companies was tried to be observed. In order
for the companies to act according to their country goals or deficiencies, they must first
be able to make their own self-assessments. With this study, LPI offers an exemplary
decision making structure for different companies with its global to LPI local approach.
At the same time, technically fuzzy AHP method is almost not encountered. Fuzzy
decision-making techniques used in two studies. In the study of Bayır, AHP and
VIKOR methods from Multi Criteria Decision Making Techniques (CCPVT) were
used to measure the logistics performance of 20 European countries within the
framework of LPI data for 2016 [3]. Oğuz and colleagues in this study, the data of 7
636 E. Şahin et al.
Asian countries in the World Bank, they were ranked and evaluated their logistics
performance with the TOPSİS method [8].
LPI achieves general points about countries. Thus, the criteria that countries are
strong and weak are determined. However, it is relative for companies to determine
their advantages and weaknesses individually in terms of logistics. In order to eliminate
this relativity and analyze according to LPI, Fuzzy AHP is an application that enables
companies to identify areas where they are superior and weak by performing individual
analysis. In this paper, the LP survey scores based on answers given in each category,
applying the sample business groups in Turkey have been created.
Based on these analyzes, the most important criterion was found. Survey results
and LPI Turkey 2018 year scores were compared statistically and similarities and
differences were found. Fuzzy AHP and AHP rankings were analyzed. These results
provide a practical application that shows which areas they are behind the LPI index
and which criteria can improve any firm performance by making individual analyses of
the enterprises. The outline of this paper organized as follows. Section 2 contains
problem definition, some conceptual aspects of LPI and the methodology of the survey
on Turkish companies’ procedure. The numerical results under the proposed methods
and their comparisons are in Sect. 3. Finally, some concluding remarks are provided in
Sect. 4.
2 Problem Definition
According to the Logistics Performance Index (LPI) published every two years, the
surveys and reports of the countries consist of scores and evaluations in different
categories. However, in terms of the general scores of the countries, it can be difficult to
make individual firm based comments about logistics performance. The score of a
company in the country may be below or above the general score of that country score.
Any firm self-evaluation or knowing this difference between company and country
score, it is anticipated that a company, which evaluates itself, can contribute to
increasing the logistics performance of the country.
During the application process, AHP decision making method was used by ana-
lyzing the data obtained from the questionnaires sent to the companies. In this method,
the most important criteria affecting the logistic performance index were found and
other criteria were ranked according to their importance. In line with these criteria, the
company that uses logistics performance most effectively has been chosen. Fuzzy AHP
decision making method was used to compare the results from AHP. In this method,
the most important criteria that affect the logistics performance were found and other
criteria were ranked according to their importance levels. In line with these criteria, the
company that uses logistics performance most effectively has been chosen. The results
obtained from AHP and Fuzzy AHP methods were compared.
2.1 LPI Survey on Turkish Companies

The answers given by the companies to the questions with a range of 17–22 are
included in the Logistics Performance Index. In the evaluation made over 5, especially
the values taken from 4 and above can be evaluated as partially better. It can be said
that threats have started to emerge for companies with a value close to 1 and 1. The
following Table 1 gives a summary of all Turkish Manufacturing Industry LPI survey
results. In the study, a survey was sent to 120 firms, and 20 firms replied to the survey.
These companies are coded with the letters in the top row in Table 1.
Table 1. Turkish manufacturing companies’ LPI survey results
Company A B C D E F G H I J K L M N O P R S T U
A 1 0,932 1,165 2,009 1,273 0,876 0,876 0,876 1 0,876 0,876 1 0,538 0,823 0,876 0,56 2,33 2,33 0,777 0,876
B 1,073 1 1,25 2,155 1,366 0,94 0,94 0,94 1,073 0,94 0,94 1,073 0,577 0,883 0,94 0,601 2,5 2,5 0,833 0,94
C 0,858 0,8 1 1,724 1,093 0,752 0,752 0,752 0,858 0,752 0,752 0,858 0,462 0,707 0,752 0,481 2 2 0,667 0,752
D 0,498 0,464 0,58 1 0,634 0,436 0,436 0,436 0,498 0,436 0,436 0,498 0,268 0,41 0,436 0,279 1,16 1,16 0,387 0,436
E 0,785 0,732 0,915 1,578 1 0,688 0,688 0,688 0,785 0,688 0,688 0,785 0,423 0,647 0,688 0,44 1,83 1,83 0,61 0,688
F 1,142 1,064 1,33 2,293 1,454 1 1 1 1,142 1 1 1,142 0,614 0,94 1 0,639 2,66 2,66 0,887 1
G 1,142 1,064 1,33 2,293 1,454 1 1 1 1,142 1 1 1,142 0,614 0,94 1 0,639 2,66 2,66 0,887 1
H 1,142 1,064 1,33 2,293 1,454 1 1 1 1,142 1 1 1,142 0,614 0,94 1 0,639 2,66 2,66 0,887 1
I 1 0,932 1,165 2,009 1,273 0,876 0,876 0,876 1 0,876 0,876 1 0,538 0,823 0,876 0,56 2,33 2,33 0,777 0,876
J 1,142 1,064 1,33 2,293 1,454 1 1 1 1,142 1 1 1,142 0,614 0,94 1 0,639 2,66 2,66 0,887 1
K 1,142 1,064 1,33 2,293 1,454 1 1 1 1,142 1 1 1,142 0,614 0,94 1 0,639 2,66 2,66 0,887 1
L 1 0,932 1,165 2,009 1,273 0,876 0,876 0,876 1 0,876 0,876 1 0,538 0,823 0,876 0,56 2,33 2,33 0,777 0,876
M 1,858 1,732 2,165 3,733 2,366 1,628 1,628 1,628 1,858 1,628 1,628 1,858 1 1,53 1,628 1,041 4,33 4,33 1,443 1,628
N 1,215 1,132 1,415 2,44 1,546 1,064 1,064 1,064 1,215 1,064 1,064 1,215 0,654 1 1,064 0,68 2,83 2,83 0,943 1,064
O 1,142 1,064 1,33 2,293 1,454 1 1 1 1,142 1 1 1,142 0,614 0,94 1 0,639 2,66 2,66 0,887 1
P 1,785 1,664 2,08 3,586 2,273 1,564 1,564 1,564 1,785 1,564 1,564 1,785 0,961 1,47 1,564 1 4.16 4.16 1,387 1,564
R 0,429 0,4 0,5 0,862 0,546 0,376 0,376 0,376 0,429 0,376 0,376 0,429 0,231 0,353 0,376 0,24 1 1 0,333 0,376
S 0,429 0,4 0,5 0,862 0,546 0,376 0,376 0,376 0,429 0,376 0,376 0,429 0,231 0,353 0,376 0,24 1 1 0,333 0,376
T 1,288 1,2 1,5 2,586 1,639 1,128 1,128 1,128 1,288 1,128 1,128 1,288 0,693 1,06 1,128 0,721 3 3 1 1,128
U 1,142 1,142 1,33 2,293 1,454 1 1 1 1,142 1 1 1,142 0,614 0,94 1 0,639 2,66 2,66 0,887 1
3 AHP and Fuzzy AHP Methods
AHP and Fuzzy AHP application will take place after obtaining the survey results.
AHP is a decision-making method that relies on evaluating and identifying criteria and
alternatives to achieve the specified goal. Their aim is to determine the main criteria
and alternatives (if any, its sub-objectives) and to define them in a hierarchical order, by
comparing the criteria and alternatives [11]. It is possible to eliminate the uncertainty in
decision making process with fuzzy logic Fuzzy AHP. Instead of precise results in
AHP application, the data are evaluated by determining certain intervals [13].
In accordance with the AHP decision making method, a comparison matrix for 6
categories was prepared separately and the values were normalized in Table 2.
According to these data, 0.74833 values were written from customs to infrastructure.
This value is the division of the customs score by the infrastructure score. This table is
made to sort the categories according to their importance. Normalized values were
obtained by dividing the totals by each value and then the weight scores of the cate-
gories were obtained (Table 3).
Table 2. Unified matrix of categories of AHP

Customs Infrastructure International Logistics Tracking & Timeliness
shipments competence tracing
Customs 1 0,74833 0,742822787 0,741930641 0,76978193 0,78531702
Infrastructure 1,336306 1 0,992634902 0,991442726 1,02866044 1,04941999
International 1,346216 1,00742 1 0,998798979 1,03629283 1,05720642
shipments
Logistics 1,347835 1,00863 1,001202465 1 1,03753894 1,05847767
competence
Tracking & 1,299069 0,97214 0,964978205 0,963819246 1 1,02018115
tracing
Timeliness 1,273371 0,95291 0,945889073 0,94475304 0,98021807 1
Table 3. Weight scores of categories

Customs Infrastructure International Logistics Tracking & Timeliness
shipments competence tracing
Weights 2,471 3,302 3,3265 3,3305 3,21 3,1465
In AHP decision making method, the results are shown in Table 4. The degrees of
importance were obtained by multiplying the weight ranges and line spacing.
Accordingly, M firm has been chosen as the most important. Firm M is in front of the
other 20 firms. In other companies, they can decide how to draw a road map by
examining the status of the companies they set as competitors or targets. Evaluation
with fuzzy AHP method is given below.
Table 4. AHP scores and ranking

AHP Scores Ranking
A 0,0483 14
B 0,0499 13
C 0,0525 7
D 0,0510 9
E 0,0545 6
F 0,0471 15
G 0,0565 3
H 0,0553 4
I 0,0448 18
J 0,0501 11
K 0,0567 2
L 0,0449 17
M 0,0569 1
N 0,0500 12
O 0,0355 16
(continued)
AHP Scores Ranking
P 0,0503 10
R 0,0355 20
S 0,0440 19
T 0,0549 5
U 0,0514 8
A comparison matrix of categories was created to be used in the BAHP application.

While creating the matrix, attention was paid to the comparisons of companies.
Unlike AHP, Low (L), Medium (M), Upper (U) fuzzy values are determined here.
M values were taken from previous AHP matrices. L and U values were calculated to
give the median value. Fuzzy ranges are created instead of exact values.
The sum of the values l, m, and u in the comparison matrix is taken and then
column totals are obtained. The total portion will be used when performing the nor-
malization process. Then, a formulation was used when proceeding to the final decision
matrix stage.
Table 5. Fuzzy AHP scores and ranking

Decision matrix Ranking
Customs 0.1136 6
Infrastructure 0.1703 5
International shipments 0.1769 4
Logistics competence 0.1822 1
Tracking & tracling 0.1782 3
Timeliness 0.1785 2
The final decision matrix has been created in Table 5 above; logistics competency
has been selected as the most important category. Timing follows logistics competence.
There are cargo tracking, international transportation, infrastructure and customs
respectively.
Customs is observed as the least important category. BAHP application was ini-
tiated among companies. Comparisons were made for 20 companies. The company
ranking from the AHP results is intended to be compared with the results in Table 6.
By removing the necessary matrices, the L-M-U fuzzy values are obtained.
Table 6. Comparisons with AHP and Fuzzy AHP

Firms Ranking with AHP Ranking with Fuzzy-AHP
A 14 1
B 13 3
C 7 18
D 9 9
E 6 19
F 15 17
G 3 7
H 4 8
I 18 6
J 11 4
K 2 14
L 17 2
M 1 12
N 12 11
O 16 20
P 10 16
R 20 5
S 19 13
T 5 15
U 8 10
4 Results and Conclusion
Fuzzy AHP result has a strategic importance for companies. The best companies’
Chosen should continue by advancing their goals and strategies, while other companies
should find their deficiencies and take steps to develop their logistics strategies and
plans if necessary. Firm A is the best firm as a result of Fuzzy AHP. He has the firm
with the least points. In order to increase their share in this competitive environment in
the world, all companies should try to make a difference in customs, infrastructure,
international transportation, logistics competence, load tracking and timing. A com-
parison of AHP and Fuzzy AHP methods in terms of firms is given in Table 5 above.
M firm, which is the most effective in AHP, ranks 12th in the Fuzzy AHP method.
Company A, which is the most effective in Fuzzy AHP, is 14th in the AHP method.
The reason for this difference is that the importance of the 6 categories in the methods
are different and the differences in the way of thinking. With this study, it can be seen
that LPI’s AHP and Fuzzy AHP techniques can be used to reveal individual firm
performance analysis. Because of the Fuzzy AHP ranking with the same criteria, is
different from the LPI Turkey rankings. In future studies, the criteria can be detailed by
adding sub-criteria determined by the companies.
References
1. Acar, A.Z., Gürol, P.: Historical development of logistics literature in Turkey. İşletme
Araştırmaları Dergisi 5(3), 289–312 (2013)
2. Akdoğan, M.Ş., Durak, A.: Logistic and marketing performances of logistics companies: a
comparison between Germany and Turkey. Proc. Soc. Behav. Sci. 235, 576–586 (2016)
3. Bayır T., Yılmaz Z.: AB Ülkelerinin Lojistik Performans Endekslerinin Ahp ve Vikor
Yöntemleri İle Değerlendirilmesi. Middle East J. Educ. (MEJE), 3, 273–92 (2017)
4. Bozkurt, C., Mermertaş, F.: Türkiye ve G8 Ülkelerinin Lojistik Performans Endeksine Göre
Karşılaştırılması. İşletme ve İktisat Araştırmaları Dergisi 7(2), 107–117 (2019)
5. Çatuk, C., Aydın, K., Atalay, E.: Uluslararası ticarette karayolunun lojistik performansına
etkisi. Al-Farabi Int. J. Soc. Sci. 3(4), 120 (2019)
6. Güngör, Ş., Dursun, E., Karaoğlan, A.: Lojistik Faaliyetlerin Akdeniz Ülke Ekonomileri ile
İlişkisi. İktisat Politikası Araştırmaları Dergisi 6(2), 91–106 (2019)
7. Kılınç, E., Fidan, O., Mutlu, H.M.: Türkiye, Çin ve Rusya Federasyonu’nun Lojistik
Performans Endeksine Göre Karşılaştırılması. Uluslararası Ekonomik Araştırmalar Dergisi 5
(2), 17–34 (2019)
8. Oguz S., Alkan G., Yılmaz B.: Seçilmiş Asya Ülkelerinin Lojistik Performanslarının
TOPSIS Yöntemi ile Değerlendirilmesi. IBAD, (Özel Sayı), pp. 497–507 (2019)
9. Rashidi, K., Cullinane, K.: Evaluating the sustainability of national logistics performance
using data envelopment analysis. Transp. Policy 74, 35–46 (2019)
10. Roy, J.: Logistics and the competitiveness of canadian supply chains. In: Sydor, A. (ed.)
Global Value Chains: Impacts and Implications. Trade Policy Research, pp. 313–333.
Minister of Public Works and Government Services, Canada (2011). Section 4
11. Saaty, T.L.: The Analytical Hierarchy Process. Mc Graw Hill, New York (1980)
12. Uca, N., Civelek, M.E., Çemberci, M.: Yolsuzluk Algısının Gayrisafi Yurt İçi Hasıla
Üzerine Etkisinde Lojistik Performans ile Küresel Rekabetin Ara Değişken Rolü: Türkiye
Değerlendirmesi. Uluslararası Toplum Araştırmaları Dergisi 10(17), 1229–1261 (2019)
13. Zadeh, L.A.: Fuzzy Sets. Inf. Control, 8, 338–353 (1965)
14. Zeybek, H.: Uluslarası Ticarette Demiryolunun Lojistik Performansa Etkisi. Demiryolu
Mühendisliği 9, 79–90 (2019)
15. https://lpi.worldbank.org/domestic/environment_institutions. Accessed 15 Dec 2019
Sustainable Transportation Service Provider
Evaluation Utilizing Fuzzy MCDM Procedure
Mehtap Dursun1(&) and Ergun Ari2

1
Ortakoy, Istanbul, Turkey
[email protected]
2
Industrial Engineering Department, Bilgi University,
Eyupsultan, Istanbul, Turkey
[email protected]
Abstract. Nowadays transportation systems are a vital piece of daily lives of

human beings. It was evaluated that almost 40% of the world’s population
spends nearly one hour out every day. As individuals have turned out to be
substantially more subject to transportation systems lately, transportation sys-
tems confront new opportunities and new difficulties. Sustainable transportation
has also an important role in supply chain and logistics management. It con-
structs the basis of sustainable supply chain management. Determining the most
suitable sustainable transportation service provider requires considering multiple
conflicting criteria, which are organized in a hierarchical structure. In this study,
a hierarchical fuzzy multi-criteria decision making methodology is utilized in
order to evaluate sustainable transportation service provider alternatives.
Keywords: Decision support systems Fuzzy sets Multi-criteria decision

making Sustainable transportation service provider
1 Introduction
Recently, due to technological advancements, changes in public needs, and growing in

urban population, sustainable development received great interest. Sustainable devel-
opment is described as the concept of meeting the present requirements without
compromising the capacity of future generations to meet their requirements [1]. Dif-
ferent financial, natural, and social factors ought to be taken as a great importance to
meet basic objectives of sustainability while seeking after financial, environmental, and
social objectives.
Sustainability has also gained attention in the transportation field. The development
of approaches for sustainable transportation frequently includes cross disciplines par-
ticipation, as well as regional and central governmental coordinated effort [2]. Sus-
tainable transportation is characterized as “the transportation that addresses mobility
issues while additionally saving and improving human and environmental wellbeing,
financial advancement and social equity now and in the future” [3]. It is a challenging
area of study with a number of problems to handle.
Sustainable Transportation Service Provider Evaluation Utilizing Fuzzy MCDM 643
The Council of the European Union [4] suggested the necessities for a sustainable
transportation system as
(1) Ensuring essential access and improving needs of people, organizations and the
public securely and reliably by regarding public wellbeing, environment and
generational equity;
(2) Being moderate, productive and reasonable that gives choices to transport modes
and advances focused nearby improvement;
(3) Controlling waste and harmful outflows with the goal that they stay under the
world’s assimilation limit of them, use renewable sources beneath their renewal
rate, decrease noise and effect on land and use resources in less risk until
appropriate renewable replacements are improved.
Sustainable transportation has an important role in supply chain and logistics
management. Supply chain management (SCM) can be seen as a strategic partnership
between retailers and suppliers. SCM connects firms to their suppliers, manufacturers,
retailers, and customers. The effective management of supply chains has a positive
impact on the overall performance of the organizations. To transfer goods and mate-
rials, firms usually outsource transportation services. Sustainability is one of the key
aspects in selecting the most suitable transportation service provider, which requires to
consider multiple criteria. Sustainable transportation is also the basis of sustainable
SCM.
Although sustainable transportation service provider evaluation is an important
problem in SCM, in the literature, there are only a few studies on the subject. Paul et al.
[5] integrated expert opinion, best-worst approach, and VIKOR method to valuate
transport service providers using sustainable criteria. Mavi et al. [6] combined fuzzy
SWARA and fuzzy MOORA for evaluating the third-party reverse logistic providers in
the plastic industry by considering sustainability and risk factors. Yayla et al. [7]
employed Buckley’s fuzzy AHP and fuzzy TOPSIS methods for third-party logistic
providers evaluation.
This paper employs hierarchical fuzzy MCDM methodology for the evaluation of
sustainable transportation service providers. The methodology considers the conflicting
evaluation criteria under fuzzy environment, which are yielded in a hierarchy. More-
over, it ranks the alternatives by considering both the distances to ideal and anti-ideal
solutions. Furthermore, in the literature, the number of studies that handle the sus-
tainable transportation evaluation problem are not sufficient. This paper fills the gap on
the subject. The rest of the study is organized as follows. In Sect. 2, the employed
hierarchical fuzzy MCDM approach is delineated. The case study is illustrated in
Sect. 3. Conclusions are given in the final section.
2 Hierarchical Fuzzy MCDM Approach
This paper uses the hierarchical distance-based fuzzy MCDM algorithm developed by
Karsak and Ahiska [8] for determining the most suitable sustainable transportation
service provider. The methodology is explained as follows:
644 M. Dursun and E. Ari
Step 1. Organize a committee of experts and describe the alternatives and evalu-
ation criteria.
Step 2. Form the decision matrix that gives the importance weights of criteria and
sub-criteria, and the ratings of alternatives.
Step 3. Normalize the evaluation matrix as
8 y y
>
< y yjk ;
ijk
k 2 CBj ; i ¼ 1; 2; . . .; m; j ¼ 1; 2; . . .; n
y0ijk ¼ ð1Þ
jk jk
> yjk yijk

: yjk y ; k 2 CCj ; i ¼ 1; 2; . . .; m; j ¼ 1; 2; . . .; n
jk
where y0ijk gives the normalized value of yijk , which is the rating of alternative i via the
sub-criterion k of criterion j, m is the number of alternatives, n is the number of criteria,
CBj is the set of benefit-related criteria and CCj is the set of cost-related criteria,
yjk ¼ max yijk and yjk ¼ min yijk .
i i
Step 4. Aggregate the ratings of alternatives at the sub-criteria level to criteria level
via the following equation
P
~ 1 ~yijk
w
~xij ¼ ðxaij ; xbij ; xcij Þ ¼ Pjk 1
k
; 8i; j ð2Þ
kw~ jk
where ~xij is the aggregate rating of alternative i with respect to criterion j, w

~ 1jk indicates
the average importance weight assigned to sub-criterion k of criterion j.
Step 5. Normalize the aggregate ratings as follows:
!
xaij x
aj xbij xaj xcij xaj
~rij ¼ ðraij ; rbij ; rcij Þ ¼ ; ; ð3Þ
xcj x
aj xcj xaj xcj xaj

where xcj ¼ max xcij , x

aj ¼ min xaij .
i i
Step 6. Determine the ideal solution A ¼ ðr1 ; r2 ; . . .; rn Þ and the anti-ideal solution
A ¼ ðr1 ; r2 ; . . .; rn Þ, where rj ¼ ð1; 1; 1Þ and rj ¼ ð0; 0; 0Þ for j ¼ 1; 2; . . .; n.

Step 7. Compute the weighted distances from ideal solution and anti-ideal solution
(Di and Di , respectively) for each alternative as
X
1 1 1
Di ¼ j
1=2fmaxðwaj r aij 1 ; w cj rcij 1jÞ þ wbj rbij 1jg; i ¼ 1; 2; . . .; m ð4Þ
X
1 1 1
D
i ¼ j
1=2fmaxðw aj r aij 0 ; w cj r cij 0 jÞ þ w bj rbij 0jg; i ¼ 1; 2; . . .; m ð5Þ
Step 8. Compute the proximity of the alternatives to the ideal solution, Pi , as
;i¼1;2;...;m:
pi ¼ D
i =ðDi þ Di Þ ð6Þ
Step 9. Rank the alternatives according to Pi values in descending order.

3 Case Study
The case study is performed in a dye manufacturer in Turkey. Hierarchical fuzzy

MCDM method is adopted for transportation service provider selection problem for the
related case. This method has been proposed since the problem includes a hierarchical
structure of the criteria, uncertainty in evaluating the relative importance of criteria/sub-
criteria and grading of alternative projects.
The firm has 6 potential transportation service provider. First, evaluation criteria are
determined by reviewing the literature as in Table 1.
Table 1. Evaluation criteria [5].
Criteria
Economic Issues (C1)
Cost performance (C11)
Financial performance (C12)
Experience (C13)
Market share (C14)
Environmental Issues (C2)
Green technology utilization (C21)
Environment awareness (C22)
Recycling policy (C23)
Environmental legal and policy framework (C24)
Energy-efficient transportation utilization (C25)
Social Issues (C3)
Relationships with clients (C31)
Ease of communication (C32)
Labor relations (C33)
Ethical awareness (C34)
Health and safety (C35)
Employee welfare (C36)
Human rights (C37)
Operational Issues (C4)
Service quality (C41)
Location (C42)
Reputation (C43)
Responsiveness (C44)
Optimization capability (C45)
Delivery (C46)
Information technology utilization (C47)
Risk management policy (C48)
Information sharing (C49)
Capacity (C410)
Flexibility (C411)
The evaluation is performed by four experts and they give their opinions by con-
structing a consensus utilizing the fuzzy linguistic scale given in Table 2.
Table 2. Linguistic scale.

Linguistic term Fuzzy number
VH (0, 0, 0.25)
H (0, 0.25, 0.50)
M (0.25, 0.50, 0.75)
L (0.50, 0.75, 1)
VL (0.75, 1, 1)
The evaluations are given in Table 3.
Table 3. Evaluation of the alternatives and criteria.

Criteria Weights of criteria A1 A2 A3 A4 A5 A6
C1 H
C11 H M H L VH L M
C12 M M VH VL H M L
C13 VH H M M H H M
C14 M M H M M H H
C2 VH
C21 H L M M M H M
C22 H L L M L M H
C23 H M M L M M M
C24 M M L L M H L
C25 VH L VL VL L L VL
C3 H
C31 M H H H VH VH H
C32 H H M H M H VH
C33 M M L M M M H
C34 M L M VL L L M
C35 VH H VH H M H M
C36 M M H H M M H
C35 H M H L L M M
C4 VH
C41 VH M L H M H VH
C42 H H M VH H M L
C43 VH M M H M L M
C44 VH H VH H VH M H
C45 M L M VL L M L
(continued)
Criteria Weights of criteria A1 A2 A3 A4 A5 A6
C46 VH H H VH VH M M
C47 H L VL L H H M
C48 H M H VH M L H
C49 M M M H H M VH
C410 M H M VH M M H
C411 VH VL M M H VL L
Ratings of alternatives at the sub-criteria level are aggregated to criteria level

employing Eq. (2). Then, aggregated ratings, which are given in Table 4, are nor-
malized via Eq. (3).
Table 4. Normalized aggregated ratings.

Criteria A1 A2 A3 A4 A5 A6
C1 0.27 0.56 0.84 0.36 0.70 0.98 0 0.25 0.53 0.49 0.78 1 0.22 0.53 0.82 0.13 0.44 0.76
C2 0 0.36 0.72 0.03 0.29 0.67 0.03 0.29 0.67 0.07 0.43 0.80 0.24 0.62 1 0.17 0.43 0.82
C3 0.21 0.56 0.92 0.35 0.69 0.97 0.17 0.50 0.86 0 0.39 0.73 0.24 0.60 0.92 0.28 0.67 1
C4 0.1 0.43 0.81 0.14 0.49 0.81 0.41 0.77 1 0.35 0.71 1 0 0.34 0.74 0.18 0.56 0.88
The weighted distances from ideal solution and anti-ideal solution are computed
using Eqs. (4) and (5). The alternatives are ranked as in Table 5.
Table 5. Ranking of the alternatives.

Alternatives Pi* Ranking
A1 0.5486 5
A2 0.5865 3
A3 0.5419 6
A4 0.6197 1
A5 0.5838 4
A6 0.5875 2
4 Concluding Remarks
Sustainable transportation systems object to make contribution to economy and reduce

environmental damages by providing a number of benefits namely road safety, efficient
city management, energy efficiency, and reduced travel time and fuel consumption.
They are also critical to maintain sustainable supply chain and logistics management.
Sustainable transportation service provider evaluation is an important MCDM problem,
which requires considering multiple conflicting criteria yielding in a hierarchical
structure. Hence, this paper employs hierarchical fuzzy MCDM methodology to select
the most appropriate sustainable transportation service provider in a dye manufacturer
in Turkey.
Future researches may focus on calculating the weights of the evaluation criteria
employing an analytical technique. Moreover, a group decision making framework can
be utilized for the evaluation.
References
1. United Nations General Assembly: Report of the world commission on environment and
development: Our common future. Oslo, Norway: United Nations General Assembly,
Development and International Co-operation: Environment (1987)
2. Chang, H.L., Chen, P.C.: Exploring senior officials’ policy beliefs regarding sustainable
transportation. Transp. Res. Part D Transp. Environ. 14, 249–254 (2009)
3. Deakin, M., Curwell, S., Lombardi, P.: Sustainable urban development: the framework and
directory of assessment methods. J. Environ. Assess. Policy Manage. 4, 171–197 (2002)
4. CEU, C. of the E.U: Council Resolution on the Integration of Environment and Sustainable
Development into the Transport Policy (2001)
5. Paul, A., Moktadir, M.A., Paul, S.K.: An innovative decision-making framework for
evaluating transportation service providers based on sustainable criteria. Int. J. Prod. Res.
(2020). https://doi.org/10.1080/00207543.2019.1652779
6. Mavi, R.K., Goh, M., Zarbakhshnia, N.: Sustainable third-party reverse logistic provider
selection with fuzzy SWARA and fuzzy MOORA in plastic industry. Int. J. Adv. Manuf.
Technol. 91, 2401–2418 (2017)
7. Yayla, A.Y., Oztekin, A., Gumus, A.T., Gunasekaran, A.: A hybrid data analytic
methodology for 3PL transportation provider evaluation using fuzzy multi-criteria decision
making. Int. J. Prod. Res. 53(20), 6097–6113 (2015)
8. Karsak, E.E., Ahiska S.S.: Fuzzy multi-criteria decision making approach for transport
projects evaluation in Istanbul. In: Lecture Notes in Computer Science, vol. 3483, pp. 301–
311 (2005)
Development of Assessment Model
for Selection of Sustainable Energy Source
in India: Hybrid Fuzzy MCDM Approach
S. K. Saraswat(&), Abhijeet Digalwar, and S. S. Yadav
Department of Mechanical Engineering,

Birla Institute of Technology and Science, Pilani, Pilani 333031, India
[email protected]
Abstract. The development of India is continuously affected due to some

severe issues of the energy crisis and greenhouse gas emission. To overcome or
minimize these problems, India should increase the share of sustainable energy
sources in an overall generation. Therefore, this work aims to develop a model
for the selection of the most sustainable energy sources in India. The model is
developed using integrated multi-criteria decision-making (MCDM) approach to
deal with the number of conflicting and uncertain criteria. The developed model
has a unique combination of seven energy alternatives, six criteria, and their
related twenty-six sub-criteria. The weights were collected using linguistic ter-
minology to avoid any kind of incomplete or vague information. Fuzzy Analytic
Hierarchy Process (F-AHP) was employed to make a pairwise comparison and
to obtain the weights of the considered criteria and sub-criteria. Fuzzy Weighted
Aggregated Sum Product Assessment (F-WASPAS) used for the ranking of the
energy alternatives. Result shows that economic criterion has the highest weight,
followed by the environmental and technical criterion. Solar energy obtained as
the most sustainable alternative energy source in India. Wind energy was chosen
as the second-most sustainable alternative energy source, followed by the hydro
and biomass energy. Sensitivity analysis was performed by changing the values
of k coefficient. Results were compared and validated with three other well-
known MCDM approaches of VIKOR, TOPSIS, and PROMETHEE - II.
Keywords: Fuzzy AHP Fuzzy WSM MCDM Solar energy India
1 Introduction
Energy is a prime factor for the social and economic growth of the country. Energy can
be obtained from coal, natural gas, oil, nuclear, hydrogen, hydro, solar, wind, wave,
biomass, and geothermal. The energy sources nuclear, natural gas, oil, and coal are the
fossil fuel energy sources, and remaining wind, biomass, hydro, solar, hydrogen, and
geothermal are the renewable energy sources [1]. Nearly 81% of the world and 64% of
India’s energy generated from fossil fuels [2–4]. India is the third-largest country in the
energy generation and its consumption [5].
650 S. K. Saraswat et al.
Total installed power capacity of India till January 2020 is 368929.82 MW which
covered by the coal thermal (204664.50 MW), hydro (45699.2 MW), wind (37607.70
MW), solar (34035.66 MW), bioenergy (10001.11 MW), nuclear (6780 MW) and gas
power (24955.36 MW) [4]. For power generation, India mainly depends on fossil fuel-
based power plants, but it suffers from the problems of fossil fuel depletion, greenhouse
gas emission, and climate change. Renewable energy sources are environmentally
friendly, but they are uncertain and intermittent [6, 7]. So, there is a need to select the
most sustainable energy alternatives among renewable and non-renewable energy
sources covering technical, economic, environmental, social, political, and flexible
criteria. Therefore, in future energy policies and generation scenario, increases the
share of sustainable energy alternatives.
The problem with the selection of most sustainable alternative energy sources is not
such simply since we are dealing with several factors or aspects. To solve the multi
aspects problem, the MCDM approach is widely used. MCDM approach is a suitable
technique for highly uncertain conditions and conflicting criteria issues. It is also
commonly used in sorting and ranking of the alternatives [8]. In this work, a fuzzy
AHP approach is used for criteria as well as sub-criteria weights calculation due to its
simplicity and robustness. Similarly, WASPAS approach used for the ranking of the
energy alternatives because of a unique combination of WSM and WPM two MCDM
approaches.
The paper is organized in the six sections, namely introduction, literature review,
research methodology, results and discussion, validation of findings, and conclusions.
2 Literature Review
In the recent decade, a lot of research work has been done for the selection or iden-
tification of the most suitable energy source or to prioritize the different energy alter-
natives. In 2017, Colak and Kaya [1] prioritized the renewable energy sources (RES) in
Turkey using integrated fuzzy AHP and fuzzy TOPSIS (Technique for Order Prefer-
ence by Similarity to the Ideal Solution) MCDM approach. Haddad et al. [9] evolved
an AHP approach to rank different RES for the Algerian electricity system. They found
solar energy as the best RES, order followed by the wind, geothermal, biomass, and
hydropower. Lee and Chang [10] prioritized the five RES based on the four criteria and
their related ten sub-criteria. They also gave some recommendations for the develop-
ment of RES in Taiwan. Kaya and Kahraman [11] developed an MCDM approach for
the selection of the best energy technology in Turkey. They selected wind energy as the
best energy technology in Turkey, followed by biomass and solar energy. Kahraman
et al. [12] determined the most appropriate renewable energy alternative for Turkey.
They employed fuzzy AHP and fuzzy axiomatic design approach through expert choice
software. Sreeimikiene et al. [13] developed an MCDM approach to select the most
sustainable electricity production technology in Lithuania. Kahraman and Kaya [14]
developed an integrated AHP-VIKOR approach for the selection of suitable renewable
energy sources and appropriate production sites in Istanbul, Turkey. Ahmad and Tahar
[15] identified the most appropriate RES in Malaysia for electricity generation. Based
on their analysis, they highlighted that solar energy is the most appropriate RES in
Development of Assessment Model for Selection of Sustainable Energy Source 651
Malaysia. Grilli et al. [16] adopted the MCDM approach to identify the best RES for
energy production in the Alps. Especially for India, there is no such literature available
for assessment of the most sustainable alternative energy sources covering multi-
dimensional aspects. The work sets a milestone of consideration of seven energy
alternatives with highest six criteria and twenty six sub-criteria.
Research methodology explains the steps of fuzzy AHP and fuzzy WASPAS
approaches, and introduction of the considered experts. In the research work, three
experts were selected based on their work field experience. First expert is an energy
resource allocation expert, second expert is an academia, and third expert is a man-
agerial post environment expert. The judgments were collected using linguistic ter-
minology to avoid any kind of incomplete and vague information.
3.1 Fuzzy Analytic Hierarchy Process (F-AHP)

AHP approach is developed by prof. Thomas L. Saaty in the 1970s. For assessment of
most sustainable energy sources, a hierarchal network of the goal, criteria, and alter-
natives has been developed. Then, perform pairwise comparison to obtain relative
importance of the criteria as well as sub-criteria. Weights were calculated using fuzzy
geometric mean values. Here n is the number of criteria in Eq. 1.
x ¼ ðp1 p2 . . . pn Þ1=n ; y ¼ ðq1 q2 . . . qn Þ1=n ; z ¼ ðr1 r2 . . . rn Þ1=n

ð1Þ
Fuzzy weights were obtained using Eq. 2.
wl ¼ al ða1 þ a2 þ . . . þ an Þ1 ; wm ¼ am ða1 þ a2 þ . . . þ an Þ1 ;

ð2Þ
wu ¼ au ða1 þ a2 þ . . . þ an Þ1
De-fuzzified crisp numeric values were obtained by the average of all three values of
fuzzy lower, middle, and upper.
3.2 Fuzzy WSM, WPM, and WASPAS

Weighted Sum Model (WSM)
Xn
AWSM
i ¼ j¼1
Wj xij ð3Þ
Where Wij = weight of each criteria, xij = performance value of alternative for that
sub-criteria, Ai = weighted normalized matrix
Weighted Product Model (WPM)

Yn
AWPM
i ¼ xW
j¼1 ij
ð4Þ
Weighted Aggregated Sum Product Assessment (WASPAS) is a unique technique,

which is a combination of two well-known MCDM approaches of WSM and WPM.
Qi ¼ kQ1i þ ð1 kÞQ2i ð5Þ
Where, Qi = joint generalized criterion of WASPAS, k should be (0 k 1), if

k = 0 then Qi = Q2i (WPM), Qi = Q1i (WSM)
4.1 Application of Fuzzy AHP for Weights Calculation

Fuzzy AHP is the most widely used MCDM approach because it breaks the problem in
the hierarchical network tree and then solves it [9, 17]. In the Fuzzy AHP approach, a
pair-wise comparison is made among the criteria and sub-criteria to obtain the weights.
The expert’s linguistic weights were converted into the fuzzy weight by using the
discussed methodology. Table 1 shows the fuzzy geometric mean value, fuzzy weights,
crisp weights, and ranks of the criteria.
Table 1. Determination of criteria weightage using fuzzy AHP MCDM approach.

Criteria A fuzzy geometric Fuzzy weights COA (center of Ranking
mean value area)
Economic (1.026, 1.098, 1.380) (0.199, 0.186, 0.194 1
0.198)
Technical (0.920, 1.076, 1.215) (0.178, 0.183, 0.178 3
0.174)
Social (0.800, 0.987, 1.089) (0.155, 0.1696, 0.160 5
0.156)
Environmental (0.987, 1.076, 1.328) (0.191, 0.183, 0.188 2
0.191)
Political (0.891, 0.981, 1.127) (0.173, 0.166, 0.167 4
0.162)
Flexible (0.531, 0.671, 0.826) (0.103, 0.114, 0.113 6
0.118)
Consistency ratio = 0.0208 < 0.10
By performing pair-wise comparison among the criteria, it is obtained that eco-

nomic criterion has the highest weight. Environmental is the second-highest weight
criterion, the order followed by the technical, political, social, and flexible criterion, as
shown in Table 1.
By performing the pair-wise comparison, obtain the local weights of the criteria as
well as sub-criteria. Finally, to get the sub-criteria global weights, local weights of
criteria and sub-criteria are multiplied, as shown in Table 2.
Table 2. Calculation of fuzzy AHP approach for global weights of the sub-criteria.
Criteria Local Sub-criteria Local Global
Economic 0.194 Capital cost/investment cost (C1) 0.176 0.0341
0.194 Operation and maintenance cost, 0.158 0.0307
(C2)
0.194 Levelized cost of energy, (C3) 0.135 0.0262
0.194 Payback Period, (C4) 0.125 0.0243
0.194 Operational life, (C5) 0.109 0.0211
0.194 Fuel cost, (C6) 0.143 0.0277
0.194 Availability of funds and incentives, 0.154 0.0299
(C7)
Technical 0.178 Technology maturity, (C8) 0.182 0.0324
0.178 Efficiency, (C9) 0.218 0.0388
0.178 Deployment time, (C10) 0.243 0.0433
0.178 Capacity factor, (C11) 0.202 0.0360
0.178 Reliability (C12) 0.155 0.0276
Social 0.160 Social benefits (education, science, 0.261 0.0418
and culture) (C13)
0.160 Job creation, (C14) 0.271 0.0434
0.160 Social acceptance, (C15) 0.260 0.0416
0.160 Impact on human health, (C16) 0.208 0.0333
Environment 0.188 Land requirement, (C17) 0.167 0.0314
0.188 Emission reduction, (C18) 0.350 0.0658
0.188 Impact on ecosystem, (C19) 0.268 0.0504
0.188 Climate Change, (C20) 0.215 0.0404
Political 0.167 Political acceptance, (C21) 0.186 0.0311
0.167 Foreign dependency, (C22) 0.261 0.0436
0.167 Fuel reserve years, (C23) 0.329 0.0549
0.167 Compatibility with national energy 0.224 0.0374
policy (C24)
Flexible 0.113 In integration with other sources, 0.432 0.0488
(C25)
0.113 In fulfilling the peak load demand, 0.568 0.0642
(C26)
4.2 Application of WASPAS for Ranking of the Energy Alternatives

To rank the energy alternatives, a well-known WASPAS MCDM approach is used. It is
a combination of WSM, and WPM MCDM approaches. Weights to each energy
alternatives are given according to the performance for that sub-criterion. All the three
experts were given weights using the provided fuzzy scale. Weights were collected
only in a qualitative manner, but for the quantitative criteria, the data value table was
provided to experts. Therefore, according to the data value table, they can give qual-
itative weights to the quantitative criteria.
By performing the calculation as discussed in the research methodology, solar
energy obtained the highest weighted normalized index of 0.740391 as shown in
Table 3. Therefore, solar energy is chosen as the most sustainable energy alternative
source. Wind energy is the second-most sustainable alternative energy source, the order
followed by the hydro, biomass, gas power, nuclear, and thermal energy respectively,
as shown in Fig. 1.
Table 3. The final ranking of energy alternatives.

Energy ALTERNATIVES WSM WPM WASPAS Ranking
A1 (Thermal energy) 0.42711 0.366358 0.396734 7
A2 (Hydro energy) 0.674649 0.627853 0.651251 3
A3 (Gas power) 0.496484 0.436182 0.466333 5
A4 (Wind energy) 0.755393 0.712004 0.733698 2
A5 (Nuclear energy) 0.494317 0.421905 0.458111 6
A6 (Biomass energy) 0.625129 0.583698 0.604413 4
A7 (Solar energy) 0.774325 0.706457 0.740391 1
Thermal Hydro Gas Wind Nuclear Biomass Solar

1
Preference Score
0.8
0.6
0.4
0.2
WSM WPM WASPAS
Fig. 1. Preference score of energy alternatives for the WSM, WPM, and WASPAS MCDM
approaches.
Table 4. Weighted normalized index of energy alternatives for different values of k.

k=0 k = 0.1 k = 0.2 k = 0.3 k = 0.4 k = 0.5 k = 0.6 k = 0.7 k = 0.8 k = 0.9 k=1
A1 0.366 0.079 0.159 0.238 0.317 0.397 0.476 0.555 0.635 0.714 0.793
A2 0.628 0.130 0.261 0.391 0.521 0.651 0.782 0.912 1.042 1.172 1.303
A3 0.436 0.093 0.187 0.280 0.373 0.466 0.560 0.653 0.746 0.839 0.933
A4 0.712 0.147 0.293 0.440 0.587 0.734 0.880 1.027 1.174 1.321 1.467
A5 0.422 0.092 0.183 0.275 0.366 0.458 0.550 0.641 0.733 0.825 0.916
A6 0.584 0.121 0.242 0.363 0.484 0.604 0.725 0.846 0.967 1.088 1.209
A7 0.706 0.148 0.296 0.444 0.592 0.740 0.888 1.037 1.185 1.333 1.481
Sensitivity analysis is performed by changing the values of k in Eq. 5 of the

research methodology. For sensitivity analysis, the value of k varies from k = 0 to
k = 1 with an interval of 0.1. Table 4 shows the weighted normalized index of con-
sidered seven energy alternatives for different values of k. The sensitivity results
indicated that the final ranking of the energy alternatives is not sensitive with the values
of k.
Table 5. A comparative score of energy alternatives in six MCDM approaches.

TOPSIS VIKOR WSM WPM WASPAS PROME. - II
Closeness VIKOR Weighted Weighted Weighted Net
coefficient index normalized normalized normalized outranking
index index index flow
A1 0.435466 1 0.42711 0.366358 0.396734 −0.27036
A2 0.64178 0.328876 0.674649 0.627853 0.651251 0.103229
A3 0.490411 0.702538 0.496484 0.436182 0.466333 −0.17352
A4 0.717482 0.029332 0.755393 0.712004 0.733698 0.232969
A5 0.486935 0.589991 0.494317 0.421905 0.458111 −0.16809
A6 0.595603 0.36354 0.625129 0.583698 0.604413 0.018827
A7 0.738649 0 0.774325 0.706457 0.740391 0.256949
5 Validation of Findings
There is a necessity to validate the proposed model results. The proposed model
calculated the criteria weights by the fuzzy AHP approach and energy alternatives
rankings by the WASPAS approach. In the validation process, the fuzzy AHP approach
criteria weights are used, and ranking is given by the six MCDM approaches of WSM,
WPM, WASPAS, VIKOR, TOPSIS, and PROMETHEE-II. By comparing the results,
it is found that solar energy obtained the rank first in five MCDM approaches out of six
approaches, as shown in Table 5. In VIKOR and PROMETHEE-II approaches, nuclear
energy moved one rank higher and occupied the fifth position. Similarly, in WPM
approach, wind energy switches their position with solar energy. Thermal energy got
the last rank in all the considered seven energy sources.
6 Conclusions
In this work, an integrated fuzzy AHP and fuzzy WASPAS MCDM approach were
adopted for the selection of the most sustainable energy alternative source in India. The
selection of the most sustainable energy alternative source is a complex and conflicting
issue problem because of the number of conflicting parameters. To deal with the
number of conflicting criteria and sub-criteria an MCDM approach is adopted. This
study analyzed with six criteria and their relevant twenty-six sub-criteria for seven
alternative energy sources. The fuzzy AHP approach is used for criteria and sub-criteria
weights and WASPAS approach for the ranking of the energy alternatives. By the pair-
wise comparison, it is obtained that economic criterion is the highest local weights
criterion, followed by the environmental and technical criterion. By the calculation and
analysis, solar energy is chosen as the most sustainable alternative energy source in
India. Wind energy is the second choice, and hydro energy is at the third position, the
order followed by biomass, gas power, nuclear, and thermal energy. In the WSM and
WASPAS approaches, solar energy is at the first position, but in WPM, wind energy
switches the order with solar energy.
The sensitivity analysis result shows that the ranking of energy alternatives is
consistent and stable with the change in k coefficient values.
Ranking of the energy alternatives was compared and validated with TOPSIS,
VIKOR, and PROMETHEE-II MCDM approaches results. By comparison, solar
energy is chosen as the most sustainable alternative energy source in India.
Research work has the limitations of consideration of only major renewable and
non-renewable energy sources. In future, research work can be enhanced by consid-
ering the energy mix or fuel mix scenario.
References
1. Colak, M., Kaya, I.: Prioritization of renewable energy alternatives by using an integrated
fuzzy MCDM model: a real case application for Turkey. Renew. Sustain. Energy Rev. 80,
840–853 (2017)
2. Celikbilek, Y., Tuysuz, F.: An integrated grey based multi-criteria decision making approach
for the evaluation of renewable energy sources. Energy 115, 1246–1258 (2016)
3. Yuan, J., Li, C., Li, W., Liu, D., Li, X.: Linguistic hesitant fuzzy multi-criterion decision-
making for renewable energy: a case study in Jilin. J. Clean. Prod. 172, 3201–3214 (2018)
4. Ministry of Power home page. https://npp.gov.in/dashBoard/cp-map-dashboard. Accessed
23 Jan 2019
5. Srikanth, R.: India’s sustainable development goals – glide path for India’s power sector.
Energy Policy 123, 325–336 (2018)
6. Horst, D., Jentsch, M., Pfennig, M., Mitra, I., Bofinger, S.: Impact of renewable energies on
the Indian power system: Energy meteorological influences and case study of effects on
existing power fleet for Rajasthan state. Energy Policy 122, 486–498 (2018)
7. Chandel, S.S., Shrivastva, R., Sharma, V., Ramasamy, P.: Overview of the initiatives in
renewable energy sector under the national action plan on climate change in India. Renew.
Sustain. Energy Rev. 54, 866–873 (2016)
8. Afsordegan, A., Sánchez, M., Agell, N., Zahedi, S., Cremades, L.V.: Decision making under
uncertainty using a qualitative TOPSIS method for selecting sustainable energy alternatives.
Int. J. Environ. Sci. Technol. 13(6), 1419–1432 (2016)
9. Haddad, B., Liazid, A., Ferreira, P.: A multi-criteria approach to rank renewables for the
Algerian electricity system. Renew. Energy 107, 462–472 (2017)
10. Lee, H.-C., Chang, C.-T.: Comparative analysis of MCDM methods for ranking renewable
energy sources in Taiwan. Renew. Sustain. Energy Rev. 92, 883–896 (2018)
11. Kaya, T., Kahraman, C.: Multicriteria decision making in energy planning using a modified
fuzzy TOPSIS methodology. Expert Syst. Appl. 38, 6577–6585 (2011)
12. Kahraman, C., Kaya, I., Sebi, S.: A comparative analysis for multi-attribute selection among
renewable energy alternatives using fuzzy axiomatic design and fuzzy analytic hierarchy
process. Energy 34, 1603–1616 (2009)
13. Streimikiene, D., Sliogeriene, J., Turskis, Z.: Multi-criteria analysis of electricity generation
technologies in Lithuania. Renew. Energy 85, 148–156 (2016)
14. Kahraman, C., Kaya, I.: A fuzzy multicriteria methodology for selection among energy
alternatives. Expert Syst. Appl. 37, 6270–6281 (2010)
15. Ahmad, S., Tahar, R.M.: Selection of renewable energy sources for sustainable development
of electricity generation system using analytic hierarchy process: a case of Malaysia. Renew.
Energy 63, 458–466 (2014)
16. Grilli, G., De Meo, I., Garegnani, G., Paletto, A.: A multi-criteria framework to assess the
sustainability of renewable energy development in the Alps. J. Environ. Plann. Manag. 7
(60), 1276–1295 (2017)
17. Pohekar, S., Ramachandran, M.: Application of multi-criteria decision making to sustainable
energy planning. Renew. Sustain. Energy Rev. 8, 365–381 (2004)
Evaluation of Criteria that Affect
the Sustainability of Smart Supply Chain
in a Textile Firm by Fuzzy SWARA Method
Arzu Organ , Kevser Arman(&) , and Ali Katrancı
Pamukkale University, 20160 Denizli, Turkey

[email protected]
Abstract. Sustainable supply chain management practices create a positive

effect on the image of businesses, while smart supply chain practices provide a
competitive advantage by providing flexibility, speed, risk management, cost
savings, inventory accuracy, etc. Smart supply chain management practices help
firms to achieve more sustainability in the economic, environmental and social
dimensions with innovative technologies in the supply chain. Firms have to give
importance to sustainability in order to remain competitive and enhance firm’s
long-term performance. For this reason, the aim of this study is to determine
important criteria that affect the sustainability of smart supply chain. The criteria
discussed in this study were evaluated in the scope of economic, environmental
and social practices. In order to acquire more accurate results, the criteria were
evaluated by fuzzy SWARA method. The results were discussed and the most
important main criteria and sub-criteria that affect the sustainability of the smart
supply chain have been revealed.
Keywords: Sustainability Smart supply chains Fuzzy SWARA
1 Introduction
Traditional supply chains encompass whole processes from the design phase of the
product or service to the delivery to the customer [1]. It is a flow management from
manufacturer to consumer. Industry 4.0, the fourth industrial revolution, aims to make
that production less costly, better quality, faster, more flexible, more agile and envi-
ronmentally friendly with the numerous innovative technologies [2, 3]. Supply chain
practices are increasingly becoming smart with the using of innovative technologies.
Smart supply chain management (SSCM) provides unprecedented opportunities for
cost reduction, product quality, delivery process and flexibility etc. Thanks to the smart
technologies in the supply chain perfect amount of information can be collected and
used to make various decisions [4]. The traditional supply chains faces various prob-
lems such as failure to meet the demand on time, cost, uncertainty, cost, complexity etc.
Supply chains with smart technologies overcome these problems in the supply chains
[5]. With the conversion of traditional supply chains into smart supply chains ensures
that all elements in supply chain are widely understandable, explainable, predictable
and manageable by decision units. Sustainability is generally known as the ability to
Evaluation of Criteria that Affect the Sustainability 659
meet the needs of internal and external customers without jeopardizing the resources
needed by future generations [6]. Sustainability has become a huge buzzword both in
firms trying to survıve and among the environmentally conscious society [7]. Sus-
tainability is examined economically, socially and environmentally and connections
between these three dimensions is the basic requirement for the true understanding of
sustainability. Economic, environmental and social challenges do not stop only at the
gates of companies, also considerably affect the supply chain management, therefore
these have to be considered along the supply chains management [8]. Smart tech-
nologies has revealed as indispensable tools for improving the sustainable supply chain
management practices [9]. For instance, the application of IoT can used to reduction
logistic costs (as a economic aspect) and pollution emission (as a environmental
aspect). On the other hand the application of VR/AR can provide healthy and safety
working places also can ensure that educational processes are more efficient (as a social
aspect) [3]. Sustainability of supply chain provide long-term economic benefits and
competitive advantage for the firm without damaging to environment and society [10].
Using smart practices in the supply chain helps make sustainability term more
understandable and feasible for. These two subjects, which are thought to be closely
related to each other, are composed the theme of the this paper.
This study aim to evaluate the criteria that affect the sustainability of smart supply
chain practices in a textile firm using fuzzy SWARA. The multi-criteria decision-
making approach (MCDM) allows the decision maker to choose the best alternative
among the possible solution sets in which multiple criteria are optimized [11]. New
step-wise weight assessment ratio analysis method (SWARA) one of the MCDM
methods allows the experts to provide opinions and assess about the level of signifi-
cance of criteria or in other words attributes in a rational decision making process [12].
However, it is inadequate in uncertainty conditions, that’s why fuzzy set approaches in
MCDM can be used when human’s (e.g. expert) knowledge, experiences, evaluations
are needed [13]. Fuzzy based approaches have been used by researchers due to the
uncertainty in the data and unclear expert evaluations etc. For this reason SWARA is
integrated with fuzzy set theory, which results in fuzzy SWARA. This paper aim to
determine the impact of smart supply chain practices on sustainability (as economic,
environmental and social). The novelty of this article is that it utilize the fuzzy
SWARA method for the first time to determine the impact of smart supply chain
practices on sustainability. The rest of the paper is organized as follows: In Sect. 2
presents the methodology used to evaluate sustainability, In Sect. 3 presents the
application of methods in textile firm and its conclusion. In Sect. 4 summarizes the
results, contributions and suggestion for the future research.
2 Methodology
2.1 Fuzzy SWARA
The SWARA method is a applicable technique introduced to determine the relative
weights of the criteria [14–16]. On decision making process some factors can cause
uncertainty due to deficient, unquantifiable, unobtainable information. In conventional
660 A. Organ et al.
MCDM methods, due to such uncertainties, problems cannot be handled effectively.

Therefore, fuzzy multiple attribute decision making methods help the decision maker at
this point [17]. That’s why SWARA method is integrated with fuzzy set theory in this
study. The process of determining the relative weights of the criteria using the fuzzy
SWARA method occurs as follow: [18].
Table 1. The fuzzy scale [19]

Linguistic terms Response scale
Equally important (EI) (1, 1, 1)
Moderately less important (MOL) (2/3, 1, 3/2)
Less important (LI) (2/5, 1/2, 2/3)
Very less important (VL) (2/7, 1/3, 2/5)
Much less important (MUL) (2/9, 1/4, 2/7)
Step 1. The evaluation criterias are ranked from the most significant to the least
significant considering the goal of the decision-making criteria by experts.
Step 2. Considering the fuzzy scale in Table 1, experts give a score to the factor
j interrelatedly the previous criterion (j − 1) which has higher significance. It is
repeated for the all criteria and utilizing the arithmetic means of the corresponding
scores, aggregated avarage values for evaluation criteria are obtained. It represents
comparative importance of ~sj value [19, 20].
Step 3. The coefficient ~kj is obtained as:

~1 j¼1
~kj ¼
~sj þ ~1 j[1
Step 4. The fuzzy weight ~qj is obtained as:

(
~1 j¼1
~qj ¼ xj1
kj j[1
~ j is obtained as equation
Step 5. The relative weight w
~qj
~ j ¼ Pn
w
k¼1 ~qk
where w~ j ¼ ðwlj ; wm u
j ; wj Þ is the fuzzy relative importance weight of the jth criterion
and n is the number of criteria. Fuzzy numbers A1 = (l1 , m1, u1), where l1
m1 u1 and A2 = (l2 , m2, u2) where l2 m2 u2 are presented as
follows [17]:
Fuzzy addition is obtained as:
A1 A2 ¼ ðl1 þ l2 ; m1 þ m2 ; u1 þ u2 Þ
Fuzzy subtraction is obtained as:
Fuzzy Multiplication is obtained as:
A1 A2 ¼ ðl1 l2 ; m1 m2 ; u1 u2 Þ
Fuzzy division is obtained as:
~ j ¼ ðwlj ; wm
These fuzzy weights w u
j ; wj Þ is defuzzified is obtained as:
wlj þ wm u
j þ wj
~ ¼
dðAÞ
3
In the last step, to obtain the fuzzy relative weights, the fuzzy relative weights of each
sub-criteria are multiplied by the weight of the corresponding main criterion.
3 Application
It is applied the real-life problem to evaluate the impact of smart supply chain practices
on sustainability in this paper. Three experienced decision makers (two experts and one
manager) working in the supply chain department of the textile firm evaluated the
relevant criteria. The criteria used in the study were determined as a result of com-
prehensive literature analysis. Consequently, three main criteria and total twelve sub-
criteria were determined. The criteria are C1 (economic); C11: supply flexibility [3],
C12: product quality [5], C13: cost reduction [21], C14: smart delivery [9], C2: (en-
vironment); C21: green and smart manufacturing [22], C22: green-eco design [23, 24],
C23: green and smart purchasing [24], C24: green and smart Logistics [3], C31: voice
of the customer [24, 25], C32: health and safety working environment [26], C33:
ensure the employee’s right [27] and C34: employee’s development in smart atmo-
sphere [22]. The weight of the relevant criteria was determined by the fuzzy SWARA
method.
3.1 Fuzzy SWARA Results

In this section, fuzzy SWARA method is used to determine the weight of the evaluation
criteria. The results of fuzzy SWARA are shown in Tables 2, 3 and 4.
662 A. Organ et al.
Table 2. Local weights of economic, environmental and social sub-criteria

Sj kj qj wj
Economic (C1) (1,1,1) (1,1,1) (0.46,0.49,0.54)
Environmental (C2) (0.49,0.67,0.94) (1.49,1.67,1.94) (0.51,0.60,0.67) (0.24,0.30,0.36)
Social (C3) (0.32,0.39,0.49) (1.32,1.39,1.49) (0.35,0.43,0.51) (0.16,0.21,0.27)
Table 3. Local weights of economic, environmental and social sub-criteria

Sj kj qj wj
Economic
C11 (1,1,1) (1,1,1) (0.42,0.47,0.56)
C12 (0.45,0.61,1.06) (1.45,1.61,2.06) (0.49,0.62,0.69) (0.21,0.29,0.38)
C13 (0.58,0.83,1.22) (1.58,1.83,2.22) (0.22,0.34,0.44) (0.09,0.16,0.24)
C14 (0.78,1,1.33) (1.78,2,2.33) (0.09,0.17,0.25) (0.04,0.08,0.14)
Environmental
C21 (1,1,1) (1,1,1) (0.40,0.43,0.47)
C22 (0.32,0.39,0.49) (1.32,1.39,1.49) (0.67,0.72,0.76) (0.27,0.31,0.36)
C23 (0.58,0.83,1.22) (1.58,0.83,2.22) (0.30,0.39,0.48) (0.12,0.17,0.23)
C24 (1,1,1) (2,2,2) (0.15,0.20,0.24) (0.06,0.09,0.11)
Social
C31 (1,1,1) (1,1,1) (0.44,0.48,0.53)
C32 (0.45,0.61,0.86) (1.45,1.61,1.86) (0.54,0.62,0.69) (0.24,0.30,0.36)
C33 (0.89,1,1.17) (1.89,2,2.17) (0.25,0.31,0.36) (0.11,0.15,0.19)
C34 (0.78,1,1.33) (1.78,2,2.33) (0.11,0.26,0.21) (0.05,0.07,0.11)
Table 4. Final and relative weights of evaluation criteria

Evaluation criteria Aggregated weights Defuzzifed Normalised weights Relative weights
C1 (0.46,0.49,0.54) 0.49 0.49
C11 (0.19,0.23,0.30) 0.24 0.47 0.23
C12 (0.09,0.14,0.21) 0.15 0.29 0.14
C13 (0.04,0.08,0.13) 0.08 0.16 0.08
C14 (0.02,0.04,0.07) 0.04 0.08 0.04
C2 (0.24,0.30,0.36) 0.30 0.30
C21 (0.10,0.13,0.17) 0.13 0.43 0.13
C22 (0.06,0.09,0.13) 0.09 0.31 0.09
C23 (0.03,0.05,0.08) 0.05 0.17 0.05
C24 (0.01,0.03,0.04) 0.03 0.09 0.03
C3 (0.16,0.21,0.27) 0.21 0.21
C31 (0.07,0.10,0.14) 0.11 0.47 0.10
C32 (0.04,0.06,0.10) 0.07 0.30 0.06
C33 (0.02,0.03,0.05) 0.03 0.15 0.03
C34 (0.01,0.02,0.03) 0.02 0.08 0.02
The weights of these criteria were calculated by fuzzy SWARA method. According
to the evaluation results, smart supply chain practices have the most impact on eco-
nomic (C1 = 0.49) sustainability and this is followed by environmental (C2 = 0.30)
and social (C3 = 0.21) sustainability. In addition to, C11 (supply flexibility), C12
(product quality), C21 (Green and smart manufacturing), C31 (Voice of the customer)
and C22 (Green-Eco Design) emerges as the first five criteria that affect the sustain-
ability of smart supply chain practices in textile firm.
4 Conclusion
Importance of sustainability increase day by day among today’s customers who live in
smart age and It is inevitable for companies that attach importance to sustainability in
all their operations to gain competitive advantage in the long run. This article is
prepared to reveal the effect of smart supply chain applications on sustainability. For
this reason fuzzy SWARA method were utilized. The 12 criteria identified have had a
positive impact on the economic, social and environmental sustainability of the entire
supply chain in the textile firm. According to the evaluation results, smart supply chain
practices have the most impact on economic (C1 = 0.49) sustainability and this is
followed by environmental (C2 = 0.30) and social (C3 = 0.21) sustainability. As a
result, utilizing smart technologies in a supply chain contribute more to economic and
environmental sustainability, while its impact on social sustainability is low. According
to sub-criteria results, C11 (supply flexibility) is the most important and C34 (em-
ployee’s development in smart atmosphere) the least important criterion.
Findings obtained in this study may benefit firms operating in the textile industry in
terms of the contribution of smart supply chain practices to sustainability as economic,
environmental and social. More broadly, the findings of this study can benefit any
industry that uses smart technologies in the supply chain. There are a limited number of
studies in the literature that evaluate the impact of the smart supply chain, which is a
very popular topic, on sustainability with fuzzy MCDM. Evaluation of factors affecting
sustainability in smart supply chains in the textile firm by using fuzzy SWARA has not
been used in the literature. That’s why, it is think about this paper will fill the gap in the
literature. In addition, on the future research, the results can be compared with different
sectors and these 12 criteria can be used to evaluation and select the smart-sustainable
suppliers by the other researchers with different MCDM methods.
References
1. Hugos, M.H.: Essentials of Supply Chain Management, 4th edn. Wiley, Hoboken (2018)
2. Yıldız, A.: Endüstri 4.0 ve akıllı fabrikalar. Sakarya Üniversitesi Fen Bilimleri Enstitüsü
Dergisi 22(2), 546–556 (2018)
3. Chen, Z., Ming, X., Zhou, T., Chang, Y.: Sustainable supplier selection for smart supply
chain considering internal and external uncertainty: an integrated rough-fuzzy approach.
Appl. Soft Comput. 87, 106004 (2020)
664 A. Organ et al.
4. Wu, L., Yue, X., Jin, A., Yen, D.C.: Smart supply chain management: a review and
implications for future research. Int. J. Logist. Manag. 27, 395–417 (2016)
5. Abdel-Basset, M., Manogaran, G., Mohamed, M.: Internet of Things (IoT) and its impact on
supply chain: a framework for building smart, secure and efficient systems. Future Gener.
Comput. Syst. 86, 614–628 (2018)
6. WCED (World Commission on Environment and Development): Our Common Future.
Oxford University Press, Oxford (1987)
7. Carter, C.R., Easton, P.L.: Sustainable supply chain management: evolution and future
directions. Int. J. Phys. Distrib. Logist. Manag. 41(1), 46–62 (2011)
8. Seuring, S., Sarkis, J., Muller, M., Rao, P.: Sustainability and supply chain management an
introduction to the special issue. J. Cleaner Prod. 16(14), 1545–1551 (2008)
9. Büyüközkan, G., Göçer, F.: Digital supply chain: literature review and a proposed
framework for future research. Comput. Ind. 97, 157–177 (2018)
10. Carter, C.R., Rogers, D.S.: A framework of sustainable supply chain management: moving
toward new theory. Int. J. Phys. Distrib. Logist. Manag. 38(5), 360–387 (2008)
11. Yıldırım, B.F., Önder, E..: İşletmeciler, mühendisler ve yöneticiler için operasyonel,
yönetsel ve stratejik problemlerin çözümünde çok kriterli karar verme yöntemleri. Dora
Yayınları. 2. Baskı, Bursa (2015)
12. Khan, S.U., Niazi, M., Ahmad, R.: Factors influencing clients in the selection of offshore
software outsourcing vendors: an exploratory study using a systematic literature review.
J. Syst. Softw. 84(4), 686–699 (2011)
13. Kahraman, C.: Fuzzy Multi-criteria Decision Making: Theory and Applications with Recent
Developments. Springer, Heidelberg (2008)
14. Keršuliene, V., Zavadskas, E.K., Turskis, Z.: Selection of rational dispute resolution method
by applying new step-wise weight assessment ratio analysis (SWARA). J. Bus. Econ.
Manag. 11(2), 243–258 (2010)
15. Stanujkic, D., Karabasevic, D., Zavadskas, E.K.: A framework for the selection of a
packaging design based on the SWARA method. Inzinerine Ekonomika – Eng. Econ. 26(2),
181–187 (2015)
16. Çakır, E., Bilge, E.: Bütünleşik SWARA – MOORA Yöntemi ile Kurumsal Müşterilerin
Banka Tercihlerinin Belirlenmesi: Aydın İlinde Bir Uygulama. Anemon Muş Alparslan
Üniversitesi Sosyal Bilimler Dergisi 7(6), 269–289 (2019)
17. Mavi, R.K., Goh, M., Zarbakhshnia, N.: Sustainable third-party reverse logistic provider
selection with fuzzy SWARA and fuzzy MOORA in plastic industry. Int. J. Adv. Manuf.
Technol. 91(5–8), 2401–2418 (2017)
18. Perçin, S.: An integrated fuzzy SWARA and fuzzy AD approach for outsourcing provider
selection. J. Manuf. Technol. Manag. 30(2), 531–552 (2019)
19. Zarbakhshnia, N., Soleimani, H., Ghaderi, H.: Sustainable third-party reverse logistics
provider evaluation and selection using fuzzy SWARA and developed fuzzy COPRAS in the
presence of risk criteria. Appl. Soft Comput. 65, 307–319 (2018)
20. Ghorabaee, M.K., Amiri, M., Zavadskas, E.K., Antucheviciene, J.: A new hybrid fuzzy
MCDM approach for evaluation of construction equipment with sustainability considera-
tions. Arch. Civil Mech. Eng. 18(1), 32–49 (2018)
21. Tiwari, S., Wee, H.M., Daryanto, Y.: Big data analytics in supply chain management
between 2010 and 2016: Insights to industries. Comput. Ind. Eng. 115, 319–330 (2018)
22. Frank, A.G., Dalenogare, L.S., Ayala, N.F.: Industry 4.0 technologies: ımplementation
patterns in manufacturing companies. Int. J. Prod. Econ. 210, 15–26 (2019)
23. Foo, P.Y., Lee, V.H., Tan, G.W.H., Ooi, K.B.: A gateway to realising sustainability
performance via green supply chain management practices: a PLS–ANN approach. Expert
Syst. Appl. 107, 1–14 (2018)
24. Diabat, A., Kannan, D., Mathiyazhagan, K.: Analysis of enablers for implementation of
sustainable supply chain management—a textile case. J. Clean. Prod. 83, 391–403 (2014)
25. Hsu, C.-C., Liou, J.J.H., Chuang, Y.-C.: Integrating DANP and modified grey relation
theory for the selection of an outsourcing provider. Expert Syst. Appl. 40(6), 2297–2304
(2013)
26. Ajayi, A., Oyedele, L., Delgado, J.M.D., Akanbi, L., Bilal, M., Akinade, O., Olawale, O.:
Big data platform for health and safety accident prediction. World J. Sci. Technol. Sustain.
Dev. 16, 2–21 (2019)
27. Gold, S., Awasthi, A.: Sustainable global supplier selection extended towards sustainability
risks from (1 + n)th tier suppliers using fuzzy AHP based approach. IFAC-Papers OnLine 48
(3), 966–971 (2015). https://doi.org/10.1016/j.ifacol.2015.06.208
Green Supplier Selection Using Intuitionistic
Fuzzy AHP and TOPSIS Methods: A Case
Study from the Paper Mills
Ezgi Demir1(&) and Gözde Koca2(&)

1
Faculty of Business and Administrative Sciences, Piri Reis University,
İstanbul, Turkey
[email protected]
2
Faculty of Business and Administrative Sciences,
Bilecik Şeyh Edebali University, Bilecik, Turkey
[email protected]
Abstract. The incorporation of environmental criteria into the conventional

supplier selection practices is essential for organizations seeking to promote
green supply chain management. Challenges associated with green supplier
selection have been broadly recognized by procurement and supplier manage-
ment professionals. The development and implementation of practical decision
making tools that seek to address these challenges are rapidly evolving. This
article contributes to this knowledge area by comparing the application of
intuitionistic multi-criteria supplier selection methods in a fuzzy environment.
The intuitionistic fuzzy AHP and TOPSIS methods are thoroughly discussed.
The methods are then utilized to complete a green supplier evaluation and
selection study for an actual company from the paper mill. For this purpose, 3
alternative green suppliers have been identified. To evaluate the alternatives, 10
criteria in three dimensions including supplier performance, environmental
protection and supplier risk have been selected. Since the Multi-Criteria Deci-
sion Making (MCDM) methods are often used in problem of supplier selection
from past to present. The Intuitionistic Fuzzy Sets (IFS) are used to reflect
approval, rejection, and hesitation of decision makers by dealing with real life
uncertainty, imprecision, vagueness, and linguistic human decisions. The main
purpose of this study is to define the weights of criteria with Intuitionistic
Fuzzy AHP. After that, the Intuitionistic Fuzzy TOPSIS and AHP method were
used to rank alternatives from best to worst. At the end of the study the results
related with alternatives were compared and interpreted.
Keywords: Green supplier selection Intuitionistic fuzzy set theory AHP

TOPSIS
1 Introduction
In today’s conditions where competition has increasing day by day, the purchasing
function for companies is one of the business functions that will ensure profitability and
resistance of the company to competitive conditions. It is best sought while making the
Green Supplier Selection Using Intuitionistic Fuzzy AHP and TOPSIS 667
selection; however, the best mentioned cannot always be determined by quantitative

qualities, sometimes qualitatively best sought. This situation makes it difficult to
choose. The methods used by scientists in supplier selection are summarized [5]. Chen
[5] has separated the supplier selection models as singular models and multiple models.
Although it is possible to vary multiple models with different combinations, it is aimed
to give a general framework about the methods used. The technological developments,
globalization, innovation and like this items have resulted in market demands
becoming more dynamic. It is important to meet the demands simultaneously in order
to keep the businesses in a competitive environment. Rapid response to dynamic
market demands can only be achieved by integrating suppliers, manufacturers, dis-
tributors, service providers and customers into the supply chain. And the coordination
of the product, money and information flow on the supply chain need to be managed
successfully. A decision problem in which multiple qualitative and quantitative criteria
are effective in the decision process is important for enterprises. This decision problem,
which can be called the supplier selection problem as the most general expression, has
been examined by many researchers in the literature. And solutions have been proposed
with different methods. Environmental management has becoming increasingly
important by the organizational stakeholders, including governments, customers,
employees and competitors, pay more attention to environmental protection. Today, if
businesses want to survive in global markets, they should not ignore environmental
issues.
Green supply chain management has been defined as the integration of environ-
mental management and supply chain management. Green supply chain management
has also been expressed as a social and environmental strategy for the coordination of
important inter-organizational processes to improve the long-term economic perfor-
mance of the business. Terminology of the concept and many elements of green supply
chain management has changed over the years. A comprehensive list of this termi-
nology can be listed as follows:
– Sustainable supply network management,
– Corporate social responsibility networks,
– Environmental supply chain management,
– Green purchasing and supply,
– Environmental purchasing,
– Green logistics and environmental logistics,
– Sustainable supply chain.
Green supply chain management is a concept that has been developed with the
synthesis of two important strategies, especially environmental protection and effi-
ciency. While environmental protection has provided the basis for sustainable devel-
opment; efficiency has provided a framework for continuous improvement. In this
organizational application; green supplier selection has included many factors and
activities to expand the environmental life cycle. In the supply process, suppliers help
the implementation and development of the environmental management system within
the organizational structure of the product so that the product can reach the consumer.
668 E. Demir and G. Koca
Today’s competitive markets have focused on environmental issues by directing

businesses with critical factors such as cost, quality and service level to increase their
supply chain profits. In other words, Green Supply Chain Management takes ecological
issues and economy into consideration while traditional supply chain management only
takes into account the economic factor.
Supplier selection has also included activities such as the supplier’s definition,
analysis and selection to strengthen each layer of the supplier.
When we evaluate the green supply chains, the purchasing function has been seen
as the most important activity. In order to manage this function correctly, it is important
to reach the right suppliers. Otherwise, the image of the company may have been
damaged. Although the company itself does not have an environmentally harmful
attitude, the attitude of its suppliers also affects the image of the firm. For this reason,
the selection of suppliers for companies that want to take a green attitude is even more
important.
In this paper at first, 10 sub-criteria under 3 main criteria for green supplier
selection have been considered according to the literature review. Firstly, the criteria
were weighted according to the judgments of 3 decision makers via Intuitionistic
Fuzzy AHP. After that, the Intuitionistic Fuzzy TOPSIS and AHP method were used to
rank alternatives from the best to the worst. At the end of the study, the results related
with alternatives were compared and interpreted. In accordance with this purpose, in
the introduction part, the differences between supplier selection and green supplier
selection have been explained. In the second part, the criteria and application areas used
in the selection of green suppliers were discussed in previous studies. In this study,
genuine criteria have been discussed. Another genuine aspect of the study is its
application area in paper mills. In the third part, methodology has been mentioned. It
has been for the first time Intuitionistic Fuzzy AHP and Intuitionistic Fuzzy TOPSIS
methods have been used for green supplier selection. In the fourth part, the steps of the
analysis have been examined. In the last part, the results, the ordering of the alterna-
tives and the future studies have been mentioned. For this reason, the study aims to
contribute to the literature with its criteria and methodology.
2 Literature Review
In this study, the criteria and usage areas used in green supplier selection since 1997
have been examined. Searches have been made on the basis of studies, Web of Science,
Science Direct and Google Scholar. There are articles, book chapters and papers on this
subject. First, Noci [22] used pollution control, cost, green competence, green image of
the supplier and collaboration criteria for decreasing solid wastes, water pollution and
air emissions. Humpherys et al. [17] used pollution control, environmental manage-
ment, green competence, green image of the supplier, environmental design and
cooperation of management criteria to the decision support tool. Chiou et al. [6] used
cost, quality, environmental management, delivery, green competence, corporate social
responsibility and risk criteria to electronic industry with Fuzzy Analytic Hierarchy
Process method. Lee et al. [21] used pollution control, cost, quality, environmental
management, green competence, green image of the supplier, green product and
technological competence criteria for decreasing solid wastes, water pollution and air
emissions. Kuo and Tien [20] used cost, quality, environmental management, delivery,
service and corporate social responsibility criteria to find the most important criteria for
the green supplier selection with ANN and ANP methods. Kannan et al. [18] used
pollution control, cost, quality, environmental management, delivery, service and
technological competence criteria to automobile manufacturing company with
Fuzzy AHP and Fuzzy TOPSIS methods. Bali et al. [3] used pollution control, delivery,
service, green image of the supplier, green design, green product and reverse logistics
to automobile lighting system with Grey relational analysis. Yazdani [24] used pol-
lution control, cost, quality, environmental management, green competence, green
design, green product, environmental design and reverse logistics criteria to automobile
manufacturing company with AHP and Fuzzy TOPSIS methods. Dobos and Vör-
ösmarty [12] used pollution control, cost, quality, delivery and reuse criteria for
building the composite indicators and the DEA efficiency measuring DEA efficiency.
Hashemi et al. [15] used pollution control, cost, quality and collaboration criteria to
automotive industry with integrated ANP-improved GRA approach methods. Hu et al.
[16] used pollution control, quality, environmental management, delivery and services
criteria to low carbon economy with Linguistic 2-tuple method. Kuo et al. [19] used
pollution control, quality, environmental management, green design, cooperation of
management and risk criteria to electronics company with DANP, VIKOR methods.
Freeman and Chen [13] used cost, quality, environmental management, delivery and
green competence to electronic machinery manufacturer with AHP/Entropy-TOPSIS
methods. Denizhan et al. [10] used pollution control, quality, cost, delivery, green
competence, services and finance criteria to machinery manufacturing sector with
AHP/Fuzzy AHP methods. Gupta and Barua [14] used pollution control, quality, green
competence, green image of the supplier, finance, collaboration, green product and
technological competence criteria to decrease solid wastes, water pollution and air
emissions with BWM and Fuzzy TOPSIS methods. Yu et al. [25] used pollution
control criterion to carbon footprint based incentive supplier selection. Banaeian et al.
[4] used cost, quality, environmental management, services and collaboration criteria to
agricultural industry with Fuzzy TOPSIS, Fuzzy VIKOR and Fuzzy GRA methods.
Daldır and Tosun [8] used green storage, green recycling, green production capacity,
green packaging, consuming resources, pollution control, logistics costs, production
cost, delivery time, error rate in production, product warrant and environmental
competences criteria to manufacturing industry with Fuzzy AHP and Fuzzy WASPAS
methods. Çelik and Ustasüleyman [7] used environmental competences, quality, ser-
vices and cost criteria to a fitted kitchen producer company with Fuzzy AHP and
Fuzzy TOPSIS methods. Deshmukh and Vasudevan [11] used cost, quality, delivery
and risk criteria to plastic products manufacturing with AHP/Fuzzy AHP methods.
Akcan and Taş [2] used recyclable packaging, sharing information about GSCM,
environment management system, waste management, greenhouse gases emissions
control in manufacturing, green image, collaborate with costumers for green supply
chain, green transportation, efficient resource utilization, senior management’s support
and commitment, hazardous and toxic substances using in production to international
yachting company operating with SWARA and TOPSIS methods. Đalić et al. [9] used
environmental image, recycling, pollution control, environmental management system,
environmentally friendly products, resource consumption and green competencies

criteria with integrated fuzzy PIPRECIA – interval rough SAW methods.
3 Methodology
3.1 Intuitionistic Analytical Hierarchy Process

AHP Method was created by Saaty [23] and afterward was accepted in the decision
making process. In intuitionistic fuzzy AHP the pairwise comparison matrix is used.
Intuitive fuzzy numbers are used to determine weights. Weight determination is done
with simple arithmetic operations of triangular intuitionistic fuzzy numbers. To develop
IF-AHP an example for green supplier selection problem is provided.
3.2 Intuitionistic Fuzzy TOPSIS

The TOPSIS was first introduced in 1981 by Hwang and Yoon. TOPSIS method is the
most common method of sorting alternatives according to selection criteria. TOPSIS
method sorts alternatives from best to worst. Fuzzy numbers, unlike other methods, are
used in real life uncertainty, inaccuracy and uncertainty. Individual decisions are
represented by linguistic expressions. Intuitionistic fuzzy sets better express people’s
approval, rejection, and indecision on the situations.
4 Analysis
In this study 10 criteria in three dimensions related to green supplier selection were
considered for measuring the 3 alternative green supplier in Paper Mills firms. The
criteria are evaluated by 3 experts that have different expert field. The work has
practiced in a real paper mill. Paper factory located in Istanbul. The paper mill has two
different warehouses. This study has been carried out for the green supplier selection
project during the transition to the green supply chain. In the study, a group of 3 experts
has been involved as decision makers. “Project manager”, one of the experts, has been
involved in various projects and has 20 years of experience in business practices.
“Supply chain manager” in the expert group has 15 years of experience. Supply chain
manager has previously taken an active role in the supply chain process of different
businesses and green supply chain processes. “Factory Manager” in the expert group
has 25 years of work experience and previously worked as manager in many factories.
Before the meeting, the foresights were taken by the “supply chain manager” based on
their past experiences and a literature study was made by making use of the previous
studies. A comprehensive literature study and the criteria used are given in the previous
sections. In the first meeting, the list of criteria was examined with the project team, the
number of criteria was reduced and some new criteria were added. In the second
meeting, data on the criteria were entered by experts. After evaluating the entered data,
final values were agreed on in the last meeting. Accordingly, the criteria are as in
Table 1. The meetings lasted an average of 75 min and were completed on a 4-week
calendar. In this study, 3 main criteria affecting the work flow and 10 sub-criteria were
discussed in order to carry out the procurement processes in the paper mill quickly and
without interruption. All criteria are maximizing. The criteria are passed at the paper
mills firm, unlike other literature studies. For this reason, the specific criteria in
regarding to this business were originally determined. Paper mills factory has 3
alternative suppliers. First, the criteria were evaluated within themselves according to
the Abdullah and Najib [1] article. The criteria weights were then evaluated by the IF-
AHP method by each decision maker. In the last case, the local and global weights of
the criteria can be seen in Table 1. After evaluating the criteria objectively, the alter-
natives were evaluated objectively. Firstly, IF-AHP method was used to evaluate
alternatives. Intuitionistic Fuzzy Sets were not used in green supplier selection in
previous studies. At this point, it is aimed to add originality to the study. The Intu-
itionistic fuzzy set is an expanded version of the classical Fuzzy Set Theory (FST) to
solve the uncertainty issue. AHP has been applied for parsing and binary comparisons
for various levels of the hierarchy. The aim is to set priority criteria to rank the best
green supplier. Then IF-TOPSIS method was made in Table 2. Then, TOPSIS method
was applied to sort the alternatives. The aim is to make a comparison by selecting two
methods. A numerical example illustrates our method in Table 2. The reason for
choosing Intuitionistic in two methods is because of the homogeneity.
Table 1. The matrix of weights for main criteria and sub-criteria with Intuitionistic fuzzy AHP
Main criteria Local Sub-criteria Local Global Ranking
Supplier 0.374 Product quality 0.342 0.128 3
performance Green product 0.337 0.126 4
Elasticity 0.252 0.094 5
Environmental 0.456 Environmental 0.199 0.041 8
protection performance
Innovation 0.398 0.181 1
capability
Green logistics 0.362 0.165 2
Supplier risk 0.178 Labor intensity 0.122 0.021 10
Financial stability 0.281 0.050 7
Reputation of the 0.354 0.063 6
supplier
Information 0.195 0.034 9
security
Table 2. Ranking of green supplier in regarding of IF-AHP and IF-TOPSIS methods

Alternatives Ranking for IF-AHP Ranking for IF-TOPSIS
A1 1 1
A2 2 3
A3 3 2
5 Conclusion and Further Studies
In this study, a specific sector selection was made by selecting suppliers of paper mills.
In the literature, green supplier selection was made for the first time with the intu-
itionistic fuzzy set. In this way, a difference is presented in terms of both subject and
application technique. According to intuitionistic Fuzzy AHP the most important cri-
teria is innovation capability. And the other criteria can be listed as follows green
logistics, product quality, green product, elasticity, reputation of the supplier, financial
stability, environmental performance, information security and labor intensity respec-
tively. After that there is a bit difference of ranking with intuitionistic Fuzzy AHP and
intuitionistic Fuzzy TOPSIS. For future studies, the areas of application might be
extended to apply other private sectors rather than paper mills. Also, other integrated
weighting and sorting techniques like intuitionistic fuzzy VIKOR and COPRAS
methods could be used. Additionally, the success of firms could be measured from
other extended criteria.
References
1. Abdullah, L., Najib, L.: A new type-2 fuzzy set of linguistic variables for the fuzzy analytic
hierarchy process. Expert Syst. Appl. 41, 3297–3305 (2014)
2. Akcan, S., Taş, M.A.: Green supplier evaluation with SWARA-TOPSIS integrated method
to reduce ecological risk factors. Environ. Monit. Assess. 191, 1–22 (2019)
3. Bali, O., Kose, E., Gumus, S.: Green supplier selection based on IFS and GRA. Grey Syst.:
Theory Appl. 3, 158–176 (2013)
4. Banaeian, N., Mobli, H., Fahimnia, B., Nielsen, I.E., Omid, M.: Green supplier selection
using fuzzy group decision making methods: a case study from the agri-food industry.
Comput. Oper. Res. 89, 337–347 (2018)
5. Chen, Y.J.: Structured methodology for supplier selection and evaluation in a supply chain.
Inf. Sci. 181, 1651–1670 (2011)
6. Chiou, C.Y., Hsu, C.W., Hwang, W.Y.: Comparative investigation on green supplier
selection of the American, Japanese and Taiwanese electronics industry in China. In:
Industrial Engineering and Engineering Management, pp. 1909–1914 (2008)
7. Çelik, P., Ustasüleyman, T.: Bulanik Çok Kriterli Karar Verme Yöntemleri ile Yeşil
Tedarikçilerin Değerlendirilmesi. Akademik Sosyal Araştırmalar Dergisi, pp. 375–390
(2018)
8. Daldır, I., Tosun, Ö.: Bulanık WASPAS ile Yeşil Tedarikçi Seçimi. Uludağ Üniversitesi
Mühendislik Fakültesi Dergisi, pp. 193–207 (2018)
9. Đalić, I., Stević, Ž., Karamasa, C., Puška, A.: A novel integrated fuzzy PIPRECIA–interval
rough saw model: green supplier selection. Decis. Making: Appl. Manag. Eng. 3, 126–145
(2020)
10. Denizhan, B., Yalçıner, A.Y., Berber, Ş.: Analitik hiyerarşi proses ve bulanık analitik
hiyerarşi proses yöntemleri kullanılarak yeşil tedarikçi seçimi uygulaması. Nevşehir Bilim ve
Teknoloji Dergisi 6, 63–78 (2017)
11. Deshmukh, A.J., Vasudevan, H.: Supplier selection in plastic products manufacturing
MSMEs using a combined traditional and green criteria based on AHP and fuzzy AHP. In:
Proceedings of International Conference on Intelligent Manufacturing and Automation,
pp. 593–600. Springer, Singapore (2019)
12. Dobos, I., Vörösmarty, G.: Green supplier selection and evaluation using DEA-type
composite indicators. Int. J. Prod. Econ. 157, 273–278 (2014)
13. Freeman, J., Chen, T.: Green supplier selection using an AHP-Entropy-TOPSIS framework.
Supply Chain Manag.: Int. J. 20, 327–340 (2015)
14. Gupta, H., Barua, M.K.: Supplier selection among SMEs on the basis of their green
innovation ability using BWM and fuzzy TOPSIS. J. Cleaner Prod. 152, 242–258 (2017)
15. Hashemi, S.H., Karimi, A., Tavana, M.: An integrated green supplier selection approach
with analytic network process and improved grey relational analysis. Int. J. Prod. Econ. 159,
178–191 (2015)
16. Hu, Z., Rao, C., Zheng, Y., Huang, D.: Optimization decision of supplier selection in green
procurement under the mode of low carbon economy. Int. J. Comput. Intell. Syst. 8, 407–
421 (2015)
17. Humphreys, P.K., Wong, Y.K., Chan, F.T.S.: Integrating environmental criteria into the
supplier selection process. J. Mater. Process. Technol. 138, 349–356 (2003)
18. Kannan, D., Khodaverdi, R., Olfat, L., Jafarian, A., Diabat, A.: Integrated fuzzy multi
criteria decision making method and multi-objective programming approach for supplier
selection and order allocation in a green supply chain. J. Cleaner Prod. 47, 355–367 (2013)
19. Kuo, T.C., Hsu, C.W., Li, J.Y.: Developing a green supplier selection model by using the
DANP with VIKOR. Sustainability 7, 1661–1689 (2015)
20. Kuo, R.J., Wang, Y.C., Tien, F.C.: Integration of artificial neural network and MADA
methods for green supplier selection. J. Cleaner Prod. 18, 1161–1170 (2010)
21. Lee, A.H., Kang, H.Y., Hsu, C.F., Hung, H.C.: A green supplier selection model for high-
tech industry. Expert Syst. Appl. 36, 7917–7927 (2009)
22. Noci, G.: Designing ‘green’ vendor rating systems for the assessment of a supplier’s
environmental performance. Eur. J. Purch. Supply Manag. 3, 103–114 (1997)
24. Yazdani, M.: An integrated MCDM approach to green supplier selection. Int. J. Ind. Eng.
Comput. 5(3), 443–458 (2014)
25. Yu, F., Yang, Y., Chang, D.: Carbon footprint based green supplier selection under dynamic
environment. J. Cleaner Prod. 170, 880–889 (2018)
InterCriteria Analysis Applied to Emerging
Europe and Central Asia University Rankings
Veselina Bureva(&)
Laboratory of Intelligent Systems, Prof. Dr. Assen Zlatarov University,

Prof. Yakimov Blvd., 8010 Burgas, Bulgaria
[email protected]
Abstract. InterCriteria analysis (ICA) is a new tool for decision making. This
method can help to make decision according to the similarities or dissimilarities
between the criteria. In the current research work an investigation of the
Emerging Europe and Central Asia (EECA) university rankings is discussed.
The aim of the ICA application over the datasets from Emerging Europe and
Central Asia (EECA) university rankings is to identify the behavior of the
universities in different countries. The datasets are downloaded by QS World
University Rankings website. The aim of ICA application is to discover possible
dependencies, correlations or opposite behavior between the universities in
different countries. The indicators’ correctness is estimated by second ICA
application. The resulting sets contain a segmentation of the universities
activities according to the selected indicators. The investigations are made using
ICrAData software.
Keywords: InterCriteria analysis Intuitionistic fuzzy pairs University

rankings
1 Introduction
The proposed procedure investigates the relationships between the universities and
indicators. The ICA applications determine the universities with similar, independent or
opposite behavior or work progress. The dependencies between the indicators will be
determined. The step of indicator’s correlation identifying presents the correctness of
the selected criteria.
1.1 InterCriteria Analysis (ICA)

ICA [5, 6] is a decision making method based on the theory of intuitionistic fuzzy sets
[4, 7] and index matrices [3]. The input for ICA is two-dimensional index matrix repre-
senting the set of universities on the rows and the set of the indicators on the columns.
Depending on the desired investigation the datasets can be transposed. In the next time
moment two counters are generated: counter for similar relations between the values and
counter for different relations between the values. These counters are used for calculating
the degree of memberships and the degree of non-membership. Different algorithms
for calculation of degrees of correspondence and degrees of non-correspondence are
InterCriteria Analysis Applied to Emerging Europe 675
proposed according to the uncertainty representation [19]. The resulting index matrix
contains the degrees of correspondence and non-correspondence between the estimated
objects represented as intuitionistic fuzzy pairs ha; bi where a þ b 1 and a; b 2 ½0; 1.
The ICA method is extended in series of papers [1, 2, 9, 11, 24]. Many develop-
ments of the ICA analysis in the areas of university rankings [10, 12, 14, 16, 22, 23],
economics [8], neural networks [21] and genetic algorithms [17, 20] are presented.
1.2 QS Emerging Europe and Central Asia University Rankings 2019

In the current research work an investigation of QS Emerging Europe and Central Asia
University Rankings 2019 is presented [18]. The rankings include 350 universities
from Emerging Europe and Central Asia. The methodology of Emerging Europe and
Central Asia (EECA) university rankings is presented in [15]. The universities are
estimated by 10 indicators:
1) Academic reputation;
2) Employer reputation;
3) Faculty student;
4) International faculty;
5) International students;
6) Citations per paper;
7) Papers per faculty;
8) Faculty staff with PhD;
9) Web impact;
10) International research network.
The indicators “academic reputation” and “employer reputation” are received by
the surveys. The “faculty student” ratio is based on the number of students enrolled per
full-time academic staff member. The indicator “papers per faculty” reflects research
productivity, based on papers published per academic faculty members using data from
Elsevier’s Scopus. The indicator “international research work” assesses the degree of
international openness in terms of research collaboration for each evaluated university.
The “web impact” ratio estimates the institution online presence. The “staff with PhD”
is calculated as proportion of faculty members with PhD degree. The “citations per
paper” ratio is based on Scopus database. The last two indicators “international faculty”
and “international students” are based on the proportion of faculty members and stu-
dents who are international [18].
The top 50 universities are selected from all the 350 universities. The dataset is
extracted from the website of QS World University Rankings. The input dataset contain
the following universities:
1) Lomonosov Moscow State University (RUSSIA)
2) Novosibirsk State University (RUSSIA)
3) Charles University (CZECH REPUBLIC)
4) Saint Petersburg State University (RUSSIA)
5) University of Tartu (ESTONIA)
6) University of Warsaw (POLAND)
676 V. Bureva
7) Jagiellonian University (POLAND)

8) Middle East Technical University (TURKEY)
9) Czech Technical University in Prague (CZECH REPUBLIC)
10) Bogaziçi Üniversitesi (TURKEY)
11) Masaryk University (CZECH REPUBLIC)
12) Koç University (TURKEY)
13) Tomsk State University (RUSSIA)
14) Bilkent University (TURKEY)
15) Warsaw University of Technology (POLAND)
16) Moscow Institute of Physics and Technology (MIPT/Moscow Phystech) (RUSSIA)
17) Vilnius University (LITHUANIA)
18) Sabanci University (TURKEY)
19) Al-Farabi Kazakh National University (KAZAKHSTAN)
20) Eötvös Loránd University (HUNGARY)
21) Istanbul Technical University (TURKEY)
22) Tallinn University of Technology (TalTech) (ESTONIA)
23) Belarusian State University (BELARUS)
24) National Research University Higher School of Economics (HSE University)
(RUSSIA)
25) Brno University of Technology (CZECH REPUBLIC)
26) National Research Nuclear University MEPhI (Moscow Engineering Physics
Institute) (RUSSIA)
27) University of Szeged (HUNGARY)
28) Istanbul University (TURKEY)
29) Budapest University of Technology and Economics (HUNGARY)
30) University of Debrecen (UKRAINE)
31) Taras Shevchenko National University of Kyiv (HUNGARY)
32) University of Ljubljana (SLOVENIA)
33) Hacettepe University (TURKEY)
34) Babes-Bolyai University (ROMANIA)
35) National Research Tomsk Polytechnic University (RUSSIA)
36) Bauman Moscow State Technical University (RUSSIA)
37) University of Bucharest (ROMANIA)
38) University of Zagreb (CROATIA)
39) Vilnius Gediminas Technical University (LITHUANIA)
40) L.N. Gumilyov Eurasian National University (ENU) (KAZAKHSTAN)
41) Peter the Great St. Petersburg Polytechnic University (RUSSIA)
42) Kazan (Volga region) Federal University (RUSSIA)
43) Ural Federal University – UrFU (RUSSIA)
44) Wroclaw University of Science and Technology (WRUST) (POLAND)
45) Comenius University in Bratislava (SLOVAKIA)
46) Palacký University in Olomouc (CZECH REPUBLIC)
47) Sofia University “St. Kliment Ohridski” (BULGARIA)
48) Corvinus University of Budapest (HUNGARY)
49) University of Wroclaw (POLAND)
50) Kaunas University of Technology (LITHUANIA)
The paper has the following structure: An introduction is presented in Sect. 1.

Section 1.1 and Sect. 1.2 introduce the InterCriteria analysis basics and Emerging
Europe and Central Asia (EECA) universities rankings 2019. The ICA applications are
discussed in Sect. 2. In Sect. 2.1 an ICA application for indicators identifying is pre-
sented. In Sect. 2.2 a comparison of universities similarities is made.
2 InterCriteria Analysis Applied to Emerging Europe

and Central Asia University Rankings
In the current section two investigations are presented: ICA application determining the
dependencies between the indicators and ICA application identifying the behavior of
the universities. The ICrAData software is used for the developments [13].
2.1 ICA Application for Indicators Behavior Identification

The first application of ICA over Emerging Europe and Central Asia University
Rankings is developed according to the necessity to determine the relationships
between the indicators. The procedure found 45 pairs of indicators in dissonance (weak
dissonance, dissonance, strong dissonance). Received results determine the correctness
of the selected indicators. The selected criteria are independent (Table 1).
Table 1. Relationships between the indicators

Correlation Type of consonance Number of pairs
[0.25; 0.33) Weak Dissonance (WD) 1
[0.33; 0.43) Dissonance (D) 3
[0.43; 0.57) Strong Dissonance (SD) 26
[0.57; 0.67) Dissonance (D) 11
2.2 ICA Application for University Relationships Identification

The aim of the second application of the ICA over Emerging Europe and Central Asia
University Rankings is to identify the possible relationships between different insti-
tutions. The results of ICA are the following: 27 pairs of universities in weak negative
consonance, 986 pairs of universities in dissonance, 139 pairs of universities in weak
positive consonance and 15 pairs of universities in positive consonance (Table 2).
The pairs of universities in positive consonance are presented in the Table 3. These
institutions have close relationships according to their functioning and working pro-
cess. The dependencies between the universities in different countries represent similar
level of the education process.
678 V. Bureva
Table 2. Relationships between the universities

Correlation Type of consonance Number of pairs
[0.15; 0.25) Weak negative consonance 27
[0.33; 0.43) Dissonance (D) 171
[0.43; 0.57) Strong Dissonance (SD) 327
[0.57; 0.67) Dissonance (D) 259
[0.75; 0.85) Weak Positive Consonance (WPC) 139
[0.85; 0.95) Positive Consonance (PC) 15
Table 3. Pairs of universities in positive consonance

Positive consonance
hl; mi University 1 University 2
h0:93; 0:04i Eötvös Loránd University University of Wroclaw (POLAND)
(HUNGARY)
h0:91; 0:07i University of Warsaw (POLAND) Jagiellonian University (POLAND)
h0:91; 0:09i Koç University (TURKEY) Sabanci University (TURKEY)
h0:91; 0:07i University of Tartu (ESTONIA) University of Ljubljana (SLOVENIA)
h0:89; 0:07i Warsaw University of Budapest University of Technology and
Technology (POLAND) Economics (HUNGARY)
h0:89; 0:11i Hacettepe University (TURKEY) Babes-Bolyai University (ROMANIA)
h0:89; 0:11i Czech Technical University in Budapest University of Technology and
Prague (CZECH REPUBLIC) Economics (HUNGARY)
h0:87; 0:09i Czech Technical University in Warsaw University of Technology
Prague (CZECH REPUBLIC) (POLAND)
h0:87; 0:13i Middle East Technical University Czech Technical University in Prague
(TURKEY) (CZECH REPUBLIC)
h0:87; 0:13i Tallinn University of Technology Budapest University of Technology and
(TalTech) (ESTONIA) Economics (HUNGARY)
h0:87; 0:13i Babes-Bolyai University Wroclaw University of Science and
(ROMANIA) Technology (WRUST) (POLAND)
h0:87; 0:13i Comenius University in Sofia University “St. Kliment Ohridski”
Bratislava (SLOVAKIA) (BULGARIA)
h0:87; 0:13i Vilnius Gediminas Technical Corvinus University of Budapest
University (LITHUANIA) (HUNGARY)
h0:87; 0:13i Comenius University in University of Wroclaw (POLAND)
Bratislava (SLOVAKIA)
h0:87; 0:13i Sofia University “St. Kliment University of Wroclaw (POLAND)
Ohridski” (BULGARIA)
Results from the ICA applications from Sect. 2.1 and Sect. 2.2 are presented in the
intuitionistic fuzzy triangle on the Fig. 1. The pairs of universities in positive conso-
nance are situated close to right bottom angle of the triangle (green points). The pairs of
universities with opposite behavior are visualized in the upper angle of the triangle
(blue points).
Fig. 1. Results in intuitionistic fuzzy triangle: ICA application by indicators (left) and ICA
application by universities (right).
The ICA is applied over the datasets of Emerging Europe and Central Asia
University Rankings to determine possible tendencies, dependencies or relationships
between the selected indicators and universities. The ICA application for indicators
correlations identification gives us intuitionistic fuzzy pairs in dissonance as a result.
Obviously, the indicators are independent. Their selection is correct and they are not
dependent. The second application of the ICA over the datasets of Emerging Europe
and Central Asia University Rankings has the aim to investigate the relationships
between the universities. The most pairs of universities are in dissonance. Small
selection of universities has similar behavior. A little part of universities have deter-
mined opposite tendencies.
3 Conclusion
The ICA applications over the datasets of Emerging Europe and Central Asia
University Rankings are investigated. The indicators independency is confirmed. The
universities with close relationships are determined. In the end of the investigation a
little part of universities with opposite behavior are found. The bigger part of the
universities is independent. In the next research work a comparison between the uni-
versities rankings of different years will be investigated. The procedure will present the
correlations between the universities in the time.
by DCM # 577/17.08.2018.
680 V. Bureva
References
1. Atanassov, K., Marinov, P., Atanassova, V.: InterCriteria Analysis with interval-valued
intuitionistic fuzzy evaluations. In: Cuzzocrea, A., Greco, S., Larsen, H., Saccà, D.,
Andreasen, T., Christiansen, H. (eds.) Flexible Query Answering Systems 2019. Lecture
Notes in Computer Science, vol. 11529, pp. 329–338. Springer, Cham (2019)
2. Atanassov, K., Ribagin, S., Sotirova, E., Bureva, V., Atanassova, V., Angelova, N.:
Intercriteria Analysis using special type of intuitionistic fuzzy implications. Notes Intuit.
Fuzzy Sets 23(5), 61–65 (2017)
3. Atanassov, K.: Index matrices: towards an augmented matrix calculus. Studies in
Computational Intelligence, vol. 573. Springer, Cham (2014)
5. Atanassov, K., Atanassova, V., Gluhchev, G.: InterCriteria analysis: ideas and problems.
Notes Intuit. Fuzzy Sets 21(1), 81–88 (2015)
6. Atanassov, K., Mavrov, D., Atanassova, V.: InterCriteria decision making: a new approach
for multicriteria decision making, based on index matrices and intuitionistic fuzzy sets.
Issues Intuit. Fuzzy Sets Gen. Nets 11, 1–8 (2014)
7. Atanassov, K., Szmidt, E., Kacprzyk, J.: On intuitionistic fuzzy pairs. Notes Intuit. Fuzzy
Sets 19(3), 1–13 (2013)
8. Atanassova, V., Doukovska, L., Krawczak, M.: Intercriteria Analysis of countries in
transition from factor-driven to efficiency-driven economy. Notes Intuit. Fuzzy Sets 24(2),
84–96 (2018)
9. Atanassova, V., Doukovska, L., Michalíková, A., Radeva, I.: Intercriteria Analysis: from
pairs to triples. Notes Intuit. Fuzzy Sets 22(5), 98–110 (2016)
10. Bureva, V., Michalíková, A., Sotirova, E., Popov, S., Riečan, B., Roeva, O.: Application of
the InterCriteria Analysis to the universities rankings system in the Slovak Republic. Notes
Intuit. Fuzzy Sets 23(2), 128–140 (2017)
11. Bureva, V., Sotirova, E., Atanassova, V., Angelova, N., Atanassov, K.: Intercriteria Analysis
over intuitionistic fuzzy data. In: Lirkov, I., Margenov, S. (eds.) Large-Scale Scientific
Computing 2017. Lecture Notes in Computer Science, vol. 10665. Springer, Cham (2018)
12. Bureva, V., Sotirova, E., Sotirov, S., Mavrov, D.: Application of the InterCriteria decision
making method to Bulgarian universities ranking. Notes Intuit. Fuzzy Sets 21(2), 111–117
(2015)
13. Ikonomov, N., Vassilev, P., Roeva, O.: ICrAData – software for InterCriteria Analysis. Int.
J. Bioautom. 22(1), 1–10 (2018)
14. Krawczak, M., Bureva, V., Sotirova, E., Szmidt, E.: Application of the InterCriteria decision
making method to universities ranking. In: Atanassov, K.T., Castillo, O., Kacprzyk, J.,
Krawczak, M., Melin, P., Sotirov, S., Sotirova, E., Szmidt, E., Tré, G.D., Zadrożny, S. (eds.)
Novel Developments in Uncertainty Representation and Processing. AISC, vol. 401,
15. Methodology of QS World University Rankings. https://www.topuniversities.com/eeca-
rankings/methodology. Accessed 18 Mar 2020
16. Parvathi, R., Atanassova, V., Doukovska, L., Yuvapriya, C., Indhurekha, K.: InterCriteria
Analysis of rankings of Indian universities. Notes Intuit. Fuzzy Sets 24(1), 99–109 (2018)
17. Pencheva, T., Angelova, M., Atanassova, V., Roeva, O.: InterCriteria Analysis of genetic
algorithm parameters in parameter identification. Notes Intuit. Fuzzy Sets 21(2), 99–110
(2015)
18. QS World university rankings-Emerging Europe and Central Asia (EECA) university
rankings. https://www.topuniversities.com/university-rankings/eeca-rankings/2019. Acces-
sed 19 Mar 2020
19. Roeva, O., Vassilev, P., Ikonomov, N., Angelova, M., Su, J., Pencheva, T.: On different
algorithms for InterCriteria relations calculation. In: Hadjiski, M., Atanassov, K.T. (eds.)
Intuitionistic Fuzziness and Other Intelligent Theories and Their Applications. Studies in
Computational Intelligence, vol. 757, pp. 143–160. Springer, Cham (2019)
20. Roeva, O., Vassilev, P., Fidanova, S., Paprzycki, M.: InterCriteria Analysis of genetic
algorithms performance. In: Fidanova, S. (ed.) Recent Advances in Computational
Optimization. Studies in Computational Intelligence, vol. 655, pp. 235–260. Springer,
Cham (2016)
21. Sotirov, S., Atanassova, V., Sotirova, E., Doukovska, L., Bureva, V., Mavrov, D., Tomov,
J.: Application of the intuitionistic fuzzy InterCriteria Analysis method with triples to a
neural network preprocessing procedure. Comput. Intell. Neurosci. 2017, 9 (2017). https://
doi.org/10.1155/2017/2157852. Article ID 2157852
22. Sotirova, E., Bureva, V., Chountas, P., Krawczak, M.: An application of intercriteria
decision making method to the rankings of universities in the United Kingdom. Notes Intuit.
Fuzzy Sets 22(3), 112–119 (2016)
23. Sotirova, E., Shannon, A.: Application of InterCriteria Analysis to the rankings of Australian
universities. Notes Intuit. Fuzzy Sets 21(4), 136–142 (2015)
24. Traneva, V., Tranev, S., Szmidt, E., Atanassov, K.: Three dimensional InterCriteria Analysis
over intuitionistic fuzzy data. In: Kacprzyk, J., Szmidt, E., Zadrożny, S., Atanassov, K.,
Krawczak, M. (eds.) Advances in Fuzzy Logic and Technology 2017. Advances in
Intelligent Systems and Computing, vol. 641, pp. 442–449. Springer, Cham (2018)
Determining Significant Factors Affecting
Vaccine Demand and Factor Relationships
Using Fuzzy DEMATEL Method
İkbal Ece Dizbay(&) and Ömer Öztürkoğlu
Yasar University, Bornova, İzmir, Turkey

[email protected]
Abstract. Even though deadly effects of outbreaks such as SARS, H1N1,

EBOLA and COVID-19 took the attention of the community, generating 100%
vaccination uptake from people who are expected to be affected by such out-
breaks is almost impossible. Hence, determining the actual vaccine demand for
typical viruses and reaching this population are becoming important issues to
prevent from spreading such viruses like wildfire. Many countries have been
forecasting vaccine demand according to birth rates. However, factors such as
district-level per capita income, literacy rate, urbanization, and some other
factors should also be considered for more accurate forecasts. There is another
factor affecting countries’ vaccine demand; some of the vaccines that WHO
recommends are included in the national immunization programs, while others
are only recommended. Thus, this study first aims to explore the most significant
factors that affect the demand of vaccines that are not included in national
immunization campaigns then to present the cause and effect relationships
among the factors using the fuzzy DEMATEL method to provide insights to
managers for better vaccine demand forecast and to increase vaccine uptake.
According to the analysis results, immunization related beliefs is the most
important factor among others included in this research.
Keywords: Vaccine demand Factor relationship Fuzzy DEMATEL
1 Introduction
According to WHO [1], vaccines protect 2.5 million lives from deadly diseases every
year. Not only children but also pregnant women and elderly people are vaccinated
regularly against some diseases. In addition to these, people from every age can be
vaccinated against some outbreaks of vaccine-preventable diseases such as influenza
pandemic. Millions of people are vaccinated every year but also WHO [2] estimates
that nearly 19.4 million babies, under the age one, could not receive scheduled basic
vaccines. By efficiently vaccinating the target population, eliminating deadly diseases is
possible. As an example of the results of deadly outbreaks, in the 18th century, 400,000
people were dying annually in Europe from smallpox disease and thousands of people
went blind. Even if the development of the smallpox vaccine began about 1774, in
1967, the illness still was not under control. In 1967, a global campaign began and in
1980 WHO announced that the world was free of this disease [3]. Another example of a
Determining Significant Factors Affecting Vaccine Demand 683
deadly disease outbreak of this age is the novel Coronavirus (COVID-19). By the
beginning of 2020, COVID-19 escalated into a real epidemic. In just 3.5 months, more
than 1.3 million people around the world were infected and more than 126,500 people
died from the outbreak [4]. In addition to health-related risks, OECD has announced
that the world economy is at risk [5]. To decrease the expected impact of this crisis,
scientists are working on the COVID-19 vaccine, but unfortunately releasing a new
vaccine to the market takes a long time. Even if the vaccine is released to the market,
estimating the demand for this vaccine and meeting this huge demand in a short period
is difficult.
Children are generally vaccinated against some of the childhood diseases through
national immunization programs, conducted by governments or non-governmental
organizations. However, some vaccines are not included in national immunization
programs even if they are suggested by WHO or other health organizations. As an
example of Turkey, most of the vaccines recommended by WHO are included in the
National Immunization Program, except Rotavirus, Influenza and Meningococcal
conjugate. The main contribution of this study is presenting the macro-level insights
about the effects of several factors on demand of such vaccines and their relations. The
structure of the remainder of this paper is as follows: In Sect. 2, the literature review of
the vaccine demand factors is given. In Sect. 3 and 4 Fuzzy-DEMATEL Method and
its application is presented. Finally, managerial insights, limitations and concluding
remarks are presented in Sect. 5.
2 Literature Review
There are several classifications made by researchers to classify determinants of vac-

cine uptake. Ghosh et al. [6] classify immunization coverage factors in two categories
according to their literature review, supply-side and demand-side. According to their
research demand-side factors covers birth order, sex, maternal education, and
employment status, occupation type, immunization related beliefs, mother’s health
knowledge, health-seeking behavior, household wealth index, religion, and caste.
Phillips et al. [7] categorized vaccine coverage determinants for low and middle-
income countries; facility readiness, decision to vaccinate and mother able.
For this study, six factors that affect vaccine demand were determined; district-level
per capita income, literacy rate of the district, immunization-related-beliefs, healthcare
workers’ knowledge, urbanization, and special immunization campaigns or strategies.
Rammohan et al. [8] concluded that district-level per capita income is an important
factor to have better immunization outcomes in India. Another study showed that
children living in urban areas in India and the lowest wealth quartile have lower
coverage [9]. For the effect of illiteracy rate, Kibreab et al. [10] conclude that mothers’
education level has a positive effect on immunization outcomes. Contrary to this result,
some researchers concluded that mothers’ education level does not affect immunization
coverage [11]. Arokiasamy et al. [12] concluded that having an illiterate mother or
father affects immunization coverage. In Turkey Akıs et al. [13] concluded that parents
with a higher educational level are more positive about vaccination. Immunization-
related beliefs or disbeliefs are another reason that decreases immunization coverage.
684 İ. E. Dizbay and Ö. Öztürkoğlu
These beliefs can be due to the adverse effects of vaccination, recommendations of

healthcare professionals, knowledge, social influences, information about the vaccine,
etc. For the influenza vaccination in Turkey, Ciblak et al. [14] surveyed the pharmacist
and occupational physicians and they indicate that disbelief in the effectiveness of the
influenza vaccine is effecting vaccine uptake negatively. Sissoko et al. [15] indicated
that the place of residence (living in urban or rural areas) is significantly associated with
non-vaccination. There are some contrary opinions about this relation; some
researchers concluded that children living in urban residences are vaccinated less than
children living in rural areas [16]. Health workers’ knowledge about vaccines is another
factor for vaccine uptake. If there is not enough reliable data about a new vaccine,
health workers hesitate about vaccine uptake. When the Rotavirus vaccine was first
released, healthcare workers were reluctant to recommend it due to its side effects.
After several papers published about this vaccine, some of the health workers begun to
recommend it. Furthermore, several studies indicated that vaccination campaigns affect
vaccine demand. Shea et al. [17] concluded that mass media campaigns can be effective
but this effect is due to access to media.
This study specifically focuses on vaccines that are suggested by WHO but are not
mandatorily applied to children. Thus, the parents trigger the demand for these types of
vaccines. Because of this characteristic, forecasting demand for these types of vaccines
is erratic. Therefore, the healthcare managers or healthcare product distributors face a
critical decision about how much vaccines should be ordered to avoid from stock out
and minimizing wastes. Hence, this study aims to present macro-level insights into the
effects of several factors on vaccine demand.
As mentioned in the previous section, six factors were determined from the liter-
ature those are expected to affect vaccine demand: Immunization-related beliefs (F1),
District-level per capita income (F2), Urbanization (F3), Healthcare workers’ knowl-
edge about the vaccines (F5), and Special immunization campaigns or strategies (F6).
As mentioned in the literature that each factor may have a different impact on demand.
Moreover, they may also affect each other. Therefore, in this study, (1) the significance
of each factor’s influence on the demand and (2) the cause and effect relationship
among the factors was identified. Hence, one of the most commonly used MCDM
(Multi-Criteria Decision-Making) methods called DEMATEL (Decision Making Trial
and Evaluation Laboratory) was used in this study because of it both ranks the factors
according to their importance and causal relationships between them. Due to the
vagueness of human thoughts and language in decision-making, several studies pre-
sented the strength of using fuzzy variables in the assessments in the DEMATEL
approach; Wu and Lee [18]; Lin and Wu [19]; Feng and Ma [20]. In this study, the
Fuzzy-DEMATEL approach was implemented, which was first proposed by Wu and
Lee [18] and Lin and Wu [19] and later corrected by Mokhtarian [21]. The basic steps
of the implemented Fuzzy-DEMATEL approach are as follows:
Step 1: Determine the influencing factors. In this study, the factors are F1, F2, …,
F6.
Step 2: Design a semantic scale that shows the degree to what extent a factor
influences another factor through pairwise-comparison. Similar to Wu and Lee [18],
five fuzzy linguistic scales with triangular fuzzy members was implemented as
demonstrated in Table 1. The invited experts are asked to evaluate the influence of
each factor using this scale.
Table 1. The fuzzy linguistic scale

Linguistic terms Triangular fuzzy numbers Crisp numbers
Very high influence (VH) (0.75, 1.00, 1.00) 4
High influence (H) (0.50, 0.75, 1.00) 3
Low influence (L) (0.25, 0.50, 0.75) 2
Very low influence (VL) (0, 0.25, 0.50) 1
No influence (No) (0, 0, 0.25) 0
Step 3: Obtain the initial relation matrices consisting of linguistic terms for each
expert.
Step 4: The linguistic terms are converted into the triangular fuzzy numbers given in
Table 1.
Step 5: Defuzzify the fuzzy numbers using the CFCS (Converting Fuzzy date into
the Crisp Scores) method proposed by Opricovic and Tzeng [22]. Then combine
experts’ evaluations using their crisp values.
Step 6: Calculate the initial direct relation (Z), normalized direct relation (X) and
comprehensive impact matrices (T). The comprehensive impact matrix shows the
total impact relationship among the factors.
Step 7: Calculate the influence (Di), the affected (Ri), the center (Pi) and the cause
(Ei) degrees for each factor i to identify the significance of the factors and the cause
and effect relationship between them, where;
Pi ¼ Di þ Ri ð1Þ
Ei ¼ Di Ri ð2Þ
4 Implementation and Results
For the implementation phase, five experts were invited to determine the influences of
each factor. Four of them have a medical background and the other has a business
background. All of the experts that have medical background have at least five-year
experience in vaccination, immunization or public health. The non-medical expert has
experience in forecasting, inventory management, logistics, and healthcare engineering
studies. To provide insights into the influence of factors on vaccine demand forecasting,
the thoughts of these experts can be used. The characteristics of the experts are shown
in Table 2. Because of confidentiality, their names and institutions were not shared in
this study.
Table 2. The characteristics of the experts

Expert ID Degree Tittle Academic/Practice Experience year
Expert #1 Medical Doctor Professor Both 41 years
Expert #2 Medical Doctor – Practice 30 years
Expert #3 Midwifery Asst. Prof. Both 26 years
Expert #4 Medical Doctor – Practice 12 years
Expert #5 Industrial Eng. Assoc. Prof. Academic 16 years
Experts’ pairwise comparisons were collected through a face-to-face interview.

After the goal of the study was explained, each factor was introduced in detail. Next,
the assessment scale and the procedure were explained. Table 3 shows the Expert #3’s
linguistic assessment.
Table 3. The linguistics assessment of Expert #3

Expert #3 F1 F2 F3 F4 F5 F6
F1 NO NO NO NO VL VH
F2 VH NO VH H VH H
F3 VH NO NO H H VH
F4 VH NO NO NO VL VL
F5 VH NO NO VH NO VH
F6 VH NO NO H NO NO
After implementing the fuzzy-DEMATEL procedure described briefly in the pre-

vious section, the initial direct relation (Z) and comprehensive impact matrices (T) were
obtained and given in Tables 4 and 5.
Table 4. The direct relation matrix

Z F1 F2 F3 F4 F5 F6
F1 0.033 0.051 0.033 0.161 0.370 0.409
F2 0.543 0.033 0.759 0.339 0.802 0.586
F3 0.409 0.418 0.033 0.339 0.718 0.759
F4 0.915 0.077 0.087 0.033 0.389 0.279
F5 0.847 0.062 0.169 0.669 0.033 0.707
F6 0.915 0.116 0.077 0.759 0.316 0.033
Table 5. The comprehensive impact matrix with the influence and the affected degrees.
T F1 F2 F3 F4 F5 F6 Di
F1 0.2026 0.0446 0.0475 0.1822 0.2169 0.2475 0.9414
F2 0.7041 0.1023 0.3394 0.4605 0.5698 0.5626 2.7385
F3 0.6120 0.2029 0.1101 0.4223 0.4976 0.5489 2.3938
F4 0.5291 0.0644 0.0770 0.1712 0.2757 0.2723 1.3897
F5 0.6344 0.0794 0.1180 0.4353 0.2271 0.4521 1.9463
F6 0.5978 0.0843 0.0861 0.4175 0.2937 0.2231 1.7025
Ri 3.2800 0.5779 0.7781 2.0890 2.0808 2.3064
Using the Di and Ri values obtained in Table 5, the center (Pi) and the cause (Ei)
degrees were calculated for each factor i using Eqs. 1 and 2. Hence, the centrality
degree Pi indicates the importance of the factors where the highest Pi shows the most
important factors. Moreover, the cause degree Ei is used to determine which factors are
cause or effect. Hence, the factors that have negative Ei values are the effect and the
positive ones are the cause factors. Table 6 demonstrated the centrality and the cause
degrees for each factor as well as their importance and groups.
Table 6. The importance order and the groups of factors according to the centrality and cause
degrees.
F1 F2 F3 F4 F5 F6
Pi (Di+ Ri) 4.221 3.316 3.172 3.479 4.027 4.009
Ei (Di− Ri) −2.339 2.161 1.616 −0.699 −0.135 −0.604
Ei (+) Cause Group: F2, F3
Ei (−) Effect Group: F1, F6, F4, F5
Importance order (descending): F1, F6, F5, F4, F2, F3
2.5
2.0 F2
1.5 F3
1.0
0.5
0.0 F5
Ei
-0.5 3 4 F6 5
F4
-1.0
-1.5
-2.0
-2.5 F1
-3.0
Pi
Fig. 1. The causal diagram of the factors

As seen in Table 6, immunization-related beliefs are determined as the most

influential factor of vaccine demand while urbanization is the least influential. More-
over, district-level per capita income and urbanization are indicated as the causes and
the other factors as the effects. Figure 1 demonstrates cause and effect factors in a
causal diagram.
5 Managerial Insights and Concluding Remarks
This study investigated the factors affecting the demand for vaccines that are recom-
mended by WHO but not included in childhood immunization programs of countries.
Because such vaccines are not included in national immunization programs, parents
need to decide whether to get their children vaccinated. This research provides some
managerial insights to decision-makers into demand factors for these types of vaccines,
which are;
– Immunization-related beliefs have the greatest influence on demand forecasting
among the other factors. Based on this, it can be suggested to survey parents to
determine misbeliefs about these vaccines to increase the demand for such vaccines.
Thus, the spread of these misbeliefs can be prevented.
– According to analysis results, district-level per capita income and urbanization are
determined as the cause group. These criteria should be controlled and decision-
makers should focus more on these.
This study contains some limitations. The first limitation is the shortage of
respondents. For future research, conducting more surveys to a larger number of
experts is planned to ensure the validity of the research. Secondly, macro factors
affecting vaccine demand is included in this analysis; future research should include
macro and micro factors.
References
1. World Health Organization. https://vaccine-safety-training.org/Importance-of-immunization-
programmes.html. Accessed 02 Mar 2020
2. World Health Organization. https://www.who.int/news-room/fact-sheets/detail/immuni-
zation-coverage. Accessed 05 Mar 2020
3. Riedel, S.: Edward Jenner and the history of smallpox and vaccination. Baylor Univ. Med.
Center Proc. 18(1), 21–25 (2005)
4. Worldometers. https://www.worldometers.info/coronavirus/. Accessed 14 Apr 2020
5. OECD. http://oecd.org/economic-outlook. Accessed 05 Mar 2020
6. Ghosh, A., Laxminarayan, R.: Demand- and supply-side determinants of diphtheria-
pertussis-tetanus nonvaccination and dropout in rural India. Vaccine 35(7), 1087–1093
(2017)
7. Phillips, D.E., Dieleman, J.L., Lim, S.S., Shearer, J.: Determinants of effective vaccine
coverage in low and middle-income countries: a systematic review and interpretive
synthesis. BMC Health Serv. Res. 17, 681 (2017)
8. Rammohan, A., Awofeso, N.: District-level variations in childhood immunizations in India:

the role of socio-economic factors and health infrastructure. Soc. Sci. Med. 145, 163–172
(2015)
9. Agarwal, S.: The state of urban health in India; comparing the poorest quartile to the rest of
the urban population in selected states and cities. Environ. Urban. 23(1), 13–28 (2011)
10. Kibreab, F., Lewycka, S., Tewelde, A.: Impact of mother’s education on full immunization
of children aged 12–23 months in Eritrea: population and health survey 2010 data analysis.
BMC Public Health 20(1), 1–10 (2020)
11. Trivedi, R., Singh, S., Adhikari, P., Jatav, D.P.: Coverage evaluation of primary
immunization and the associated determinants in an urban slum of Rewa, Indian.
J. Community Health 26, 37–40 (2011)
12. Arokiasamy, P., Jain, K., Goli, S., Pradhan, J.: Health inequalities among urban children in
India: a comparative assessment of Empowered Action Group (EAG) and South Indian
states. J. Biosoc. Sci. 45(2), 167–185 (2013)
13. Akıs, S., Velıpasaoglu, S., Camurdan, A.D., Beyazova, U., Sahın, F.: Factors associated with
parental acceptance and refusal of pandemic influenza A/H1N1 vaccine in Turkey. Eur.
J. Pediatr. 170, 1165–1172 (2011)
14. Ciblak, M.A., Platformu, G.: Influenza vaccination in Turkey: prevalence of risk groups,
current vaccination status, factors influencing vaccine uptake and steps taken to increase
vaccination rate. Vaccine 31(3), 518–523 (2013)
15. Sissoko, D., Trottier, H., Malvy, D., Johri, M.: The influence of compositional and
contextual factors on non-receipt of basic vaccines among children of 12-23-month old in
India: a multilevel analysis. PLoS ONE 9, e106528 (2014)
16. Uddin, M.J., Larson, C.P., Oliveras, E., Khan, A.I., Quaiyum, M.M.A., Saha, N.C.: Child
immunization coverage in rural hard-to-reach Haor areas of Bangladesh: possible alternative
strategies. Asia Pac. J. Public Health 21(1), 8–18 (2009)
17. Shea, B., Andersson, N., Henry, D.: Increasing the demand for childhood vaccination in
developing countries: a systematic review. BMC Int. Health Hum. Rights 9, 5 (2009)
18. Wu, W.W., Lee, Y.T.: Developing global managers’ competencies using the fuzzy
DEMATEL method. Expert Syst. Appl. 32(2), 499–507 (2007)
19. Lin, C.J., Wu, W.W.: A causal analytical method for group decision-making under fuzzy
environment. Expert Syst. Appl. 34(1), 205–213 (2008)
20. Changli, F., Ma, R.: Identification of the factors that influence service innovation in
manufacturing enterprises by using the fuzzy DEMATEL method. J. Cleaner Prod. 253,
120002 (2020)
21. Mokhtarian, M.N.: A note on developing global manager’s competencies using the fuzzy
DEMATEL method. Expert Syst. Appl. 38(7), 9050–9051 (2011)
22. Opricovic, S., Tzeng, G.H.: Defuzzification within a multicriteria decision model. Int.
J. Uncertainty Fuzziness and Knowl.-Based Syst. 11(5), 635–652 (2003)
23. World Health Organization. https://www.who.int/immunization/newsroom/newsstory_
rotavirus_vaccines_immunization_programmes/en/. Accessed 09 Mar 2020
24. World Health Organization. https://www.who.int/gho/immunization/rotavirus/en/. Accessed
09 Mar 2020
25. Wiysonge, C.S., Uthman, O.A., Ndumbe, P.M., Hussey, G.D.: Individual and contextual
factors associated with low childhood immunisation coverage in sub-Saharan Africa: a
multilevel analysis. PLoS ONE 7, e37905 (2012)
Benchmarking and Visualization
of the Ergonomic Risks in a Manufacturing
Plant: Joint Application of Fuzzy AHP
and Treemap
Samira Keivanpour(&)
Department of Mathematics and Industrial Engineering,

Polytechnique Montreal, Montreal, QC, Canada
[email protected]
Abstract. Occupational risk assessment in manufacturing plants is essential.

However, the simplicity of the methods and the capability to visualize the results
play a critical role in their applications. The competitor benchmarking is
involved with comparing the risk assessment of different manufacturers com-
peting in the same sector or region. This paper proposes a model for bench-
marking and visualization of the ergonomic risks in manufacturing. The joint
application of the fuzzy analytical hierarchical process (AHP) and treemap is
applied for benchmarking and mapping the risks. A numerical example is
provided to illustrate the effectiveness of the proposed approach.
Keywords: Fuzzy AHP Ergonomic risks Treemap Visualization
1 Introduction
Ergonomic risk assessment is a vital part of the safety program for manufacturing
plants. Multi-criteria approaches are frequently used in this context. Several studies
addressed the occupational risk assessment with multi-criteria approaches [1–7]. Fuzzy
logic is a promising approach in the probabilistic context and considering the lack of
numerical data and the presence of judgment and opinions of the experts. The com-
plexity and the cost of the quantitative risk assessment approaches play an essential role
in their applications particularly in small and medium enterprises [8, 9]. Among studies
applied Multi-criteria decision making (MCDM) approaches in occupational health and
safety context, ergonomic risk factors have not received much attention in the literature
[3]. Moreover, the ability of the models in sensitivity analysis and providing the
decision-making dashboard is essential. Visualization is widely applied in the industrial
context to handle analyzing multi-attributes databases and dealing with the complexity
of the interpretation of the modeling outcomes [12, 13, 15].
In this paper, a joint application of fuzzy AHP and treemap is proposed for
benchmarking and visualization of the ergonomic risks in a manufacturing plant. The
contributions of this study are as follows
Benchmarking and Visualization of the Ergonomic Risks 691
• Fuzzy AHP is applied for benchmarking the ergonomic risk assessment of a

manufacturing plant via a simple method. AHP express is recently proposed by Leal
as a simple way for the application of the AHP method [10]. A fuzzy version of this
approach is proposed to avoid the complexity of applying fuzzy AHP method
operations.
• A visual interface with a treemap is proposed for facilitating the interpretation of the
results for decision-makers.
• This joint application of fuzzy AHP and treemap could be used as a comparative
tool for evaluating the ergonomic risk mitigating strategies of manufacturers in a
horizon plan.
The rest of the paper is organized as follows: Sect. 2 provides a brief literature
review, Sect. 3 illustrates the method, Sect. 4 shows the method application via a
numerical example and finally Sect. 5 concludes with some remarks and the per-
spectives of future research.
2 Literature Review
Multicriteria decision making (MCDM) is used widely in the context of occupational

health and safety risk assessment. Here, a brief review of these approaches is provided.
According to Pinto et al., occupational safety is the key stage for meeting the safety
target in workplaces. Furthermore, it is essential for supporting decision-makers. The
authors explained that a few methods for risk assessment are adopted due to the
complexity and the costs of these methods [2].
Delice et al. applied the fuzzy MCDM model for ranking the risk factors for
healthcare workers [4]. Gul et al. also used a fuzzy AHP approach for health risk
assessment in a hospital in Turkey [5]. Ghosh et al. used the Delphi method and AHP
for developing safety indicators. The authors stated that MCDM is an appropriate
method for avoiding groupthink behavior [6]. Ozdemir et al. applied a hybrid MCDM
approach for risk assessment in university laboratory operations [7]. Gul performed a
literature review on the risk assessment approaches in occupational health and risk
assessment [3]. The author emphasized the role of the MCDM approach in this context.
Moreover, considering the lack of numerical data in the process, the application of
fuzzy MCDM approaches with the capability of considering the judgments and
opinions of decision-makers instead of quantitative data is preferable. The author also
highlighted the effectiveness of fuzzy AHP and (analytic network process) ANP for
assigning non-crisp risk factors and dealing with the inconsistency of decision pro-
cesses. The author also stated that ergonomic factors are more neglected in this context.
Another gap addressed by the author is the lack of sensitivity analysis and the decision
support systems for facilitating the assessment process and applying the results. This
study aims to fill the gap by proposing a simple fuzzy AHP model for benchmarking
the ergonomic risks of manufacturing facilities and visualizing the results via the
treemap approach to expedite the decision-making process for the mangers and
policymakers.
692 S. Keivanpour
3 Method
Here, the key steps of the proposed approach are explained.
3.1 Step 1: Identifying the Ergonomic Risk Factors for Each Task
in Each Work Cell in a Manufacturing Plant and Building the AHP
Model
The AHP approach is developed by Saaty [11] and is widely used in the business and
industrial context. Identifying the risk criteria and alternatives and forming the hier-
archical structure is the first step of this approach [15, 16]. In this step for each work
cell, the tasks should be identified. For each task, the relevant ergonomic risk factors
based on each posture should be addressed. The frequency, time of exposure, and the
current method approach used for mitigating this risk should be addressed. The
alternatives of the model will be the target plant and the selected enterprises as the
representative in the industrial sector that are ranked based on their ergonomic risk
performance.
3.2 Step 2: Performing Pairwise Comparison and Forming the Matrices

with Assigning the Triangular Fuzzy Number to Criteria
and Alternatives
Leal proposed a simplified version of the AHP approach for reducing the number of
pairwise comparisons from (n(n − 1)/2) to (n − 1) [10]. In this method for each
component in the hierarchy structure the elements that are more important in com-
parison to the criteria in the higher level should be identified. Then, the importance of
the criteria should be calculated. The priorities of each alternative should be calculated
based on each criteria ascending in the hierarchal structure toward the main objective.
The triangular fuzzy number is used for pairwise comparison in this study. So, the
equation proposed by Leal [10] is modified for fuzzy comparison as Eq. 1
f ¼ 1
prj P ð1Þ
f
Cij k 1
ceik
Where C fij ¼ ðxij; yij; zijÞ is the fuzzy number for comparison of criterion i to j and prj
f
is the priority of criteria i to j.
3.3 Step 3: Calculating the Priorities Matrix and Ranking

the Alternatives
e of priorities of subcriteria j in criterion i is defined as Eq. 2:
A fuzzy vector ki
i
e
ki ¼ e
a1 ... e
a im ð2Þ
A fuzzy matrix ec i of priorities of each alternative in each subcriteria j of criteria i is:

2 3
ec i1;1 . . . ec i1;t
ec i ¼ 4 : : : 5 ð3Þ
ec i1;m . . . ec t;m
i
e
Each fuzzy matrix ec i can be ordered by each criterion i grouped in a fuzzy matrix X:
2 3
ec i1;1 . . . ec i1;t
6 : : : 7
6 i 7
6 ec 1;m . . . ec it;m 7
6 7
e ¼ 6 ...
X ... ... 7 ð4Þ
6 n 7
6 ec 1;1 . . . ec n1;t 7
6 7
4 : : : 5
ec n1;mn . . . ec nt;mn
Each fuzzy vector e e of priorities of subcriteria,

ki can be inserted in a fuzzy matrix X
which is divided to correspond to the priorities of alternatives e
h:
2 3
ea 11 0 0
e
h¼4 0 : 0 5 ð5Þ
0 0 e
a nmn
The multiplication of e
h by ec i gives the matrix e
l of priorities for each alternative for
criteria i.
2 3
e
l 1;1 ... e
l t;1
e 4
l ¼ ... ... ... 5 ð6Þ
e
l 1;n ... e
l t;n
Then, a vector of priorities of each criterion for the objective will be calculated. The
product of priority of criteria into e
l gives the final priorities of alternatives and the final
ranking.
½e
p1 ... pn e
e l ¼½e
q1 ... e
qt ð7Þ
Where n; m; t is the number of criterion, subcriterion, and alternatives respectively.
3.4 Step 4: Mapping the Risk of Each Work Cell, Tasks, and Postures
Based on the Overall Ranking of the Criteria with the Treemap
Approach
Treemap provides visualization of hierarchical data using the space-filling algorithms
[14]. There are several applications of treemap in the stock market, performance
measurement, and operations [12–14]. Figure 1 shows a simple example of mapping
694 S. Keivanpour
the risk profile of a work cell with three tasks. Each rectangle represents a posture. The
size and color could be used to demonstrate the risk and the change in the risk
alleviation strategy.
Fig. 1. A simple example of using treemap approach to mapping the ergonomic risk profile of a
work cell
4 A Numerical Example
In this section, a numerical example is illustrated to show the application of the pro-
posed method. The target manufacturer X is considered in this example. The objective
is to provide the benchmark with two manufacturers Y, Z which are the representatives
of low and medium ergonomic risk companies in the sector. The hierarchical structure
of the model is shown in Fig. 2. A manufacturing plant with two work cells is
considered.
Plant
WorkstaƟon WorkstaƟon
1 2
W1_Task1 W1_Task2 W2_Task1
W1_T1_Arm W1_T1_Back W1_T2_Arm W1_T2_Leg W2_T1_Arm W2_T1_Back
W1_T1_A_P1 W1_T1_A_P2 W1_T1_B_P1 W1_T1_B_P2 W1_T2_A_P1 W1_T2_A_P2 W1_T2_L_P1 W1_T2_L_P2 W2_T1_A_P1 W2_T1_A_P2 W2_T1_B_P1 W2_T1_B_P2
Frequency
Time-
Exposure
Avoiding-
Risk-opp. .....
Fig. 2. The manufacturing plant with two work cells, AHP criterion and subcriterias
Work cell 1 encompasses two tasks. There are two types of ergonomic risk for task
1 including arm and back and task 2, arm and leg risk factors. For arm and back risk in
task 1, two postures impose the risk. Work cell 2 includes one task with arm and back
risk. For each posture, three sub-criteria are defined: frequency of the exposure, time to
exposure, and any strategy for alleviating the risk.
Table 1 and 2 shows the examples of priority matrices for comparison of the sub-
criteria at the fifth and fourth level of the hierarchy and alternatives for task 1 respectively.
Table 1. Pairwise comparison of Frequency concerning time-exposure and avoiding risk

strategy for two postures of the arm and two postures of back in Task 1
Frequency Time-exposure Avoiding-risk-opp.
W1_T1_A_P1
Frequency – (3,4,5) (2,3,4)
W1_T1_A_P2
Frequency – (3,4,5) (1,2,3)
W1_T1_B_P1
Frequency – (3,4,5) (1,2,3)
W1_T1_B_P2
Frequency – (2,3,4) (2,3,4)
Table 2. Pairwise comparison of alternatives concerning sub-criteria at the fifth level in Task 1
X Y Z
FR_W1_T1_A_P1
X – (2,3,4) (4,3,5)
T.E_W1_T1_A_P1
X – (1,2,3) (2,3,4)
A.R.O_W1_T1_A_P1
X – (1,2,3) (2,3,4)
Each fuzzy vector resulted in Table 1 can be inserted in a fuzzy matrix X e of

priorities of subcriteria to form the matrix eh as shown in Table 3.
The results of risk ranking will be mapped via treemap approach. For example,
Fig. 3 shows the risk profile of 12 postures in work cells 1 and 2. The three rectangles
in each part represent the frequency, time to exposure, and risk-avoiding strategies in
the benchmark of plant X. The size and color of each rectangle could be defined based
on the objective of the manufacturer in the risk evaluation process. This approach
provides flexibility in the presentation of the results and facilitate the decision making
for the policymakers and managers. WfngTfmgfA; B; LgPflg represents the posture
number, risk type, task number, and the work cell. Hence, for example, posture 2, risk
related to the leg, in task 2, related to work cell 1 requires an immediate corrective
action plan for reducing the risk (Fig. 3). By integrating the final weights and the
product of priority of criteria into the priorities of alternatives, the final priorities of
alternatives and the ranking are calculated. The final score of plant X is 0.1674, for Y
0.5723 and for Z, 0.2671. The result placed Plant X after Y and Z. Therefore, the
overall risk of this plant is more than a competitor in the sector with a medium
ergonomic risk profile.
696
S. Keivanpour
Table 3. Matrix e
h for work cell 1 and 2
W1_T1_A_P1 W1_T1_A_P2 W1_T1_B_P1 W1_T1_B_P2 W1_T2_A_P1 W1_T2_A_P2 W1_T2_L_P1 W1_T2_L_P2 W2_T1_A_P1 W2_T1_A_P2 W2_T1_B_P1 W2_T1_B_P2
W1_T1_Arm (0.5,0.75,1) (0,025,0.5) 0 0 0 0 0 0 0 0 0 0
W1_T1_Back 0 0 (0.5,0.75,1) (0,0.15,0.3) 0 0 0 0 0 0 0 0
W1_T2_Arm 0 0 0 0 (0.5,0.75,1) (0,0.15,0.3) 0 0 0 0 0 0
W1_T2_Leg 0 0 0 0 0 0 (0.5,0.75,1) (0,025,0.5) 0 0 0 0
W2_T1_Arm 0 0 0 0 0 0 0 0 (0.5,0.75,1) (0,025,0.5) 0 0
W2_T1_Back 0 0 0 0 0 0 0 0 0 0 (0.5,0.75,1) (0.15,0.3,0.45)
Fig. 3. Treemap of 12 postures resulted from fuzzy AHP benchmark
5 Conclusion
The ergonomic risk evaluation is an important part of the safety plan for the manu-
facturing units. The small and medium manufacturers require a simple and reasonably
priced method for evaluation of the ergonomic risks. They also need to compare their
performance with the other manufactures in the sector or region. This paper provides a
model for ergonomic risk assessment in a manufacturing environment. The joint
application of fuzzy AHP and treemap provides benchmarking and visualization of the
risk for manufacturers. In this study, the dependency of the risk factors is ignored.
However, the fuzzy ANP model could be proposed as future research to consider
dependency. For using AHP express method, the assumption of consistency of pairwise
comparison matric is essential. Addressing the consistency of the comparison matrix
for group decision making should be studied.
References
1. Gul, M., Guneri, A.F.: Use of FAHP for occupational safety risk assessment: an application
in the aluminum extrusion industry. In: Fuzzy Analytic Hierarchy Process, pp. 249–271
(2018)
2. Pinto, A., Nunes, I.L., Ribeiro, R.A.: Occupational risk assessment in construction industry–
overview and reflection. Saf. Sci. 49(5), 616–624 (2011)
3. Gul, M.: A review of occupational health and safety risk assessment approaches based on
multi-criteria decision-making methods and their fuzzy versions. Hum. Ecol. Risk Assess.:
Int. J. 24(7), 1723–1760 (2018)
698 S. Keivanpour
4. Delice, E.K., Zegerek, S.: Ranking occupational risk levels of emergency departments using
a new fuzzy MCDM model: a case study in Turkey. Appl. Math. Inf. Sci. 10(6), 2345–2356
(2016)
5. Gul, M., Ak, M.F., Guneri, A.F.: Occupational health and safety risk assessment in hospitals:
a case study using two-stage fuzzy multi-criteria approach. Hum. Ecol. Risk Assess.: Int.
J. 23(2), 187–202 (2017)
6. Ghosh, N., et al.: Use of multi criteria decision making methods in assessment of
occupational safety and health. J. Ind. Saf. Eng. 6(3), 25–34 (2019)
7. Ozdemir, Y., Gul, M., Celik, E.: Assessment of occupational hazards and associated risks in
fuzzy environment: a case study of a university chemical laboratory. Hum. Ecol. Risk
Assess.: Int. J. 23(4), 895–924 (2017)
8. Yuan, F.-C.: The use of a fuzzy logic-based system in cost-volume-profit analysis under
uncertainty. Expert Syst. Appl. 36(2), 1155–1163 (2009)
9. Keivanpour, S.: An integrated approach to value chain analysis of end of life aircraft
treatment. Dissertation, Université Laval (2015)
10. Leal, J.E.: AHP-express: a simplified version of the analytical hierarchy process method.
MethodsX 7, 100748 (2020)
11. Saaty, T.L.: Decision making—the analytic hierarchy and network processes (AHP/ANP).
J. Syst. Sci. Syst. Eng. 13(1), 1–35 (2004)
12. Keivanpour, S.: Adapting treemaps to student academic performance visualization. In:
Science and Information Conference. Springer, Cham (2019)
13. Keivanpour, S., Kadi, D.A.: Strategic eco-design map of the complex products: toward
visualisation of the design for environment. Int. J. Prod. Res. 56(24), 7296–7312 (2018)
14. Walker, D.J.: Visualisation with treemaps and sunbursts in many-objective optimisation.
Genet. Program Evol. Mach. 19(3), 421–452 (2018). https://doi.org/10.1007/s10710-018-
9329-0
15. Kimiagari, S., Keivanpour, S.: An interactive risk visualisation tool for large-scale and
complex engineering and construction projects under uncertainty and interdependence. Int.
J. Prod. Res. 57(21), 6827–6855 (2019)
16. Keivanpour, S., Haleh, H., Ganjavi, H.S.: A decision model for selecting the optimum oil
production profile using multi criteria decision making and social choice theory. In: ASME
2014 International Mechanical Engineering Congress and Exposition. American Society of
Mechanical Engineers Digital Collection (2014)
Multicriteria Decision Making-Theory
An Integrated Fuzzy DEMATEL
and Fuzzy Cognitive Mapping
Methodology for Prioritizing Smart
Campus Investments
Ali Karasan1,2(B) and Cengiz Kahraman2

1
[email protected]
2
[email protected]
Abstract. Based on the development of the technologies, traditional edu-

cation systems are in a transformation to catch the era. This transfor-
mation is mainly based on 4 dimensions: Strategic plan of the university,
functionality for the users, financial aspects, and physical infrastructure.
Based on these dimensions, there are 4 most attracted investment areas
in the literature review: smart buildings, parking systems, energy savings,
e-learning systems. In this study, a preliminary work for a smart cam-
pus project of a public university is constructed to prioritize the invest-
ment alternatives based on the dimensions’ sub-criteria. For the evaluation
process, fuzzy DEMATEL and cognitive mapping methods are applied.
Fuzzy DEMATEL method is applied to obtain influences of the criteria
and investments based on the expert judgments. Then, cognitive mapping
is conducted to prioritize the investments together with the sub-criteria.
Keywords: Smart campus · Fuzzy DEMATEL · Fuzzy cognitive

maps · Decision making
1 Introduction
Smart concept consists of many categories which serve to achieve several pur-
poses based on the constructed context. Smart cities, smart parking systems,
smart campuses, smart buildings, smart grids, etc. are the most common and
studied projects for years. Among them, a smart campus concept links devices,
applications, and people to improve operational efficiency by enhancing basic
services of the universities such as buildings, parking areas, education channels
and environments, library, etc. [1].
There are many criteria to consider for the transition of a traditional cam-
pus into smart one while evaluating the dimensions which should be prioritized.
702 A. Karasan and C. Kahraman
The prioritizing problem requires operational, environmental, social, and eco-

nomics criteria to be conducted in the process. These criteria are mostly intan-
gible and requires subjective evaluations. Also, most of the criteria have depen-
dencies and conflicting with each other. Through that, these issues create a
system that has vague and imprecise data. To handle the data, usage of fuzzy
set theory which has ability to reflect uncertainty of the data by incorporating
linguistic terms into mathematical formulations is one of the most appropriate
way.
Fuzzy set theory was introduced by Zadeh to capture the uncertainties in
human thoughts through the degree of membership functions which are employed
to the elements in a set to represent their belongings [2]. From the introduction of
the fuzzy sets, there are many extensions which try to enhance their advantages
[3–8]. Through them, there is no absolute superiority for one to another and the
decision of which one is the most appropriate extension can be decided based on
the available data for an application.
Since the prioritization problem consists of multi evaluation criteria and
many alternatives together with human knowledge on the operation of the sys-
tem, the structure of the problem can be constructed with an integrated decision
making methodology to handle its complexity. The cognitive maps were firstly
introduced by Axelrod to represent interrelationships and their corresponded
binary values [9]. After that, Kosko constructed fuzzy cognitive maps to increase
its ability to represent the data [10]. In the system, construction methodologies
rely on the exploitation of experts’ experience on system’s model and behavior
[11]. On the other hand, DEMATEL is an efficient technique that involves the
relationship data of the parameters/criteria to determine the inner and outer
dependencies of them [12,13].
In this study, an integrated methodology consists of fuzzy DEMATEL and
cognitive mapping is employed to prioritize the investment alternatives with
respect to 4 dimensions’ criteria. DEMATEL method is conducted to obtain
interrelationships and their corresponded crisp values. Then, cognitive mapping
is constructed to reveal the prioritization of the alternatives. Rest of the paper
is organized as follows: Sect. 2 presents the fuzzy sets and steps of the applied
methodology. In Sect. 3, the investment prioritization for the smart campus tran-
sition process problem is given together with its results and analyses. The paper
ends with conclusions and further suggestion.
2 Methodology
In this section, the integrated approaches and their steps are presented. Fuzzy
DEMATEL method is introduced by Chang et al. to handle linguistic fuzzy infor-
mation while determining the interrelationships. Steps of the proposed method-
ology are presented as follows:
Step 1. Determine the structure of the problem.
Step 2. Construct the direct relation matrices based on expert judgments by
using the linguistic scale given in Table 1.
An Integrated Fuzzy DEMATEL and Fuzzy Cognitive Mapping 703
Step 3. Convert linguistic terms to their corresponded triangular fuzzy numbers

in the total relation matrix.
Table 1. Linguistic scale for fuzzy DEMATEL
Linguistic Corresponded Corresponded

term survey value fuzzy number
No Influence 0 (0, 0, 0)
Very Low Influence 1 (0, 0.25, 0.5)
Low Influence 2 (0.25, 0.5, 0.75)
High Influence 3 (0.5, 0.75, 1)
Very High Influence 4 (1, 1, 1)
Step 4. Transform triangular fuzzy numbers to construct generalized direct-

relation matrix (S) by using Algorithm 1.
Algorithm 1. Transformation of fuzzy numbers

Input : n: number of compared ones, (n = 1, · · · , k),
xij : comparison value assigned by the expert, (i = 1, · · · , k), (j = 1, · · · , k)
Output: aij : corresponded crisp evaluation
1 for ∀ xij ← 1 to k do
2 Step 4.1: Normalization
(r −min l )
3 xrij = ijΔmax ij
min
(mij −min lij )
4 xmij = Δmax
min
(lij −min lij )
5 xlij = Δmax
min
6 where Δmax
min = max rij − min lij and xij = (xlij , xmij , xrij )
7 Step 4.2: Compute right (rs) and left (ls) normalized values
xrij
8 xrsij =
(1+xrij −xmij )
xmij
9 xlsij =
(1+xmij −xlij )
10 Step 4.3: Compute total normalized crisp values
[xlsij (1−xlsij )+xrsij ×xrsij ]
11 xij =
[1−xlsij +xrsij ]
12 Step 4.4: Compute crisp values
13 aij = min lij + xij × Δmaxmin
14 end
Step 5. Normalize (S) matrix.

Step 6. Obtain total-relation matrix (M ).
Step 7. Calculate the threshold value.
Step 8. Obtain the interrelationships and their values. The position of the values
can be negative, neutral, and positive. They are calculated as follows:
While assigning the influences in Step 2, expert also indicates the position
of the assigning number which can be negative or positive as a matrix form
(D). After Step 8, the obtained values are scalarly multiplied by matrix D to
demonstrate the direction of the values in cognitive map.
Step 9. Represent the cognitive map based on the outputs of the Fuzzy DEMA-
TEL.
Step 10. Construct the weight matrix.
Step 11. Compute the dimensions and sub-criteria weights by using Eq. 1.
⎛ ⎞
N
= f ⎝Ai +
(k+1) (k) (k) μ (k) π ⎠
Ai Aj wji − Aj wji (1)
j=i
Since the nature of concepts can be negative, their values belong to the
interval [−1, 1], f (x) = tanh(x) function is used. The iteration is run until all
the weights are converged or a certain number of iterations is repeated.
3 Application
Smart campus concept investments are evaluated with respect to 4 dimensions

and 13 sub-criteria. The dimensions are adopted from the study of Den Heijer
[14]. The structure of the problem is given in Fig. 1.
Fig. 1. Structure of the problem

Table 2. Direct relation matrix
I1 I2 I3 I4 C11 C12 C13 C14 C15 C21 C22 C23 C31 C32 C33 C41 C42
I1 0 0 4 2 1 1 4 0 3 2 2 3 2 1 3 0 0
I2 0 0 4 0 1 2 4 0 2 1 2 2 0 2 2 2 4
I3 1 2 0 0 2 3 2 0 2 0 0 1 2 1 4 1 1
I4 1 0 0 0 3 2 1 0 2 2 1 3 1 2 1 2 3
C11 3 3 3 3 0 4 4 0 1 0 1 2 3 2 2 1 1
C12 0 0 0 1 1 0 0 3 1 4 2 0 0 0 0 0 0
C13 0 0 0 0 1 0 0 0 3 1 2 2 0 1 1 1 1
C14 1 0 1 1 0 3 0 0 3 3 2 0 0 0 0 0 0
C15 0 0 0 2 1 1 0 2 0 2 4 0 3 1 1 0 0
C21 1 3 1 4 2 2 4 1 1 0 0 2 1 2 2 1 2
C22 1 3 1 4 2 2 4 1 1 2 0 2 1 2 2 1 2
C23 2 3 1 3 1 1 2 0 0 1 3 0 1 2 2 0 0
C31 3 3 4 1 1 0 2 0 1 0 1 1 0 1 3 0 0
C32 1 2 1 2 0 0 0 0 0 0 0 2 3 0 3 0 0
C33 2 3 4 2 0 0 1 0 0 1 2 2 4 1 0 2 2
C41 0 2 2 0 0 0 1 0 1 0 1 0 0 2 0 0 1
C42 0 3 3 0 1 0 2 0 1 0 1 0 0 2 0 0 0
Through the proposed methodology, direct relation matrix (see Table 2) is

constructed based on one expert judgment by using the scale given in Table 1.
After that, corresponding triangular fuzzy numbers are assigned and defuzzi-
fication algorithm which is given in Algorithm 1 is applied. Through that, gener-
alized direct-relation matrix is obtained. As a next step, normalization procedure
is applied. Normalized values are used to obtain total relation matrix (M (mij ))
which is given in Table 3.
Threshold value (TSV) is calculated by using Eq. 2 as follows:
maxi,j=1,k mij + mini,j=1,k mij

T SV = (2)
2
Through the TSV, interrelationships and their value are obtained. After that,
by using the position of the values, inputs of the fuzzy cognitive map are calcu-
lated as in Table 4.
The structure of the cognitive map is given in Fig. 2 as follows:
After 84 iterations, weights are converged and the results are given in Table 5.
Based on the results, energy saving investment is the first alternative. Also,
the most important sub-criterion is obtained as Investment costs.
Table 3. Total relation matrix
I1 0.093 0.128 0.254 0.187 0.146 0.142 0.273 0.037 0.189 0.155 0.183 0.200 0.181 0.180 0.225 0.081 0.096
I2 0.077 0.125 0.247 0.111 0.135 0.149 0.262 0.034 0.163 0.124 0.174 0.164 0.106 0.197 0.188 0.143 0.208
I3 0.110 0.162 0.127 0.097 0.136 0.153 0.183 0.029 0.145 0.078 0.106 0.133 0.155 0.152 0.226 0.109 0.116
I4 0.116 0.114 0.126 0.112 0.164 0.143 0.176 0.032 0.150 0.138 0.147 0.177 0.138 0.186 0.155 0.131 0.156
C11 0.171 0.210 0.236 0.206 0.115 0.223 0.287 0.040 0.166 0.111 0.178 0.190 0.200 0.206 0.213 0.125 0.145
C12 0.046 0.061 0.063 0.115 0.101 0.063 0.084 0.104 0.099 0.172 0.118 0.058 0.055 0.065 0.064 0.039 0.050
C13 0.047 0.072 0.074 0.073 0.091 0.053 0.083 0.021 0.131 0.087 0.126 0.118 0.064 0.111 0.109 0.078 0.088
C14 0.081 0.058 0.101 0.113 0.059 0.137 0.082 0.030 0.138 0.139 0.123 0.059 0.059 0.066 0.069 0.038 0.048
C15 0.067 0.088 0.097 0.156 0.111 0.113 0.111 0.088 0.076 0.129 0.197 0.083 0.156 0.127 0.132 0.053 0.066
C21 0.127 0.201 0.185 0.236 0.164 0.162 0.274 0.073 0.158 0.100 0.131 0.182 0.153 0.199 0.197 0.122 0.161
C22 0.135 0.213 0.196 0.250 0.175 0.172 0.291 0.077 0.167 0.167 0.139 0.193 0.162 0.211 0.209 0.129 0.171
C23 0.136 0.184 0.166 0.186 0.130 0.127 0.201 0.029 0.097 0.123 0.180 0.110 0.139 0.177 0.182 0.073 0.090
C31 0.152 0.181 0.243 0.137 0.124 0.087 0.195 0.024 0.133 0.076 0.141 0.144 0.102 0.150 0.203 0.070 0.085
C32 0.093 0.130 0.130 0.122 0.056 0.052 0.090 0.013 0.058 0.049 0.068 0.128 0.145 0.074 0.160 0.046 0.056
C33 0.144 0.204 0.265 0.167 0.101 0.094 0.198 0.026 0.109 0.119 0.169 0.173 0.221 0.170 0.142 0.138 0.154
C41 0.029 0.108 0.113 0.042 0.039 0.038 0.099 0.013 0.082 0.033 0.085 0.045 0.043 0.114 0.057 0.031 0.079
C42 0.039 0.138 0.146 0.054 0.088 0.052 0.136 0.016 0.094 0.041 0.096 0.059 0.056 0.127 0.073 0.041 0.051
Table 4. Inputs of the fuzzy cognitive map
I1 0 0 0.732 0.252 0 0 0.867 0 0.265 0.020 0.225 0.349 0.209 −0.200 0.526 0 0
I2 0 0 0.686 0 0 0 0.789 0 0.081 0 0.158 −0.089 0 −0.323 0.259 0 −0.403
I3 0 0.073 0 0 0 0.012 0.226 0 0 0 0 0.000 0.024 0.001 −0.535 0 0
I4 0 0 0 0 0.090 0 0.175 0 0 0 0 0.178 0 0.244 0.021 0 −0.027
C11 0.140 0.415 0.607 0.392 0 0.511 0.971 0 0.099 0 0.190 0.274 0.344 0.389 0.437 0 0
C12 0 0 0 0 0 0 0 0 0 0.145 0 0 0 0 0 0 0
C13 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
C14 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
C15 0 0 0 0.027 0 0 0 0 0 0 0.322 0 0.030 0 0 0 0
C21 0 −0.351 −0.239 0.602 0.091 0.071 0.880 0 0.041 0 0 0.213 0.005 0.340 0.325 0 0.067
C22 0 0.439 −0.320 0.705 0.163 0.142 1.000 0 0.110 0.111 0 0.293 −0.072 0.427 0.412 0 0.137
C23 0 0.233 0.104 0.245 0 0 0.357 0 0 0 0.201 0 0 0.182 0.217 0 0
C31 0.004 −0.209 0.655 0 0 0 0.310 0 0 0 0 0 0 0 −0.371 0 0
C32 0 0 0 0 0 0 0 0 0 0 0 0 0.000 0 0.057 0 0
C33 0 −0.376 0.810 −0.108 0 0 −0.335 0 0 0 −0.127 −0.155 −0.495 −0.132 0 0 −0.018
C41 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
C42 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Table 5. Results of the application
I1 0.629 C11 0.807 C21 0.769 C33 0.987

I2 0.888 C12 0.909 C22 0.937 C41 0.132
I3 0.988 C13 1.000 C23 0.943 C42 −0.749
I4 0.992 C14 0.132 C31 −0.601
C15 0.867 C32 0.946
Fig. 2. Structure of the cognitive map
4 Conclusion
In this study, an integrated decision making methodology consisting of fuzzy
DEMATEL and fuzzy cognitive mapping is applied to prioritize investment
alternatives for a smart campus project. Fuzzy DEMATEL method is applied
to demonstrate interrelationships and their values of the compared ones. The
results of the method is conducted as inputs of the cognitive mapping. Through
the results, the transition process starts with the energy saving investments.
Based on the results, the most important sub-criteria is determined as invest-
ment costs. Also, culture and footprint(m2) sub-criteria can be negligible during
the transition process.
For further study, determination of the threshold value in fuzzy DEMATEL
method can be simulated to check its effects on the final results. Based on the
results, energy saving alternatives can be prioritized firstly. Also, a multi expert
system can be constructed to improve the results. Since a multi expert system
may have hesitant evaluations, different type of fuzzy sets such as neutrosophic,
intuitionistic can be applied to handle the indeterminacy.
References
1. Cimino, V., Rombo, S.E.: Design and prototyping of a smart university campus.
In: Krishnan, S., Golden, J., Harold, R. (eds.) Handbook of Research on Implemen-
tation and Deployment of IoT Projects in Smart Cities, pp. 228–252. IGI Global,
New York (2019)
reasoning-1. Inf. Sci. 8, 199–249 (1975)
4. Atanassov, K.: Intuitionistic fuzzy sets. Fuzzy Sets Syst. 20(1), 87–96 (1986)
5. Smarandache, F.: Neutrosophic set-a generalization of the intuitionistic fuzzy set.
Int. J. Pure Appl. Math. 24(3), 287–297 (2005)
7. Yager, R.R.: Pythagorean fuzzy subsets. In: 2013 joint IFSA world congress and
NAFIPS annual meeting (IFSA/NAFIPS), pp. 57–61. IEEE, Edmonton (2013)
8. Kutlu Gündoğdu, F., Kahraman, C.: Spherical fuzzy sets and spherical fuzzy TOP-
SIS method. J. Intell. Fuzzy Syst. 36(1), 337–352 (2019)
9. Axelrod, R.: Schema theory: an information processing model of perception and
cognition. Am. Polit. Sci. Rev. 67(4), 1248–1266 (1973)
10. Kosko, B.: Fuzzy cognitive maps. Int. J. Man Mach. Stud. 24(1), 65–75 (1986)
11. Groumpos, P.P.: Fuzzy cognitive maps: basic theories and their application to com-
plex systems. In: Groumpos, P.P. (ed.) Fuzzy Cognitive Maps, pp. 1–22. Springer,
Berlin (2010)
12. Fontela, E., Gabus, A.: The DEMATEL observer, DEMATEL 1976 report. Battelle
Geneva Research Center, Geneva (1976)
13. Chang, B., Chang, C.W., Wu, C.H.: Fuzzy DEMATEL method for developing
supplier selection criteria. Expert Syst. Appl. 38(3), 1850–1858 (2011)
14. Den Heijer, A.C.: Managing the university campus: information to support real
estate decisions. Doctorate Thesis, Eburon Uitgeverij BV, Germany (2011)
Different Approaches to Fuzzy Extension
of an MCDA Method and Their
Comparison
Boris Yatsalo(B) and Alexander Korobov
Institute of Intelligent Cybernetic Systems, National Research Nuclear University

MEPHI (IATE NRNU MEPHI), Obninsk-Moscow, Russian Federation
[email protected]
Abstract. Fuzzy extension of an Multi-Criteria Decision Analysis

(MCDA) method implies a choice of an approach to assessing corre-
sponding functions of fuzzy variables and the use of a method for rank-
ing of alternatives based on ranking of fuzzy quantities. In this paper,
three key approaches to assessing functions of Fuzzy Numbers (FNs) are
considered: approximate computations based on propagating triangular
FNs through all computations, Standard Fuzzy Arithmetic (SFA), and
Transformation Methods (TMs). In addition, three methods are used for
ranking of FNs: Centroid Index, Integral of Means, and Yuan’s ranking
method. Combination of an approach to assessing functions of FNs along
with a fuzzy ranking method forms a Fuzzy MCDA (FMCDA) model.
Distinctions in ranking alternatives by FMCDA models, which are dif-
ferent fuzzy extensions of an MCDA method, are considered for Fuzzy
TOPSIS (FTOPSIS) models as an example. It is demonstrated with
the use of Monte Carlo simulation that distinctions in ranking alter-
natives by different FTOPSIS models may be considered as significant.
In such circumstances, the problem of choosing a model for application
within multi-criteria decision analysis in fuzzy environment is relevant
and requires further research.
Keywords: Fuzzy number · Extension principle · Ranking method ·

Standard fuzzy arithmetic · Transformation method · Overestimation ·
Fuzzy MCDA · Fuzzy TOPSIS
1 Introduction
Fuzzy Multi-Criteria Decision Analysis (FMCDA) [2] is a demanded method-
ological approach to evaluating a wide range of multicriteria problems in uncer-
tain/fuzzy environment and has been used in thousands of applications [7].
The primary objective of this contribution is the development and compari-
son of Fuzzy MCDA (FMCDA) models, which represent different approaches to
fuzzy extension of a classical MCDA method.
Supported by the Russian National research project RFBR-19-07-01039.
710 B. Yatsalo and A. Korobov
The two key stages of any FMCDA model are assessing functions of fuzzy
variables/Fuzzy Numbers (FNs) and implementing a method for ranking of alter-
natives/FNs. There exist four main approaches to determining functions of FNs,
which are used in FMCDA:
– Implementation of the Zadeh’s Extension Principle (EP) [13,20]; the EP is a

foundation (definition) for functions of fuzzy variables (if an assessment of a
function of FNs coincides with that according to EP, it is called the proper
value); however, EP is ineffective even for implementing simple arithmetic
operations;
– approximate calculating functions based on the basic type of input FNs (Tri-
angular FNs, TrFNs, or/and Trapezoidal FNs, TpFNs) and their propagating
through all computations; it is a generally accepted approach [6,8];
– implementation of the Standard Fuzzy Arithmetic (SFA) [3,4] based on oper-
ations with alpha-cuts; it is an effective approach, however, it can not lead to
proper assessing functions of FNs in some cases, including those when in an
expression under consideration there are dependent variables, and can lead
to an overestimation [4] of the output value;
– Transformation Methods (TMs) [4]: correct implementation of a TM leads to
a proper assessing functions of FNs (with increasing the alpha-cuts number);
however, the use of TMs requires a significant effort with the algorithms and
computing time.
There are more than 50 methods for ranking of FNs [8,11,12]. Combination
of an approach to assessing functions of FNs and a ranking method forms a
model of fuzzy extension for a chosen classical MCDA method.
In this work, the authors explore distinctions in ranking alternatives by dif-
ferent FMCDA models for fuzzy extensions of TOPSIS [5] as an example. The
surveys of FTOPSIS [2] models and their applications have been presented in
various publications, e.g., [7,17].
This paper is structured as follows. Section 2 briefly revises concepts of FNs
and fuzzy ranking methods. In Sect. 3, different FTOPSIS models are intro-
duced, and distinctions in ranking alternatives by FTOPSIS models are explored.
Finally, Sect. 4 concludes this paper.
2 Preliminaries
This section briefly revises basic concepts of FNs and ranking methods used in
this contribution.
Definition 1 [10]. A fuzzy number Z is a normal bounded fuzzy set on R with

the following property: for each α ∈ (0, 1], α-cut Zα is a closed interval.
Hereinafter, F denotes the set of FNs in accordance with Definition 1. The follow-
ing standard notions are also used in this paper: for Z ∈ F, its support, supp(Z),
is defined as supp(Z) = {x ∈ R : μZ (x) > 0}; as FN Z is bounded and taking
Different Extensions of an MCDA Method 711
into account the property of α-cuts [10,13], the closure of support, is a closed
interval (segment), denote it as [c1 , c2 ]; FN Z is considered as positive, Z > 0, if
c1 > 0, and non-negative, Z ≥ 0, if c1 ≥ 0. According to Definition 1, each FN Z
can be identified with the family of segments [10,13]: Z = {[Aα , Bα ], 0 ≤ α ≤ 1},
where Zα = [Aα , Bα ], 0 < α ≤ 1, are α-cuts of FN Z, and [A0 , B0 ] = [c1 , c2 ].
Fuzzy singleton Z, denoted as Z = c, c ∈ R, is a special type of FNs in F with
the membership function μZ (x) = {0, x < c; 1, x = c; 0, x > c}; for singleton
Z = c, Aα = Bα = c, α ∈ [0, 1].
In FMCDA, two classes of fuzzy ranking methods are mainly used: defuzzi-
fication based and pairwise comparison methods [8,11,12]. In this paper, two
defuzzification based ranking methods are used: Centroid Index (or Yager-1),
and Integral of Means (or Yager-2) [11,14,15].
Defuzzification of FN Z according to Centroid Index (CI) method is defined as

CI(Z) = xμZ (x)dx/ μZ (x)dx, (1)
Integral of Means (IM) for FN Z = {[Aα , Bα ]} is assessed by the following

formula: 1
IM (Z) = (Aα + Bα )/2 dα. (2)
0
FN with higher value of CI (IM ) has higher rank according to CI (IM) ranking
method.
In this paper, Yuan’s method (Y), from the class of pairwise comparison
ranking methods, is considered [19]. This method is based on Yuan’s fuzzy pref-
erence relation and can be briefly described as follows. Let Zi = {[Aiα , Bαi ]}, Zj =
{[Ajα , Bαj ]} ∈ F and Zij = Zi − Zj = {[Aα , Bα ]}. Within Yuan’s fuzzy preference
relation [12,19], the area SY+ is considered as a “distance” of the positive part
of Zij = {[Aα , Bα ]} to the axis OY and is determined as [18]:
1
SY+ (Zij ) = (Bα θ(Bα ) + Aα θ(Aα ))dα, (3)
0
here θ(x) is the Heaviside function: θ(x) = {1, x ≥ 0; 0, x < 0}. The total
adjusted area under FN Zij is determined [18,19] as
1
SY (Zij ) = SY+ (Zij ) + SY+ (Zji ) = (|Bα | + |Aα |)dα. (4)
0
Definition 2 [19]. Let Zi , Zj ∈ F and Zij = Zi − Zj . The Yuan’s fuzzy pref-

erence relation, RY = ((Zi , Zj ), μY (Zi , Zj )), where μY (Zi , Zj ) represents the
degree of preference of Zi over Zj , is defined as:
μY (Zi , Zj ) = SY+ (Zij )/SY (Zij ) if SY (Zij ) > 0, (5)
μY (Zi , Zj ) = 0.5 if SY (Zij ) = 0 (6)

It should be stressed, Definition 2 is valid also for singletons.

Yuan’s (Y) ranking is implemented as follows: FN Zi is equivalent to FN Zj ,
Zi ∼Y Zj , iff μY (Zi , Zj ) = 0.5; Zi exceeds Zj , Zi Y Zj , iff μY (Zi , Zj ) > 0.5,
and Zi ≺Y Zj iff μY (Zi , Zj ) < 0.5. Yuan’s fuzzy preference relation is transitive
[19] and reciprocal: μY (Zi , Zj ) + μY (Zj , Zi ) = 1.
The basic requirements (axioms) to fuzzy ranking methods and their fulfill-
ment by CI, IM, Y and other methods have been presented in [11,12].
The following Proposition has been proved in [16].
Proposition 1. Ranks of independent FNs by Integral of Means and Yuan’s
ranking methods coincide.
Here, without additional definitions, arbitrary selected FNs or those without
any functional connections/relations are considered as independent. In Sect. 3,
values of the generalized criterion are dependent.
3 Different FTOPSIS Models and Their Comparison
3.1 Fuzzy Extensions of TOPSIS

Classical TOPSIS method has been extended to fuzzy versions in many publica-
tions, e.g., [1,9]. Comprehensive survey with statistical data of Fuzzy TOPSIS
(FTOPSIS) usage has been presented in [7,17]. A general approach to fuzzy
extension of TOPSIS including the steps of FTOPSIS implementation has been
presented in [17]. In this contribution, the authors consider FTOPSIS models,
which differ by approaches to fuzzy extension of TOPSIS when calculating func-
tions of FNs and by methods for ranking of output values of the generalized
criterion (closeness coefficient).
The general FTOPSIS model can be presented as follows:
Di−
Di = D(ai ) = , (7)
Di− + Di+
where Di is a value of the generalized FTOPSIS criterion (coefficient of closeness

for alternative i), Di+ and Di− are correspondingly the (weighted) distances to
ideal point, I + , and to anti-ideal point, I − (within this general model, I + =
(1, ...., 1) and I − = (0, ..., 0)):

m
m
Di+ = ( wk2 (1 − xik )2 )1/2 ; Di− = ( (wk2 (xik )2 )1/2 , (8)
k=1 k=1
here, wk is a (fuzzy) weight coefficient for criterion k, and xik is a normalized

(fuzzy) value of alternative i for criterion k, k = 1, ...m, i = 1, ..., n [17].
For FTOPSIS, the following models can be considered depending on the
approach to assessing functions of FNs when computing (7), (8) and Dij =
Di − Dj , and on the used ranking method.
– Models FTTrCI, FTTrIM; here, FT means FTOPSIS; Tr: approximate

computations are implemented with propagating TrFNs (Triangular FNs)
through all calculated formulas of the model; CI and IM are methods for
ranking of FNs I(Di ), i = 1, ..., n, I = CI, IM . It can be added, the use of
approximated methods with the propagating triangular or/and trapezoidal
FNs is the most popular approach in FMCDA;
– Models FTSCI, FTSIM, FTSY: assessing Di (7) is based on SFA (Standard
Fuzzy Arithmetic) with subsequent direct evaluating Dij as Dij = Di − Dj
when using Y ranking method. It should be stressed, implementation of SFA
in this case leads to overestimation [4] of the output result in comparison with
the proper value of Di and Dij [17] (as in formula (7) taking into account
(8), values wk and xik are in numerator and denominator; the latest means
determining functions with dependent FNs that is the reason of overestima-
tion).
– Models FTRIM, FTRCI, FTRSY: determining Di (7) is based on RTM
(Reduced Transformation Method) [4,17]; for Y ranking method, the differ-
ence Dij , i, j = 1, ..., n, is assessed directly as Dij = Di − Dj . For this case,
models FTRIM and FTRCI result in proper values Di (7); at the same time,
value Dij is calculated with overestimation (as FNs wk are simultaneously in
expressions for assessing Di and Dj , and overestimation has the place in such
a case as FNs Di and Dj are dependent).
– Models FTRY and FTEY: within model FTRY, RTM is implemented directly
for expression Dij = Di −Dj , where on the places of Di and Dj corresponding
expressions (7) and (8) are used; within model FTEY, for determining Dij ,
RTM is implemented for FNs {xik } and {xjk }, and GTM (General TM) [4]
is implemented for variables {wk }, i, j = 1, ..., n; k = 1, ..., m [17].
3.2 Comparison of Different FTOPSIS Models
For comparison of different FTOPSIS models, presented in Subsect. 3.1, mul-

ticriteria problems with 4 criteria and 4 alternatives are considered. A short
description of an approach to the comparison is presented below.
Scenarios for exploring distinctions in ranking alternatives by these mod-
els are generated by Monte Carlo simulation of input criteria values and weight
coefficients. For this, random number generator with uniform distribution in seg-
ment [0,1] is used for generating three types of scenarios with TrFNs: TrFNs of
the general shape (asymmetric FNs), symmetric TrFNs, and fuzzy linguistic vari-
ables of the standard (presented by TrFNs) 7-terms scale (using discrete uniform
7-points distribution). The algorithm of generating asymmetric and symmetric
TrFNs is as follows: three points are generated at each Monte Carlo iteration:
the left of them, say a, and the right, b, are the marginal points of a forming
TrFN, the 3-rd point in the middle, c, is the vertex of an asymmetric triangular
FN (a, c, b); at this iteration, a symmetric TrFN is also formed as (a, (a+b)/2, b).
The number of Monte Carlo iteration is N2 = 5000 with intermediate output of
the results for number of iterations N1 = 1000.
Two groups of ranks {r} are considered: g1 = {r = 1}, i.e., distinctions in

ranking alternatives by FTOPSIS models for rank 1 are considered, and group
g2 = {r = 1, ..., 4}: all ranks r, 1 ≤ r ≤ 4 are taken into account.
Model M0 = FTRY is considered as a base one, and distinctions in ranking
alternatives by M0 and a model Mi , i = 1, ..., M , for each of group gl , l = 1, 2,
are counted. At each iteration (t + 1), number of distinctions is assessed as
DISg (t + 1; Mi ) = DISg (t; Mi ) + 1 if there is at least 1 distinction in ranks
alternatives by models M0 and Mi in the group g, g = g1 , g2 ; otherwise, DISg (t+
1; Mi ) = DISg (t; Mi ).
In this paper, the following FTOPSIS models have been included in Tables 1,
2 and 3: FTTrCI, FTTrIM, FTSCI, FTSIM, FTRCI, FTRIM. Models FTSY
and FTRSY are not included in Tables below as ranks of these models are
identical correspondingly to ranks by FTSIM and FTRIM in accordance with
Proposition 1. The following should also be stressed here. Depending on the
properties of real function of many variables, algorithms for implementing TMs,
and the number of alpha-cuts, the results of using the transformation methods
ETM and RTM can be very close. In this work, no differences in ranks between
the FTEY and FTRY models were found. Therefore, in the indicated tables,
instead of the proper FTEY model, which requires significant resources for its
implementation, the FTRY model is used.
In Tables 1, 2 and 3, relative frequencies of distinctions (%) (100 ∗
DISg (Nk , Mi )/Nk , g = 1, 2, i = 1, .., 4; k = 1, 2) are presented for indicated
models and three types of scenarios for input data: symmetric TrFNs, asymmet-
ric TrFNs, and fuzzy linguistic variables.
Table 1. Distinctions (%) in ranking alternatives by indicated models in comparison

with basic model FTRY for ranks 1/(1−4) and scenario with symmetric TrFNs.
N/Model FTTrCI FTTrIM FTSCI FTSIM FTRCI FTRIM

N1 = 1000 74.5/91.3 72.4/89.7 71.5/90 63.1/84.5 3.4/11 2.1/8.3
N2 = 5000 73.2/91.6 71.4/90.2 70.8/90 62.7/84.4 3.44/12.1 2.5/9.3

with basic model FTRY for ranks 1/(1−4) and scenario with asymmetric TrFNs.

N1 = 1000 73.1/91.6 70.8/90.2 70.6/90.7 58/83.1 10.26.4 2.4/8.7
N2 = 5000 71/91 68.5/89.3 68.6/89.6 57.2/82.6 9/26.1 2.9/8.9

with basic model FTRY for ranks 1/(1−4) and scenario with linguistic TrFNs.

N1 = 1000 10.2/27.8 7.7/21.4 8.1/22.6 4.6/15 0.7/3/3 0.4/2.2
N2 = 5000 11/29 8.5/22.7 9/23.6 5.4/15.3 1.1/2.7 0.4/1.8
The results, presented in these Tables, are discussed in the Conclusions.
4 Conclusions
There are several issues concerning FMCDA/FTOPSIS that should be stressed.
According to Monte Carlo simulation, the distinctions in ranking alterna-
tives within both problematiques of ranking (all ranks, r = 1 − 4, are taken into
account) and choice (rank r = 1 is considered) should be considered as signifi-
cant. According to Tables 1, 2 and 3, the distinctions in ranking alternatives by
FTRY model and approximate models with propagating TrFNs, FTTrCI and
FTTrIM, exceed 70% within the choice problematique and are about 90% for
ranking problematique. For models with implementing Standard Fuzzy Arith-
metic (SFA), FTSCI and FTSIM, distinctions are about 60–70% for choice and
80–90% for ranking problematiques. For models with the “proper” assessing
functions of FNs within fuzzy extension of TOPSIS, FTRCI and FTRIM, dis-
tinctions are between acceptable and hardly acceptable (for asymmetric sce-
nario) for FTRCI, and acceptable for FTRIM. Additional graphical analysis of
the output values for generalized criterion demonstrates big and asymmetrical
overestimation for models, which are based on propagating TrFNs and SFA that
is the main reason of significant distinctions both in Tables 1 and 2.
It should be pointed out, as IM and Y ranking methods are equivalent when
ranking independent FNs, the distinctions between FTRY and FTRIM models
in 2–3% for choice and 8–9.5% for ranking problems reflect an influence of the
dependency of ranked FNs. The smallest distinctions have the place when using
linguistic variables, Table 3, however, they are still unacceptable within ranking
problematique for models FTTrCI, FTTrIM, and FTSCI, and between accept-
able and hardly acceptable for FTSIM. The distinctions for “proper” models
when using linguistic TrFNs may be considered as negligible. Additional evalua-
tions of the described in Subsect. 3.2 approach with 5 criteria and 5 alternatives
demonstrate the increase of distinctions for all scenarios under consideration.
Distinctions in ranking alternatives by different MCDA methods (and, cor-
respondingly, FMCDA models, e.g., FTOPSIS and FMAVT) can be justified
by conceptually different models for decision-making. However, in our case, the
source method is the same, TOPSIS. To the best of our knowledge, only sim-
plified/approximate FTOPSIS models have been developed and implemented
in applications based on standard operations with Tr/TpFNs. Models with
an approach to “proper assessing” functions of FNs (FTRCI, FTRIM, and
FTEY/FTRY) can be considered as a consistent fuzzy extension of the source

FTOPSIS method. However, according to Tables 1, 2, distinctions in ranking
alternatives by these models may also be considered as unacceptable for ranking
problems and, first of all, distinctions between models FTRCI and FTRY.
A natural questions in such conditions is as follows. Taking into account the
level of distinctions in ranking alternatives, which of FTOPSIS models can be
recommended for multicriteria decision analysis of the real case studies in the
fuzzy environment!?
Correspondingly, the problem of distinctions in ranking alternatives by dif-
ferent FMCDA/FTOPSIS models requires further research. Another problem,
which is not considered in this short paper, is exploring the level or measure of
distinctions in ranking alternatives by different models.
References
1. Chen, C.T.: Extensions of the TOPSIS for group decision-making under fuzzy
environment. Fuzzy Sets Syst. 114(1), 1–9 (2000)
2. Chen, S., Hwang, C.: Fuzzy Multiple Attribute Decision Making: Methods and
Applications, vol. 375. Springer, Berlin (1992)
3. Dubois, D., Prade, H.: Operations on fuzzy numbers. Int. J. Syst. Sci. 9(6), 613–626
(1978)
4. Hanss, M.: Applied Fuzzy Arithmetic. Springer, Heidelberg (2005)
5. Hwang, C.L., Yoon, K.: Multiple Attribute Decision Making: Methods and Appli-
cations. Lecture Notes in Economics and Mathematical Systems, vol. 186. Springer,
Berlin (1981)
6. Kahraman, C., Kaya, I., Cebi, S.: A comparative analysis for multiattribute selec-
tion among renewable energy alternatives using fuzzy axiomatic design and fuzzy
analytic hierarchy process. Energy 34(10), 1603–1616 (2009)
7. Kahraman, C., Onar, S.C., Oztaysi, B.: Fuzzy multicriteria decision-making: a
literature review. Int. J. Comput. Intell. Syst. 8(4), 637–666 (2015)
8. Kahraman, C., Tolga, A.: An alternative ranking approach and its usage in multi-
criteria decision-making. Int. J. Comput. Intell. Syst. 2(3), 219–235 (2009)
9. Kaya, T., Kahraman, C.: Multicriteria decision making in energy planning using a
modified fuzzy TOPSIS methodology. Expert Syst. Appl. 38(6), 6577–6585 (2011)
10. Klir, G., Yuan, B.: Fuzzy Sets and Fuzzy Logic: Theory and Applications. Simon
& Schuster, Upper Saddle River (1995)
11. Wang, X., Kerre, E.: Reasonable properties for the ordering of fuzzy quantities (I).
Fuzzy Sets Syst. 118(3), 375–385 (2001)
12. Wang, X., Kerre, E.: Reasonable properties for the ordering of fuzzy quantities
(II). Fuzzy Sets Syst. 118(3), 387–405 (2001)
13. Wang, X., Ruan, D., Kerre, E.: Mathematics of Fuzziness Basic Issues. Springer,
Heidelberg (2009)
14. Yager, R.: On choosing between fuzzy subsets. Kybernetes 9(2), 151–154 (1980)
15. Yager, R.: A procedure for ordering fuzzy subsets of the unit interval. Inf. Sci.
24(2), 143–161 (1981)
16. Yatsalo, B., Korobov, A., Martı́nez, L.: From MCDA to Fuzzy MCDA: violation
of basic axioms and how to fix it. Neural Comput. Appl. (2020, accepted)
17. Yatsalo, B., Korobov, A., Oztaysi, B., Kahraman, C., Martı́nez, L.: A general
approach to fuzzy TOPSIS based on the concept of fuzzy multicriteria acceptability
analysis. J. Intell. Fuzzy Syst. 38, 979–995 (2020)
18. Yatsalo, B., Martı́nez, L.: Fuzzy rank acceptability analysis: a confidence measure
of ranking fuzzy numbers. IEEE Trans. Fuzzy Syst. 26, 3579–3593 (2018)
19. Yuan, Y.: Criteria for evaluating fuzzy ranking methods. Fuzzy Sets Syst. 44,
139–157 (1991)
20. Zadeh, L.: The concept of a linguistic variable and its applications to approximate
reasoning. Inf. Sci., Part I, II, III (8,9), 199–249, 301–357, 43–80 (1975)
Multi-objective Task Scheduling in Cloud
Computing Environment by Hybridized
Bat Algorithm
Timea Bezdan , Miodrag Zivkovic , Eva Tuba , Ivana Strumberger ,

Nebojsa Bacanin(B) , and Milan Tuba
Singidunum University, Danijelova 32, 11000 Belgrade, Serbia

{tbezdan,mzivkovic,istrumberger,nbacanin}@singidunum.ac.rs,
{etuba,tuba}@ieee.org
Abstract. Cloud computing is a relatively new computing technology,

which provides online on-demand computing services to cloud users. Task
scheduling plays a crucial role in the cloud model. An efficient task alloca-
tion method, results with better resource utilization, have an impact on
the quality of service, the overall performance, and user experience. The
task scheduling should be carried out on multiple criteria, which is a diffi-
cult optimization problem and belongs to the class of NP-hard optimiza-
tion problem. As the complexity of the problem increases, the exhaustive
search becomes enormous. Consequently, an optimization technique is
needed that can find the approximate solution in less amount of time.
In this paper, we propose a hybridized bat optimization algorithm for
multi-objective task scheduling. The simulations are performed in the
CloudSim toolkit using standard parallel workloads, and the obtained
results show that the proposed technique gives better results than other
similar methods.
Keywords: Cloud computing · Task scheduling · Bat algorithm ·

Hybridized algorithm · Multi-objective optimization
1 Introduction
Cloud computing is a relatively new computing model, which manages and deliv-
ers software applications and hardware as resources over the Internet. A criti-
cal characteristic of the cloud system that the resources are in virtual form.
Cloud platforms with virtualization technology enable users to lease computing
power in the form of virtual machines (VM). The cloud users submit tasks that
should be allocated to the available virtual machines while increasing resource
utilization and improve task execution. The task scheduling problem represents
an NP-hard optimization problem. Thus optimization methods can be applied
to find the optimal or near-optimal solution in a reasonable time while taking
into account other performance parameters, such as resource utilization, cost,
Multi-objective Task Scheduling in Cloud Computing Environment 719
completion time, and others. The objective of this paper is to propose a swarm-
intelligent-based multi-objective task scheduling in a cloud computing environ-
ment. Some of the famous examples of swarm-intelligence algorithms and their
successful practical implementations are artificial bee colony (ABC) algorithm
[4,7], fireworks algorithm (FWA) for RFID network planning [8], convolutional
neural network design by artificial flora (AF) algorithm [5,6], implementation of
the brainstorm optimization algorithm [11]. Swarm intelligent algorithms have
plenty of applications in the domain of cloud computing, such as [2,3,9,10].
The rest of the paper is organized as follows: we present the problem formu-
lation in Sect. 2; the procedure of the proposed algorithm is described in Sect. 3.
Section 4 presents the simulation results and Sect. 5 concludes the paper and
describes future work.
2 Problem Formulation
This section formulates the task scheduling problem based on the multi-objective
model. The objectives during the process of task scheduling are the makespan
minimization and financial cost reductions in the infrastructure as a service
(IaaS) cloud computing service model. Cloud service providers (CSP) offer dif-
ferent configurations of virtual machine instances that enable the cloud user to
select the CPU, memory, storage space, and the capacity of networking needed
to process the request. The set of instances in cloud platform are described as
follows: I = {I1 , I2 , I3 , ..., In }. A series type offered by CSP, can be defined as a
set: V = {V1 , V2 , V3 , . . . , Vs , . . . , VS }. Each type of series Vs consists of instance
types defined by the set: Vs = {vs1 , vs2 , vs3 , . . . , vsk , . . . , vsK }. The CPU capacity is
expressed in compute units (CU), which is indicated by pks in an instance type
(vsk ). The measure of CU is represented in a million floating-point operations
per second (MFLOPS), and the cks denotes the cost per time unit. The task
submitted by end-users are described by the set t = {t1 , t2 , t3 , . . . , tn }. The aim
is to assign the tasks to virtual instance types by the scheduler algorithm while
optimizing the makespan and execution cost under task execution constraint of
the deadline.
The task execution time is calculated as follows:
si
e(ti , vsk ) = k , (1)
ps
where the execution time of the ith task on the instance type vsk is denoted by
e(ti , vsk ); si denotes the length of the task and pks indicates to the compute unit.
The price model, offered by IaaS provider can be described by the set: P =
{P1 , P2 , P3 , . . . , Pt , . . . , Pr }, and the cost of instance type usage is calculated by
the bill function, which takes three inputs, the pricing model (Pt ), instance series
type (Vs ), and instance type vsk . Hence, the IaaS cloud service is described by
C = (V, Vs , P ).
The makespan is calculated as follows:
makespan = max{Fti : ti ∈ T } (2)

720 T. Bezdan et al.
where Fti denotes the finish time of the task i.

The cost is calculated by the following formula:
n

cost = cks × [F (ti , vsk ) − S(ti , vsk )], (3)
i=1
where F (ti , vsk ) and S(ti , vsk ) represent the finish time and start time, respec-
tively.
The objective is to minimize the objective function that is formulated as
follows:
f = (makespan, cost)T . (4)
3 Proposed Scheduling Algorithm

This section first describes the original bat algorithm [13], followed by the
hybridized version.
3.1 Bat Algorithm

Bat algorithm (BA) [13] is a swarm-intelligence-based algorithm, it was proposed
by Xin-She Yang in 2010. The search process in the algorithm is motivated by
the bats, how they are searching and locating preys by the echolocation ability.
The bats broadcast noise, and based on the reflected sound waves; they can make
a difference between the foods and preys, as well as to estimate the distance.
During the optimization process, the location update of bats (solution) is
performed by the following equation:
xti = xt−1
i + vit , (5)
where the current solution is denoted by xt−1

and the new, updated position at
i
iteration t of the i-th solution is denoted by xti . The velocity is denoted by vit .
The velocity at time step t is calculated as follows:
vit = vit−1 + (xt−1

i − x∗ )fi , (6)
where the current global best position is denoted by x∗ , and fi indicates to the
frequency of i-th bat.
The frequency of the solution is uniformly drawn from the defined range
between the minimum and maximum frequency, and it is evaluated as follows:
fi = fmin + (fmax − fmin )β, (7)
where fmin and fmax are the minimum and maximum frequency, respectively
and β is a random number, β ∈ [0, 1].
The random walk modifies the current fittest solution, and it directs the
exploitation process of the algorithm, which is defined as follows:
xnew = xold + At , (8)

where the average loudness value of all bats are denoted by At , is a scaling
factor with a random value between 0 and 1.
Once the prey is found by a bat, the loudness is updated according to the
following equation:
Ati = αAt−1
i , rit = ri0 [1 − exp(−γt)] (9)
Ati → 0, rit → ri0 , while t → ∞ (10)

where Ati indicates to the loudness of i-th bat, at iteration t, and r is the pulse
emission rate. The values of α and γ are constant.
3.2 Hybridized Bat Algorithm

The exploitation process is potent in the original bat algorithm; however, in the
exploration of the whole search space, it gets trapped in the local optima. In
this work, we adopted the hybridized bat algorithm [12], where the exploration
phase of BA is enhanced by the onlooker bee search from the artificial bee colony
(ABC) algorithm [7], in this way, getting stuck in the local optima is avoided.
The initial population of solutions is generated randomly within the lower
and upper bounds.
xi,j = lbj + rand(ubj − lbj ), (11)
where lbj and ubj indicates to the lower and upper bound, respectively; rand
is a random number from the uniform distribution, rand ∈ [0, 1]; and xi,j is
a solution, where the elements are denoted by j and i refers to the number of
solutions.
After generating the initial population, the fitness is evaluated for each indi-
vidual in the population; the fitness value is calculated as follows:

1
if fxi ≥ 0
Fxi = fxi (12)
1 + |fxi | otherwise,
where the fitness function of i-th individual is denoted by Fxi , and fxi is the
objective function of the i-th individual.
Two different procedures improve search space exploration. The iteration
counter (t) decides whether the bat algorithm’s search procedure will be utilized
or the onlooker bee mechanism. At every even time step, the individuals move
and update the position according to the Eq. (5); on the other hand, if the value
of t is odd, the onlooker procedure is employed by using the following equation:
xti,j = xt−1 t−1 t−1

i,j + rand · (xi,j − xk,j ), (13)
where the new solution at time step t is denoted by xti,j . The j-th element of the
i-th individual is denoted by xt−1 t−1
i,j and xk,j denotes the k-th neighbor solution,
rand ∈ [0, 1].
The exploitation of the promising area is performed by the random walk of
the solution according to the Eq. (8).
Whether the newly generated solution is being accepted or discarded is

decided the selection probability, that is formulated as follows:
Favg
p i = n (14)
i=1 Fxi
where pi denotes the selection probability and F is the fitness value.

The pseudocode of the hybridized BA is presented in the Algorithm 1.
Algorithm 1. Pseudocode of the hybridized bat algorithm

Objective function f (x)
Initialize the population of bats, the values of vi , ri and Ai , define the frequency of
pulse (fi ) at xi , the value of the maximum iteration (M axIter), and set the iteration
counter (t) to 0
while t < M axIter do
for i = 1 to N (each N individuals in the population) do
if t is even then
Calculate the velocity and frequency value by using Eq. (6) and Eq. (7)
Perform the bat search procedure using Eq. (5)
else
Perform the onlooker search procedure by using Eq. (13)
end if
if rand > ri then
Select the fittest solution
Perform the random walk process by using Eq. (8)
end if
if pi < Ai and f (xi ) < f (x∗ ) then
The newly generated solution is accepted
Reduce Ai and increase ri by utilizing Eq. (9)
end if
end for
Find and save the current best solution x∗
end while
Return the best solution
4 Experimental Results
The experiments of the multi-objective task scheduling by hybridized bat opti-

mization algorithm are performed on the CloudSim toolkit; researchers widely
use that for cloud computing environment modeling and simulations. In this
study, the experiments are performed on one price option and one instance type.
The model and the experimental setup is based on [1]. For this experiment,
a single datacenter with two hosts is selected; virtual machine instance with
2048 MB RAM, 1 TB storage capacity, 10 Gbps bandwidth, Linux operating sys-
tem, x86 architecture, Xen VMM; the number of virtual machines set to 20. The
length of a task is in the range between 5000 GB and 50000 GB, the file size is
between 10 GB and 100 GB, and the memory is between 10 GB and 100 GB. The
virtual machine types and configuration, along with the pricing, are presented
in Table 1. The control parameters of the scheduler algorithms are depicted in
Table 2.
Table 1. VM type and configuration
Name vCPU SSD Memory Processing Cost/hour

storage (GB) capacity ($)
(GB) (MFLOPS)
c3.large 2 2 × 16 3.75 8800 0.105
c3.xlarge 4 2 × 40 7.5 17600 0.210
c3.2xlarge 8 2 × 80 15 35200 0.420
c3.4xlarge 16 2 × 160 30 70400 0.840
c3.8xlarge 32 2 × 320 60 140800 1.680
Table 2. Hybridized bat algorithm control parameters
Parameter Notation Value

Population size N 20
Maximum iteration MaxIter 30
Constant minimum loudness Amin 1
Maximum initial loudness A0 100
Minimum frequency fmin 0
Maximum frequency fmax 1
Constant parameter α 0.9
Constant parameter γ 0.9
The performance of the proposed approach is evaluated on the HPC2N, and

NASA Ames iPSC/860 standard parallel workload traces by using different met-
rics, the Hyervolume indicator, makespan, and cost. In [1] the authors used the
same standard parallel workloads, and besides that, they also used synthetic
workloads, which is not used in this work. In the referred paper, for tackling
multi-objective optimization, the concept of Pareto optimality was used. In our
approach, we used the weighted sum technique, where, for both objectives, we set
an equal weight coefficient of 0.5. The task executions are non-preemptive, and
there is no precedence constraint among tasks. The algorithm is tested by per-
forming 30 independent runs. The obtained results are compared to other meta-
heuristic approaches, EMS-C, ECMSMOO, BOGA, and CMSOS. The results of
the comparable techniques are taken from the [1]. Figures 1 and 2 present the
Hypervolume improvement. Based on the graph, it can be concluded that the
BA outperforms EMS-C, ECMSMOO, BOGA, and CMSOS where the task size
is 1000, 2000 and 4000. Moreover, the hybridized BA algorithm outperforms all
the compared algorithms on both NASA and HSPC2N datasets. The relation-
ship between the cost and makespan is presented in Figs. 3 and 4, where the
proposed scheduler continuously results with better cost and makespan against
the other metaheuristic approaches.
Fig. 1. Hypervolume improvement of the Fig. 2. Hypervolume improvement of the

proposed method-NASA proposed method-HSPC2N
Fig. 3. Relationship of the cost and Fig. 4. Relationship of the cost and
makespan-NASA makespan-HSPC2N
5 Conclusion
In the cloud computing environment, the task scheduling problem is an essen-
tial issue because it has a direct influence on cloud performance. To address
this issue, in this paper, we propose a scheduling algorithm based on the meta-
heuristic approach, the hybridized bat algorithm. The problem is modeled as
a multi-objective optimization problem, the objectives during the process of
task scheduling are the makespan minimization and financial cost reductions
in the IaaS cloud computing service model. The experiments are conducted
on the HPC2N and NASA Ames iPSC/860 standard parallel workload traces.
The obtained results are compared to other metaheuristic approaches, EMS-C,
ECMSMOO, BOGA, and CMSOS, and the result analysis shows that the pro-
posed BA-ABC technique produces better makespan and reduces the cost. In
future work, we intend to include more objectives and to do experiments with
other hybridized swarm intelligence algorithms.
Acknowledgment. The paper is supported by the Ministry of Education, Science

and Technological Development of Republic of Serbia, Grant No. III-44006.
References
1. Abdullahi, M., Ngadi, M.A., Dishing, S.I., Abdulhamid, S.M., Ahmad, B.I.: An
efficient symbiotic organisms search algorithm with chaotic optimization strategy
for multi-objective task scheduling problems in cloud computing environment. J.
Netw. Comput. Appl. 133, 60–74 (2019)
2. Bacanin, N., Bezdan, T., Tuba, E., Strumberger, I., Tuba, M., Zivkovic, M.: Task
scheduling in cloud computing environment by grey wolf optimizer. In: 2019 27th
Telecommunications Forum (TELFOR), pp. 1–4. IEEE (2019)
3. Bacanin, N., Tuba, E., Bezdan, T., Strumberger, I., Tuba, M.: Artificial flora opti-
mization algorithm for task scheduling in cloud computing environment. In: Inter-
national Conference on Intelligent Data Engineering and Automated Learning, pp.
437–445. Springer (2019)
4. Bacanin, N., Tuba, M.: Artificial bee colony (ABC) algorithm for constrained opti-
mization improved with genetic operators. Stud. Inform. Control 21(2), 137–146
(2012)
5. Bezdan, T., Tuba, E., Strumberger, I., Bacanin, N., Tuba, M.: Automatically
designing convolutional neural network architecture with artificial flora algorithm.
In: ICT Systems and Sustainability, pp. 371–378. Springer (2020)
6. Cheng, L., Wu, X.H., Wang, Y.: Artificial flora (AF) optimization algorithm. Appl.
Sci. 8, 329 (2018). https://doi.org/10.3390/app8030329
7. Karaboga, D., Akay, B.: A modified artificial bee colony (ABC) algorithm for
constrained optimization problems. Appl. Soft Comput. 11(3), 3021–3031 (2011)
8. Strumberger, I., Tuba, E., Bacanin, N., Beko, M., Tuba, M.: Bare bones fireworks
algorithm for the RFID network planning problem. In: 2018 IEEE Congress on
Evolutionary Computation (CEC), pp. 1–8, July 2018. https://doi.org/10.1109/
CEC.2018.8477990
9. Strumberger, I., Tuba, E., Bacanin, N., Tuba, M.: Dynamic tree growth algorithm
for load scheduling in cloud environments. In: 2019 IEEE Congress on Evolutionary
Computation (CEC), pp. 65–72, June 2019. https://doi.org/10.1109/CEC.2019.
8790014
10. Strumberger, I., Tuba, M., Bacanin, N., Tuba, E.: Cloudlet scheduling by
hybridized monarch butterfly optimization algorithm. J. Sens. Actuator Netw.
8(3), 44 (2019). https://doi.org/10.3390/jsan8030044
11. Tuba, E., Strumberger, I., Bezdan, T., Bacanin, N., Tuba, M.: Classification and
feature selection method for medical datasets by brain storm optimization algo-
rithm and support vector machine. Procedia Comput. Sci. 162, 307–315 (2019).
(7th International Conference on Information Technology and Quantitative Man-
agement (ITQM 2019): Information Technology and Quantitative Management
Based on Artificial Intelligence)
12. Tuba, M., Bacanin, N.: Hybridized bat algorithm for multi-objective radio fre-
quency identification (RFID) network planning. In: 2015 IEEE Congress on Evo-
lutionary Computation (CEC), pp. 499–506, May 2015. https://doi.org/10.1109/
CEC.2015.7256931
13. Yang, X.S.: A New Metaheuristic Bat-Inspired Algorithm, pp. 65–74. Springer,
Heidelberg (2010)
Qualitative Factors in the Fuzzy
Earned Value Method
Dorota Kuchta(&)
Department of Management Systems and Organisation Development,

Wrocław University of Science and Technology,
Wybrzeże Wyspańskiego 27, 50-370 Wrocław, Poland
[email protected]
Abstract. The Earned Value Method is in the first place a tool of project cost
control. In consecutive control moments the attempt is made to reestimate the
total cost of the whole project, individual activities or groups of activities, taking
into account the information available in the control moment. Because the
method is future oriented, the fuzzy approach is here more than appropriate.
There exist several fuzzy versions of the method. None of them, however,
allows to take into account important qualitative factors (inducing emotions,
rumours about changes in the project, negotiations going on etc.). This paper
proposes such a fuzzy version of the method which allows considering such
factors. Additionally, the proposed method focuses on individual activities (and
not on the whole project, like the existing methods). The most critical activities
in terms of cost should be selected for analysis.
Keywords: Earned value Project activity cost Project budgetary risk
1 Introduction
In project management it is important not only to plan the project accordingly, but also
to track its progress in an efficient way. The project tracking system should provide
early warning as to serious problems which can be judged as possible to occur in the
future. Of course, no “secret” knowledge is available for anybody, but the tracking
system should take into account all the knowledge that is somehow accessible.
The Earned Value Method (Fleming and Koppelman 2000) provides the possibility
for an efficient project tracking system. It gives the opportunity to reestimate, in each
project control moment, the project total cost and duration. The method is most
appropriate with respect to cost and we will concentrate here on this project parameter.
The reestimation of the cost of the project or an activity or a group of activities is called
in the Earned Value Method Estimate at Completion.
However, the basic form of determining Estimate at Completion is too limited to be
efficient, especially in case of projects where the uncertainty degree is high (Kuchta
2005). That is why various extensions of the method have been proposed – crisp (Cabri
and Griffiths 2006; Kantor et al. 2016; Koke and Moehler 2019; Kuchta 2019;
Sulaiman et al. 2006; Vanhoucke 2010) and fuzzy (Eshghi et al. 2019; Kuchta 2005;
Qualitative Factors in the Fuzzy Earned Value Method 727
Moradi et al. 2018; Mortaji et al. 2013; Moslemi Naeni and Salehipour 2011; Naeni
et al. 2011; Salari et al. 2014; Salari et al. 2015; Zohoori et al. 2019) ones.
The above mentioned fuzzy extensions of the Earned Value Method are especially
important, as they take into account the fact that the knowledge about the future (here
about the future development of the project) is uncertain and prone to unknown
changes. However, there are still uncovered gaps in the Earned Value Method fuzzy
extensions. They concern the problem of the factors that should be taken into account
in each control moment in the reestimation of the total project cost. These factors
should focus on the future quantity and cost of resources to be still used by the project.
The existing literature applies here an approach which in some cases will not reflect
well enough the project situation. We mean here the cases when all or selected project
activities, because of their changeability and/relatively high cost, should be reestimated
in a direct way in order to grasp the considerable influence they may have on the total
project cost. In today’s world we experience numerous situation of sudden, unexpected
changes. If they concern relatively expensive individual activities, they should be taken
into account in the Earned Value Method. The existing fuzzy extensions of the method
do not make it possible, as they look at projects as the whole and focus on the history
till the control moment.
The paper objective is thus to propose a fuzzification of one formula used in the
Earned Value method for the reestimation of the cost of individual activities. It will be
assumed that the activity, in planning stage of the project, is seen as consisting of
processing a units at the unitary cost c (a generalization to several types of units and
several types of unitary cost would be straightforward). The fuzzification would take
into account several qualitative factors which may influence a or c.
The structure of the paper is as follows: in Sect. 2 the existing approaches, also the
fuzzy ones, to the calculation of the Estimate at Completion are analyzed. In Sect. 3 the
proposed approach is described and Sect. 4 presents some conclusions.
2 State of Art of the Calculation of Estimate at Completion
2.1 Definition and Application of Estimate at Completion

Estimate at Completion (abbreviated as EAC) is one of essential elements of the Earned
Value Method. It is defined for a control moment t, thus we deal here in fact with a
function EAC ðtÞ, with the domain composed of all potential control moments t.
EAC ðtÞ stands for the value of the total cost of a project activity, project activity
groups or a whole project estimated (with the possible and necessary care) in the
control moment t. The most accurate way of calculating EAC ðtÞ is to calculate it
separately of each activity and then find the total for the project or groups of activities.
It is not always possible from the practical point of view (oftentimes there are too many
activities to treat each of them individually), but gives a higher cost control efficiency,
thus here we assume that we treat each, or at least some selected (most expensive)
activities individually. From now on we will thus speak of EAC ðtÞ as referring to an
individual activity.
728 D. Kuchta
EAC ðtÞ – in the control moment t - is compared with BAC ðtÞ (Budget at Com-
pletion), which in its turn stands for the whole budget available for the realization of the
respective activity. The difference BAC ðtÞ EAC ðtÞ is denoted as VAC ðtÞ (Variance at
Completion). If it is negative, it constitutes a warning that if nothing is changed in the
way the respective activity has been performed so far (till the control moment t), the
available activity budget will not be kept. Of course, the ratio VAC ðtÞ=BAC ðtÞ has to be
analysed in order to assess the severity of the situation, but in each case VAC ðtÞ is an
important information. On its basis decisions whether to undertake budget or price
negotiations or activity scope changes will be taken. The smaller the t, the more time is
left for this type of decisions and actions.
2.2 Existing Methods of Calculations of Estimate at Completion

for Individual Project Activities
The basic, obvious formula for EAC ðtÞ is as follows:
EAC ðtÞ ¼ ACWPðtÞ þ ETC ðtÞ ð1Þ
where ACWPðtÞ (Actual Cost of Work Performed) is the actual total cost of the
respective activity incurred since the beginning of its realization till the control moment
t and ETC ðtÞ (Estimate to Complete) is the cost which, according to the best knowledge
in the control moment t, is still to be incurred in order to terminate the activity in
question. ACWPðtÞ is always crisp, ETC ðtÞ and EAC ðtÞ can be fuzzy.
In the control moment t we can also define the following magnitudes:
• ar ðtÞ: the (crisp) actual number of units processed until the control moment t;
• cr ðtÞ: the (crisp) actual average unitary cost of processing the units until moment t;
• aðtÞ: the total number of the units to be processed by the activity, estimated in the
control moment t: this magnitude can be unknown in the control moment and thus
fuzzy;
In the literature on Earned Value Method listed in the introduction it is assumed that
ETC ðtÞ and thus EAC ðtÞ depend on:
• the actual progress of the activity, thus the ratio ar ðtÞ=aðtÞ, which may be a fuzzy
number;
• the initial estimate of the activity cost, which is the product ac, which may be fuzzy
because of the possible fuzziness of a or/and c;
• the quality or tardiness indicators of the work done at the activity till the control
moment t.
In the methods available in the literature there is no possibility to address directly
the ETC ðtÞ, which can be expressed as
ETC ðtÞ ¼ ðaðtÞ ar ðtÞÞcðtÞ ð2Þ
where cðtÞ stands for the unitary cost of processing units of the activity after the control
moment t. The main literature gap regards the impossibility to directly address cðtÞ. In
extreme cases cðtÞ may be completely or largely independent on the past and no
indicators regarding the past would be relevant for the calculation of ETC ðtÞ. We
would have such a situation in case of the change of the human resources (the project
team or the external enterprise which processes the units) as well as in many other
situations of a decisive change in the conditions an activity is performed. Also the work
remaining to be performed, aðtÞ ar ðtÞ, may change due to factors independent of the
situation up to the control moment: decisions of the top management of the organi-
sation on the project scope, customers changing objectives etc. Our proposal concerns
such situations when formula (3) will have to be applied and its components may have
values which would depend on other factors than those known or relating to the
situation before control project t.
3 Proposed Method of Calculating the Estimate

at Completion of an Individual Activity
Let us consider the problem of determining, for a given project activity, the value
ETC ðtÞ, according to formula (3). Generally the idea of the method is to reconsider
both possibly non-crisp parameters in formula (3), cðtÞ and aðtÞ: Let us start with the
parameter cðtÞ. Apart for the “easy” cases, when cðtÞ is equal or directly related to cðtÞ,
in which the formulae from the literature can be applied, there are cases when in the
moment t essential possible changes in the future unitary cost of performing the activity
are to be suspected. This may happen e.g. in the following situations:
a) the supplier of the materials or services or the project team is changed in the
moment t and a new contract is being negotiated;
b) there are serious problems with the present supplier of the materials or services and
thus a risk that a new one will have to be searched for;
c) the present supplier of the materials and services may bankrupt;
d) there are indicators that the present supplier of materials or services is not satisfied
with the present contract and may want to renegotiate it;
e) economic situation of the country is unbalanced and the value of the local currency
may deteriorate;
f) there has been changes in the project team und unexperienced team members have
replaced more experienced ones. The unitary cost of processing one unit in the
activity risks to increase, as less experienced team members, even if less paid, will
need more time to correctly process one unit (they are likely to use reworking in
their work, as they will commit mistakes).
In all such and similar cases cðtÞ will be a fuzzy value, because the above factors
are rather of a qualitative nature: we are facing some risks, the possibility of some
changes, but no crisp information is available yet. And still, for the warning system to
be efficient, it is necessary to take into account all information which may be relevant
for the future actual cost, also the qualitative, imprecise and unconfirmed one.
Let us assume here trapezoidal fuzzy numbers, i.e. cðtÞ will have the form
with the corresponding trapezoidal membership function,
730 D. Kuchta
expressing the possibility that the respective argument will be the actual unitary cost of
processing activity units after the control moment t. cðtÞ will be constructed in such a
way that for each of the situations a) – f) possible values of the future unitary cost will
be generated and then the trapezoidal number will be selected as
an approximation of the results, according to (Ban et al. 2011).
A similar analysis should be applied to the parameter aðtÞ from formula (3), which
stands for the scope of the activity, assessed in moment t. This value may be fuzzy
because of the following phenomena, which can be characterised only qualitatively:
a) there are rumours that the customer has changed his or her mind and wants to have
more or less units processed than it was planned;
b) there are indicators that the quality of the units processed so far is not sufficient and
rework will be necessary, which is equivalent to increasing the scope of the
activity;
c) there are rumours that the project will be broken up for reasons outside the project;
d) there are rumours that a modified version of units will have to processed form the
moment t onwards.
aðtÞ can be modelled as a trapezoidal fuzzy number in similar way as cðtÞ. Then
ETC ðtÞ and thus EAC ðtÞ (formula (2)) will be a trapezoidal fuzzy number, when
appropriate arithmetic operations will be used. This trapezoidal fuzzy number will then
be compared with BAC(t), which will allow to assess the risk of an inacceptable amount
over budget when the activity is completed.
4 Conclusions
A fuzzy version of the Earned Value Method has been proposed which differs from the
existing fuzzy Earned Value Method versions in that in focuses on individual activities,
the crucial ones, which consume the highest portion of project budget (and not on the
whole project, where important details may be lost), and on qualitative factors relating
to the future with respect to the control moment (and not to the past).
In the proposal the future (with respect to the control moment) number of units to
be processed in the activity and the future unitary cost are modelled as trapezoidal
fuzzy numbers. Then the total cost of the activity (estimated in the control moment) is
also a trapezoidal fuzzy number, which can be compared with the budget of the activity
(or rather its value according to the knowledge of the decision maker in the control
moment). This allows to assess the risk of the activity to exceed the budget in an
unacceptable way.
For the comparison any existing methods of comparison between fuzzy numbers
can be selected, the choice being the responsibility of the project manager. Of course,
the trapezoidal fuzzy numbers can be replaced by other forms of fuzzy numbers,
according to the preferences of the decision maker. Also the choice of fuzzy arithmetic
is left to the decision maker.
Real world case studies are needed to validate the proposal. Such case studies are
likely to extend the list of situations (from Sect. 3) which have to be taken into account
while applying the proposal.
Acknowledgments. This research was supported by the National Science Centre (Poland),
under Grant 394311: Selected methods supporting project management, taking into consideration
various stakeholder groups and using type-2 fuzzy numbers.
References
Ban, A., Brândaş, A., Coroianu, L., Negruţiu, C., Nica, O.: Approximations of fuzzy numbers by
trapezoidal fuzzy numbers preserving the ambiguity and value. Comput. Math Appl. 61(5),
1379–1401 (2011)
Cabri, A., Griffiths, M.: Earned value and Agile reporting. In: Proceedings - AGILE Conference,
pp. 17–22 (2006). https://doi.org/10.1109/AGILE.2006.21
Eshghi, A., Mousavi, S.M., Mohagheghi, V.: A new interval type-2 fuzzy approach for analyzing
and monitoring the performance of megaprojects based on earned value analysis (with a case
study). Neural Comput. Appl. 31(9), 5109–5133 (2019)
Fleming, Q., Koppelman, J.: Earned Value Project Management (2000)
Kantor, J., Long, K., Becla, J., Economou, F., Gelman, M., Juric, M., Wu, X.: Agile software
development in an earned value world: a survival guide. In: Proceedings of SPIE - The
International Society for Optical Engineering, vol. 9911 (2016)
Koke, B., Moehler, R.C.: Earned Green Value management for project management: A
systematic review. J. Cleaner Prod. 230, 180–197 (2019)
Kuchta, D.: Fuzzyfication of the earned value method. WSEAS Trans. Syst. 4, 12 (2005)
Kuchta, D.: Combination of the Earned Value method and the Agile approach - a case study of a
production system implementation. In: Burduk, A. (Ed.) Intelligent Systems in Production
Engineering and Maintenance, pp. 87–96 (2019)
Moradi, N., Meysam Mousavi, S., Vahdani, B.: An interval type-2 fuzzy model for project-
earned value analysis under uncertainty. J. Multiple-Valued Logic Soft Comput. 30(1), 79–
103 (2018)
Mortaji, S.T.H., Bagherpour, M., Noori, S.: Fuzzy earned value management using L-R fuzzy
numbers. J. Intell. Fuzzy Syst. 24(2), 323–332 (2013)
Moslemi Naeni, L., Salehipour, A.: Evaluating fuzzy earned value indices and estimates by
applying alpha cuts. Expert Syst. Appl. 38(7), 8193–8198 (2011)
Naeni, L.M., Shadrokh, S., Salehipour, A.: A fuzzy approach for the earned value management.
Int. J. Proj. Manag. 29(6), 764–772 (2011)
Salari, M., Bagherpour, M., Kamyabniya, A.: Fuzzy extended earned value management: a novel
perspective. J. Intell. Fuzzy Syst. 27(3), 1393–1406 (2014)
Salari, M., Bagherpour, M., Reihani, M.H.: A time-cost trade-off model by incorporating fuzzy
earned value management: a statistical based approach. J. Intell. Fuzzy Syst. 28(4), 1909–
1919 (2015)
732 D. Kuchta
Sulaiman, T., Barton, B., Blackburn, T.: AgileEVM - Earned value management in scrum
projects. In: Proceedings - AGILE Conference, pp. 7–16 (2006). https://doi.org/10.1109/
AGILE.2006.15
Vanhoucke, M.: Measuring Time - Improving Project Performance using Earned Value
Management (2010)
Zohoori, B., Verbraeck, A., Bagherpour, M., Khakdaman, M.: Monitoring production time and
cost performance by combining earned value analysis and adaptive fuzzy control. Comput.
Ind. Eng. 127, 805–821 (2019)
The Novel Integrated Model of Plithogenic Sets
and MAIRCA Method for MCDM Problems
Abdullah Özçil(&) , Ayşegül Tuş , Gülin Zeynep Öztaş ,

Esra Aytaç Adalı , and Tayfun Öztaş

[email protected]
Abstract. The most of the decision making processes contain uncertainty due
to the real world problems. In order to deal with this uncertainty, different set
theories have been introduced in the literature. Plithonegic set is the relatively
newest one among them. The concept of plithogeny which was developed by
Smarandache [4] is the generalizations of neutrosophic sets, logic, probability,
and statistics. Two main elements, contradiction and appurtenance degrees, of
plithogenic set help to improve the accuracy of the results under the uncertain
environment. In this study, it is aimed to propose a model based on plithogenic
sets and the MAIRCA (Multi-Attributive Ideal-Real Comparative Analysis)
method. To the best of our knowledge plithonegic sets are integrated for the first
time in this study with the MAIRCA method which aims to minimize the gap
between ideal and empirical values. A case study about green supplier selection
problem in the automotive industry adapted from Gupta et al. [18] is handled to
show the applicability of the proposed model. The plithogenic aggregation
operations are performed to aggregate different decision makers’ opinions on
criteria performance values with respect to each green supplier. Results have
showed that the MAIRCA method could be integrated with plithonegic sets
efficiently.
Keywords: MCDM MAIRCA Plithogenic sets Green supplier selection
1 Introduction
The processes of suppliers’ selections have become more complicated and difficult
because of adding green criteria that take into account declining resources and envi-
ronmental pollution [1]. The green criteria are part of the Green Supply Chain Man-
agement (GSCM). Working with the right green supplier ensures that pollutants are
eliminated at the source and increases the environmental performance of the business
[2]. In the literature, many researchers have drawn attention to green supplier selection
issues and many sets namely fuzzy, type-2, intuitionistic, hesitant and neutrosophic,
have been performed to model vagueness and uncertainty in the structure of the
problem [3].
In this study, a novel green supplier selection decision model based on plithogenic
sets and the MAIRCA (Multi-Attributive Ideal-Real Comparative Analysis) method is
proposed. Plithogenic set theory which is relatively new has been developed by
734 A. Özçil et al.
Smarandache [4]. It can be said that plithogenic sets are the generalization of neu-
trosophic sets and they are utilized to process uncertainty and inconsistency in the
problem [5]. On the other hand MAIRCA method which is one of the MCDM methods
is performed to minimize the total gap between theoretical and real evaluations of green
supplier alternatives considering all criteria [6].
The rest of this paper is organized as follows. In the next section, the concept of
plithogenic sets and some basic operations are introduced. In the same section,
MAIRCA method explained in detail. Then the real case study is illustrated to select
the best green supplier. Finally, the conclusion is given in the last section.
2 Integration of Plithogenic Sets and MAIRCA Method
In this section, we introduce the related concepts that are utilized in the proposed
method and provide the steps of the Plithogenic MAIRCA method. The plithogenic
sets allow more real life inputs to be included in the decision making process than other
sets in the literature. MAIRCA method gives the possibility to be preferred equally to
every alternative. Unlike other MCDM methods, a common value is assigned to
alternatives and alternative scores are determined by distances from this common
value. On top of that, the proposed method is illustrated with a numerical example.
2.1 Plithogenic Sets

Smarandache [4] defined the plithogeny as “the genesis, creation, formation, devel-
opment and evolution of new entities from dynamics and organic fusions of contra-
dictory, neutrals or non-contradictory multiple old entities.” The entity word in the
definition is used for knowledge in various fields [4]. The most important tool in
transferring the concept of plithogeny to the applications is plithogenic set. It is the
generalization of the crisp, fuzzy, intuitionistic fuzzy, and neutrosophic sets [7].
A plithogenic set (P, A, V, d, c) is a set that includes many elements as a number of
attributes A ¼ fa1 ; a2 . . .am g, m 1, attributes’ values V ¼ fv1 ; v2 . . .vn g, n 1, an
appurtenance degree d(x,v) of the element x, with respect to some given criteria and
finally contradiction degree values c(vj,vD) between each jth attribute value (vj) and the
dominant (most important) attribute value vD [8]. Contradiction degrees help to provide
more accurate results under the uncertain environment [5].
Plithogenic sets, which are new in the literature, have been used together with
TOPSIS-CRITIC [5], BWM (Best-Worst Method) [7], and TOPSIS with quality
function deployment methods especially in supply chain management problems. In this
study, plithogenic set is used to combine the opinions of the decision makers. The
triangular neutrosophic numbers are proposed by Abdel-Baset and Mohamed [5] are
used to express the views of decision makers.
2.2 MAIRCA Method

The Multi Attributive Ideal-Real Comparative Analysis (MAIRCA) method has been
firstly proposed by the Center for Logistics Research of Defense University in Belgrade
The Novel Integrated Model of Plithogenic Sets 735
[9]. The main feature of the MAIRCA method is the determination of the gap between
ideal and empirical evaluations of alternatives for each alternative. The alternative with
the lowest total gap value is the closest alternative to the ideal evaluations [6, 9, 10].
In the literature, MAIRCA method has been applied in MCDM problems such as
selection of sites for ammunition depots [6], location for the development of multi-
modal logistics center [11], suppliers [12], catering firm [13], healthcare waste treat-
ment technology [14], the optimum place for overcoming water obstacles [15]. It has
also used for bidder evaluation in the public procurement procedure [10], wastewater
management [16], sustainable manufacturing [17].
The steps of the proposed method are as follows [5, 6, 8]:
Step 1. The initial decision matrix is determined by each decision maker or expert
by using linguistic terms given in Table 1. Then these linguistic evaluations are con-
verted to triangular neutrosophic numbers and the initial decision matrix of the kth
decision maker (Xk) is defined as shown in Eq. 1.
2 3
xij x1n
6 .. 7
X k ¼ 4 ... ..
. . 5 ð1Þ
xm1 xmn
xij is the performance of ith alternative with respect to jth criterion according to kth
decision maker (i = 1,2,…,m; j = 1,2,…,n; k = 1,2,…,K).
Table 1. Linguistic variables for evaluating alternatives under each criterion

Linguistic variables Corresponding triangular neutrosophic numbers
Nothing (N) ((0.1, 0.3, 0.35), 0.1, 0.2, 0.15)
Very Low (VL) ((0.15, 0.25, 0.1), 0.6, 0.2, 0.3)
Low (L) ((0.4, 0.35, 0.5), 0.6, 0.1, 0.2)
Medium (M) ((0.65, 0.6, 0.7), 0.8, 0.1, 0.1)
High (H) ((0.7, 0.65, 0.8), 0.9, 0.2, 0.1)
Very High (VH) ((0.9, 0.85, 0.9), 0.7, 0.2, 0.2)
Absolute (A) ((0.95, 0.9, 0.95), 0.9, 0.1, 0.1)
Step 2. The dominant criterion is determined by the decision makers. The con-
tradiction degree of each criterion is defined with respect to the dominant criterion.
Then all decision makers’ initial matrices are aggregated with plithogenic operators as
follows [8]:

1 1
ðai1 ; ai2 ; ai3 Þ ^p ðbi1 ; bi2 ; bi3 Þ ¼ ai1 ^F bi1 ; ðai2 ^F bi2 Þ þ ðai2 _F bi2 Þ; ai3 _F bi3 ð2Þ
2 2
In Eq. 2 ^F and VF are fuzzy t-norm and t-conorm respectively.

Step 3. The elements of aggregated decision matrix which are in the form of
triangular neutrosophic numbers are converted into crisp values. The result is a decision
matrix (X) whose elements are crisp values. Equation 3 is used for converting trian-
gular neutrosophic numbers into crisp value [5]:
1
sð AÞ ¼ ða1 þ b1 þ c1 Þ ð2 þ a h bÞ ð3Þ
8
Step 4. The preference determination is calculated according to the alternative

selection PAi as shown in Eq. 4. It is assumed that DMs are neutral to alternatives in the
selection process and risk. The DM perceives the alternatives as if any of them has
equal chance, so the preference for choosing one of the possible alternatives;
1 Xm
PAi ¼ ; PA ¼ 1; i ¼ 1; 2; 3; . . .; m ð4Þ
m i¼1 i
Step 5. Theoretical evaluation matrix (Tp) is formed in Eq. 5. tpij is the element of
the theoretical evaluation matrix and calculated by multiplying the preferences
according to the alternatives (PAi) and the criteria weights (wi,i = 1,2,…,n). Criteria
weights are acquired from one of the weighting methods.
ð5Þ
Step 6. The real evaluation matrix (Tr) is formed and shown in Eq. 6.
ð6Þ
The real evaluation matrix elements (trij) are calculated by multiplying the elements
of the initial decision making matrix (X) and real evaluation matrix elements (Tp).
During this process, operation types of the criteria are taken into consideration.
Equation 7 and 8 are performed for the benefit type criteria where bigger values are
desired and the cost type criteria where smaller values are desired respectively:

xij x
i
trij ¼ tpij ð7Þ
xiþ x i

xij xiþ
trij ¼ tpij þ ð8Þ
x
i xi
The xij, x+i [max(x1,x2,…,xm)] and x−j [min(x1,x2,…,xm)] values are the elements of the
initial decision making matrix, the maximum values of the marked criterion by its
alternatives and the minimum values of the marked criterion by its alternatives,
respectively.
Step 7. The total gap matrix (G) is calculated by taking the differences between
theoretical (tpij) and real evaluations (trij) and is shown in Eq. 9.
2 3
g11 g12 g1n
6 g21 g22 g2n 7
G ¼ Tp Tr ¼ 4 ð9Þ
5
gm1 gm2 gmn
The total gap matrix elements (gij) take the value as shown in Eq. 10 in the range
[0, (tpij − trij)].

0 if tpij ¼ trij
gij ¼ ð10Þ
tpij trij if tpij [ trij
It is desirable that there is no gap (tpij = trij) between theoretical (tpij) and real (trij)
evaluation. If there is a gap (tpij ˃ trij), this indicates the distance from the ideal
alternative.
Step 8. The final value of the criteria functions (Qi ) for the alternatives are cal-
culated with Eq. 11. Alternatives are listed in descending order which means that with
the lowest value is the best one.
X
n
Qi ¼ gij ; i ¼ 1; 2; 3; . . .; m ð11Þ
j¼1
2.3 Numerical Example

In this part a case study about green supplier selection problem in the automotive
industry adapted from Gupta et al. [18] is handled to show the applicability of the
proposed model. In this problem nine criteria [resource consumption (C1), staff envi-
ronment training (C2), service level (C3), eco-design (C4), green image (C5), envi-
ronmental management system (C6), price/cost (C7), pollution control (C8), quality
(C9)] are determined for the selection process. Among them C1 and C7 are assumed as
minimization criteria whereas others are assumed as maximization criteria. Then 4
green supplier alternatives (GS1, GS2, GS3, and GS4) are identified that meet these
criteria. Since 3 decision makers (DM1, DM2, and DM3) participate in the decision
making process, this case study can be thought as a group decision making. They are
asked to evaluate green supplier alternatives with respect to each criterion. During the
evaluation the decision makers use the linguistic variables in Table 1. The results are
Table 2. Linguistic assessments of alternatives by three decision makers

Alt. DM Criteria
C1 C2 C3 C4 C5 C6 C7 C8 C9
GS1 DM1 M VH M H A VH L H H
DM2 M H H VH A M M VH H
DM3 VH M H H VH VH L VH VH
GS2 DM1 H H L H VH H M H M
DM2 VH H M H VH M M VH H
DM3 VH H M VH H H VH H M
GS3 DM1 M VH VH VH M VH H M H
DM2 H H M VH VH VH H M VH
DM3 H VH H M H M VH VH H
GS4 DM1 H M VH H M L VH M M
DM2 VH M VH M H M L H VH
DM3 H H H M VH VH M A M
provided in Table 2 [18]. Then each linguistic variable is converted to the neutrosophic
numbers as given in Table 1 [5].
The DMs’ evaluations are aggregated with the plithogenic operator by taking into
account contradiction degrees of each criterion. In this study two different sets of
contradiction degrees are determined to evaluate the effects of the difference in the
contradiction degree values on the results of the application. Firstly, contradiction
degrees are evaluated equally as [1/9 for all criteria]. Secondly, they are determined by
considering the dominant criterion. For this reason C6 is assumed as the dominant
criterion and contradiction degrees are identified as [8/9 6/9 7/9 5/9 3/9 0 4/9 1/9 2/9].
Because of the page constraint, plithogenic aggregation results of green supplier
alternatives which depend on the second contradiction degree set are given in Table 3.
Then these values are converted into crisp values as shown in Eq. 3 and the results
given in Table 3 are the initial data for MAIRCA method. The preference determi-
nation is calculated and the theoretical evaluation matrix is formed. For the theoretical
evaluation matrix (Tp) two different criteria sets are utilized. In the first set it is assumed
that all criteria have equal weights as [wi ¼ 0:111 i ¼ 1; . . .n] whereas the second
weight set is obtained from Gupta et al. [18] which is determined as [0.056 0.087 0.077
0.096 0.117 0.169 0.107 0.155 0.137].
The real evaluation and total gap matrices are calculated. By taking the sum of the
gaps, the final value of the criteria functions for the alternatives in other words the
scores are calculated. As a result of the two different sets for the contradiction degrees
and the criteria weights, four different scenarios are evaluated and the results are shown
in Table 4. The first and the second scenarios give the same ranking results while
different rankings are obtained from the third and fourth scenarios. Considering the
range between scores and the results of the previous study, the most appropriate
ranking is obtained in the proposed the plithogenic MAIRCA method which is the first
scenario shown in Table 4. According to the results of novel integrated model, it is
Table 3. Plithogenic aggregation results of green supplier alternatives

C1 Crisp C2 Crisp … C9
GS1 ((0.484, 0.725, 0.561 ((0.484, 0.675, 0.563 … ((0.535, 0.750, 0.599
0.963), 0.521, 0.950), 0.579, 0.150, 0.975), 0.622, 0.200,
0.150, 0.302) 0.293) 0.302)
GS2 ((0.642, 0.800, 0.591 ((0.430, 0.650, 0.598 … ((0.383, 0.613, 0.551
0.980), 0.516, 0.958), 0.773, 0.200, 0.931), 0.638, 0.125,
0.200, 0.362) 0.227) 0.227)
GS3 ((0.408, 0.638, 0.575 ((0.642, 0.800, 0.591 … ((0.516, 0.700, 0.586
0.951), 0.707, 0.980), 0.516, 0.200, 0.965), 0.642, 0.200,
0.175, 0.227) 0.362) 0.293)
GS4 ((0.516, 0.700, 0.586 ((0.388, 0.625, 0.555 … ((0.459, 0.663, 0.545
0.965), 0.642, 0.941), 0.648, 0.150, 0.943), 0.531, 0.125,
0.200, 0.293) 0.227) 0.293)
Table 4. Ranking results of alternatives with MAIRCA method

Scenarios
First Second Third Fourth
wi: Gupta wi: Equal wi: Equal
wi: Gupta
et al. [18] ci: Unequal ci: Equal
et al.
ci: Unequal ci: Equal
Score Rank Score Rank Score Rank Score Rank
0.098 1 0.107 1 0.107 2 0.094 1
0.126 3 0.122 3 0.104 1 0.111 2
0.111 2 0.111 2 0.110 3 0.114 3
0.166 4 0.160 4 0.157 4 0.167 4
obvious that plithogenic sets have an ability to overcome the vagueness and uncertainty
in evaluations of multiple alternatives.
3 Conclusion
In this study a new model that integrates plithogenic sets and MAIRCA method is
proposed for green supplier selection. The proposed model is applied by using the data
of Gupta et al. [18] obtained from the literature. With this numerical example, it has
been shown that subjectivity and contradictions in the evaluations of different decision
makers can be handled with different scenarios. Thus, it is stated that the uncertainty in
the evaluations of the alternatives can be reduced and the ranking can be made. These
findings were confirmed by comparison of results.
For future studies different sets and MCDM methods can be used in subsequent
studies and the results can be compared. The use of plithogenic sets can be expanded to
different GSCM and other MCDM problems. The effect of scale differences in pli-
thogenic sets on the results can be investigated.
References
1. Rouyendegh, B.D., Yildizbasi, A., Üstünyer, P.: Intuitionistic fuzzy TOPSIS method for
green supplier selection problem. Soft. Comput. 24(3), 2215–2228 (2020)
2. Gao, H., Ju, Y., Gonzalez, E.D.S., Zhang, W.: Green supplier selection in electronics
manufacturing: an approach based on consensus decision making. J. Clean. Prod. 245, 1–17
(2020)
3. Kahraman, C., Onar, S.C., Cebi, S., Oztaysi, B.: Extension of information axiom from
ordinary to intuitionistic fuzzy sets: an application to search algorithm selection. Comput.
Ind. Eng. 105, 348–361 (2017)
4. Smarandache, F.: Plithogeny, plithogenic set, logic, probability, and statistics. Pons,
Belgium (2017)
5. Abdel-Basset, M., Mohamed, R.: A novel plithogenic TOPSIS-CRITIC model for
sustainable supply chain risk management. J. Clean. Prod. 247, 1–15 (2020)
6. Gigović, L., Pamučar, D., Bajić, Z., Milićević, M.: The combination of expert judgment and
GIS-MAIRCA analysis for the selection of sites for ammunition depots. Sustainability 8(4),
372–402 (2016)
7. Abdel-Basset, M., Mohamed, R., Zaied, A.E.N.H., Gamal, A., Smarandache, F.: Solving the
supply chain problem using the best-worst method based on a novel plithogenic model. In:
Optimization Theory Based on Neutrosophic and Plithogenic Sets, pp. 1–19. Academic
Press (2020)
8. Abdel-Basset, M., Mohamed, R., Zaied, A.E.N.H., Smarandache, F.: A hybrid plithogenic
decision-making approach with quality function deployment for selecting supply chain
sustainability metrics. Symmetry 11(7), 1–21 (2019)
9. Pamučar, D., Vasin, L., Lukovac, L.: Selection of railway level crossings for investing in
security equipment using hybrid DEMATEL-MARICA model. In: 16th International
Scientific-Expert Conference on Railway, pp. 89–92. Faculty of Mechanical Engineering-
University of Niš, Serbia (2014)
10. Pamučar, D., Mihajlović, M., Obradović, R., Atanasković, P.: Novel approach to group
multi-criteria decision making based on interval rough numbers: hybrid DEMATEL-ANP-
MAIRCA model. Expert Syst. Appl. 88, 58–80 (2017)
11. Pamucar, D.S., Tarle, S.P., Parezanovic, T.: New hybrid multi-criteria decision-making
DEMATEL-MAIRCA model: sustainable selection of a location for the development of
multimodal logistics centre. Econ. Res. 31(1), 1641–1665 (2018)
12. Chatterjee, K., Pamucar, D., Zavadskas, E.K.: Evaluating the performance of suppliers based
on using the R’AMATEL-MAIRCA method for green supply chain implementation in
electronics industry. J. Clean. Prod. 184, 101–129 (2018)
13. Ulutaş, A.: SWARA ve MAIRCA yöntemleri ile catering firması seçimi. Bus. Manag. Stud.:
Int. J. 7(4), 1467–1479 (2019)
14. Adar, T., Delice, E.K.: New integrated approaches based on MC-HFLTS for healthcare
waste treatment technology selection. J. Enterp. Inf. Manag. 32(4), 688–711 (2019)
15. Pamučar, D.S., Ćirović, G., Božanić, D.: Application of interval valued fuzzy-rough
numbers in multi-criteria decision making: the IVFRN-MAIRCA model. Yugoslav J. Oper.
Res. 29(2), 221–247 (2019)
16. Dursun, M.: A new integrated fuzzy MCDM approach and its application to wastewater
management. Int. J. Intell. Syst. Appl. Eng. 6(1), 19–28 (2018)
17. Boral, S., Howard, I., Chaturvedi, S.K., McKee, K., Naikan, V.N.A.: A novel hybrid multi-
criteria group decision making approach for failure mode and effect analysis: an essential
requirement for sustainable manufacturing. Sustain. Prod. Consum. 21, 14–32 (2020)
18. Gupta, S., Soni, U., Kumar, G.: Green supplier selection using multi-criterion decision
making under fuzzy environment: a case study in automotive industry. Comput. Ind. Eng.
136, 663–680 (2019)
An Alternative Approach for Performance
Evaluation: Plithogenic Sets and DEA
Gülin Zeynep Öztaş(&) , Esra Aytaç Adalı , Ayşegül Tuş ,

Tayfun Öztaş , and Abdullah Özçil

[email protected]
Abstract. Hotels are tourism businesses that serve individuals or institutions

for various purposes. Thanks to today’s economic activities and technological
opportunities, the services offered by the hotels go beyond providing accom-
modation for business or touristic travelers. In today’s technology, these services
are evaluated by customers on the internet. These evaluations may provide
subjective information about the performance of a hotel regarding service
delivery. One of the methods used for performance evaluation is Data Envel-
opment Analysis (DEA) in literature. It is a nonparametric analysis for mea-
suring and evaluating the relative efficiencies of a set of decision making units in
terms of various inputs and outputs. In this study, the customers’ subjective
evaluations are defined as plithogenic sets derived from plithogeny by
Smarandache [1]. They are the generalization of neutrosophic sets and systems
that consider uncertainty and inconsistency of information in the evaluation
process. Subjective evaluations aggregated by plithogenic operators are utilized
as data in DEA to determine the performance of each hotel. The performances
are compared and interpreted.
Keywords: Plithogenic sets DEA Performance evaluations Tourism
1 Introduction
Many travelers trust online reviews and use them as a reference point while making
decisions about their journey namely hotel selection [2]. There are many websites
including almost all details of hotels in a wide range. These websites provide numerical
and linguistic reviews of the available hotels offered by groups. The evaluations
assigned by the groups to the criteria of the different hotel alternatives contain
uncertainty because of the frictions inherent to the evaluations and the subjective
quality of the perception [3]. In this paper the elimination of uncertainty in multi-
criteria decision process is performed by considering plithogenic sets in performance
analysis. Plithogenic set can be defined shortly as the generalization of neutrosophic set
[1]. For the performance analysis, DEA is applied to evaluate the performances of the
hotels located in Çeşme/Turkey under uncertain environment. This study contributes to
the literature by integrating plithogenic sets and DEA for the first time. We assert that
using plithogenic sets in DEA paves the way for the other performance evaluation
methods to be more realistic.
An Alternative Approach for Performance Evaluation 743
The organization of the paper is as follows: Sect. 2 and 3 provide the methodology
used for the alternative approach; Sect. 4 presents the related literature and the gap;
Sect. 5 highlights the proposed method with a hotel performance application. Lastly,
the results are discussed and the future directions are given in Sect. 6.
2 Data Envelopment Analysis
Data Envelopment Analysis (DEA) is a nonparametric frontier method for calculating

relative efficiencies of decision-making units (DMUs) by considering the variables as
multiple inputs and outputs in a linear programming model.
The initial DEA model (CCR) developed by Charnes et al. [4] provides total
efficiency scores of DMUs. However, in 1984 Banker et al. [5] modified the model
(BCC) which calculates technical efficiency scores of each DMU. Both models can be
reconfigured as either input-oriented or output-oriented. In this study, we preferred
input oriented model because in hotel performance analysis, to reach the best overall
score we investigate the efficient amount of inputs.
The input oriented CCR model is given in Eq. 1 [6].
X
s
Max Ek ¼ ur yrk
r¼1
X
m
vi xik ¼ 1 ð1Þ
i¼1
X
s X
m
ur yrj vi xij 0; j ¼ 1; 2; . . .; n
r¼1 i¼1
ur ; vi 0; r ¼ 1; 2; . . .; s; i ¼ 1; 2; . . .; m
Hereby, consider that there are n numbers of DMU to evaluate. The notations given
in Eq. 1 are explained respectively: Ek is the efficiency score of kth DMU; yrj is the rth
output amount of jth DMU; xij is the ith input amount of jth DMU; ur is the weight of rth
output; vi is the weight of ith input; m is the number of inputs; s is the number of
outputs.
3 Plithogenic Sets
Plithogenic sets are the generalizations of neutrosophic sets and the attribute values are
the elements of plithogenic set [7]. In this manner a plithogenic set (P, A, V, d, c) is a
set that includes various elements defined by a number of attributes A = {a1, a2, …,
am}, m 1 and where V = {v1, v2, …, vn}, n 1 are the values for each attribute
[8]. Contradiction [c(v,D)] and appurtenance degrees functions [d(x,v)] are character-
istic elements of a plithogenic set. Especially contradiction degree determines the
744 G. Z. Öztaş et al.
difference between the values of each attribute and the dominant attribute. With this
feature it helps the problem solution in terms of obtaining more realistic results [9].
In this study, the plithogenic aggregation operation is applied to aggregate different
decision making groups’ assessments on input and output performance values with
respect to each hotel alternative. This procedure includes the following steps:
Step 1: The assessments of different decision making groups which are based on
online reviews are used as inputs and outputs in this study. The original data consist of
numerical values that lie between 0–10 and their linguistic equivalents. The numerical
values of the data are used and transformed into triangular neutrosophic numbers by
using linguistic variables which are given in Table 1 to reflect uncertainty in this study.
Table 1. Linguistic scale for aggregation of DMGs

Linguistic variable Triangular neutrosophic scale
Nothing (N) ((0.10, 0.30, 0.35), 0.1, 0.2, 0.15)
Very Low (VL) ((0.15, 0.25, 0.10), 0.6, 0.2, 0.3)
Low (L) ((0.40, 0.35, 0.50), 0.6, 0.1, 0.2)
Medium (M) (0.65, 0.60, 0.70), 0.8, 0.1, 0.1)
High (H) ((0.70, 0.65, 0.80), 0.9, 0.2, 0.1)
Very High (VH) ((0.90, 0.85, 0.90), 0.7, 0.2, 0.2)
Absolute ((0.95, 0.90, 0.95), 0.9, 0.10, 0.10)
Source: [9]
Step 2: The dominant criterion among inputs and outputs is determined and the
contradiction degrees of each input and output are defined with respect to the dominant
criterion. Then all decision making groups’ assessments are aggregated with plitho-
genic operator ð^p Þ as follows [8]:

1 1
ðai1 ; ai2 ; ai3 Þ ^p ðbi1 ; bi2 ; bi3 Þ ¼ ai1 ^F bi1 ; ðai2 ^F bi2 Þ þ ðai2 _F bi2 Þ; ai3 _F bi3
2 2
ð2Þ
^F and VF are fuzzy t-norm and t-conorm respectively in Eq. 2.

Step 3. The aggregated assessments are triangular neutrosophic numbers and they
are transformed crisp values using Eq. 3 [9]:
1
s ð A Þ ¼ ð a 1 þ b 1 þ c 1 Þ ð 2 þ a h bÞ ð3Þ
8
The crisp aggregated assessments are utilized as data in DEA to determine the
performance of each hotel.
4 Literature Review
In real world problems it is almost impossible to obtain data with exact numbers. For this
reason, in the literature many methods like DEA have been integrated with methods that
allow working with uncertain numbers to gain more accurate results [10–16] are the
some of the studies developed fuzzy DEA to obtain more realistic solutions by
considering various approaches. The details of the fuzzy DEA topic can be seen from
[17, 18]. Moreover, fuzzy extensions have been introduced to the literature as well.
Some studies using intuitive fuzzy sets in DEA are as follows: [19–23]. Also, DEA
under pythagorean fuzzy environment which is another extension, has been contributed
to literature as well, see [24, 25]. Lastly, [26, 27] are the most recently published studies
that integrate DEA and neutrosophic sets.
Since the concept of plithogeny is relatively new, scholars have not studied too
much yet. It is possible to make integrations and/or applications in a similar way as in
neutrosophic sets. The most recent studies combine plithogenic sets and regular exact
methods are as follow: Plithogenic VIKOR [7], plithogenic quality function deploy-
ment [8], plithogenic best-worst method [28], plithogenic TOPSIS-CRITIC [9].
However, DEA has not been studied yet under plithogenic environment. For that
reason, we claim that this study will contribute to the literature.
5 Application
In this paper, we evaluate hotels’ performances which are located in Çeşme, Turkey
with plithogenic sets and DEA. Performance indicators of hotels are retrieved from
online reviews of [29]. Due to raw data have missing values, an elimination operation
is performed and consequently three decision making groups (couples - DMG1,
families - DMG2, and friends – DMG3), four inputs (food, room, service and swim-
ming), two outputs (recommendation and price-performance criteria) and sixteen hotels
(DMU1, DMU2, …, DMU16) have remained. Then the input oriented CCR model is
used to evaluate the performance of the hotel.
In Table 1 linguistic variables and their mathematical equivalents are given. These
linguistics variables are used in the aggregation phase of three DMGs. The linguistic
assessments of DMUs are given in Table 2. For the plithogenic aggregation of DMGs’
assessments, swimming is assumed as the dominant criterion and contradiction degrees
are identified as [4/6 2/6 2/6 0 3/6 5/6]. DMGs’ assessments are aggregated by taking
into consideration contradiction degrees. To evaluate these DMUs’ performances with
DEA, single and crisp valued inputs and outputs are needed for each DMU. Therefore,
aggregated assessments in the form of triangular neutrosophic numbers are transformed
into a crisp number with Eq. 3.
In Table 3, between 2nd and 7th columns crisp valued inputs and outputs are given.
The last column of the table shows the efficiency scores of DMUs according to the
DEA model. 8 of 16 hotels are evaluated as efficient. DMU 16 has the lowest efficiency
score so that it is concluded that this DMU is the furthest away from its rivals in terms
of efficiency.
Table 2. Assessments of DMGs’

DMU DMG Food Room Service Swim. Recom. Price. Prf.
DMU1 DMG 1 VH VH VH A VH VH
DMG 2 A VH A A A A
DMG 3 A A A A A A
DMU2 DMG 1 VH A A A VH VH
DMG 2 VH A A A A VH
DMG 3 A A A A A VH
DMU3 DMG 1 VH VH VH A VH H
DMG 2 A VH A A A H
DMG 3 VH VH VH A VH VH
DMU4 DMG 1 VH VH A VH VH VH
DMG 2 VH VH A VH VH VH
DMG 3 VH VH A VH A VH
DMU5 DMG 1 VH VH VH VH A VH
DMG 3 VH VH A VH A VH
DMU6 DMG 1 VH VH A VH A VH
DMG 2 VH VH VH VH VH VH
DMG 3 VH VH VH A VH H
DMU7 DMG 1 VH VH VH A VH VH
DMG 2 VH VH VH VH VH H
DMG 3 VH VH A A VH VH
DMU8 DMG 1 VH VH VH A A VH
DMG 2 VH H VH A VH VH
DMG 3 VH VH A A A VH
DMU9 DMG 1 VH VH VH A VH H
DMG 3 VH VH A VH VH VH
DMU10 DMG 1 H VH VH VH VH H
DMG 2 H VH VH VH VH H
DMU11 DMG 1 VH VH VH VH VH VH
DMG 3 VH A A VH VH VH
DMU12 DMG 1 VH VH VH VH VH VH
DMG 3 A VH VH A H A
DMU13 DMG 1 H A VH A H H
DMG 2 M A VH A H H
DMG 3 M A H A H H
(continued)
DMU DMG Food Room Service Swim. Recom. Price. Prf.
DMU14 DMG 1 H VH VH A VH VH
DMG 2 H VH VH VH H VH
DMU15 DMG 1 H VH VH VH VH H
DMG 2 H VH VH H VH H
DMG 3 A A A VH A VH
DMU16 DMG 1 H VH H A VH H
DMG 2 M H H VH H H
DMG 3 M H H A L L
Table 3. Transformed values and DEA scores (%)

DMU Food Room Service Swim. Recom. Price Prf. Score
1 0.9357 0.7883 0.8462 0.8127 0.8978 0.9594 96.28
2 0.8981 0.9092 0.9092 0.8127 0.8978 0.8607 91.02
3 0.8558 0.6952 0.7475 0.8127 0.8063 0.7995 95.5
4 0.8174 0.6952 0.9092 0.5333 0.8516 0.8607 100
5 0.8174 0.6952 0.7883 0.6059 0.8978 0.8607 100
6 0.8174 0.6952 0.7475 0.617 0.8063 0.8145 95.95
7 0.8174 0.6952 0.7883 0.7052 0.7619 0.8325 94.94
8 0.8174 0.6887 0.7883 0.8127 0.8978 0.8607 100
9 0.8174 0.6952 0.7883 0.6059 0.7619 0.8325 95.14
10 0.7742 0.6952 0.6952 0.617 0.7619 0.7995 98.49
11 0.8174 0.7883 0.7883 0.5333 0.7619 0.8607 100
12 0.8981 0.6952 0.6952 0.617 0.735 0.9269 100
13 0.6784 0.9092 0.6764 0.8127 0.6988 0.742 100
14 0.7742 0.6952 0.6952 0.6059 0.7496 0.8607 100
15 0.8469 0.7883 0.7883 0.5372 0.8516 0.7995 100
16 0.6784 0.6684 0.6615 0.7052 0.5418 0.5915 79.47
6 Conclusion
In this paper a performance evaluation approach that contains plithogenic sets and DEA
for the first time. In this context, performances of 16 hotels that operate in Çeşme,
Turkey are investigated with the help of online reviews. Subjective assessments of
different customer segments gathered with plithogenic set operators thus, it is possible
to integrate contradiction degree and appurtenance degree of these assessments. Then,
aggregated assessments are transformed into crisp numbers and DEA is implemented.
According to the results, it is found out that 8 of 16 hotels are efficient whereas the rest
of them are inefficient.
For the further studies, the proposed approach can be applied in different sectors.
To acquire more accurate information more decision makers can be incorporated into
the analysis. With the theoretical progresses, DEA can be modeled with plithogenic
numbers and different DEA approaches can be compared.
References
1. Smarandache, F.: Plithogeny, Plithogenic Set, Logic, Probability, and Statistics. Pons,
Belgium (2017)
2. Yu, S.M., Wang, J., Wang, J.Q., Li, L.: A multi-criteria decision-making model for hotel
selection with linguistic distribution assessments. Appl. Soft Comput. 67, 741–755 (2018)
3. Di Caprio, D., Santos-Arteaga, F.J.: A novel perception-based DEA method to evaluate
alternatives in uncertain online environments. Comput. Ind. Eng. 131, 327–343 (2019)
4. Charnes, A., Cooper, W.W., Rhodes, E.: Measuring the efficiency of decision making units.
Eur. J. Oper. Res. 2(6), 429–444 (1978)
5. Banker, R.D., Charnes, A., Cooper, W.W.: Some models for estimating technical and scale
inefficiencies in data envelopment analysis. Manag. Sci. 30(9), 1078–1092 (1984)
6. Cooper, W.W., Seiford, L.M., Zhu, J.: Handbook on Data Envelopment Analysis, vol. 164.
7. Abdel-Basset, M., El-Hoseny, M., Gamal, A., Smarandache, F.: A novel model for
evaluation Hospital medical care systems based on plithogenic sets. Artif. Intell. Med. 100,
101710 (2019)
8. Abdel-Basset, M., Mohamed, R., Zaied, A.E.N.H., Smarandache, F.: A hybrid plithogenic
decision-making approach with quality function deployment for selecting supply chain
sustainability metrics. Symmetry 11(7), 903 (2019)
9. Abdel-Basset, M., Mohamed, R.: A novel plithogenic TOPSIS-CRITIC model for
sustainable supply chain risk management. J. Cleaner Prod. 247, 119586 (2020)
10. Sengupta, J.K.: A fuzzy systems approach in data envelopment analysis. Comput. Math
Appl. 24(8–9), 259–266 (1992)
11. Kahraman, C., Tolga, E.: Data envelopment analysis using fuzzy concept. In: Proceedings of
28th IEEE International Symposium on Multiple-Valued Logic, pp. 338–343. IEEE, Japan
(1998)
12. Kao, C., Liu, S.T.: Fuzzy efficiency measures in data envelopment analysis. Fuzzy Sets Syst.
113(3), 427–437 (2000)
13. Guo, P., Tanaka, H.: Fuzzy DEA: a perceptual evaluation method. Fuzzy Sets Syst. 119(1),
149–160 (2001)
14. Saati, S.M., Memariani, A., Jahanshahloo, G.R.: Efficiency analysis and ranking of DMUs
with fuzzy data. Fuzzy Optim. Decis. Making 1(3), 255–267 (2002)
15. Lertworasirikul, S., Fang, S.C., Joines, J.A., Nuttle, H.L.: Fuzzy data envelopment analysis
(DEA): a possibility approach. Fuzzy Sets Syst. 139(2), 379–394 (2003)
16. Wen, M., Li, H.: Fuzzy data envelopment analysis (DEA): model and ranking method.
J. Comput. Appl. Math. 223(2), 872–878 (2009)
17. Hatami-Marbini, A., Emrouznejad, A., Tavana, M.: A taxonomy and review of the fuzzy
data envelopment analysis literature: two decades in the making. Eur. J. Oper. Res. 214(3),
457–472 (2011)
18. Emrouznejad, A., Tavana, M.: Performance Measurement with Fuzzy Data Envelopment
Analysis. Springer, Heidelberg (2014)
19. Hajiagha, S.H.R., Akrami, H., Zavadskas, E.K. Hashemi S.S.: An intuitionistic fuzzy data
envelopment analysis for efficiency evaluation under uncertainty: Case of a finance and
credit institution. E a M: Ekonomie a Manag. 16(1), 128–137 (2013)
20. Puri, J., Yadav, S.P.: Intuitionistic fuzzy data envelopment analysis: an application to the
banking sector in India. Expert Syst. Appl. 42(11), 4982–4998 (2015)
21. Otay, I., Oztaysi, B., Onar, S.C., Kahraman, C.: Multi-expert performance evaluation of
Knowl.-Based Systems 133, 90–106 (2017)
22. Singh, S.: Intuitionistic fuzzy DEA/AR and its application to flexible manufacturing
systems. RAIRO-Oper. Res. 52(1), 241–257 (2018)
23. Arya, A., Yadav, S.P.: Development of intuitionistic fuzzy data envelopment analysis
models and intuitionistic fuzzy input–output targets. Soft. Comput. 23(18), 8975–8993
(2018)
24. Fan, J., Liu, X., Wu, M., Wang, Z.: Green supplier selection with undesirable outputs DEA
under Pythagorean fuzzy environment. J. Intell. Fuzzy Syst. 37(2), 2443–2452 (2019)
25. Wu, M.Q., Zhang, C.H., Liu, X.N., Fan, J.P.: Green supplier selection based on DEA model
in interval-valued pythagorean fuzzy environment. IEEE Access 7, 108001–108013 (2019)
26. Kahraman, C., Otay, İ., Öztayşi, B., Onar, S.Ç.: An integrated AHP & DEA methodology
with neutrosophic sets. In: Kahraman, C., Otay, İ. (eds.) Fuzzy Multi-criteria Decision-
Making Using Neutrosophic Sets. Studies in Fuzziness and Soft Computing, vol. 369,
27. Abdelfattah, W.: Data envelopment analysis with neutrosophic inputs and outputs. Expert
Syst. 36(6), e12453 (2019)
28. Abdel-Basset, M., Mohamed, R., Zaied, A.E.N.H., Gamal, A., Smarandache, F.: Solving the
supply chain problem using the best-worst method based on a novel plithogenic model. In:
Optimization Theory Based on Neutrosophic and Plithogenic Sets, pp. 1–19. Academic
Press (2020)
29. Otelpuan Homepage. https://otelpuan.com/tr/. Accessed 05 Mar 2020
Multi-criteria Decision Making Problem
with Intuitionistic Fuzzy Sets: A Novel Ranking
Method
Hatice Ercan-Tekşen(&)
Eskisehir Osmangazi University, Eskisehir, Turkey

[email protected]
Abstract. Multi-criteria decision making is a state of decision making that we

encounter in all areas of life. Many decisions we make in daily life contain more
than one criterion. Decision making becomes more difficult as the number of
criteria increases. Therefore, decision making in business and/or real life is
complex and difficult. One reason for difficulty in a decision making is the sub-
jective opinion of the decision maker. In addition, not only numerical values are
used but also linguistic data can be taken into account when making a decision.
In order to reduce subjectivity and digitize linguistic expressions, fuzzy logic
approach, which has been used in many fields in recent years, can be used. In the
scope of the study, intuitionistic fuzzy sets, which are expressed as extensions of
fuzzy sets, will be used. Intuitionistic fuzzy sets explain that membership degree
of set is important, but non-membership degree of set is also important as well.
In this study, intuitionistic fuzzy sets will be used for an example decision-
making problem, and the results will be compared by proposing a novel ranking
method. Thus, a novel ranking method for intuitionistic fuzzy sets will be
introduced to the literature.
Keywords: Multi-criteria decision making Intuitionistic fuzzy sets Ranking

method
1 Introduction
Multi-criteria decision making problem is a common problem in daily life. When trying
to make decisions, there is very little case where only one criterion is evaluated. For
this reason, there are many studies in this field.
It is developed in different methods to solve these problems as well as different
multi-criteria decision making problems. Furthermore, recent studies on the expression
of the decision-making with a flexible number of digitizing rather than a single value is
available. At this point, it is seen that fuzzy set theory is applied for multi-criteria
decision making.
The fuzzy logic approach was first introduced by Zadeh (1965) [1]. According to
the classical set understanding, an element either belongs to the set or not. According to
the fuzzy logic, the inclusion of a number in the set can be expressed with certain
membership degrees. This flexibility is particularly beneficial in decision making
problems, because the decision-maker’s opinion is needed in many decision-making
Multi-criteria Decision Making Problem with Intuitionistic Fuzzy Sets 751
problems and it is difficult to express these views with a single value. Therefore, fuzzy
sets have been used in recent years in multi-criteria decision making problems.
Fuzzy sets have evolved over the years with different versions. These are called
extensions of fuzzy sets. One of these is intuitionistic fuzzy sets. Unlike ordinary fuzzy
sets, intuitionistic fuzzy sets also deal with the non-membership degree. In other words,
they have degrees of both membership and non-membership. Intuitionistic fuzzy sets
were first discovered by Atanasssov (1986) and then used in many areas [2]. One of
them is multi-criteria decision making problems.
Boran et al. (2009) solved an intuitionistic fuzzy multi-criteria decision making
problem with the help of TOPSIS method for supplier selection [3]. It is a study that
uses the score function for intuitionistic fuzzy sets used in multi criteria decision
making problems [4]. Li (2013) used intuitionistic fuzzy sets for knowledge map
design [5]. Devi and Yadav (2013) used the ELECTRE method used for multi-criteria
decision making for “plant location selection” and used intuitionistic fuzzy sets for this
[6]. İntuitionistic fuzzy VIKOR method has been used in strategic decision making [7].
Rouyendegh (2014) used the TOPSIS method for the intuitionistic multi-criteria
decision making problem [8]. Liao and Xu (2014) analyzed the analytic hierarchy
process method with intuitionistic fuzzy numbers [9]. Abdullah and Najib (2014) used
the intuitionistic analytical hierarchy process in their work on energy planning [10]. In
another study of Abdullah and Najib (2014), they used a new priority table for intu-
itionistic fuzzy sets for multi-criteria decision making problems [11]. İntuitionistic
fuzzy TOPSIS method was used for ergonomic evaluations [12]. The study used
another AHP technique and supports it with intuitionistic fuzzy sets [13]. The intu-
itionistic fuzzy multi-criteria decision making method has been used for effective
hazardous waste management [14]. Das and Guha (2016) solved the intuitionistic fuzzy
multi-criteria decision making problem by using the sorting method of trapezoidal
fuzzy sets [15].
In this study, a new ranking method has been developed for intuitionistic triangular
numbers. The developed method will be applied and compared with the intuitionistic
fuzzy multi-criteria decision making problem. The following sections of the study are
as follows: In Sect. 2, intuitionistic fuzzy sets will be mentioned. In Sect. 3, ranking
methods for intuitionistic triangular sets will be mentioned, and the ranking method
developed in this section will be explained. In Sect. 4, a numerical example and
ranking methods will be compared. In the last section, a general summary of the study
will be made.
The intuitionistic fuzzy numbers first introduced by Atanassov (1986) are actually
extensions of type-1 fuzzy numbers [2]. Two degree functions, namely degree of
membership and degree of non-membership, are defined for intuitionistic fuzzy
numbers.
~ is an intuitionistic fuzzy set which derived from X universal set and l ~ ð xÞ :
A A
X ! ½0; 1; vA~ ð xÞ : X ! ½0; 1 and 0 lA~ ð xÞ þ vA~ ð xÞ 1 are defined. In this case, the
752 H. Ercan-Tekşen

~ set is defined as A
intuitionistic fuzzy A ~ ¼ x; l ~ ð xÞ; v ~ ð xÞ : x 2 X where l ~ ð xÞ is
A A A
membership function of intuitionistic fuzzy set and vA~ ð xÞ is non-membership function
of intuitionistic fuzzy set.
Sets whose function of membership and non-membership are concave and convex
functions are called intuitionistic triangular fuzzy sets. Some operators are needed for
calculations and are shown as follows for intuitionistic triangular fuzzy sets A ~ 1 and
~
A2 [16]:

~1 þ A
A ~ 2 ¼ ða1 þ a2 ; b1 þ b2 ; c1 þ c2 Þ; a0 þ a0 ; b1 þ b2 ; c0 þ c0 ; minðl1 ; l2 Þ; maxðv1 ; v2 Þ
1 2 1 2
ð1Þ

~1 A
A ~ 2 ¼ ða1 a2 ; b1 b2 ; c1 c2 Þ; a0 a0 ; b1 b2 ; c0 c0 ; minðl1 ; l2 Þ; maxðv1 ; v2 Þ
1 2 1 2
ð2Þ

~1 A
A ~ 2 ¼ ða1 a2 ; b1 b2 ; c1 c2 Þ; a0 a0 ; b1 b2 ; c0 c0 ; minðl1 ; l2 Þ; maxðv1 ; v2 Þ
1 2 1 2
ð3Þ

A ~ 2 ¼ ða1 =a2 ; b1 =b2 ; c1 =c2 Þ; a0 =a0 ; b1 =b2 ; c0 =c ; minðl1 ; l2 Þ; maxðv1 ; v2 Þ
~ 1 =A 0
1 2 1 2
ð4Þ
3 Ranking Methods for Intuitionistic Fuzzy Sets
In this section, a study in the literature will be mentioned, and then the ranking method
developed within the scope of this study will be described.
3.1 Li et al.’s Ranking Method

Li et al. (2012) calculated value and ambiguities for both membership and non-
membership functions [17]. These are shown as follows:
ð a þ 4 b þ c Þ l2
Vl ¼ ð5Þ
6
ða0 þ 4 b þ c0 Þ ð1 vÞ2
Vv ¼ ð6Þ
6
ð c a Þ l2
Al ¼ ð7Þ
6
ðc0 a0 Þ ð1 vÞ2
Av ¼ ð8Þ
6
Equation (5) and (6) are for determining values, Eq. (7) and (8) are for determining
ambiguities. The decision is made after the average values in Eq. (9) and (10) are
obtained.
Vl þ Vv
V ¼ ð9Þ
2
Al þ Av
A ¼ ð10Þ
2
If value of two sets is VA [ VB , then set A is greater than set B. If values of two
sets are VA ¼ VB and ambiguities of two sets are AA [ AB then set A is smaller than
set B.
3.2 A New Ranking Method for Intuitionistic Fuzzy Sets

For intuitionistic fuzzy sets, a method has been developed in which calculation is also
easy. Ranking can be done using a single formula, so decision making will be easy. In
addition, a ranking formula has been developed that takes the availability of negative
data into account. The developed method is given in the equation below.

aþc ð 1 þ l vÞ aþ2 bþc
RIFS ¼ þ Si þ þ Si ð11Þ
2 2 4
The expression shown as Si is the smallest negative value that allows data to rise to
a positive value.
4 Comparison of Ranking Methods with Numerical

Examples
In order to find the solution of multi-criteria decision making problem by using intu-
itionistic triangle fuzzy numbers, the proposed method with the method used in the
literature is compared. A numerical example will be used for this, making it easier to
track and understand the calculations.
Table 1 shows the values given to alternatives and criteria. Then, Table 2 shows
the weighted values for the same data.
~S1 ¼ \ð0:592; 0:774; 0:910Þ; 0:6; 0:4 [
~S2 ¼ \ð0:769; 0:903; 1Þ; 0:4; 0:5 [
~S3 ¼ \ð0:653; 0:849; 0:956Þ; 0:5; 0:2 [

Table 1. Triangular intuitionistic fuzzy sets decision matrix

Alternatives
A1 A2 A3
Criteria C1 <(5.7, 7.7, 9.3); 0.7, 0.2> <(6.5, 8.6, 10); 0.4, 0.5> <(6.5, 8.2, 9.3); 0.8, 0.1>
C2 <(5, 7, 9); 0.6, 0.3> <(8, 9, 10); 0.6, 0.3> <(7, 9, 10); 0.7, 0.2>
C3 <(5.7, 7.7, 9); 0.8, 0.1> <(8.3, 9.7, 10); 0.7, 0.2> <(0, 9, 10); 0.5, 0.2>
C4 <(8.33, 9.67, 10); 0.6, 0.4> <(8, 9, 10); 0.6, 0.3> <(6, 8, 9); 0.6, 0.2>
C5 <(3, 5, 7); 0.6, 0.3> <(7, 9, 10); 0.6, 0.2> <(6.3, 8.3, 9.7); 0.7, 0.2>
Table 2. Weighted normalized triangular intuitionistic fuzzy sets decision matrix for example 1
Alternatives
A1 A2 A3
Criteria C1 <(0.083, 0.111, 0.134); 0.7, 0.2> <(0.091, 0.12, 0.14); 0.4, 0.5> <(0.091, 0.115, 0.13); 0.8, 0.1>
C2 <(0.15, 0.21, 0.27); 0.6, 0.3> <(0.24, 0.27, 0.3); 0.6, 0.3> <(0.21, 0.27, 0.3); 0.7, 0.2>
C3 <(0.068, 0.092, 0.108); 0.8, 0.1> <(0.1, 0.116, 0.12); 0.7, 0.2> <(0.084, 0.108, 0.12); 0.5, 0.2>
C4 <(0.249, 0.291, 0.3); 0.6, 0.4> <(0.24, 0.27, 0.3); 0.6, 0.3> <(0.18, 0.24, 0.27); 0.6, 0.2>
C5 <(0.042, 0.07, 0.098); 0.6, 0.3> <(0.098, 0.126, 0.14); 0.6, 0.2> <(0.088, 0.116, 0.136); 0.7, 0.2>
The results of values and ambiguities are found according to Eqs. (5)–(10).

Vl S~1 ¼ 0:276, Vl ~S2 ¼ 0:144, Vl ~S3 ¼ 0:209, Vv ~ S1 ¼ 0:276, Vv ~S2 ¼ 0:244,

Vv ~S3 ¼ 0:534.

V ~S1 ¼ 0:276, V ~S2 ¼ 0:184 and V ~S3 ¼ 0:372 are calculated. Accordingly,
the alternative rankings are as follows: A3 [ A1 [ A2 .
Ranking values are calculated with the developed method using Eq. (11). Since
there is no negativity in numerical data, Si value is taken as 0.
R1 ¼ 1:030, R2 ¼ 1:092 and R3 ¼ 1:203.
Sorting the alternatives according to these ranking values is as follows:
A3 [ A1 [ A2 .
The proposed method has been shown to give the same results compared to the
existing method.
By giving another numerical example, the two methods will be compared.
Weighted normalized data are given in Table 3.
Table 3. Weighted normalized triangular intuitionistic fuzzy sets decision matrix for example 2
Criteria
C1 C2 C3 C4
Alternatives A1 <(0.07, 0.12, 0.14); <(0.11, 0.17, 0.23); <(0.24, 0.27, 0.3); <(0.05, 0.12, 0.15);
0.75, 0.15> 0.7, 0.1> 0.6, 0.3> 0.65, 0.2>
A2 <(0.15, 0.21, 0.27); <(0.16, 0.19, 0.22); <(0.18, 0.24, 0.26); <(0.16, 0.2, 0.25);
0.7, 0.3> 0.7, 0.1> 0.65, 0.2> 0.65, 0.3>
A3 <(0.12, 0.16, 0.19); <(0.19, 0.25, 0.30); <(0.15, 0.20, 0.23); <(0.19, 0.21, 0.23);
0.7, 0.1> 0.7, 0.2> 0.7, 0.2> 0.75, 0.2>
A4 <(0.08, 0.10, 0.12); <(0.09, 0.12, 0.15); <(0.11, 0.13, 0.16); <(0.11, 0.14, 0.15);
0.5, 0.2> 0.6, 0.3> 0.65, 0.15> 0.7, 0.1>
A5 <(0.24, 0.27, 0.3); <(0.1, 0.11, 0.12); <(0.24, 0.27, 0.3); <(0.09, 0.12, 0.14);
0.6, 0.3> 0.7, 0.2> 0.6, 0.3> 0.6, 0.2>
A6 <(0.09, 0.11, 0.13); <(0.21, 0.27, 0.3); <(0.09, 0.11, 0.12); <(0.18, 0.24, 0.27);
0.8, 0.1> 0.7, 0.2> 0.8, 0.2> 0.7, 0.3>
The score values obtained from Table 3 are calculated as:
~S1 ¼ \ð0:47; 0:68; 0:82Þ; 0:6; 0:3 [
~S2 ¼ \ð0:65; 0:84; 1Þ; 0:65; 0:3 [
~S3 ¼ \ð0:65; 0:85; 0:95Þ; 0:7; 0:2 [
~S4 ¼ \ð0:39; 0:49; 0:58Þ; 0:5; 0:3 [
~S5 ¼ \ð0:67; 0:77; 0:86Þ; 0:6; 0:3 [
~S6 ¼ \ð0:57; 0:73; 0:82Þ; 0:7; 0:3 [
For these data, the ranking values obtained using the method developed by Li et al.

are as follows: Vl ~S1 ¼ 0:241, Vl ~S2 ¼ 0:353, Vl ~ S3 ¼ 0:408, Vl ~ S4 ¼ 0:122,

Vl ~S5 ¼ 0:277, Vl ~S6 ¼ 0:352, Vv ~S1 ¼ 0:327, Vv ~ S2 ¼ 0:409, Vv ~ S3 ¼ 0:533,

Vv ~S4 ¼ 0:239, Vv ~S5 ¼ 0:376, Vv ~S6 ¼ 0:352.

V ~S1 ¼ 0:284, V ~S2 ¼ 0:381, V ~S3 ¼ 0:471, V ~ S4 ¼ 0:181, V ~ S5 ¼ 0:327

and V ~S6 ¼ 0:352 are calculated. Accordingly, the alternative rankings are as follows:
A3 [ A2 [ A6 [ A5 [ A1 [ A4 .
When the new method is applied to the same data, the following sequence values
are obtained: R1 ¼ 0:858, R2 ¼ 1:249, R3 ¼ 1:279, R4 ¼ 0:529, R5 ¼ 1:086 and
R6 ¼ 0:994.
Sorting the alternatives according to these ranking values is as follows: A3 [ A2 [
A5 [ A6 [ A1 [ A4 .
When two methods are compared, a general similarity is seen. Only rankings of A5
and A6 are found different.
5 Conclusion
Multi-criteria decision making is used in many fields. The most important reason for
this is that a single criterion is not paid attention to more than one criterion in a decision
making process. In real life problems, a decision making process is achieved by
considering many features.
In the evaluation of the criteria, a single value is given for the criteria in the
classical approach and the values close to that value are not taken into consideration or
it may be necessary to reduce subjectivity when seeking for expert opinions. In these
cases, fuzzy logic is used.
In this study, a new ranking method developed for intuitionistic fuzzy sets is
mentioned. This new ranking method is adapted to the intuitionistic fuzzy multi-criteria
decision making problem. The results of the proposed method were compared with a
ranking method used in the literature. The rankings of the results are similar.
In future studies, the developed method can be compared with different comparison
methods. In addition, the number of samples can be increased and/or compared to a
larger sample.
References
3. Boran, F.E., Genc, S., Kurt, M., Akay, D.: A multi-criteria intuitionistic fuzzy group
decision making for supplier selection with TOPSIS method. Expert Syst. Appl. 36, 11363–
11368 (2009)
4. Chen, T.Y.: A comparative analysis of score functions for multiple criteria decision making
in intuitionistic fuzzy settings. Inf. Sci. (Ny) 181, 3652–3676 (2011)
5. Li, M.: Extension of axiomatic design principles for multicriteria decision making problems
in intuitionistic fuzzy environment. In: Mathematical Problems in Engineering (2013)
6. Devi, K., Yadav, S.P.: A multicriteria intuitionistic fuzzy group decision making for plant
location selection with ELECTRE method. Int. J. Adv. Manuf. Technol. 66, 1219–1229
(2013)
7. Chatterjee, K., Kar, M.B., Kar, S.: Strategic decisions using intuitionistic fuzzy Vikor
method for information system (IS) outsourcing. In: International Symposium on
Computational and Business Intelligence, pp. 123–126 (2013)
8. Rouyendegh, B.D.: Developing an integrated AHP and intuitionistic fuzzy TOPSIS
methodology. The. Vjesn. Gaz. 21, 1313–1320 (2014)
9. Xu, Z., Liao, H.: Intuitionistic fuzzy analytic hierarchy process. IEEE Trans. Fuzzy Syst. 22,
749–761 (2014)
10. Abdullah, L., Najib, L.: Sustainable energy planning decision using the intuitionistic fuzzy
analytic hierarchy process: choosing energy technology in Malaysia. Int. J. Sustain. Energy
35, 1–18 (2014)
11. Abdullah, L., Najib, L.: A new preference scale of intuitionistic fuzzy analytic hierarchy
process in multi-criteria decision making problems. J. Intell. Fuzzy Syst. 26, 1039–1049
(2014)
12. Maldonado-Macías, A., Alvarado, A., García, J.L., Balderrama, C.O.: Intuitionistic fuzzy
TOPSIS for ergonomic compatibility evaluation of advanced manufacturing technology. Int.
J. Adv. Manuf. Technol. 70, 2283–2292 (2013)
13. Liao, H., Xu, Z.: Consistency of the fused intuitionistic fuzzy preference relation in group
intuitionistic fuzzy analytic hierarchy process. Appl. Soft Comput. 35, 812–826 (2015)
14. Büyüközkan, G., Gocer, F.: An intuitionistic fuzzy MCDM approach for effective hazardous
waste management. In: Intelligence Systems in Environmental Management: Theory and
Applications. Springer, Cham (2016)
15. Das, S., Guha, D.: A centroid-based ranking method of trapezoidal intuitionistic fuzzy
numbers and its application to MCDM problems. Fuzzy Inf. Eng. 8, 41–74 (2016)
16. Büyüközkan, G., Göcer, F.: Application of a new combined intuitionistic fuzzy MCDM
approach based on axiomatic design methodology for the supplier selection problem. Appl.
Soft Comput. 52, 1222–1238 (2017)
17. Li, D.F., Nan, J.X., Zhang, M.J.: A ranking method of triangular intuitionistic fuzzy numbers
and application to decision making. Int. J. Comput. Intell. Syst. 3(5), 522–530 (2010)
Preference-Oriented Fuzzy
TOPSIS Method
Alicja Mieszkowicz-Rolka and Leszek Rolka(B)
The Faculty of Mechanical Engineering and Aeronautics,

Rzeszów University of Technology, 35-959 Rzeszów, Poland
{alicjamr,leszekr}@prz.edu.pl
Abstract. This paper proposes an approach to multi-criteria decision-

making which is based on a modified fuzzy TOPSIS method. We assume
that the process of determining the ranking of alternatives should depend
on subjective preferences for the linguistic values of fuzzy criteria. Pref-
erence degrees, which have to be given by a decision-maker, will be used
for constructing the vectors A+ , and A− , representing the positive, and
the negative ideal solutions, respectively. The vectors A+ and A− will be
described by applying the notion of fuzzy linguistic label. Final ranking
will be obtained by evaluating the distance between particular alter-
natives and the vectors A+ , and A− . The presented approach will be
illustrated by a computational example.
Keywords: Multi-criteria decision-making · TOPSIS · Linguistic

labels
1 Introduction
The problem of finding the best alternative (object), with respect to several cri-
teria, is an important activity in everyday life of human beings. This is often
a non-trivial task, especially when the number of alternatives, and the set of
criteria are becoming large. Moreover, because the criteria are contradictory in
general, even an expert can hardly manage such a complex procedure or get
insight into the solution obtained. In order to solve multi-objective optimiza-
tion problems in a systematic manner, many sophisticated methods have been
proposed, such as SAW, TOPSIS, AHP, ELECTRE, and PROMETHE [7,8,15].
Over recent decades, novel bio-inspired methods, including evolutionary com-
puting, ant colony, and particle swarm optimization, were developed.
Standard approach to multiple-criteria optimization assumes that only
strictly defined criteria are considered. However, since subjectivity and vague-
ness are inevitably present in many real-world situations, they must be taken
into account in the optimization algorithms. To express subjective evaluation
or acceptance, linguistic terms are commonly used, for example, “very high” or
“moderately good”. The fuzzy set theory, founded by Zadeh [17], is a well-known
paradigm suitable to deal with phenomena which are vague in nature. Formal
Preference-Oriented Fuzzy TOPSIS Method 759
expression of linguistic terms by means of fuzzy sets was further developed and
successfully used by many researchers, who introduced several extensions, e.g.,
intuitionistic, type-2, and hesitant fuzzy sets [1,16]. As a consequence of apply-
ing linguistic terms instead of numbers in decision-making [2,5,14], fuzzy rep-
resentation was put into the most popular multi-criteria optimization methods
[3,4,6,9,10,13].
In this paper, we introduce a modification of the standard TOPSIS approach,
with the aim to change its pure arithmetic character, and to make it more human-
oriented. To this end, the notion of fuzzy linguistic label will be applied, which
was introduced recently [11,12]. We assume that an expert does not construct his
or her decision model by a detailed comparison of all objects to each other. The
expert rather tries to find out subsets of representative objects which are similar
with respect to linguistic values of fuzzy attributes. The groups of similar objects
possess the same description in the form of a linguistic label, being a tuple of
dominating linguistic values of attributes.
In contrast to the standard TOPSIS method, we do not determine the vectors
A+ , and A− , which represent the positive and negative ideal solutions, by select-
ing best (or worst) values of criteria in the whole decision table. Instead, we base
on subjective preferences for linguistic values which are specified by a decision-
maker for all fuzzy criteria. The vectors A+ , and A− will be expressed by the
ideal linguistic labels, which correspond to preferences of a decision-maker.
We start by recalling the formalism of the standard TOPSIS method, then we
present notions used in fuzzy information systems, and finally, our fuzzy TOPSIS
method will be introduced and illustrated by a computational example.
2 Crisp and Fuzzy TOPSIS Method in Decision-Making
TOPSIS (Technique for Order Preferences by Similarity to an Ideal Solution)

belongs to the most popular methods of multi-criteria decision-making. It finds
the best alternative which should be as far as possible from the negative ideal
solution and as close as possible to the positive ideal solution [8]. Before we
introduce a modified TOPSIS method, which is suitable for dealing with subjec-
tive fuzzy criteria, we should present the details of the classic approach used in
optimization task with objective crisp criteria.
2.1 TOPSIS Method with Numerical Criteria
Let us recall the steps of the standard TOPSIS method. A set of solution can-
didates is given in the form of a decision matrix D [m × n], where the rows
correspond to m alternatives, and the columns to n criteria. In order to express
importance of every criterion, a vector of weights W = [w(a1 ), w(a2 ), . . . , w(an )]
should be given by a decision-maker.The weights must satisfy the standard
n
requirement of summing up to unity: i=1 w(ai ) = 1.
760 A. Mieszkowicz-Rolka and L. Rolka
In the first step, the decision matrix D is transformed into a normalized

matrix V [m × n] with elements vij , (i = 1, 2, . . . , m , j = 1, 2, . . . , n)

m

vij = wj × dij / d2ij . (1)
i=1
In the second step, the vector A+ = [v1+ , v2+ , . . . , vj+ , . . . , vn+ ], which represents
a positive ideal solution, is determined as follows

+ maxi=1,...,m vij for j ∈ J+ ,
vj = (2)
mini=1,...,m vij for j ∈ J− ,
and the vector A− = [v1− , v2− , . . . , vj− , . . . , vn− ], which represents a negative ideal
solution, is obtained as follows

− mini=1,...,m vij for j ∈ J+ ,
vj = (3)
maxi=1,...,m vij for j ∈ J− .
where J+ denotes the subset of indices of the benefit

criteria, and J− the subset
of indices of the cost criteria, respectively, (J+ J− = ∅).
In the third step, Euclidean distances between each alternative and the vec-
tors A+ , and A− are calculated as

n
di = (vij − vj+ )2 , i = 1, . . . , m,
+
(4)
j=1

n
di = (vij − vj− )2 ,
−
i = 1, . . . , m. (5)
j=1
Finally, a rating factor is determined for every alternative
d−
i
ri = , i = 1, . . . , m. (6)
d+
i + d−
i
The alternative with the highest value of the rating factor ri is the best solution.
2.2 Fuzzy Information Systems
For describing the process of decision-making, we utilize the notion of fuzzy

information system [11] that is defined as a 4-tuple
ISF = U, A, V, f , (7)

where:
U – is a nonempty set of objects (alternatives), called the universe,
A – is a finite set of fuzzy attributes (criteria),
V – is a set of fuzzy (linguistic) values of attributes, V = a∈A Va , Va is the
set of linguistic values of an attribute a ∈ A,
f – is an information function, f : U ×V → [0, 1], f (x, V ) ∈ [0, 1], for all x ∈ U ,
and V ∈ V.
Every fuzzy attribute aj ∈ A, where j = 1, 2, . . . , n, can take linguistic
values from a family denoted by Aj = {Aj 1 , Aj 2 , . . . , Aj nj }. Membership degree
of every element x ∈ U , in all linguistic values of each fuzzy attribute, should
be assigned by an expert. We impose the following requirements to be satisfied
during the assignment of membership:

∃Ajk Ajk ∈ Aj , µAjk (x) ≥ 0.5 , (8)
nj

power (Aj (x)) = µAjk (x) = 1. (9)
k=1
The requirements (8), and (9) constitute a generalization of the characteristic

properties of crisp decision systems.
Analysis of a fuzzy information systems can become easier, if we use a simpli-
fied method of classifying its elements. People neither compare all objects with
each other, nor do they use any kind of similarity relation. Instead, they assess
similarity of a new object to already known groups of representative patterns.
Any such group can be described with the help of a label which is a tuple of lin-
guistic values of fuzzy attributes. To identify linguistic labels used by an expert,
we must find out which linguistic values of attributes are the most important in
the decision-making process.
2.3 TOPSIS Method with Fuzzy Criteria

In order to adapt the TOPSIS method to deal with subjective criteria, we apply
the notion of fuzzy information system presented previously.
We request a decision-maker to provide his or her preferences for the linguistic
values of every fuzzy attribute. Let us denote by pref jk , the preference degree
for the linguistic value Ajk of a fuzzy attribute aj ∈ A, where j = 1, 2, . . . , n,
and k = 1, 2, . . . , nj .
In the case of fuzzy TOPSIS method, decision matrix D [m × n × o] is created
from the decision table that contains membership degrees of alternatives in the
linguistic values of all fuzzy attributes. The third dimension o is the maximal
number of linguistic values used for fuzzy attributes: o = maxj=1,...,n nj .
In the first step, we transform the decision matrix D into the compatibility
matrix V + [m × n]
nj

vij = wj ×
+
pref jk × dijk , (10)
k=1
and the compatibility matrix V − [m × n]

nj

−
vij = wj × (1 − pref jk ) × dijk . (11)
k=1
+
The element vij expresses the accordance of the i-th alternative with preferences
−
for the linguistic values of the j-th fuzzy attribute, whereas the element vij
expresses the accordance with the anti-preferences (negated preferences).
We assume that an (imaginary) positive ideal alternative should have the
maximal membership degrees (equal to 1) in such linguistic values which have the
highest preference for the decision-maker. Conversely, a negative ideal alternative
should have the maximal membership degree in the linguistic values that have
the lowest preference. Both the positive ideal and the negative ideal alternatives
can be represented by their corresponding ideal linguistic labels.
We denote by L+ , and by L− the positive, and the negative ideal linguis-
tic labels, respectively. Every linguistic label can be expressed in the form of
an ordered tuple of selected (most important) linguistic values of all fuzzy
attributes.
In the case of multiple ideal alternatives, we should perform separate calcu-
lations for all possible pairs of the positive and negative ideal alternatives.
In the second step, vector A+ = [v1+ , v2+ , . . . , vj+ , . . . , vn+ ], which represents
the positive ideal solution, is determined as
vj+ = wj × pref j , (12)
and vector A = −
[v1− , v2− , . . . , vj− , . . . , vn− ], which represents the negative ideal
solution, is determined as
vj− = wj × (1 − pref j ), (13)
where pref j denotes the maximal preference degree specified for the linguistic
values of a selected attribute aj ∈ A.
Without loss of generality, we assume that there is only one most preferred
linguistic value of every fuzzy attribute. The vector A+ represents the positive
ideal solution obtained by determining the accordance of the positive ideal alter-
native with preferences for the linguistic values. The vector A− represents the
negative ideal solution obtained by determining the accordance of the negative
ideal alternative with anti-preferences for the linguistic values.
In the third step, Euclidean distance between each alternative and the vectors
A+ , and A− are calculated as

n +
d+i = (vij − vj+ )2 , i = 1, . . . , m, (14)
j=1

n −
di = (vij
−
− vj− )2 , i = 1, . . . , m. (15)
j=1
The ranking of alternatives is determined by the function (6).

3 Example
Let us consider a fuzzy information system with eight alternatives, and three
attributes (criteria): a1 , a2 , and a3 . The attribute a1 has five linguistic values,
whereas the attributes a2 , and a3 have three linguistic values.
Table 1. Decision table.
a1 a2 a3
A11 A12 A13 A14 A15 A21 A22 A23 A31 A32 A33
x1 0.00 0.70 0.30 0.00 0.00 0.00 0.80 0.20 0.90 0.10 0.00
x2 0.00 0.00 0.70 0.30 0.00 0.60 0.40 0.00 0.80 0.20 0.00
x3 0.00 0.50 0.50 0.00 0.00 0.00 0.80 0.20 1.00 0.00 0.00
x4 0.25 0.75 0.00 0.00 0.00 0.00 0.90 0.10 0.75 0.25 0.00
x5 0.00 0.00 1.00 0.00 0.00 0.00 0.40 0.60 0.00 0.70 0.30
x6 0.00 0.00 0.40 0.60 0.00 0.30 0.70 0.00 0.00 0.80 0.20
x7 0.00 0.00 0.00 0.75 0.25 0.00 1.00 0.00 0.80 0.20 0.00
x8 0.40 0.60 0.00 0.00 0.00 0.25 0.75 0.00 0.55 0.45 0.00
The information system is given in the form of a decision table which contains
membership degrees in the linguistic values of all attributes (Table 1).
Preferences for the linguistic values of all attributes are given in Table 2. We
see, for example, that the linguistic value A13 is the most important one for the
decision-maker, in the case of the attribute a1 .
We assume all attributes to be are equally important: W = [0.33, 0.33, 0.34].
Table 2. The degrees of preference for the linguistic values of attributes.
a1 a2 a3
A11 A12 A13 A14 A15 A21 A22 A23 A31 A32 A33
0.25 0.75 1.00 0.75 0.25 0.50 1.00 0.25 1.00 0.50 0.00
As the positive ideal alternative should reflect the preferences of the decision-
maker, it will be represented by the positive linguistic label L+
1 = (A13 , A22 , A31 ).
There are two negative ideal alternatives, having the corresponding negative
linguistic labels L− −
1 = (A11 , A23 , A33 ), and L2 = (A15 , A23 , A33 ), respectively.
−
In the following, all calculations will be performed for the pair (L+ 1 , L1 ) of the
ideal linguistic labels.
Next, we determine the compatibility matrices V + and V − , according to
formulae (10), and (11), which are presented in Table 3.
Table 3. Compatibility matrices.
V+ V−
a1 a2 a3 a1 a2 a3
x1 0.2722 0.2805 0.3230 0.0577 0.0495 0.0170
x2 0.3052 0.2310 0.3060 0.0248 0.0990 0.0340
x3 0.2888 0.2805 0.3400 0.0413 0.0495 0.0000
x4 0.2063 0.3053 0.2975 0.1237 0.0248 0.0425
x5 0.3300 0.1815 0.1292 0.0000 0.1485 0.2108
x6 0.2805 0.2805 0.1428 0.0495 0.0495 0.1972
x7 0.2063 0.3300 0.3060 0.1237 0.0000 0.0340
x8 0.1815 0.2888 0.2635 0.1485 0.0413 0.0765
Table 4. Ranking of alternatives.
ri Position ri Position
x1 0.8364 2 x5 0.5233 8
x2 0.7801 3 x6 0.5894 7
x3 0.8667 1 x7 0.7515 4
x4 0.7335 5 x8 0.6532 6
With the help of formulae (12), we get the vector of positive ideal solutions
A+1 = [0.3300, 0.3300, 0.3400].
There are two vectors od negative ideal solutions A− 1 = [0.2475, 0.2475,
0.3060], and A−2 = [0.2475, 0.2475, 0.3060].
Finally, we compute the distances (14), and (15) between each alternative
and the vectors A+ , and A− , respectively. The ranking of alternatives, calculated
using the function (6), is given in Table 4.
Further calculations were made for different vectors of weights of fuzzy cri-
teria. A change of weights by 20% has only a slight influence on the results.
An interesting case can be observed for the vector W = [0.20, 0.45, 0.35], for
which we get the following ranking order: x3 , x7 , x1 , x4 , x8 , x2 , x6 , x5 . We
can conclude that the alternative x3 is always the best in all considered cases,
and this in accordance with our expectations, as this alternative has a very high
compatibility with the linguistic values A31 and A22 .
4 Conclusions
Two kinds of subjectivity are considered in the presented approach. Firstly, we

apply fuzzy criteria instead of numerical ones. Membership degree of alterna-
tives in the linguistic values of fuzzy criteria is a subjective choice of an expert.
Secondly, subjective preferences of particular linguistic values of criteria have to
be specified by a decision-maker. Vectors of positive, and negative ideal solutions

are constructed with respect to preferences of the decision-maker and by using
the concept of ideal linguistic label. In future research, the method should be
developed to take into account both subjective and objective criteria.
References
2. Cabrerizo, F., Pedrycz, W., Perez, I., Alonso, S., Herrera-Viedma, E.: Group deci-
sion making in linguistic contexts: an information granulation approach. Procedia
Comput. Sci. 91, 715–724 (2016). http://www.sciencedirect.com/science/article/
pii/S1877050916312455
3. Chen, C.T.: Extensions of the TOPSIS for group decision making under fuzzy
environment. Fuzzy Sets Syst. 114, 1–9 (2000)
4. Chou, S.Y., Chang, Y.H., Shen, C.Y.: A fuzzy simple additive weighting sys-
tem under group decision-making for facility location selection with objec-
tive/subjective attributes. Eur. J. Oper. Res. 189, 132–145 (2008)
5. Chuu, S.J.: Interactive group decision-making using a fuzzy linguistic approach for
evaluating the flexibility in a supply chain. Eur. J. Oper. Res. 213(1), 279–289
(2011)
6. Deni, W., Sudana, O., Sasmita, A.: Analysis and implementation fuzzy multi-
attribute decision making SAW method for selection of high achieving students in
faculty level. Int. J. Comput. Sci. 10(1), 674–680 (2013)
7. Greco, S., Ehrgott, M., Figueira, J.R.: Multiple Criteria Decision Analysis: State
of the Art Surveys. Springer, New York (2016)
8. Hwang, C.L., Lai, Y.J., Liu, T.Y.: A new approach for multiple objective decision
making. Comput. Oper. Res. 20(8), 889–899 (1993)
9. Kacprzyk, J., Zadrożny, S., Fedrizzi, M., Nurmi, H.: On group decision making,
consensus reaching, voting and voting paradoxes under fuzzy preferences and a
fuzzy majority: a survey and a granulation perspective. In: Pedrycz, W., Skowron,
A., Kreinovich, V. (eds.) Handbook of Granular Computing, pp. 906–929. Wiley,
Chichester (2008)
10. Kahraman, C., Onar, S.C., Oztaysi, B.: Fuzzy multicriteria decision-making: a
literature review. Int. J. Comput. Intell. Syst. 8(4), 637–666 (2015)
11. Mieszkowicz-Rolka, A., Rolka, L.: A novel approach to fuzzy rough set-based anal-
ysis of information systems. In: Wilimowska, Z., et al. (eds.) Information Systems
Architecture and Technology. Knowledge Based Approach to the Design, Control
ad Decision Support. Advances in Intelligent Systems and Computing, vol. 432,
12. Mieszkowicz-Rolka, A., Rolka, L.: Fuzzy linguistic labels in multi-expert decision
making. In: Martı́n-Vide, C., et al. (eds.) Theory and Practice of Natural Comput-
ing. Lecture Notes in Computer Science, vol. 10687, pp. 126–136. Springer, Cham
(2017)
13. Pedrycz, W., Ekel, P., Parreiras, R.: Fuzzy Multicriteria Decision-Making: Models,
Methods and Applications. Wiley, Chichester (2011)
14. Skorupski, J.: Interactive group decision-making using a fuzzy linguistic approach
for evaluating the flexibility in a supply chain. Expert Syst. Appl. 41, 7406–7414
(2014)
15. Vaidya, O.S., Kumar, S.: Analytic hierarchy process: an overview of applications.
Eur. J. Oper. Res. 169(1), 1–29 (2006)
16. Xu, Z.: Hesitant Fuzzy Sets Theory. Springer, Cham (2014)
Intelligent Statistical Analyses
A New Approach to Analyze Perceived Design
Quality of Mobile Phone Using Fuzzy
Hierarchical Conjoint Analysis
Yusra Erdem1, Selcuk Cebi1(&), and Esra Ilbahar1,2

1
Besiktas, Istanbul, Turkey
[email protected]
2
Besiktas, Istanbul, Turkey
Abstract. The smartphones entered the market in the early 2000s and the
market share of these products has been drastically increased with the devel-
opment of technologies in telecommunication and the internet. Nowadays, in the
market, there are various phone models including different features produced by
companies in order to reach each customer. However, it is virtual to understand
the expectations of the customer during the development phase to introduce a
product that satisfies the customer. This process is complex, and it is also hard to
achieve this target because of uncertainty in understanding the perceived quality
of customers. Therefore, in the scope of this study, a new approach has been
proposed including the short cut pairwise comparison, conjoint analysis, and
Fuzzy Set Theory. The short cut pairwise comparison is a modified version of
the analytic hierarchical process is used to measure the importance degrees of
mobile phone design parameters. Then, Conjoint analysis is utilized to deter-
mine customer values on the level of the design parameters. In this study, Fuzzy
set theory has been used to analyze of linguistic judgments of customers and to
understand perceived design quality.
Keywords: Mobile phone design parameters Short cut pairwise comparison

method Conjoint analysis Fuzzy set theory
1 Introduction
The smartphones entered the market in the early 2000s and the market share of these
products has been drastically increased with the development of technologies in
telecommunication and the internet. Therefore, the number of mobile applications has
been increasing steadily in the growing smartphone market. End-users who only used
computers at home to keep track of their social media accounts and to surf on the
internet has preferred smartphones instead of computers since their practical usage.
Therefore, it was the first time the number of mobile phone users exceeded the number
of desktop computer users in 2014. For this reason, businesses give priority to
developing mobile applications and they allocate a big budget for this purpose to bring
their products to the forefront and reach their customers in the market. This also affects
770 Y. Erdem et al.
the production of mobile phones. Therefore, there are various phone models including
different features produced by companies in order to reach each customer. However, it
is virtual to understand the expectations of the customer during the development phase
to introduce a product that satisfies the customer. This process is complex, and it is also
hard to achieve this target because of uncertainty in understanding the perceived quality
of customers. Therefore, in the scope of this study, a new approach has been firstly
proposed including the short cut pairwise comparison, conjoint analysis, and Fuzzy Set
Theory. The short cut pairwise comparison is a modified version of the analytic
hierarchical process is used to measure the importance degrees of mobile phone design
parameters. Although the analytical hierarchical process (AHP) uses a reciprocal
matrix, the required computation is enormous. In another word, the increasing number
of the hierarchy level causes an exponential increase in the number of pairwise com-
parisons, and this means that more time is required to complete the survey. Further-
more, the consistency of the matrix is calculated after pairwise comparisons are ended
by participants. This process causes loss of time and money since the inconsistent
survey cannot be used for the research. Therefore, in order to shorten the pairwise
comparison process and provide a dynamic structure for the calculation of consistency,
a new algorithm has been developed in this study. Then, Conjoint analysis is utilized to
determine customer values on the level of the design parameters. In the scope of this
study, Fuzzy set theory has been used to analyze of linguistic judgments of customers
and to understand perceived design quality.
The rest of this paper is organized as follows; Sect. 2 presents the literature on
smartphone design parameters. The structure of the proposed approach is given in
Sect. 3. An illustrative example is given in Sect. 4. Finally, concluding remarks are
2 Literature Review on Smartphone Design
In large-scale literature research, it has been seen that there are many studies on
smartphone design. Some of these studies, which give an idea about smartphone design
based on customer preferences, were made by using multi-criteria decision-making
methods. A modified AHP approach has been applied to understand customer pref-
erences by Ho et al. [1]. They have obtained results by examining the attribute pref-
erences and price preferences separately. Tan and Adli [2] proposed a methodology to
select a smartphone. They used Data Envelopment Analysis (DEA) by considering 19
attributes and 25 smartphones. AHP has been applied to enhance DEA results. Isiklar
and Buyukozkan [3] used two multi-criteria decision making methods to evaluate the
mobile phone selections from users’ preferences order. AHP method is applied to
obtain criteria weights and importance order, TOPSIS is applied to rank the alterna-
tives. Also, the Fuzzy ELECTRE I method used to analyze data about smartphone
brand choices by Belbag et al. [4]. They have been carried out a pre-examination to
A New Approach to Analyze Perceived Design Quality 771
determine the most favorite 5 brands and smartphone choosing criteria. A two-phased
multi-criteria decision making approach has been used to make the best smartphone
selection by Yildiz and Ergul [5]. In the first phase, 28 smartphone alternatives were
ranked according to 3 criteria and 17 sub-criteria by using analytic network process
(ANP). In the second phase, a model which includes the top four alternatives of ANP
analysis was created and the generalized Choquet integral (GCI) method were used
finally. Salhieh and Al-Harris’ article includes an integrated approach using DEA and
conjoint analysis (CA) [6]. Also, they have made an application about smartphone
design by determining the criteria which are weight, talk time, screen size, and price.
Conjoint analysis has been applied in the first step. In addition, their methodology was
also concern with the development burden. Then DEA was applied according to the
CA results. Another study including the Technology Acceptance Model (TEM) and the
Structural Equation Model (SEM) was applied to empirical analysis by Kang et al. [7].
Cebi and Ilbahar [11] proposed Fuzzy Kano Model to develop smart phone design.
As described above, there are many studies on smartphone design in the literature.
This study, different from other approaches, aims to achieve reliable results by short-
ening the data collection process and detecting inconsistent data before the analysis.
3 Proposed Approach: Fuzzy Hierarchical Conjoint Analysis
In this study, a new approach has been proposed which includes a short cut fuzzy
pairwise comparison (SCFPC), conjoint analysis and fuzzy logic.
3.1 A Short Cut Fuzzy Pairwise Comparison Method

Creating a pairwise comparison matrix takes a long time in AHP analysis applications.
After the criteria weights are determined, AHP analysis can be applied and the con-
sistency rate can be calculated. This means that if the study is inconsistent, a long-term
effort is wasted. In addition, since the questions, which are asked to expert, are always
in the form of pairwise comparisons, there is a high probability that questions will be
answered incorrectly in cases where the number of criteria is too much.
Within the scope of this study, the mentioned problems were tried to be eliminated
and an algorithm was developed by making some additions to AHP analysis. A pro-
gram has been created with the Python 3.6.0 programming language to provide easy
usage in applications of the proposed SCFPC method. The pseudocodes of the pro-
posed algorithm is as follows:
772 Y. Erdem et al.
Start
Define the pairwise comparison matrix and variables: Data
= [n][n], i=0, b=0, count=0, int loop
Define function: MatrixCreation
Print “Enter the # of criteria:”, n
For i in range n
Print “Enter the first row of the comparison matrix:”,
data[0][i]
Create the pairwise comparison matrix from the first row:
For i in range n
For b in range n
Data[i][b] = Data[0][b]/Data[0][i]
End of the function MatrixCreation
Call function MatrixCreation
Convert the element of matrix into the triangular fuzzy
numbers
Define function: Verification
loop = n/4. If loop is a float, it will be rounded to the
next number
While count less than loop:
If n greater than or equals to 5
Create two random numbers between 1 and (n-3), m and j1
Else
M equals to 0, j1 equals to 1
Read (m,j1) index from data matrix, q1
If q1 is greater than 9, then q1up is equal to 9, q1down
is equal to 9
Else if q1 is less than or equal to 9, q1up is equal to
nearest and greater integer number. q1down is equal to
nearest and smaller integer number
Print “Enter (m,j1) index of the matrix”, q2
If q2 is between q1up and q1down, count = count+1 and
print “Correct answer!”
Else, count = count+1
Read (j1, m) index from data matrix, s3
Read (m, j1+1) index from data matrix, s1
Print “Enter (j1, m) index of the matrix:”, s4
Print “Enter (m, j1+1) index of the matrix:”, s2
Calculate the up and down values of s1 and s3 as step 13
and 14
If s2 is between s1up and s1down values and s4 is be-
tween s3up and s3down values
Print “"Value of the index (m,j1) is correct!"
Else if s2 or s4 is out of range
Print “"The program will be terminated due to the data
are inconsistent. Check the data and try again!"
End of the while loop
Print “Comparison matrix has been confirmed. You can con-
tinue with FAHP analysis.”
End of the function Verification
Call function Verification
At first, the researcher collects all criteria for the handled problem and then con-
structs the hierarchy of the criteria. Then, this hierarchy is transformed into a dynamic
survey by the developed algorithm and is posted to participants. While the preferences
of the participants are collected, their consistency is tested simultaneously. If the
consistency of the preferences is satisfied, then the pairwise comparison matrix has
been created and verified. The fundamental scale, definitions, and triangular fuzzy
numbers used in the matrix are given in Fig. 1. Then, the importance degrees of the
parameters are obtained by using Buckley’s fuzzy AHP approach. The steps of the
approach is as follows [8, 9];
Fig. 1. Fuzzy importance scale
2 3
1 e
a 12 e
a 13 e
a 1n
6e e a 2n 7
e
6 a 21 1 a 23 7
6 a 31 a 3n 7
A ¼ 6e
e e
a 32 1 e 7 ð1Þ
6 .. .. .. .. .. 7
4. . . . . 5
e
a n1 e
a n2 e
a n3 1
where Ae and aij are comparison matrix and the triangular fuzzy numbers (TFNs). The
Eq. 2 is used to calculate the fuzzy geometric mean. The fuzzy weight of each criterion
is obtained by using the Eq. 2 and Eq. 3.
p ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
rei ¼ n
a i1 e
e a i2 e a i3 . . . e a in ð2Þ
wi ¼ rei ðer 1 er 2 er 3 . . . er n Þ1

f ð3Þ
where fwi is the fuzzy weight of ith criterion. Then the weight of each criterion and the
order of importance have been obtained by defuzzification of fuzzy weights.
In the conventional AHP method, an advanced analysis has been conducted to
prove whether the pairwise comparison matrix provides consistency. Equation 4–6 is
used for consistency analysis [10].
774 Y. Erdem et al.
e w ¼ kmax w
A ð4Þ
ðkmax nÞ
CI ¼ ð5Þ
ðn 1Þ
CI
CR ¼ ð6Þ
RI
where CI, RI, and CR are consistency index, random consistency index given in
Table 1, and consistency ratio. The CR of a pairwise matrix should be less than 0.1
Table 1. Random index

n 1 2 3 4 5 6 7 8 9 10
RI 0.00 0.00 0.58 0.90 1.12 1.24 1.32 1.41 1.45 1.49
3.2 Conjoint Analysis

Conjoint analysis, a multivariate data analysis technique, is adopted to identify prior-
ities and to determine preferences of consumers by providing them with various
hypothetical profiles of products or services. These hypothetical profiles are generated
by combining different levels of each product or service attribute. Then, these hypo-
thetical profiles are presented to the consumers to rank them, rate them or choose one of
them. More significant attributes of products or services can be identified based on the
choices or the rank provided by the potential consumers [12]. As a result of this
analysis, more crucial attributes for consumers can be identified, and consequently this
information can be considered in product design process to stay competitive in the
market [13].
4 Illustrative Example
In this section, an illustrative example of smartphone design has been given. The
design characteristics of the smartphone are given in Fig. 2. The hierarchical structure
consists of two levels, 6 criteria, and 45 sub-criteria. The analysis is started by
determining the importance of the criteria and creating pairwise comparison matrix
with the algorithm given in Sect. 3.1.
Fig. 2. Hierarchical structure of the smartphone design characteristics.
Table 2. Sub-criteria and final weights of the smartphone design characteristics.

Criteria Sub-criteria Weights Final kmax , CI, CR
weights RI
Operational Processor 0.497 0.035 kmax = 7 CR = 0.0
hardware (0.07) RAM 0.079 0.006 CI = 0.0
Display card 0.083 0.006 RI = 1.32
Battery capacity 0.070 0.005
Internal memory capacity 0.120 0.008
Expandable memory capacity 0.062 0.004
Upload and download speed 0.088 0.006
Multimedia Quality of rear camera 0.520 0.255 kmax = 7 CR = 0.0
hardware (0.49) Quality of front camera 0.083 0.040 CI = 0.0
Optical zoom 0.068 0.033 RI = 1.32
Led flash 0.058 0.028
Stereo speaker 0.083 0.040
Multiple microphone 0.124 0.061
Radio 0.065 0.032
(continued)
776 Y. Erdem et al.
Criteria Sub-criteria Weights Final kmax , CI, CR
weights RI
Software features (0.40) Smart assistant 0.039 0.016 kmax = 9 CR = 0.0
Sports and health 0.006 0.003 CI = 0.0
package RI = 1.45
Diary 0.005 0.002
Personal 0.009 0.003
notifications/directions/reminders
Media player 0.005 0.002
Ability to communicate with 0.295 0.118
other devices
Multi-user features 0.010 0.004
Free navigation software 0.295 0.118
Advanced security tools 0.335 0.134
Dimensional Design Weight 0.025 0.001 kmax = 7 CR = 0.0
Features (0.02) Screen size 0.211 0.004 CI = 0.0
Shape 0.202 0.004 RI = 1.32
Material 0.163 0.003
Plug-in battery 0.003 0.000
Water and dust resistance 0.211 0.004
Impact resistant glass 0.184 0.004
Environmental Mechanical home screen button 0.020 0.001 kmax = 7 CR = 0.0
design (0.01) Features Dual SIM card slot 0.154 0.002 CI = 0.0
Standard headphone jack 0.170 0.002 RI = 1.32
Fast charging 0.213 0.002
Wireless charging 0.176 0.002
Screen pen 0.112 0.001
Keys that facilitate the use of 0.154 0.002
media tools
Communication Bluetooth version 0.023 0.000 kmax = 8 CR = 0.0
channels (0.01) Wi-Fi version 0.176 0.002 CI = 0.0
Infrared sensor 0.003 0.000 RI = 1.41
NFC 0.213 0.002
USB type 0.129 0.001
Gyroscope 0.129 0.001
Accelerometer 0.129 0.001
GPS 0.199 0.002
The outputs of the proposed SCFPC algorithm by using the python programming
language are given in Table 2. According to the obtained outcomes in Table 2, the CR
values of the comparison matrix is lower than 0.1. It means that the pairwise com-
parison matrices are consistent.
Then, the conjoint analysis is applied to sublevels of the design parameters given in
Fig. 2. The numbers of levels of the variables for the Operational Hardware factor are
3, 3, 3, 3, 3, 2, and 2 respectively. There two levels for each sub criterion of Multimedia
Hardware, Software Features, Dimensional Design Features, Environmental Design

Features, Communication Channels factors, respectively. At first, a survey has been
structured including design alternatives of features are unique by using orthogonal
design and applied it. Then, the obtained data are analyzed by using SPSS.
5 Conclusion
AHP method which is based on the pairwise comparisons of the factors, is an effective
and also widely used method in the literature to obtain the importance degrees of the
factors based on subjective judgments. However, as the number of pairwise compar-
isons increases, ensuring matrix consistency becomes a difficult task. In addition, the
number of mathematical operations required for the calculation increases exponentially.
Furthermore, the consistency ratio of the pairwise comparison matrix is calculated after
survey data collected and if any one of the pairwise comparison matrices is obtained as
inconsistency, the whole effort will be wasted. Conjoint Analysis is also another widely
used method to determine the importance degrees of the parameters. However, as the
number of parameters and the number of levels for each parameter increases, it
becomes difficult to compare product features produced by the methodology. There-
fore, in the scope of this study, we propose a new approach including the short cut
fuzzy pairwise comparison method, conjoint analysis, and Fuzzy Set Theory. The
developed short cut fuzzy pairwise comparison (SCFPC) method is a modified version
of fuzzy AHP method and it used to obtain importance degrees of the main factors. The
advantages of the proposed SCFPC method, are as follows; (i) It shortens the pairwise
comparison process and (ii) it provides a dynamic structure for the calculation of
consistency. The Conjoint analysis is used to determine customer values on the level of
the design parameters while the fuzzy set theory has been used to analyze of linguistic
judgments of customers. The proposed approach applied to measure the importance
degrees of mobile phone design parameters. In the scope of the study, 6 main criteria,
45 sub-criteria, and two levels of each sub-criterion have been analyzed. Compared
with the conventional AHP and Conjoint method, it is proven that the proposed
approach reduces the process time and effort to reach the solution much more. For
further study, the same algorithm may be developed for the analytical network process.
In addition, the proposed algorithm may be extended to intuitionistic fuzzy sets.
References
1. Ho, F., Wang, C.N., Ho, C.T., Chiang, Y.C., Huang, Y.F.: Evaluation of smartphone feature
preference by a modified AHP approach. In: IEEE International Conference on Industrial
Engineering and Engineering Management 2015, pp. 591–594 (2015)
2. Tan, L.P., Adli, M.: Incorporating AHP in DEA analysis for smartphone comparisons. In:
Regional Conference on Mathematics, Statistics and Applications, pp. 307–315 (2006)
3. Işıklar, G., Büyüközkan, G.: Using a multi-criteria decision making approach to evaluate
mobile phone alternatives. Comput. Stand. Interfaces 2(29), 265–274 (2007)
778 Y. Erdem et al.
4. Belbag, S., Gungordu, A., Yumusak, T., Yilmaz, K.G.: The evaluation of smartphone brand
choice: an application with The Fuzzy Electre I Method. Int. J. Bus. Manag. Invent. 5(3),
2319–8028 (2016)
5. Yildiz, A., Ergul, E.U.: A two-phased multi-criteria decision-making approach for selecting
the best smartphone. S. Afr. J. Ind. Eng. 26(3), 194–215 (2015)
6. Salhieh, S.M., Al-Harris, M.Y.: New product concept selection: an integrated approach using
data envelopment analysis (DEA) and conjoint analysis (CA). Int. J. Eng. Technol. 3(1), 44
(2014)
7. Kang, Y.M., Cho, C., Lee, S.: Analysis of factors affecting the adoption of smartphones. In:
Proceedings of the 1st International Technology Management Conference 2011, pp. 919–
925 (2011)
8. Buckley, J.J.: Fuzzy hierarchical analysis. Fuzzy Sets Syst. 17(1985), 233–247 (1985)
9. Gül, M., Çelik, E., Güneri, A.F., Gümüş, A.T.: Simülasyon ile bütünleşik çok kriterli karar
verme: Bir hastane acil departmanı için senaryo seçimi uygulaması. İstanbul Ticaret
Üniversitesi Fen Bilim. Dergisi 11(22), 1–18 (2012)
10. Saaty, T.L., Vargas, L.G.: Estimating technological coefficients by the analytic hierarchy
process. Soc.-Econ. Plann. Sci. 13(6), 333–336 (1979)
11. Çebi, S., Ilbahar, E.: Analysis of parameters affecting the smart phone design by using fuzzy
Kano model. Alphanumeric J. 6(1), 83–92 (2018)
12. Hair, J.F., Anderson, R.E., Tatham, R.L., Black, W.C.: Multivariate Data Analysis, 5th edn.
Prentice-Hall International Inc. (1999)
13. Ilbahar, E., Cebi, S., Guler, M.G.: Using fuzzy AHP and conjoint analysis to identify
importance degrees of parameters on academic performance. In: Proceedings of the 4th
International Fuzzy Systems Symposium (FUZZYSS 2015), Turkey, pp. 380–384 (2015)
Congestion Trajectories Using Fuzzy Gaussian
Travel Time Based on Mesoscopic and Cellular
Automata Traffic Model
A. Boulmakoul1(&), L. Karim1,2, M. Nahri1, and A. Lbath3

1
Computer Science Department, FSTM, Hassan II University of Casablanca,
Casablanca, Morocco
[email protected],
2
LISA Lab., ENSA Berrechid, Hassan 1st University, Settat, Morocco
3
LIG/MRIM, CNRS, University Grenoble Alpes, Grenoble, France
[email protected]
Abstract. In the literature on engineering and traffic theory, numerous measures

have been proposed to represent the state of traffic conditions urban networks
sections. Again, the classical approach distinguishes two approaches to describe
traffic congestion. One is based on the correlation between traffic supply and
demand. The other is supported by a quality of service measures. In both cases,
the congestion measures are influenced by uncertainty due to the imprecision of
the measurement, the road user’s perception and the data variation depending on
the weather, etc. To take the uncertainty into account, a process for generating
fuzzy Gaussian variables has been proposed. This process requires the average
speed of movement on a link, the speed of free traffic flow and the dispersion of
speed measurements on the road. In this work we design a travel time model by
the Gaussian fuzzy numbers. We propose modeling of the traffic in urban net-
work. The generating ideas are based on the integration of mesoscopic modeling
taking into account macroscopic variables and microscopic traffic behaviors by
means of cellular automata. We develop a simple model on links that we gen-
eralize on networks. We then derive two processes to compute travel time on a
link. Thereafter, the travel time on a segment is modeled by a Gaussian fuzzy
variable. Developments resulting from this work contribute to generating the
fuzzy congestion trajectories and to simulate congestions propagation in urban
network and specially to analyze travel time in a smart city.
Keywords: Fuzzy numbers Urban computing Travel time Traffic

modeling Intelligent transportation systems
1 Introduction
The Traffic jam is one of the main societal and economic problems related to trans-
portation in industrialized countries. In this regard, managing traffic in congested urban
networks requires a clear understanding of traffic flow operations; how does the con-
gestion propagate through the network. There is an increasing interest in traffic
780 A. Boulmakoul et al.
simulation as a tool for evaluation of traffic systems. Several simulation models pro-
posed in the literature. Traffic congestion can be described by the presence of heavy
traffic at a road level. It can also be the difficulty of accessing a route when it exceeds
its limits or the inability to reach a defined destination within an easy timeframe. On a
section of track, the consideration of the fundamental diagram gives a simple definition
of congestion: traffic is said to be congested if it evolves in the right part of the diagram,
which is to say if the concentration is greater than the concentration critical (concen-
tration corresponding to the maximum flow) [15, 16]. The rules for producing con-
gestion events are simple to deploy. The fundamental point in this calculation is based
on the used measurement technologies. The aspect related to traffic sensors is not
considered in this work, see [3]. However, models from traffic flow theory are useful
tools to support the policy development and the assessment of traffic flow. In this work,
we propose a solution to integrate two simulation models: Nagel model [14] for inner
of the junctions and the LwR model [16] which is more appropriate for the input and
the output links of the junctions (see Fig. 1). We develop a simple model on links that
we generalize on networks. Thus, we then derive two processes to compute travel time
on a link. Thereafter, the travel time on links is modeled by a Gaussian fuzzy variable.
The rest of this article is organized as follows: the elements of mesoscopic simulation
and problem statements are presented first in the following section. In Sect. 3, a model
for computing travel time indicator is developed, to integrate the spatio-temporal
process of fuzzy travel time based on fuzzy Gaussian numbers. In the same section is
described the meta-models of fuzzy congestion trajectories. Section 4 presents pro-
posed analytic architecture and describes fuzzy congested trajectories meta-model.
Section 5 propose the conclusion and future improvement of this work.
Fig. 1. Traffic models integration

Congestion Trajectories Using Fuzzy Gaussian Travel Time 781
2 Traffic Modeling and Simulation
In this section, we present the preliminary elements for the definition and computation
of the fuzzy Gaussian travel time. First, we will recall cellular automata model followed
by LWR model [16]. Next, we describe a simple algorithm for developing travel time
on links, according to the fundamental traffic diagram based on the LWR model.
Finally, we describe a fuzzification process of this indicator by mean of a fuzzy
Gaussian number.
2.1 Cellular Automata Model

The flows computation of a section depends on its location and its topology (rela-
tionship with neighboring sections). We will recall, density and flow computing for-
mulas. The cellular automata model presents a discrete stochastic model for simulating
highway traffic. The model proposed in [14] is defined as a one-dimensional array of L
sites with open conditions, each site is either occupied by a vehicle or empty, each
vehicle has a speed v between 0 and vmax. An update is performed in parallel
according to the following steps for all vehicles.
1. Acceleration: if the vehicle speed v is less than vmax, and if the distance from the
next vehicle is greater than v + 1, then the speed increases by 1 (v v + 1)
2. Deceleration: if a vehicle at site i considers a vehicle at site i + j (with j v), it
decreases its speed to j − 1 (v j − 1)
3. Randomization: with a probability p, the speed v of each vehicle (if v 1) is
decremented by 1 (v v − 1).
4. Car motion: each vehicle advances v sites.
This model takes into account natural speed fluctuations due to human behavior and
variations in external conditions. The time average density q and time average flow q
are described in the following:
1 Xt0 þ T 1 Xt0 þ T
q¼ ni ðtÞ; q ¼ ni;i þ 1 ðtÞ ð1Þ
T t0 T t0
Where T: time period, ni ðtÞ = 0 if site i is empty, 1 if occupied, at time step t. ni;i þ 1 ðtÞ
if a car motion is detected between site i and i + 1.
2.2 Semi-macroscopic Traffic Model

Traffic flow may be considered as the flow of a continuous stream. Just assume that
every moment and every spatial point is an infinitesimal portion of the vehicle. The
fundamental traffic diagram describes the most important quantitative characteristic of
traffic flows. Typically, it defines a relationship between density and traffic flow.
/ðqÞ ¼ q v. The genetic forms of the fundamental diagram are given in the figures. It
is obvious that traffic flow phenomena depend strongly on the road occupancy rates. As
the density q is sufficiently small, the average speed v is practically independent of q,
this is explained by the fact that the vehicles are too far apart to interact with each other.
Therefore, at low density of vehicles, traffic is practically free. However, from practical
experience, vehicles must move more slowly with increasing density, when the forward
movement of vehicles is severely hampered by others due to the reduction in the
average distance between them. Various analytical forms have been suggested to the
fundamental diagram of traffic. These forms take into account the basic features consist
of two linear branches and has a triangular shape, historically, one of the earliest forms
is the Greenshields model [7]

q
/ðqÞ ¼ vf q 1 ð2Þ
qmax
This corresponds to the linear relationship speed-density

q
vðqÞ ¼ vf 1 ð3Þ
qmax
The model Lighthill Whitham-Richards (LWR) is based on the assumption that the
flow is a function of the density [16]:
/ðqÞ ¼ q v: ð4Þ
The model equations are summarized as follows:
/ðq2 Þ /ðq1 Þ
vw ¼ ð5Þ
q2 q1
8
< þ @ ðqv
@q
@t
Þ
@x ¼ 0
d ¼ /ðqÞ ¼ q v ð6Þ
:
v ¼ wðqÞ
When two characteristic lines intersect, a shock wave is created. By definition, a

shock wave is a mathematical discontinuity in density and speed.
3 Travel Time Index
The travel time index (TTI) given by the Texas Transportation Institute compares the
travel time rates in the peak period and travel time rate during free flow. The TTI is
calculated as given below [5, 6, 9]:
Actual Travel Time Rate

TTI ¼ : ð7Þ
Travel Time Rate during free flow conditions
A speed-density function is used to estimate the travel time index (TTI) given
above.
Notation:
vaf : the free-flow speed on link a, La : length of link a, xa ðtÞ: number of vehicles on link a
in time interval t. TTI(a): Travel Time Index of link a.

qa
vðqÞ ¼ vaf 1 a ð8Þ
qmax
The travel time is index is given by

, !
La La vaf
TTI ðaÞ ¼ ¼ ð9Þ
vðqÞ a
vf vðqÞ

qa 1 qa qa
¼ 1 a 1þ a iff a 0 qa qamax ð10Þ
qmax qmax qmax
vaf qamax q La
if q qamax TTIðaÞ ¼ ¼ a max ð11Þ
vaf qamax qa ðtÞ qmax La xa ðtÞ
The measured travel time corresponds to the average travel time on a segment.
When the volume is very low, the predicted travel time is approximately equal to free-
flow travel time.
3.1 Travel Time Fuzzification

In the literature, many works has been considered for fuzzy travel time modelling [8,
10]. In our approach, modeling is simple and considers the statistical parameters of this
indicator. Instead of modeling the travel time with laws of probabilities such as the
gamma law type; Gaussian fuzzy numbers (see Fig. 2) are better adapted and allow to
make simplified calculations. To define travel time as a fuzzy number in the form of
Gaussian distribution [17], the membership function required for building the Gaussian
expressed as follows: Ta gfnðl ; r Þ.
a a
shaped fuzzy set must be
ðtlÞ2
lTa ðtÞ ¼ exp r2 , where la and ra represent the mean and the standard devi-
ation of travel time on link a respectively.
Travel time on road: In this construction we assume that the speed on each link
varies between 0 and vaf . The fuzzy travel time index for a path H is calculated as
follows:
X X qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X ffi
eH ¼
T gfnðla ; ra Þ ¼ gfn la ; ð r a Þ2 ð12Þ
a2H a2H a2H
The development of travel time index could be generic. Several alternatives are
inferable using triangular fuzzy numbers, which can be used in other computational
models. It is also possible according to this model to produce this indicator either by
simulation or by the direct exploitation of the sensor data.
Fig. 2. (a) Travel time on a link as Gaussian Fuzzy Number function membership, (b) Fuzzy
travel time’s process description
4 Analytic Architecture and Fuzzy Congested Trajectories
In this section, we describe hierarchical analytic architecture for real time traffic situ-
ation detection in each road network segment. Recently, advances in communication
technologies and ubiquitous computing devices as well as in-vehicle computers are
shaped intelligent transportation systems showing up the so-called internet of vehicles.
A big advance in managing architecture has accompanied these advances. However,
the classical architectures based only on the cloud computing managing connected
devices as well as supporting services such as traffic safety and management and other
services addressed to road users’ comfort, show several limits touching latency as well
as efficiency. Recent several works suggest incorporating new layer close to connected
devices in the name of fog computing. Thus, the global IoV managing architecture is
composed from the edge, the fog and the cloud layers [4, 18, 19]. Similarly, we follow
the same path to perform real time traffic data analytics. In fact, we aim to calculate real
travel time showed in every road segment from microscopic events generated by
intelligent vehicles. Thus, we propose distributing events processing between several
fog nodes, each one of them is responsible of limited geographical area. We propose a
framework based on spark, Kafka and microservices as well as containerization [10–
13]. In fact, Moreover, recent versions of spark integrate new powerful tool for treating
continued data flows called structured streaming. Spark Structured Streaming gathers
the powerfulness of Spark Streaming and Spark SQL [1], providing the availability of
the continued flows of data for processing (see Fig. 3).
This principal consists of adding new arrival data batches to the pipeline of data
without removing the old ones. Thus, this allows processing data more efficiently and
especially taking in consideration the treatment of the delayed arrived data. To test the
framework, we used our simulator integrated with sumo as a microscopic traffic events
simulator and Traci [2] interface to retrieve data from the simulation via TCP protocol.
Each fog node calculates travel time of links belonging to the managed geographical
area managing and sends the real time results to the cloud layer for visualization and
performing deep analytics over these results. The next figure shows the congestion
situation on each link (see Fig. 4).
Fig. 3. Fuzzy congestion trajectories meta-model
Fig. 4. (a) Congestion trajectories visualization, (b) travel time index time series
This model is based on a fuzzy representation introducing various degrees of

membership of an object to the time, to its function, and to the space (Fig. 5).
Fig. 5. Fuzzy congestion trajectories

In our previous work, we established trajectories meta-model. In this paper and for
simplification purpose, we suggest an extension of this modeling to fuzzy trajectories.
The uncertainty modeled by fuzzy sets, comes from measures given by sensors, here
the fuzzy travel time, from space by the uncertainty of the localization and from the
time itself quite often defined by intervals. Fuzzy congestion trajectories are produced
from fuzzy congestion events and modeled as fuzzy space-time series. Of course, this
presentation is short; however, more presentations that are detailed will soon be
published.
5 Conclusion
The proposed approach is easy to implement. It allows representing the traffic condi-
tions on road segments and itineraries. The resulting congestion measure is used to
define fuzzy congestion trajectories. However, further refinements are needed to extend
the process to include other traffic measures. Further research is also necessary to
integrate other forms of fuzzy numbers and not remain limited to the Gaussian case.
Our approach is based on the mesoscopic traffic model, but not limited to this model.
While the development of the approach particularly addresses the modeling and sim-
ulation aspects, deployment on a real network is a planned objective in our future work.
Aspects related to data mining in a MongoDB-Spark-Hadoop ecosystem is a high
expectation for monitoring and traffic control in large and medium-sized cities.
References
1. Armbrust, M., Das, T., Torres, J., Yavuz, B., Zhu, S., Xin, R., Zaharia, M.: Structured
streaming: a declarative API for real-time applications in Apache Spark. In: Proceedings of
the ACM SIGMOD International Conference on Management of Data, pp. 601–613 (2018)
2. Behrisch, M., Bieker, L., Erdmann, J., Krajzewicz, D.: SUMO–simulation of urban mobility:
an overview. In: Proceedings of the Third International Conference on Advances in System
Simulation (SIMUL 2011) (2011)
3. Boulmakoul, A., et al.: Towards scalable distributed framework for urban congestion traffic
patterns warehousing. Appl. Comput. Intell. Soft Comput. 2015, Article ID 578601 (2015).
http://doi.org/10.1155/2015/578601
4. Butt, T.A., Iqbal, R., Shah, S.C., Umar, T.: Social internet of vehicles: architecture and
enabling technologies. Comput. Electr. Eng. 69, 68–84 (2018)
5. BPR: Traffic assignment manual. U.S. Department of Commerce, Bureau of Public Roads
(1964)
6. Daganzo, C.F.: The cell transmission model part II: network traffic. Transp. Res. 29B, 79–93
(1995)
7. Greenshieds, B.D.: A study in highway capacity. Highway Res. Board 14, 448–477 (1935)
8. Hamad, K., Kikuchi, S.: Developing a measure of traffic congestion fuzzy inference
approach. Transp. Res. Rec. 1802 (2002). https://doi.org/10.3141/1802-10
9. HCM, Highway Capacity Manual, TRB, National Research Council, Washington, D.C.
(2000)
10. Karim, L., Boulmakoul, A., Lbath, A.: Real time analytics of urban congestion trajectories
on Hadoop-MongoDB cloud ecosystem. In: ICC 2017, pp. 29:1–29:11 (2017)
11. Kreps, J., Narkhede, N., Rao, J.: Kafka: a distributed messaging system for log processing.
In: Proceedings of the NetDB, pp. 1–7 (2011)
12. Maguerra, S., Boulmakoul, A., Karim, L., Badir, H.: A distributed execution pipeline for
clustering trajectories based on a fuzzy similarity relation. Algorithms 12(2), 29 (2019)
13. MongoDB (2020). https://www.mongodb.com/hadoop-and-mongodb
14. Nagel, K., Schreckenberg, M.: A cellular automaton model for freeway traffic. J. Phys. I 2
(12), 2221–2229 (1992)
15. OECD: Managing urban traffic congestion. In: European Conference of Ministers of
Transport (OECD/ECMT 2007). ECMT (2007). (ISBN 978-92-821-0128-5)
16. Treiber, M., Kesting, A.: Traffic Flow Dynamics, Data, Models and Simulation. Springer,
Heidelberg (2013). ISBN 978-3-642-32459-8
17. Zimmermann, H.J.: Fuzzy Set Theory and Its Applications, 3rd edn. Kluwer Academic
Publishers, Norwell (2000)
18. Luckham, D.: The power of events: an introduction to complex event processing in
distributed enterprise systems. In: International Workshop on Rules and Rule Markup
Languages for the Semantic Web, p. 3. Springer (2008)
19. Zhang, W., Zhang, Z., Chao, H.C.: Cooperative fog computing for dealing with big data in
the internet of vehicles: architecture and hierarchical resource management. IEEE Commun.
Mag. 55(12), 60–67 (2017)
Data Mining Algorithms for Classification
Model of Engineering Grade
Ching-Lung Fan(&)
Department of Civil Engineering, The Republic of China Military Academy,

No. 1, Weiwu Rd., Fengshan, Kaohsiung 830, Taiwan
[email protected]
Abstract. This study uses various construction inspection data of the Public
Construction Management Information System (PCMIS) from 2003 to 2018 and
data mining (DM) to analyze the relationship among the engineering-grades,
contract prices, project category, and progress of 1,015 cases and 499 defects.
Association rule mining was used to derive 11 the rules which altogether contain
the four types of defects and project attributes. The algorithms analyze the
importance value of the attributes to gain relatively important defects. Therefore,
association rules and important defects can provide a useful reference to enable
engineering management personnel with aid to understand the relevance of
defects and project attributes to improve construction quality. In addition,
classification models are constructed using a neural network (NN), support
vector machine (SVM), and C5.0 for DM algorithms. According to the results of
accuracy and the area under the curve (AUC) show that the SVM has the best
classification benefit, followed by NN. The classification model obtained herein
can effectively predict the engineering-grade.
Keywords: Data mining Association rule Neural network Support vector

machine Decision tree Defects
1 Introduction
Studies have focused on the analysis and discussion of the causes and origins of
construction defects [1] and on the types and classification of defects [2]. These studies
used statistical analysis to classify defects and obtain a distribution of the cause of
defects. However, the statistical tools used lack the ability to automate data exploration.
Therefore, they afford limited benefits and aid to project execution teams.
Several recent studies have applied artificial intelligence and data mining
(DM) algorithms combined with big data for evaluating or estimating existing defects
[3]. These studies established decision support systems, expert systems, or case-based
reasoning systems to solve problems and quickly assess construction defects with a
certain predictive accuracy. Das and Chew [4] considered that a scientific scoring
system needs to be established through a defect database and its systematic analysis or
through defect grading to clearly define the impact of defects. Furthermore, Macarulla
et al. [5] claimed that various defect properties can be classified through attribute values
to establish a defect classification system. Large databases often contain considerable
Data Mining Algorithms for Classification Model 789
hidden knowledge patterns that can be discovered through DM. Thus far, few studies
have made the fullest use of DM in the construction field to discover a large number of
datasets [6]. Moreover, Liao and Perng [7] believe that association rules can be used to
discover the relationships and potential associations among attributes in big data. These
rules can effectively reveal unknown relationships and provide a basis for decision-
making and prediction [8].
In Taiwan, government departments systematically classified construction inspec-
tion data in the Public Construction Management Information System (PCMIS)
through the construction inspection mechanism. Then, statistical analyses were con-
ducted and defect improvement measures were implemented to improve the quality of
public constructions and project management performance. The inspection content was
divided into four categories: Quality Management System (An, 113 defects), Con-
struction Quality (Bn, 356 defects), Schedule (Cn, 10 defects), and Planning and
Design (Dn, 20 defects), totaling to 499 defect types. Since the implementation of the
construction inspection mechanism in 1993, considerable attribute data related to
public construction inspections have accumulated. Among these, defects represent the
construction quality, and the relationship between these defects and related attributes
warrants further analysis. The purposes of this study were correlations among specific
defects and related attributes that can be determined to reduce or avoid defect
occurrence.
2 Literature Review
Previous studies have used ARM for analyzing construction defects. For example,
Cheng et al. [9] proposed a genetic-algorithm-based approach that incorporated a
hierarchical concept of construction defects to discover useful information in a con-
struction defect database and to identify relationships between these defects. Lin and
Fan (2018) used the ARM and fuzzy logic to analyze the relevance and relative
importance of specific defects and assist engineering execution teams to effectively
manage critical defects [10].
Classification and regression are used to obtain a set of models through training
given a group of information that has been classified. Then, the previously trained
model can be used to predict the classification to which unclassified data belongs.
In DM, classification and regression techniques such as NN, SVM, and DT are com-
monly used. NN applications in construction engineering include the estimation of
construction costs [11] or prediction and diagnosis of related defects [12]. SVM is most
commonly used for spatial feature recognition of image data and in engineering
applications such as detection or recognition. Examples of applications include inter-
pretation of pipeline images [13], corrosion identification in steel structure bridges [14],
automatic defect detection in sewer images [15], detecting various types of construction
materials [16], and automatic detection of road patches [17]. In DT, a tree structure
diagram is used to deduct a series of classification and decision problems. It is used in
the field of engineering and management to classify and predict a wide variety of
attributes [18].
790 C.-L. Fan
3.1 Association Rule
Analyzing association rules allows researchers to define the concurrence degree of
variables according to the association frequency and dependent items among defects
(or products). This study configured the support and confidence indices of defects
according to the defects and frequencies discovered during construction inspections
and subsequently calculated the lift strength of each rule. These concepts were used to
screen key association rules, ensure associations among defects, and analyze the
strength of each association rule. The confidence index can be calculated as follows:
support: A ! B ¼ PðA \ BÞ ð1Þ
confidence: A ! B ¼ PðA \ BÞ=PðAÞ ð2Þ
lift ¼ PðA \ BÞ=PðAÞ PðBÞ ð3Þ
3.2 Neural Network (NN)

NN typically uses a set of data to build a system model from which to estimate, predict,
diagnose, and make decisions. It mainly comprises many artificial processing units
(also called artificial neurons), and the output of each processing unit serves as an input
for other processing units. The calculation function for the output and input values is
X
Yi ¼ f Wij Xi hj ð4Þ
Xi and Yi are the input signals and output signals of the artificial neuron model,
respectively, and f is the activation function of the artificial neuron model. The purpose
is to multiply the values input by other processing units with the weighting values and
then sum them to transform the output values of the processing unit. Wij is the ganglion
strength (weight) of the artificial NN, and hj is the threshold of the artificial neuron
model. When all input variable values are multiplied by the weight, the total value is
greater than the threshold value hj and then converted to an output value via the
activation function and passed to the next artificial neuron.
3.3 Support Vector Machine (SVM)

SVM is a machine learning method proposed by Vapnik [19] according to statistical
learning theory, and it is a type of classification algorithm. The basic principle is to find
the best hyperplane of the boundary in the high-dimensional feature space to classify
different binary category and minimum classification error rate. If the training data
fxi ; yi g are not linearly separable, the low-dimensional sample space fxi g can be
converted to the high-dimensional feature space Uðxi Þ through the mapping conversion
method. The optimal classification plane is found in this space, and its corresponding
classification function is given as Eq. (5). Here, ai is the Lagrange multiplier;
kðxi ; yi Þ ¼ Uðxi Þ Uðyi Þ is the kernel function, where x is the input data; n is the
dimension; Uðxi Þ is a feature space that maps x to a high dimension; f is the dimension,
and f > n; and b is the bias value.
" #
X
n
f ðxÞ ¼ sgn aiyikðxi; xjÞ þ b ð5Þ
i¼1
3.4 Decision Tree (DT)

A DT is based on the target setting to select the variables and branching methods, and it
is presented in a hierarchical structure to extract classification rules. A pruned DT
model can be used for data exploration or prediction and to find the hierarchical
relationship between a target variable and each variable. DT algorithms such as C5.0,
CART, CHAID, and QUEST are commonly used in academia and industry [20]. These
algorithms mainly differ in terms of the methods used for splitting the attribute criteria
of the root and child nodes and the numbers of split child nodes. The C5.0 algorithm,
proposed by Quinlan [21], measures the amount of information under different con-
ditions according to the information probability of different categories to find the
attribute that can obtain the maximum gain. As the node of its branch, the sum of
information of all categories is given as Eq. (6). The number of data in each category is
xj, where N is the number of all data in the dataset. The probability of occurrence of
each category can be defined as Pj ¼ xj =N, and the information of each category is
−log2 Pj.

x1 x1 x2 x2 xk xk
InfoðDÞ ¼ log 2 log 2 . . . log 2
N N N N N N
Xk
¼ j¼1
Pj log 2ðPjÞ ð6Þ
4.1 Construction Inspection Information

This study uses PCMIS construction inspection data from January 2003 to May 2018.
The total number of projects inspected was 1,015 (defect frequency: 18,246); the
inspected data included 499 defect types (Code: An, Bn, Cn, Dn), grades, contract
prices, project category, and progress. Defects are classified into four categories
792 C.-L. Fan
according to the quality management system, construction quality, progress, and

planning and design. Among these, defects of the quality management system (An)
include 113 items for procuring units, supervisory units, and contractors. The defects of
construction quality (Bn) include 356 items of strength I, strength II, and safety.
Defects in schedule (Cn) include 10 items. The defects of planning and design (Dn)
include 20 items of safety, construction, maintenance, and gender differences, with a
total of 499 defects.
The engineering-grades are divided into four levels: A, B, C, and D. For these four
grades, cluster analysis was performed with two variables: inspection score of 1,015
and defect frequency (18,246). The contract price is divided into three types: threshold
for publication (P; NT$10–50 million); threshold for supervision (S; NT$50–200
million); and threshold for large procurement (L; NT$200 million or more). The project
category is divided based on construction characteristics and complexity as new con-
struction engineering (NC), renovation engineering (RE), other engineering (OE),
hydropower and air conditioning engineering (HC), civil engineering (CE), airport
engineering (AE), and harbor engineering (HE). Finally, projects were divided
according to construction progress into 50% or less and 50% or more.
4.2 Association Rule Parameter Setting

The association rule in this study sets eight attributes X1–X4 and Y1–Y4 as decision
variables. The minimum support and minimum confidence are 10% and 90%,
respectively, and the lift must be greater than 1.5. There are 11 association rules that
satisfy the above three conditions (Table 1). Among them, X2-RE (renovation engi-
neering) and X3-P (threshold for publication) have strong association. These rules and
attributes can serve as a useful reference for management units.
For example, Rule 2: If the project category is “renovation engineering (RE)” and
the defect is “failure to review the test report of building materials, failure to compile
checklists for building materials or equipment tests, reviews or inspections, or failure to
meet the project requirements (A77),” then the contract price is “threshold for publi-
cation (P).” Rule 7: If the project category is “renovation engineering (RE)” and the
defect is “Other recording errors in building material and equipment reviews (B283)”
and “Failure to implement quality control checklist, failure to quantify inspection
standards or permissible errors, or failure to record the inspection accurately (A76),”
then the contract price is “threshold for publication (P).” Similarly, Rule 11: If the
project category is “renovation engineering (RE)” and the defect is “Failure to log the
construction journal, failure to log in the predetermined format, or incomplete logging
(A75),” then the contract price is “threshold for publication (P).” The probability of
these defects (A75, A76, A77, and B283) with renovation engineering (RE) and
threshold for publication (P) occurring together is relatively high, and special attention
should be paid when performing supervision work.
Table 1. Association rules for decision variables.

Rule Antecedent Consequent Support Confidence Lift
1 Project category = RE & Contract 12.7 97.7 1.5
engineering − grade = A prices = P
2 Project category = RE & A77 Contract 10.4 97.2 1.5
prices = P
progress = 1 prices = P
4 Project category = RE Contract 28.0 95.4 1.5
prices = P
5 Project category = RE & B283 Contract 14.1 95.1 1.5
prices = P
engineering − grade = C prices = P
7 Project category = RE & B283 & Contract 11.2 94.7 1.5
A76 prices = P
engineering − grade = C & A76 prices = P
prices = P
progress = 0 prices = P
prices = P
4.3 Effectiveness of Classification Model

A confusion matrix compares the predicted and tested model data. It can be used to
evaluate the performance of the supervised classification method. In a model of ideal
classification, the predictions match the actual values. Assume that there are two cate-
gories of the classification model, A and B, and if the prediction result of the classifi-
cation model is consistent with the actual category, it is called “True”; if it is inconsistent,
it is called “False”, which is misjudgment (erroneous judgment). There are two kinds of
misjudgments: the one that is actually A but predicted to be B (False Negative, FN) and
the one that is actually B but predicted as A (False Positive, FP). There are also two types
of correct classification predictions: the one that is actually A and the prediction is also A
(True Positive, TP) and the one that is actually B and the prediction is also B (True
Negative, TN). According to the classification results in Table 2, the model’s recall,
precision, accuracy, and the area under the curve (AUC) can be calculated.
Table 2. Confusion matrix of classification model.

Actual Predicated Recall (%) Accuracy (%) AUC
A B
A TP FN TP/TP + FN TP + TN/TP + TN + FP + FN Mean recall
B FP TN TN/FP +TN
Precision (%) TP/TP + FP TN/FN + TN
794 C.-L. Fan
The recall for a category is the number of predictions in the category divided by the
actual number of members of the category. For example, NN correctly predicted 435
entries as belonging to grade A; therefore, the corresponding recall was 96.03%, as
obtained by dividing the number of correctly predicted entries by the total number of
entries in A (453). Among the four algorithms, the SVM algorithm had the highest
prediction for grade B and D (100%); the NN algorithm had a relatively higher and
average prediction for all four grades: A, B, C, and D (Table 3).
Accuracy can be defined as the number of correct predictions across all categories
divided by the total number of entries. For example, the NN algorithm correctly pre-
dicted 435, 140, 295, and 84 entries for grades A, B, C, and D, respectively. The
summation of these values divided by the total number of entries in all levels (1,015)
yielded an accuracy of 93.99%. The accuracy rates of the SVM and C5.0 algorithms
were 89.56% and 85.81%, respectively (Table 3). Accordingly, the NN algorithm had
the highest prediction accuracy.
AUC is the average value of recall, i.e. AUC = 1/2 [(TP/TP + FN) + (TN/FP +
TN)]. AUC is used to verify the performance of the classification model, the higher the
value, the better the performance. It was found that the AUC of SVM was the best, at
0.935, followed by NN and C5.0 with the value at 0.930 and 0.846, respectively.
Table 3. Confusion matrix and accuracy for DM generation.

Algorithm Actual Predicated Recall (%) Accuracy (%) AUC
A B C D
NN A 435 9 8 1 96.03 93.99 0.930
B 7 140 2 1 93.33
C 13 3 295 8 92.48
D 0 5 4 84 90.32
Precision 95.6 89.2 95.5 89.4 n = 1,015
SVM A 420 8 2 0 97.67 89.56 0.935
B 0 114 0 0 100
C 35 35 307 26 76.18
D 0 0 0 68 100
Precision 92.3 72.6 99.4 72.3 n = 1,015
C5.0 A 408 19 39 4 86.81 85.81 0.846
B 11 132 7 5 85.16
C 28 3 255 9 86.44
D 8 3 8 76 80
Precision 89.7 84.1 82.5 80.9 n = 1,015
5 Conclusion
This article analyzes various information about PCMIS construction inspections in

2003–2018 to understand the key factors affecting defect occurrence. This information
comes from 1,015 cases including: defects, engineering-grade, contract price, project
category and progress. Association rules were used to extract valuable knowledge so
that models of defect associations could be obtained from historical databases.
Furthermore, classification and regression techniques are used to build classifica-
tion models for engineering-grades (A, B, C, and D), the main purpose of which is to
understand the characteristics of the project and use them as the basis of classification.
The classification results show that among the three data mining algorithms, SVM’s
AUC is the highest. Overall, SVM has the best performance for engineering-grade
classification; followed by NN and C5.0. This study uses the comprehensive index of
the confusion matrix to verify the effectiveness of each classification model. The results
of the classification model can effectively predict the rating grade of the construction
inspection. Future studies should analyze the time series of defects and employ
sequential data to forecast their order of occurrence and relationships at different times.
References
1. Aljassmi, H., Han, S., Davis, S.: Analysis of the complex mechanisms of defect generation
in construction projects. J. Constr. Eng. Manag. 142(2), 04015063 (2016)
2. Georgiou, J.: Verification of a building defect classification system for housing. Struct. Surv.
28(5), 370–383 (2010)
3. Sinha, S.K., Fieguth, P.W.: Neuro-fuzzy network for the classification of buried pipe defects.
Autom. Constr. 15(1), 73–83 (2006)
4. Das, S., Chew, M.Y.L.: Generic method of grading building defects using FMECA to
improve maintainability decisions. J. Perform. Constr. Facil. 25(6), 522–533 (2011). https://
ascelibrary.org/doi/abs/10.1061/%28ASCE%29CF.1943-5509.0000206
5. Macarulla, M., Forcada, N., Casals, M., Gangolells, M.: Standardizing housing defects:
classification, validation, and benefits. J. Constr. Eng. Manag. 139(8), 968–976 (2013).
https://ascelibrary.org/doi/abs/10.1061/%28ASCE%29CO.1943-7862.0000669
6. Xiao, F., Fan, C.: Data mining in building automation system for improving building
operational performance. Energy Build. 75, 109–118 (2014)
7. Liao, C.W., Perng, Y.H.: Data mining for occupational injuries in the Taiwan construction
industry. Saf. Sci. 46(7), 1091–1102 (2008)
8. Chae, Y.M., Ho, S.H., Cho, K.W., Lee, D.H., Ji, S.H.: Data mining approach to policy
analysis in a health insurance domain. Int. J. Med. Informatics 62(2–3), 103–111 (2001)
9. Cheng, Y., Yu, W.D., Li, Q.: GA-based multi-level association rule mining approach for
defect analysis in the construction industry. Autom. Constr. 51, 78–91 (2015)
10. Lin, C.L., Fan, C.L.: Examining association between construction inspection grades and
critical defects using data mining and fuzzy logic. J. Civ. Eng. Manag. 24(4), 301–317
(2018)
11. Jafarzadeh, R., Ingham, J.M., Wilkinson, S., Gonzalez, V., Aghakouchak, A.A.: Application
of artificial neural network methodology for predicting seismic retrofit construction costs.
J. Constr. Eng. Manag. 140(2), 04013044 (2014)
796 C.-L. Fan
12. Murugan, S.B., Sundar, M.L.: Investigate safety and quality performance at construction site
using artificial neural network. J. Intell. Fuzzy Syst. 33(4), 2211–2222 (2017)
13. Mashford, J., Rahilly, M., Davis, P., Burn, S.: A morphological approach to pipe image
interpretation based on segmentation by support vector machine. Autom. Constr. 19, 875–
883 (2010)
14. Chen, P.H., Shen, H.K., Lei, C.Y., Chang, L.M.: Support vector machine based method for
automated steel bridge rust assessment. Autom. Constr. 23, 9–19 (2012)
15. Halfawy, M.R., Hengmeechai, J.: Automated defect detection in sewer closed circuit
television images using histograms of oriented gradients and support vector machine.
Autom. Constr. 38, 1–13 (2014)
16. Rashidi, A., Sigari, M.H., Maghiar, M., Citrin, D.: An analogy between various machine-
learning techniques for detecting construction materials in digital images. KSCE J. Civ. Eng.
20(4), 1178–1188 (2016). https://doi.org/10.1007/s12205-015-0726-0
17. Hadjidemetriou, G.M., Vela, P.A., Christodoulou, S.E.: Automated pavement patch
detection and quantification using support vector machines. J. Comput. Civ. Eng. 32(1),
04017073 (2018)
18. Lin, C.L., Fan, C.L.: Evaluation of CART, CHAID, and QUEST algorithms: a case study of
construction defects in Taiwan. J. Asian Archit. Build. Eng. 18(6), 539–553 (2019)
19. Vapnik, V.: The Nature of Statistical Learning Theory. Springer-Verlag, New York (1995)
20. Han, J., Kamber, M.: Data Mining Concept and Technology. Morgan Kaufmann, San
Francisco (2001)
21. Quinlan, J.R.: Induction of decision trees. Mach. Learn. 1, 81–106 (1986)
A Fuzzy Strategy Analysis Simulator
for Exploring the Potential of Industry 4.0
in End of Life Aircraft Recycling
Samira Keivanpour(&)
Department of Mathematics and Industrial Engineering, Polytechnique Montreal,

Montreal, QC, Canada
[email protected]
Abstract. This paper addresses the implications of industry 4.0 in the man-
agement of the aircraft at the end of life. Considering the number of retired
aircraft in the future, the advanced and technology-driven management of this
complex product is essential for the aviation industry. A simulator based on the
joint application of fuzzy SWOT analysis, business model canvas, and fuzzy
interface systems is proposed. The simulator provides a user-friendly interface
for analyzing what-if scenarios as the results of the changes in internal and
external factors. This decision tool provides insights for executives and poli-
cymakers for the successful adoption of Industry 4.0 in the context of the
circular economy.
Keywords: End-of-life aircraft Fuzzy SWOT Industry 4.0 Recycling

Business model canvas Value dimensions Fuzzy interface
1 Introduction
End of Life (EoL) aircraft recycling is a new challenge in the aviation industry and few
scholars addressed this subject in detail via real case studies. Viability of the business
model, valorization of recovered parts and material, and decision models that support
the strategic, tactical, and operational decisions during the different sub-processes of
treatment is essential. Moreover, for the EoL aircraft, access to advanced optimization
tools and specialized machinery for reducing dismantling time and decreasing the
waste of operation is critical. Industry 4.0 is promising in advanced management of the
complex products at the end of life. Different internal and external factors should be
assessed to highlight the drivers and challenges of Industry 4.0 implications in the EoL
aircraft problem. SWOT analysis is a popular strategic tool for addressing strengths,
weaknesses (internal factors), and opportunities and threats (external factors) at the
corporate level [15, 16]. The business model canvas is also widely used for analyzing
the value creation, transferring, and capturing by the organizations [17, 18]. The nine
elements of business model canvas including value proposition, partners, resources,
activities, target customers, customers’ relationship, channels, costs, and revenue
structures highlight this value creation process [17]. Due to the imprecision of the
strategic planning tools, several studies considered the fuzzy logic concept in the
798 S. Keivanpour
strategic mapping process and via the joint application of fuzzy logic with strategic
tools [e.g. 19–23]. The fuzzy numbers and the resulted fuzzy matrix of internal and
external factors aid in dealing with the ambiguity of defining the final strategies. It
provides more flexibility in terms of strategy definition and analyzing the sensitivity of
the defined strategies with changes in decision-makers’ opinions. In this study, a
simulator based on a fuzzy SWOT matrix, business model canvas, and fuzzy interface
systems is proposed for analyzing the drivers and challenges of Industry 4.0 adoption.
In the first layer of this simulator, the fuzzy matrix of SWOT is mapped in the nine
elements of the business model canvas. In the second layer, a multistage fuzzy
Mamdani interface is developed to integrate the internal and external factors based on
three dimensions of value model: value proposition, value creation/delivery, and value
capture. Finally, the aggregated internal and external factors are mapped in a quanti-
tative SWOT matrix with scenario analysis capability based on the changes in mem-
bership functions. The contributions of this research paper are two folds: first, the joint
application of business model canvas and fuzzy SWOT analysis for strategic mapping
and second, application in adopting industry 4.0 for EoL aircraft recycling enterprises.
The rest of the paper is organized as follows: Sect. 2 discusses the literature review,
Sect. 3 describes the method. Section 4 provides a numerical example and discusses
the results and finally, Sect. 5 provides the conclusions and the direction for future
research.
2 Literature Review
Application of Industry 4.0 in the circular economy has received much attention in
recent years. However, the number of research papers are few in comparison to the
other opportunities and challenges of this new paradigm in manufacturing and product
life cycle studies. The authors in [1] emphasized the role of big data and data-driven
approach as a facilitator for decision making and optimizing industrial symbiosis. The
author explained that big data and analyzing the stakeholders’ network could reveal
more outcomes of the circular economy such as the reduction in resource consumption
and waste generation that leads to industrial sustainability. Lin stated the application of
data-driven decision-making tools for enhancing the recycling business models [2].
There are other applications such as the digital twin model for recovery of electronic
and electrical wastes [3], cloud-based inventorying monitoring of electric vehicle
battery [10], and blockchain perspective for product life cycle management [7]. The
authors in [4] proposed a research agenda in the application of Industry 4.0 in the
circular economy based on the resource-based view, stakeholder theory, institutional
theory, and ecological modernization. The authors highlighted that assessing the
resources and capabilities of the organization and identification of barriers based on
Industry 4.0 for achieving the objectives in the circular economy is a gap in the
literature. Mapping the expectation of stakeholders for technology implementation and
facilitating the mechanism for integrations of the organization and the key partners are
the essential aspects that should be studied in linking Industry 4.0 and circular econ-
omy. Studying the customers’ and competitor’s implications and the role of policy-
makers are the other items in this research agenda. Information sharing, using a
A Fuzzy Strategy Analysis Simulator for Exploring the Potential of Industry 4.0 799
data-driven approach for the diffusion of green products in the market, real-time, and
internet of things based optimization models are the other studies in the applications of
Industry 4.0 in reverse logistics and sustainable manufacturing [5, 8, 9, 11, 12]. The
authors in [6] proposed “EoL 4.0” for real-time data exchange among the key players
and implementing analytical capabilities of Industry 4.0 for maximizing the value from
EoL products. The important challenges in the circularity of utilized materials are time-
tested processes of reuse, recovery, and recycling that reinforce circular economy
initiatives [13]. Designing closed material chains for strategic and valuable materials
will be more essential than the weight percentage target of recovered material [14]. The
joint application of a strategic analysis tool like SWOT and business model canvas
provides an effective framework for addressing the drivers and barriers of Industry 4.0
in the context of the circular economy.
3 Method
The steps of the proposed methodology are described in this section. The first step is
preparing the business model canvas and determining the internal and external factors
for each element of the business model. The sub-factors are also defined in this
step. Figure 1 illustrates the business model canvas and the factors.
Fig. 1. Business model canvas, determining external and internal factors the sub-factors for each
element. For example, IP1 shows the internal factor related to partners and IP11 addresses the
related subfactor
The second step is assigning a triangular fuzzy number to each sub-factor. A tri-
angular fuzzy number is defined as follows:
IBMEij ¼ ðaij ; bij ; cij Þ ð1Þ
EBMEij ¼ ðaij ; bij ; cij Þ ð2Þ
Where IBMEij is an internal sub-factor of business model element i, j and EBMEij is

an external sub-factor.
800 S. Keivanpour
The next step is considering weight for each sub-factor considering its influence on
the business model and aggregating of sub-factors for each internal and external
factors:
Xni Xni Xni
IBMEi ¼ ð j¼1 wij aij ; j¼1
wij a ij ; w a Þ
j¼1 ij ij
ð3Þ
Xni Xni Xni

EBMEi ¼ ð j¼1
wij aij ; j¼1
wij aij ; j¼1
wij aij Þ ð4Þ
Now, the total aggregated internal score and external score for each element should
be calculated:
Xnk Xnk Xnk
TIBMEk ¼ ð i¼1
a IBMEi ; i¼1
b IBME i ; i¼1
c IBMEi Þ ð5Þ
Xnk Xnk Xnk
TEBMEk ¼ ð i¼1
a EBMEi ; i¼1
b EBMEi ; i¼1
c EBMEi Þ ð6Þ
Where k = 1, 2, … 9.
There are different defuzzification method for fuzzy numbers. The lambda-cut
method, weighted average method, maxima methods, and centroid methods. The
authors in [19] used an alpha-cut plane for addressing the final strategy in the SWOT
matrix. According to the authors, the prioritization in the SWOT matrix will be defined
based on a assigned by the strategist. Hence the fuzzy area will be defined based on the
a value. For example, Fig. 2 shows the different quadrants fuzzy area based on the
value of a.
Fig. 2. The quadrants of fuzzy areas based on a value (adopted from [19])
Based on the a-cut defuzzification method, the related quadrant of TIBMEk and
TEBMEk will be mapped on the business model canvas (Fig. 3).
When the aggregated fuzzy number for external and internal factors of each
business model element is determined, a multistage fuzzy Mamdani interface will be
used. The three value dimensions proposed by the authors in [24] are used in this study.
Customer segments and customer relationships are related to the value proposition
aspect. Key activities, key resources, channels, and key partners address the value
creation and delivery dimension, and cost and revenue are related to value capture.
Figure 4 shows the model.
Fig. 3. TIBMEk and TEBMEk of the business model elements
4 Application
4.1 EoL Aircraft Recycling Business Model

In the coming years, considering the number of aircraft at the EoL, the aviation industry
faces a new challenge. Moreover, the airlines or owners may decide to withdraw the
aircraft before the end of operational life for sustainability or reducing the operational
costs [25]. Hence, the number of initiatives that work on the recycling of the aircraft is
increasing. Most of these companies work on disassembly, parts, and recycling. The core
of the business model of the EoL aircraft recycling company is the value proposition of
recovered parts and materials. There are two revenue streams: parts and materials. There
are different parts after disassembly operation: serviceable component (after removal),
serviceable component (after shop visit), salvageable component, un-salvageable com-
ponent, rotatable component, consumable parts, life-limited components, EoL compo-
nents or subassembly parts. Estimation of the market size and the market value is essential
as the model and commonality of the recovered parts determine the value. The main
activities include decommissioning, de-polluting, disassembling, dismantling, cutting,
parts inventory management, and logistics. The key resources include skilled workers for
different sub-processes of treatment, related tools, and equipment, dismantling sites,
information management systems, parts warehouse, and repair shop. The key partners
include OEM, distributors, parts dealers, brokers, airlines, regulatory bodies, authorities,
3PL companies, shredders, and disposal sites. The customers could be the leasing com-
panies, airlines, or owners for selling the aircraft or buying the recovered parts. The
channels are also related to these two target customers. Aircraft on ground desks, part
brokers, and MRO are the main channels for selling the recovered parts.
Fig. 4. The multistage Mamdani fuzzy system for mapping SWOT matrix
802 S. Keivanpour
4.2 Perspectives of Industry 4.0 in EoL Aircraft Problem Context

There are three types of complexities for recycling of the complex products: value
chain design, optimizing operations and processes and dealing with uncertainties.
Flexibility, customization, decentralization, and efficiency are the most important
triggers of Industry 4.0 [26]. Here, the perspective of adoptions of four key elements of
Industry 4.0 including data analytics, blockchain, digital models, and automated robots
will be discussed. One of the challenging areas in aircraft recycling is optimizing the
dismantling and disassembling process. Considering the number of parts and com-
plexity of the mixed materials, a data-driven optimization model is needed to propose
the sequence of dismantling. Selling the recovered parts require the visibility and
traceability of the parts. Blockchain technology is promising in dealing with this
tracking capability. Comparing the different dismantling and disassembling strategies
require the aggregation of several dispersed databases. Digital mapping and augmented
reality provide the opportunity of flexible mapping of the skeleton based on the dif-
ferent material zones. The cutting process is one of the critical sub-processes in EoL
aircraft treatment. Cobots could improve edge preparation and optimize the cutting
process for minimizing the debris and wastes and aid the sustainability of the operation.
4.3 The Internal and External Factors and Assigning Fuzzy Numbers
In this study, the internal and external factors are derived from the recent literature of
drivers and barriers of Industry 4.0 adoption [27–32]. Then, these factors are adapted
for the EoL aircraft problem context. The triangular fuzzy number is assigned to these
factors based on the existing literature on EoL aircraft and the author’s experience. The
scores are range from −5 to 5. The internal factors related to the value proposition, key
activities, and resources are presented in Table 1. The example of a fuzzy interface for
aggregation of value proposition internal factors is shown in Fig. 5.
4.4 Results and Discussion

The results of the model in three value dimensions are shown in Fig. 6. The results
show that there are many more opportunities for value proposition in the adoption of
Industry 4.0. However, the complexity of changing the current processes and the
required resources particularly the knowledge and skills make it a challenging problem.
Moreover, the current revenue and costs model should be adapted for this new context.
The results are also aligned with the existing literature. The strengths and opportunities
of the value proposition by Industry 4.0 are considered more rather than the challenges
of application, adoptions of current processes, standards, legal framework, and
resources. These are weaknesses and threats that should be addressed precisely. Fig-
ure 7 presents the output of the model based on aggregated internal and external
factors. Three areas based on three scenarios of the most probable, worst-case, and
best-case with the changes in the membership functions are provided. This strategic
tool provides the opportunity for testing the impact of different internal and external
factors and facilitates the decision-making process considering the uncertainty,
imprecision, and complexity of multi-factors strategic planning context.
Table 1. Internal factors for three elements of the business model

Activities Value propositions Resources
Internal factors Factor Internal factors Factor Internal factors Factor
condition condition condition
Compatibility with the (−4, −2, −1) Sustainability (2, 3, 4) Need for new skills and (−3, −2,
existing competencies −1)
production/operation
processes
Compatibility with the (−4, −2, −1) Improving (2, 3, 4) Need for Creating an (−3, −2,
existing standards business open-minded and −1)
models flexible corporate culture
Ethics consideration (−3, −2, −1) Improving (2, 3, 4) Compatibility with the (−4, −2,
value chain existing infrastructure −1)
Health and safety issues (−2, 0, 1) Improving (3, 4, 5) Compatibility with the (−4, −2,
performance existing legal structure −1)
Lack of employee (−3, −2, −1) Responsiveness (3, 4, 5) Lack of knowledge (−4, −2,
awareness about Industry 4.0 −1)
Managing complexities (−3, −2, −1) Flexibility (3, 4, 5) Lack of new technology, (−4, −2,
materials, and processes −1)
Delivery time (2, 3, 4) Poor R&D on Industry (−4, −2,
4.0 adoption −1)
Traceability (2 3, 4) Poor data quality (−4, −2,
−1)
Transparency (2, 3, 4)
Fig. 5. The input variables of the value proposition (Input factors) and the rule viewer
Fig. 6. SWOT matrix for a value proposition, value creation/delivery and value capturing
804 S. Keivanpour
Fig. 7. Three fuzzy areas for SWOT analysis based on the changes in membership functions.
5 Conclusion
In this paper, a joint application of fuzzy SWOT analysis and business model canvas is
proposed for analyzing the internal and external factors that contributed to the adoption
of Industry 4.0 in the recycling of EoL aircraft. From a theory point of view, con-
sidering a fuzzy SWOT analysis with the business model canvas and using the mul-
tistage fuzzy logic interface based on the value proposition, creation/delivery, and
value capturing is a novel approach. The internal and external factors are considered
based on the elements of the business model and aggregation is addressed via the fuzzy
interface. Considering the uncertainty of perception via fuzzy numbers and mapping
the SWOT matrix based on the value creation model provides a reliable tool for
strategic planning analysis. From the practical implication point of view, recycling of
the EoL aircraft in the aviation industry and the role of Industry 4.0 in the circular
economy are the topics that require further research. A fuzzy group decision making is
suggested for future research to consider the panel of decision-makers.
References
1. Tseng, M.-L., et al.: Circular economy meets industry 4.0: can big data drive industrial
symbiosis? Resour. Conserv. Recycl. 131, 146–147 (2018)
2. Lin, K.-Y.: User experience-based product design for smart production to empower industry
4.0 in the glass recycling circular economy. Comput. Ind. Eng. 125, 729–738 (2018)
3. Wang, X.V., Lihui, W.: Digital twin-based WEEE recycling, recovery and remanufacturing
in the background of Industry 4.0. Int. J. Prod. Res. 57(12), 3892–3902 (2019)
4. De Sousa Jabbour, A.B.L., et al.: Industry 4.0 and the circular economy: a proposed research
agenda and original roadmap for sustainable operations. Ann. Oper. Res. 270(1–2), 273–286
(2018)
5. Stock, T., Günther, S.: Opportunities of sustainable manufacturing in industry 4.0.
Procedia CIRP 40, 536–541 (2016)
6. Rahman, S.M., Perry, N., Müller, J.M., Kim, J., Laratte, B.: End-of-Life in industry 4.0:
ignored as before? Resour. Conserv. Recycl. 154, 104539 (2020)
7. Liu, X.L., Wang, W.M., Guo, H., Barenji, A.V., Li, Z., Huang, G.Q.: Industrial blockchain
based framework for product lifecycle management in industry. Robot. Comput. Integr.
Manufac. 63, 101897 (2020)
8. Dev, N.K., Shankar, R., Qaiser, F.H.: Industry 4.0 and circular economy: operational
excellence for sustainable reverse supply chain performance. Resour. Conserv. Recycl. 153,
104583 (2020)
9. Garrido-Hidalgo, C., Olivares, T., Ramirez, F.J., Roda-Sanchez, L.: An end-to-end Internet
of Things solution for reverse supply chain management in industry 4.0. Comput. Ind. 112,
103127 (2019)
10. Garrido-Hidalgo, C., Ramirez, F.J., Olivares, T., Roda-Sanchez, L.: The adoption of Internet
of Things in a Circular Supply Chain framework for the recovery of WEEE: The case of
Lithium-ion electric vehicle battery packs. Waste Manag. 103, 32–44 (2020)
11. Keivanpour, S., Kadi, D.A.: Perspectives for application of the internet of things and big data
analytics on end of life aircraft treatment. Int. J. Sustain. Aviat. 4(3–4), 202–220 (2018)
12. Keivanpour, S., Kadi, D.A.: Internet of Things enabled real-time sustainable end-of-life
product recovery. IFACPapersOnLine 52(13), 796–801 (2019)
13. Tam, E., Soulliere, K., Sawyer-Beaulieu, S.: Managing complex products to support the
circular economy. Resour. Conserv. Recycl. 145, 124–125 (2019)
14. Knieke, C., et al.: Predictive and flexible Circular Economy approaches for highly integrated
products and their materials as given in E-Mobility and ICT. Mater. Sci. Forum 959 (2019).
Trans Tech Publications Ltd
15. Andrews, K.R., David, D.K.: The Concept of Corporate Strategy, vol. 101. Irwin,
Homewood (1987)
16. Porter, M.E.: Competitive Strategy: Techniques for Analyzing Industries and Competitors.
Simon and Schuster, New York (2008)
17. Osterwalder, A., Pigneur, Y.: Business Model Generation: A Handbook for Visionaries,
Game Changers, and Challengers. Wiley, Hoboken (2010)
18. Joyce, A., Paquin, R.L.: The triple layered business model canvas: a tool to design more
sustainable business models. J. Clean. Prod. 135, 1474–1486 (2016)
19. Taghavifard, M., et al.: An extension of fuzzy SWOT analysis: an application to information
technology. Information 9(3), 46 (2018)
20. Hosseini-Nasab, H., Hosseini-Nasab, A., Milani, A.S.: Coping with imprecision in strategic
planning: a case study using fuzzy SWOT analysis. IBusiness 3(01), 23 (2011)
21. Amin, S.H., Razmi, J., Zhang, G.: Supplier selection and order allocation based on fuzzy
SWOT analysis and fuzzy linear programming. Expert Syst. Appl. 38(1), 334–342 (2011)
22. Sevkli, M., et al.: Development of a fuzzy ANP based SWOT analysis for the airline industry
in Turkey. Expert Syst. Appl. 39(1), 14–24 (2012)
23. Tavana, M., et al.: An integrated intuitionistic fuzzy AHP and SWOT method for
outsourcing reverse logistics. Appl. Soft Comput. 40, 544–557 (2016)
24. Nußholz, J.L.K.: A circular business model mapping tool for creating value from prolonged
product lifetime and closed material loops. J. Clean. Prod. 197, 185–194 (2018)
25. Keivanpour, S.: An integrated approach to value chain analysis of end of life aircraft
treatment. Diss. Université Laval (2015)
26. Lasi, H., Fettke, P., Kemper, H.-G., Feld, T., Hoffmann, M.: Industry 4.0. Bus. Inf. Syst.
Eng. 6(4), 239–242 (2014)
27. Kamble, S.S., Gunasekaran, A., Sharma, R.: Analysis of the driving and dependence power
of barriers to adopt industry 4.0 in Indian manufacturing industry. Comput. Ind. 101, 107–
119 (2018)
28. Khan, A., Turowski, K.: A survey of current challenges in manufacturing industry and
preparation for industry 4.0. In: Proceedings of the First International Scientific Conference
“Intelligent Information Technologies for Industry” (IITI 2016). Springer, Cham (2016)
29. Luthra, S., Mangla, S.K.: Evaluating challenges to Industry 4.0 initiatives for supply chain
sustainability in emerging economies. Process Saf. Environ. Protect. 117, 168–179 (2018)
806 S. Keivanpour
30. Müller, J.M., Kiel, D., Voigt, K.-I.: What drives the implementation of Industry 4.0? The
role of opportunities and challenges in the context of sustainability. Sustainability 10(1), 247
(2018)
31. Ng, H.S.: Opportunities, Challenges, and Solutions for Industry 4.0. Business Management
and Communication Perspectives in Industry 4.0, pp. 32–51. IGI Global (2020)
32. Veile, J.W., et al.: Lessons learned from Industry 4.0 implementation in the German
manufacturing industry. J. Manuf. Technol. Manag. (2019)
Knowledge Data Discovery (Frequent Pattern
Growth): The Association Rules for Evergreen
Activities on Computer Monitoring
Fauzan Asrin1,2 , Saide Saide3,4(&) , Silvia Ratna5 ,

and Alex Wenda6
1
Department of Information System,
Tanjungpura University, Pontianak, Indonesia
2
Pasca Sarjana Ilmu Komputer,
Universitas Putra Indonesia YPTK Padang, Padang, Indonesia
3
Department of Information System (PRO-Knowledge Research Group),
State Islamic University of Sultan Syarif Kasim Riau (UIN SUSKA Riau),
Pekanbaru, Indonesia
[email protected]
4
Department of Information Management, National Taiwan University of
Science and Technology (NTUST), Taipei, Taiwan
5
Department of Computer Science, Islamic University of Kalimantan,
Banjarmasin, Indonesia
6
Department of Electrical Engineering, Faculty of Science and Technology,
State Islamic University of Sultan Syarif Kasim Riau (UIN SUSKA Riau),
Pekanbaru, Indonesia
Abstract. The aim of this research paper is to construct a set of guidelines that
can improve the quality and efficiency of knowledge data discovery process by
utilizing various types of knowledge domains. In addition, this paper offered the
way of how the knowledge domain could be adopted for helping the system
developer. The methodologies contain various scenarios of data exploring and
the authors used data mining approach. The paper shows evidence of important
to adopt data mining methods in the industry sector as well as the advantages
and disadvantages. Evergreen human machine interface (HMI) at PT. Chevron
Pacific Indonesia (CPI) is kind of activities to maintenance computer equipment.
Nowadays, the errors were frequently happened in the accuracy of computer
maintenance which has a profound effect on production results. Therefore, this
study focuses on the rules of association using the frequent pattern growth
algorithm (FP-growth) which is producing knowledge with trust value of 100%
and a support value is 95%. The value results of support and confidence are the
new approach and knowledge for the management level to decide decisions in
the evergreen activities process.
Keywords: Knowledge data in discovery Association rules Algorithm FP

growth Evergreen activities HMI (Human machine interface)
808 F. Asrin et al.
1 Introduction
The process of knowledge discovery in databases (KDD) was traditionally presented as

a sequence of operations which applying iteratively, lead from the raw input data to
high level, interpretable, and useful knowledge. Knowledge discovery in databases
(KDD) is an automatic, exploratory, and modeling of large data repositories [1, 2]. This
concept is the organized process of identifying validity, useful, and understandable
patterns from large and complex data sets. The major steps in KDD process typically
such as selection, preprocessing, transformation, data mining, and interpretation or
evaluation [2, 3]. Chevron Pacific Indonesia (CPI) is a company engaged in the
exploration of oil and gas [4]. CPI operates in the block area of Rokan Sumatera Riau
province. The success of exploration activities is not freelancing work with the role of
information technology that supports companies’ activities. Information technology
(IT) as an instrumental to facilitate the workers in receiving information [3, 4] and
monitoring activities of the special oil and gas explorations in the area of Duri city
(Riau Province).
Monitoring generally refers to the observation, regulation, control, and reporting of
processes, procedures, team work, and persons [6, 7]. Monitoring concerns about
employee privacy and activities; therefore, employers must find a balance between
monitoring gains and employee activities [8, 9]. One of IT’s roles is to carry out
evergreen activities. The function of evergreen activities is for HMI personal computer
(PC) or HMI treatment. HMI PC is working to facilitate the operator of monitoring
directly from control room. HMI PC used in oil exploration field in the area of Rokan
block amounted to approximately 180.
Data mining is the analysis of (often large) observational data sets to find unsus-
pected relationships and to summarize the data in novel ways that are both under-
standable and useful to the data owner [10, 11]. The two terms have different concepts
but are related to each other [12]. Data mining is part of knowledge discovery in
databases process, consisting of stages such as data selection, pre-processing, trans-
formation, data mining, and evaluation of results [13, 14]. Data mining brings together
techniques from machine learning, pattern recognition, statistics, databases, and visu-
alizations to handle information retrieval problems from large databases [13, 15]. Data
mining is not stand-alone technology. It is an important step in the KDD mainly
relating to the extraction and calculation of the patterns of the data being sliced [5, 6].
In this paper, the authors focus how to demonstrate the improvement of the entire
KDD process by using knowledge domains in several phases. Problems that occur at
Chevron Pacific Indonesia is the difficulty of CPI to explore the rules of knowledge and
see what tendencies exist in evergreen activities carried out on PC HMI. Thus, in this
study, we tested whether the Association Rule Algorithm FP-Growth method can be
used to solve problems in evergreen activity data. The results of this study are expected
to help leaders at PT. Chevron Pacific Indonesia about how to develop computer
maintenance policies more appropriate. This study is structured as follows: introduc-
tion, literature review, methods, results and analysis, and conclusion.
Knowledge Data Discovery (Frequent Pattern Growth) 809
2 Literature Review
Knowledge discovery in databases refers to the use of methodologies from machine

learning, pattern recognition, statistics, and other fields to extract knowledge from large
data, where the knowledge is not explicitly available as part of the database structure
which are includes concepts, concept interrelations, classifications, decision rules, and
other patterns of interest [2, 7–9]. The knowledge discovery procedure is consisting of
nine steps (see Fig. 1). Thus, it is required to understand the process and the different
needs and possibilities in each step.
Fig. 1. The procedure of knowledge discovery in databases [1]
(1) Developing understanding. This is the initial preparatory step to sets the scene
for understanding what should be done [16]. (2) Selecting and creating a data set. This
process is very important because the data mining learns and discovers from the
available data. (3) Pre-processing and cleansing. At this step, data reliability is
enhanced. (4) Data transformation. Methods here include dimension reduction, such as
feature selection, and extraction, and attribute transformation [14]. (5) Choosing the
appropriate data mining task. Prediction is often referred to as supervised data mining
while descriptive includes the unsupervised and visualization [17]. (6) Choosing the
data mining algorithm. This stage includes selecting the specific method for searching
patterns and multiple inducers [3, 16]. (7) Employing the algorithm. In this phase needs
to employ the algorithm several times until a satisfied result is obtained. (8) Evaluation.
In this step, evaluating and interpreting the mined patterns (rules, reliability etc.).
(9) Using the discovered knowledge. The knowledge becomes active in the sense that
we may make changes to the system and measure the effects [5, 18].
3 Methods
This research methodology describes step by step the process of this research. The
methodology was conducted with systematically which can be used as a guideline
about how to conduct the research. (1) Problem analysis. A problem analysis inves-
tigates a situation/problem in order to allow the authors to understand more fully the
problem and to recommend practical solutions for solving it. (2) Determining research
810 F. Asrin et al.
objectives. At this stage, the targets are should be achieved, especially to overcome the
existing problems. (3) Reviewing references. The search strategy is aimed at finding
references from the best scientific papers such as SSCI/SCI, Elsevier Sciencedirect,
IEEE explore and other related links [7, 8]. (4) Collecting evergreen activity data. The
more data is collected, the better rules are being able to solve the problem. Then, to test
whether the data result is able to produce the output as planned at the time of inserting
the evergreen data that has been transformed at the first stage. (5) Implement FP growth
algorithm. After that, based on the results, it will provide and generate the rules for
leadership and management level of the company about how to create a strategy for the
evergreen activities further. FP-Growth can be divided into three main parts such as
conditional pattern base, conditional FP-tree, and frequent itemset [11, 19, 20]. (6) The
test results stage. At this stage, the authors start to develop the testing process.
4 Results and Analysis
At this step, there will be several calculations, testing, processing, and analyzing data
based on existing data. The data was collected from the department of IT process
control network (PCN) of Chevron Pacific Indonesia in 2018 (chapter: Duri city). The
data is expected can help during testing and implementation of the association rule with
FP growth algorithm. Also, can provide the output (knowledge or information) and
show the trend of evergreen activities in Chevron Pacific Indonesia. Then, the data will
be grouped, compiled, and transformed into tabular data which will be processed by the
software of Rapidminer 9.0. In the following is the process:
4.1 Defining Variables

In determining the association rules, the evergreen data and activities variables were
adopted. The variables such as backup variable, update anti virus variable, inventory
variable, user variable, technical control variable, event-log variable, and restore test
variable.
4.2 Data Transformation

At this stage, the transformed data will be grouped by the variables of the evergreen
activity in every existing computing ID. The variables include backup activities,
updating anti-virus, inventory, event log, OS user, technical control, and restore test.
Grouping the data is the result of the analysis of variables. All of those variables
presented at the Table 1 below.
Table 1. Tabular data

Computing ID Evergreen activities
DRIA13NOE01 Backup, Eventlog, Inventory, OS_User, Restore, TC, AV_Update
DRIGBSPI04 Backup, Eventlog, Inventory, OS_User, Restore, AV_Update
DRIA13RSOH01 Inventory
DRIA13SOE01 Backup, Eventlog, Inventory, OS_User, Restore, AV_Update
DRIAWT4SWH01 Backup, Eventlog, OS_User, Restore, AV_Update
DRISLNMWTH01 Backup, Eventlog, Inventory, OS_User, Restore, AV_Update
BKSAMPGSH01 Backup, Eventlog, Inventory, OS_User, Restore, TC, AV_Update
DMIHCTPI01 Backup, Eventlog, OS_User, Restore, AV_Update
DRIHCTPIPEH01 Backup, Eventlog, Inventory, OS_User, Restore, TC, AV_Update
DRIHCTMONH03 Eventlog, Inventory, OS_User, TC, AV_Update
DMIHCTMTRH05 Eventlog, Inventory, OS_User, Restore, AV_Update
DRIHOSS6E01 Eventlog, Inventory, OS_User, AV_Update
LBOLBOGPH01 Eventlog, OS_User, AV_Update
BKOBNRGSH01 Backup, Eventlog, Inventory, OS_User, Restore, TC, AV_Update
PGRPGRGPI01 Backup, Eventlog, Inventory, OS_User, Restore, TC, AV_Update
PGRPGRGSE01 Backup, Eventlog, Inventory, OS_User, Restore, TC, AV_Update
PGRPGRGSH01 Backup, Eventlog, Inventory, OS_User, Restore, TC, AV_Update
BKOBLMGSH01 Backup, Eventlog, Inventory, OS_User, Restore, AV_Update
BKOBLMGSE01 Backup, Eventlog, Inventory, OS_User, Restore, AV_Update
PGRPGRGPH01 Backup, Eventlog, Inventory, OS_User, Restore, TC, AV_Update
4.3 The Item Appears

Before designing the FP growth algorithm and developing the pattern of the association
rules, it is required to do a calculation of each variable that appears in the computing ID
in order to calculate the value of support and confidence of each evergreen activities
variable such as Backup (B)(15), Eventlog (E)(19), Inventory (I)(17), OS User (OU)
(19), Restore (R)(16), Technical Control (TC)(11), AV update (AV)(19).
4.4 FP Tree Formation

The establishment of FP tree from the transaction data that transformed into tabular
data with a transaction amount of 20 evergreen. Figure 2 below presents FP tree
formation.
812 F. Asrin et al.
Fig. 2. TID 20
4.5 FP Growth Implementation

Once the development of FP tree from the set of transactions have been done, then
proceed with the FP growth stage to find a qualified and frequently of itemset. The FP
growth has 3 main steps: conditional pattern base, FP tree conditional generation, and
finding the frequent itemset.
a. The conditional pattern base step is looking for the minimum support at FP tree
according to frequency of the path ending with the smallest support value namely
Eventlog (E), OS User (OU), and AV Update (AV).
b. Based on the FP tree conditional from AV process, the resuls show that AV, OU,
and E meet minimum support.
c. Finding the frequent itemset. In Table 2, the results show that not all itemset can be
calculated because the formula rule is if A then B. Therefore, the itemset can only
calculated if they have contained two items at least. In the following is a confidence
calculation of 9 subsets i.e. (AV, OU, E):
Table 2. Frequent itemset result

Suffix Frequent itemset
AV (AV, (AV, E), (AV, OU), (AV, E, OS), (AV, OS, E), (AV, E, OS))
OU (OU, (OU, AV), (OU, E), (OU, AV, E))
E (E, (E, AV), (E, OU), (E, AV, OU), (E, AV, OU))
19 19
AV ! E ¼ 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
AV ! OU = 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
AV ! E^ OU = 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
AV^ E = 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
AV^ OU ! E ¼ 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
OU ! AV ¼ 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
OU ! E ¼ 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
OU ! AV^ E ¼ 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
E ! AV ¼ 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
E ! OU ¼ 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
E ! AV^ OU ¼ 100% ¼ 100% Support 100% ¼ 95%
19 20
19 19
E^ OU ! AV ¼ 100% ¼ 100% Support 100% ¼ 95%
19 20
814 F. Asrin et al.
Once the rules of the frequent itemset has been created, the next step is developing
knowledge with the association rules. Then, the results of the association rules for each
itemset presented as Table 3 below:
Table 3. Association rules

Itemset Association rules
AV ! E If update is Anti-Virus (AV) then definitely take an Eventlog (E) computer
AV ! OU If update Anti-Virus (AV) then definitely do a review log Operating User
(OU)
AV ! EÔU If update is Anti-Virus (AV) then definitely do a review log Operating User
(OU) and take Eventlog (E) computer
AVÊ ! OU If the Anti-Virus (AV) updates and retrieves the Eventlog (E) of the
computer, it will definitely perform an Operating User (OU) review log
AVÔU ! E If the Anti Virus (AV) updates and does the Operating User (OU) review
log, it will definitely take an Eventlog (E) computer
OU ! AV If do an Operating User (OU) review log, you will definitely update your
Anti Virus (AV)
OU ! E If do an Operating User (OU) review log, you will definitely take the
computer’s Eventlog (E)
OU ! AVÊ If do an Operating User (OU) review log, you will definitely take the
computer’s Eventlog (E) and update your Anti Virus (AV)
E ! AV If take the Eventlog (E) of the computer, you will definitely update your
Anti Virus (AV)
E ! OU If take the Eventlog (E) of the computer, you will definitely log the
Operating User (OU)
E ! AVÔU If take an Eventlog (E) computer, you will definitely update your Anti Virus
(AV) and do a log review Operating User (OU)
EÔU ! AV If take the Eventlog (E) of your computer and log the Operating User
(OU) review, you will definitely update your Anti Virus (AV)
5 Conclusions
Regarding the design analysis and test result of evergreen data with the application of
rapid miner 9.0 at Chevron Pacific Indonesia Company (Department of IT Process
Control Network), we conclude that the application and implementation of association
rules with the FP Growth algorithm showing the trend of evergreen activities both
manual calculation and using Rapid Miner 9.0 application. The evergreen activity
trends including event-log (E) computers, operating user (OU) reviews, and updating
anti-virus (AV) with a support value of 95%. Based on the results of the analysis with
the FP growth algorithm and testing with the application of rapid miner Studio 9.0, the
association rules on simulation has larger data. Lastly, we have acknowledged that
future studies could be trying this approach into manufactures sector.
Acknowledgements. The authors would love to say “Alhamdulillah” and thanks to Allah SWT
who has given grace and guidance to the authors, to the Dept. Information System, Tanjungpura
University, Indonesia and to the PRO-Knowledge Research Group, Indonesia. We also would
like to thank the Committees and Reviewers for their critiques and revision recommendations.
References
1. Maimon, O., Rokach, L.: Introduction to knowledge discovery in databases. In: Data Mining
and Knowledge Discovery Handbook, pp. 1–17 (2006)
2. Frawley, W.J., Piatetsky-Shapiro, G., Matheus, C.J.: Knowledge discovery in databases: an
overview. Am. Assoc. Artif. Intell. 13, 57–70 (1992)
3. Fayyad, U., Piatetsky-Shapiro, G., Smyth, P.: Knowledge discovery and data mining:
towards a unifying framework. In: American Association for Artificial Intelligence, pp. 82–
88 (1996)
4. Development of H.R.: Executive summary of Chevron Pacific Indonesia (2019)
5. Ding, Z., Saide, S., Siti Astuti, E., Muwardi, D., Najamuddin, N., Jannati, M., Herzavina, H.:
An adoption of acceptance model for the multi-purpose system in university library. Econ.
Res. Ekonomska Istraživanja 32, 2393–2403 (2019)
6. Nebeker, D.M., Tatum, B.C.: The effects of computer monitoring, standards, and rewards on
work performance, job satisfaction, and stress. J. Appl. Soc. Psychol. 23, 508–536 (1993)
7. Bhave, D.P.: The invisible eye? Electronic performance monitoring and employee job
performance. Pers. Psychol. 67, 605–635 (2014)
8. Moussa, M.: Monitoring Employee Behavior Through the Use of Technology and Issues of
Employee Privacy in America. SAGE Open. vol. 5, pp. 1–13 (2015)
9. Muwardi, D., Saide, S., Indrajit, R.E., Iqbal, M., Astuti, E.S., Herzavina, H.: Intangible
resources and institution performance: the concern of intellectual capital, employee
performance, job satisfaction, and its impact on organization performance. Int. J. Innov.
Manag. 24, 1–21 (2020)
10. Larose, D.T.: Data mining: Methods and models. Wiley, New Jersey (2006)
11. Hand, D.J.: Principles of data mining. In: Drug Safety, pp. 621–622 (2007)
12. Ben-Dov, M., Feldman, R.: Data mining and knowledge discovery handbook: text mining
and information extraction. In: Data Mining and Knowledge Discovery Handbook, pp. 801–
831 (2010)
13. Fayyad, U., Piatetsky-Shapiro, G., Smyth, P.: From data mining to knowledge discovery in
database. Am. Assoc. Artif. Intell. 17, 37–54 (1996)
14. Maimon, O., Rokach, L.: Data mining and knowledge discovery handbook. Springers (2010)
15. Sowmya, R., Suneetha, K.R.: Data mining with big data. IEEE Trans. Data Eng. 26, 97–107
(2014)
816 F. Asrin et al.
16. Padhy, N., Mishra, P., Panigrahi, R.: The survey of data mining applications and feature
scope. Int. J. Comput. Sci. Eng. Inf. Technol. (IJCSEIT) 2, 43–58 (2012)
17. Witten, I.H., Frank, E.: Data mining: practical machine learning tools and techniques with
java implementations. In: Data Mining: Practical Machine Learning Tools and Techniques,
vol. 31, pp. 76–77 (2002
18. Imielinski, T., Mannila, H.: A database perspective on knowledge discovery. Commun.
ACM 39, 58–64 (1996)
19. Saide, S., Astuti, E.S., Indrajit, R.E., Trialih, R., Diniaty, A., Dewi, F., Herzavina, H.: What
we give, we get back: Investigating the dimensions that influence knowledge sharing on
profit enterprise in Indonesia. J. Sci. Technol. Policy Manag. 10, 1047–1062 (2019)
20. Garg, R., Gulia, P.: Comparative study of frequent itemset mining algorithms Apriori and FP
growth. Int. J. Comput. Appl. 126, 8–12 (2015)
Enabling Big Data Analytics at Manufacturing
Fields of Farplas Automotive
Özgün Akın1(&), Halil Faruk Deniz1, Doğukan Nefis1, Alp Kızıltan1,

and Altan Çakır2
1
Farplas Otomotiv, 41400 Kocaeli, Turkey
[email protected]
2
Istanbul Technical University, 34467 Istanbul, Turkey
Abstract. Digitization and data-driven manufacturing process is needed for

today’s industry. The term Industry 4.0 stands for the today industrial digiti-
zation which is defined as a new level of organization and control over the entire
value chain of the life cycle of products; it is geared towards increasingly
individualized customer’s high-quality expectations. However, due to the
increase in the number of connected devices and the variety of data, it has
become difficult to store and analyze data with conventional systems. The
motivation of this paper is to provide an overview of understanding of the big
data pipeline, providing a real-time on-premise data acquisition, data com-
pression, data storage and processing with Apache Kafka and Apache Spark
implementation on Apache Hadoop cluster, and identifying the challenges and
issues occurring with implementation the Farplas manufacturing company,
which is one of the biggest Tier 1 automotive suppliers in Turkey, to study the
new trends and streams related to topics via Industry 4.0.
Keywords: Big data Apache Hadoop Apache Kafka Apache Spark

Apache Ni-Fi AI in manufacturing Industry 4.0 MQTT OPC
1 Introduction: Big Data in Manufacturing
In today’s modern world, companies need to grow in an increasingly competitive and

changing environment. In this manner, knowledge is a major source of competitive
advantage and value creation to the development of dynamic capabilities, and this is
reflected in organizational success. The last decade has witnessed massive progress in
the enterprises’ revenues that employed big data in their businesses. Accordingly, most
business activities take place in a big data environment, and in fact, big data are
required for many companies to survive [1].
Injection molding is a manufacturing process for producing plastic parts by injecting
plastic raw material into a mold. A wide variety of products are produced using injection
molding, which varies greatly in size, complexity and applications. According to Chen,
Nguyen, and Tai mentioned in [2] that there are two most common problems in injection
molding. The first problem is to derive the optimal process parameters for the initial
injection molding machine setting, considering several process variables to implement
effective manufacturing after product design. To derive optimal process variables, there
818 Ö. Akın et al.
is a process that causes significant losses to constantly adjust process parameters in trial
and error mode. The second problem is the difficulty in changing the value of the initial
process parameter during the injection molding process, especially when defects occur
during manufacturing. Currently, there is no method for quality enhancement of this
dilemma yet, and it is difficult to react to equipment failure and other unexpected pro-
duction problems. In addition, a fuzzy-based controlled classifier can be used in the
automotive to search, store and classify Big data.
Farplas is an automotive parts supplier company founded in 1968. The company
has been able to catch up with the technological trends and adopted the vision of
offering innovative solutions to the mobility ecosystem. Due to high-quality standards
of the company, a big data analytics project has been started recently. In the following
sections, the Farplas big data project pipeline is summarized: data ingestion, steps to
get data from the production line to a big data lake, data transformation and storage, the
method used to transform and store the data to prepare analytics and visualization, data
analytics, to extract meaningful things from this data, and data visualization, to visu-
alize these results.
2 Data Ingestion
The data collection phase is the first step to be achieved towards developing data-based
decision mechanisms and artificial intelligence-supported production systems. Col-
lecting data in various sizes and frequencies from the production points is of great
importance for more meaningful analysis [3]. Since the data sources in production are
not uniform, many difficulties are encountered during the data collection phase.
More sensors are involved, especially in the industrial field, to obtain smart
products, manufacturing equipment and production, big data applications have become
an essential element of the operation in the period of Industry 4.0 [4]. According to
data-based manufacturing, Farplas has deployed a pilot injection cell to test challenges
for intelligent production. Thus, the most suitable system solution is studied in the pilot
cell and the best solution would easily spread to all injection molding machines of the
company. Sensors are placed in various areas of the pilot injection machine at Farplas.
It is aimed to transfer the data created in the system to the big data platform, which will
be called as Flatform instantly.
The pilot injection cell at Farplas consists of the following seven components for
the analysis: a plastic injection molding machine, 6 axes robot, energy analyzer, water
collector, precision scale, ambient humidity, and temperature sensor. All these devices
use different protocols to communicate with systems, for example, Euromap63, which
is a corresponding protocol specific to plastic injection machines is used to get data
from plastic injection machines [5]. What’s more, the main obstacle of this protocol is
to make the data meaningful using complex parameter names, which are for different
brands and models of machines. The data received with the Euromap63 sensor is
carried to the main PLC board via ethernet. The energy analyzer, water collector and
ambient sensor are connected to the main PLC and parameters are defined using
machine name and model in OPC Software (Kepware). Afterwards, these parameters
are published from Kepware over MQTT protocol in Fig. 1.
Enabling Big Data Analytics at Manufacturing Fields 819
Fig. 1. The flow used for data ingestion in Farplas.
Subsequently, the data is transferred to the Flatform according to a certain rule. At

this point, Apache Ni-Fi is used in the Flatform to enable data transfer and create data
pipelines [6]. Ni-Fi, which consumes data with the Consume-MQTT processor, is
preferred to create all data pipeline in the Flatform. The conversion is made in JSON
format consumed by Ni-Fi. The transformed data is transferred to Kafka and the Kafka
distributes the data to HDFS and Elasticsearch and distributes the data to other
platforms.
3 Data Transformation and Storage
After data collection, the raw data must be transformed in data storage platforms due to
the huge volume, redundancy, vulnerability highlights of data structures. In this
respect, the application of data transformation starts with data cleaning, data integration
and data compression [7].
According to [8], Ni-Fi is utilized for integrating data from different data sources and
gives real-time control that performs ETL basic to oversee the movement of data between
sources and destinations. Using Ni-Fi, the data mobility, from HDFS and Elasticsearch,
is consistently integrated with up-to-date tools. Ni-Fi is data source with supporting
disparate and distributed sources of translating varying formats, schemas, protocols and
sizes such as machines, geo-location devices, files, log records and among others [9]. Ni-
Fi has several processors operated to data extraction, transformation and loading (ETL),
making available connectors for the file system in the Hadoop Cluster, Elasticsearch and
distributed messaging technologies such as Kafka.
Based on our prerequisites, proposed Apache Kafka, Apache Spark on Apache
Hadoop and Elasticsearch combination is the foremost appropriate big data pipeline.
The proposed system is implemented to monitor the injection molding process in real-
time as shown in Fig. 2. In this manner, data streaming and transmission are essentially
required in the proposed system. Following the results of the recent paper [10], it is
accepted that Spark and Kafka could be one of the best choices in terms of real-time
data streaming and transmitting. All this installation environment is coordinated to
work together, like a Hadoop ecosystem together with an In-Memory-Flow approach
with Apache Spark.
Fig. 2. Real-time distributed Big Data architecture and parquet transformation of raw data used
in Flatform.
Within the data storage layer, the large volume of collected and processed data
from the production line must be securely stored and integrated consistently. Fur-
thermore, mass productional data beneath the on-premise environment have high
prerequisites for data storage and preprocessing. Moreover, Hadoop cluster can support
the development and accomplish the parallel processing of large-scale data. As one
primary way to accomplish on-premise computing, Hadoop can build large-scale
cluster framework on common hardware and has advantages of low-cost, extensible,
proficient and dependable [11].
According to our tests in Farplas, about 40 TB data will be generated annually. For
this reason, data size can be a major concern. Therefore, to reduce the estimated size of
the data, Apache Parquet, is utilized. As seen in Fig. 2, Parquet is a machine-readable
columnar storage format available in the Spark + Hadoop ecosystem, which is strongly
supported by Spark, and it provides around 10–30 times more compressed than JSON
format.
4 Data Analytics
The point to be reached with the analytics layer is creating data-driven manufacturing
to make financial and soft gains. All data stored in the data lake and it is reachable from
the analytics layer in the Flatform. Analytics layer contains IDE applications, which are
Jupiter and Zeppelin to manage Spark and Python codes to build real-time data pro-
cessing and machine/deep learning models.
The output of the model is visualized in Kibana by sending it to the Elasticsearch
database via spark as often as desired. Farplas uses discrete manufacturing method with
over 80 injection molding cells and produces approximately 70 million parts per year.
Furthermore, some of the produced parts take the final shape after having assembly and
painting processes. The surface quality of the parts is very important in terms of
individual cost and company prestige. Therefore, the predictive quality model is
selected as the first use-case scenario.
Fig. 3. Correlation matrix between machine variables and sensor variables.
The first approach for the predictive quality application is the correlation matrix
method to examine production data and relations between parameters in real-time
observation, as seen in Fig. 3. Thanks to the correlation matrix relevance between
parameters which are produced in the cycle are visualized for each production cycle.
The expected situation is that the correlation matrix of each cycle is highly like the
others. Nonetheless, it is seen that some of the cycles have a different correlation matrix
when compared to the others. These cycles are easily labelled as a problematic cycle.
The second approach, each cycle is represented as vectors consisting of process
data collected for itself. Then, the distances of each cycle vector with the 10 vectors
before that cycle are calculated. When the similarities are displayed with colors, loops
with abnormalities are detected. Compared to the correlation application performed
above, it is not clearly shown which parameter the problem originates from, but since
the analyzed data is more, the detection accuracy rate of the application is higher.
The second application which is selected to detect anomalies in the production and
predict the quality of the part is the cosine similarity integration with collaborative
filtering. In this real-time data analysis approach, each vector represents the process
data collected during the cycle time. After that, the distances of each cycle vector with
the 10 cycle vectors before that cycle are calculated. When the similarities visualized
with colors, cycles that have anomaly are detected. However, this approach has not
clearly shown which parameter has the error yet.
In the third application, the correlation matrix application is combined with the
cosine similarity application and converted into a cosine matrix. In the cosine matrix,
the parameter values produced in each second for each machine cycle are defined as
vectors, and the distance of the parameter vectors within that cycle is found by the
cosine similarity method and converted into a matrix and colored. It is seen that when
the color distributions on the matrices of each cycle are examined, the color distri-
butions of the abnormal cycles are different from the others.
Fig. 4. Time series analysis of forecasting energy consumption for the next consecutive cycles.
In addition to the predictive quality application, another application that will enable
us to gain data is energy analysis. Based on the energy values consumed by the
machines, abnormalities in production can be detected, and optimization efforts can be
made to reduce energy consumption, thereby saving money from energy consumption.
The first application in Flatform is by collecting the energy data of the machine and
analyzing it with the machine learning time series methods. The abnormalities formed
by detecting the data outside these confidence intervals (see Fig. 4).
5 Data Visualization
In the flatform, open-source Elasticsearch and Kibana are used as visualization tools.
The data sent to the Elasticsearch database via Ni-Fi and then visualized in real-time
with the Kibana interface. The Elasticsearch installed on the cluster allows you to
search the data very quickly and to see the situations that occur within the desired date
range via Kibana. As seen in Fig. 5, the bottom of the picture indicates the visualization
of the selected pilot cell in Kibana. After the implementation of the pilot cell com-
pleted, the same process is applied to 10 injection machines (top of the Fig. 5) The
outputs machine learning and advanced analytics applications which are built by using
spark sent to Elasticsearch and visualized in Kibana to show valuable data to users.
Fig. 5. Every second, 400 variables are collected from 10 injection cell (top), and pilot cell
dashboard in Kibana.
6 Conclusion
In this paper, we describe the open-source big data ecosystem created in the Farplas
manufacturing company, which has a great benefit in efficiency. This system, which
can provide instant data flow from many systems, provides real-time on-premise data
acquisition. In this way, the problems can be observed instantly, and in some cases,
even a warning can be given before the problem occurs.
First, in the production layer data parameters such as machine name and model are
defined in the main PLC via OPC software and published using the MQTT protocol.
This data is then transferred to the Flatform big data platform. Apache Ni-Fi is used to
transfer published data from production layer to the Flatform and create data pipelines.
Then, data converted into JSON format on Ni-Fi is distributed to Elasticsearch, HDFS
and other platforms with the help of Kafka. Spark and Kafka could be good choices in
terms of the real-time data streaming and transmitting. All this installation environment
is coordinated to work together, like a Hadoop eco-system.
Analytics Layer of Flatform contains Jupiter Notebook and Zeppelin to manage
Spark and Python code to build artificial intelligence models. First predictive use-case
model is a predictive quality model. The output of the model is visualized in Kibana by
sending it to the Elasticsearch database via Spark as often as desired. In this way,
machine operators can see the cause of the error instantly and change the parameters of
the machine and prevent the production of more defective parts in the future.
In Farplas, using this big data infrastructure, abnormality in production are deter-
mined with the help of a cosine matrix of the energy values consumed by the machines.
Then, new parameters are obtained with optimization studies. With the help of new
machine parameters obtained with this method, energy consumption is reduced in
production.
Future research activities would tend to increase artificial intelligence applications
on this platform. AI models would be established with more parameters and solution
methods. Besides, the number of machines connected to the system can also be
increased in the next stage and connecting more system such as ERP in this platform.
References
1. Wilcox King, A., Zeithaml, C.P.: Measuring organizational knowledge: a conceptual and
methodological framework. Strat. Manag. J. 24(8), 763–772 (2003)
2. Chen, W., Nguyen, M., Tai, P.: An intelligent manufacturing system for injection molding.
Proc. Eng. Technol. Innov. 8, 9–14 (2018)
3. Mourtzis, D., Vlachou, E., Milas, N.: Industrial big data as a result of IoT adoption in
manufacturing. Procedia CIRP 55(1), 290–295 (2016)
4. Sangat, P., Indrawan-Santiago, M., Taniar, D.: Sensor data management in the cloud: Data
storage, data ingestion, and data retrieval. Concurr. Comput. 30(1), 1–10 (2018)
5. Ayer, L., Ayer, P.H.L.: Euromap 63 - Data Exchange Interface (2000)
6. Chanthakit, S., Keeratiwintakorn, P., Rattanapoka, C.: An IoT system design with real-time
stream processing and data flow integration in 2019 research. In: Invention, and Innovation
Congress (RI2C), pp. 1–5 (2019)
7. Dai, H.N., Wang, H., Xu, G., Wan, J., Imran, M.: Big data analytics for manufacturing
internet of things: opportunities, challenges and enabling technologies. Enterp. Inf. Syst. 27,
1–25 (2019)
8. Kim, S.S., Lee, W.R., Go, J.H.: A study on utilization of spatial information in
heterogeneous system based on apache NiFi. In: 10th International Conference on ICT
Convergence, pp. 1117–1119 (2019)
9. Ansari, M.H., Tabatab Vakili, V., Bahrak, B.: Evaluation of big data frameworks for analysis
of smart grids. J. Big Data 6(1), 1–14 (2019)
10. Ismail, A., Truong, H.L., Kastner, W.: Manufacturing process data analysis pipelines: a
requirements analysis and survey. J. Big Data 6(1), 1–26 (2019)
11. Da Xu, L., Duan, L.: Big data for cyber physical systems in industry 4.0: a survey.
Enterp. Inf. Syst. 13(2), 148–169 (2019)
Big Data Analytics Framework for Smart
City Real-Time Fuzzy Complex Event
Processing
Basmi Wadii1 , Azedine Boulmakoul1(B) , Lamia Karim2 , and Ahmed Lbath3

1
Casablanca, Morocco
2
LISA Lab. ENSA Berrechid, Hassan 1st University, Settat, Morocco
[email protected]
3
[email protected]
Abstract. Smart cities are urban areas that puts internet technology
in use to ease access for its citizens to its different services, using count-
less different types of electronic devices and applications that constitutes
what its known as Internet of things. It is a cyber fona where sensors and
actuators interacts to surface a virtual city of rich data, based on multi-
ple phenomenons of diverse nature - natural, industrial, urban and social
- occurring in its different areas, that emit meaningful events describing
their state as different intervals of time. We chose to call them in this
paper, fuzzy complex space time events, due to their fuzzy nature and
tight relationship with their geospacial position, time and their values
impredictible nature. Portunus is a system that aims to collect and dis-
tribute data over sensors and actuators, supporting MQTT and HTTP
to communicate over the network, additionally, it was designed as a
microservice based system, and uses NoSQL database systems to cop
up with the large data volume. However, data collection as a virtue is
not enough, but only the first step to benefit from this vague pool of rich
data.
Keywords: Big data · Fuzzy event · IOT · Smart city · Distributed

computing · Uncertainty
1 Introduction
For more than a decade, Internet of Things or IoT in short, kept and is still
growing and stretching its roots over all the industries, with uncountable number
of objects planted throughout different applications, provided by organizations,
institutions and companies. Besides, as part of plans to halt all automatable
services to increase efficiency, the daily interaction with these systems brings out
enormous chunks of data, as a result, in addition to monitoring and managing
826 B. Wadii et al.
the phenomena, the use of the collected data has helped to make them more
cost-efficient and rapidly converge to their ultimate performance.
In the literature, IoT is heavily linked to Smart City concept, as it forms
its backbone of information structure. A city defines itself as smart when all its
decision maker organizations make impactful changes in its services, by employ-
ing information technology applications to improve its citizens life quality, thus,
decreasing the necessary resources, time and improving its social and economi-
cal aspect. Moreover, it is a city that stand upfront challenges that comes with
urbanisation rate increase, natural resources strain and all sort of micro-problems
that happen within. It dates back to the 1990’s only as a phrase to promote urban
development following technology based solutions approach, then, continued to
appear and propagate till it became a paradigm that applies information commu-
nication technologies to enforce urbanization and loosen cities hard joints [10].
Also, the data collected by its IoT applications is big, versatile, rapidly generated
and juggle around veracity, thus, can be considered as Big Data [9]. However,
the fourth V i.e. veracity, is subject to many factors, such as the communication
channel, the health of the device in use, and other unexpected scenarios that can
alter data’s truth and accuracy, as we shall dig deeper in the following sections.
With the emergence of smart cities and the flourishing of IoT, the lat-
ter’s market has become full of smart devices ranging from sensors, actuators
and other complex systems combining multiple technologies. Depending on the
project, the choreography required to make them functional will vary in terms of
communication channels, operating application server that monitors the objects
interactions, health and data quality they provide over time. Additionally, the
size of the projects also keeps constant scaling and do interesting crossovers with
other technologies such as A.I., image and voice recognition, bringing to life what
could be the stepping stone to reach a 90’s sci-fi scenario. However, phenomenons
are not exclusive to natural or industrial types, but also, urban and social too.
In fact, the transformation which the world has seen in the last decades essen-
tially happened by exploiting interactions between people and services alike, as
they generate enormous data that targets its consumers themselves. We refer
to these type of events, as complex space-time events. In this paper, we aim to
present the concept of complex space-time events, an abstract layer that encap-
sulates generic - not necessarily micro-controllers - events captured by sensors,
its uncertain nature, and finally, and an architecture of a system that can be
able to manage the data flow, including its inescapable uncertainty, between all
connected objects/systems to benefit from different sources of data. The remain-
der of the paper is structured as follows. In the second section, we present the
uncertainty trait in sensors observations. Then in the third section, we present
the concept of complex space-time events, and they also showcase an uncertain
trait. In the fourth section, we confer Portunus as a framework that manages
these events and how it is possible to extend its components to overcome the
challenges that may deviate any process that use its data arsenal with the lease
imprecision and inaccuracy. Finally, in the last section, we conclude the advance-
ment and present our perspectives for our future research.
Real-Time Fuzzy Complex Event Processing 827
2 Sensor Observations Uncertainty

Smart cities are cities whom decision makers utilize technologies involving mod-
ern software applications, combined with smart micro-controller devices con-
nected to a large scale system in order to serve a specific purpose. Hence, reduc-
ing costs and time to perform some tasks which otherwise would take an eternity
or a fortune if done manually. However, the usage of electronic sensors comes
with trade offs, precisely, it is impossible to not take in account the uncertainty
that comes with the collected measurements occurring because of multitude of
reasons, namely, positioning systems, environment noises, hardware quality or
deficiency [8]. By definition, sensor’s uncertainty is defined as parameter, asso-
ciated with the result of a measurement, that characterises the dispersion of the
values that could reasonably be attributed to the measurand [5]. It originates
from the natural or industrial processes generating its data, which may suffer
from missing data o faults of sensors, errors arising in the transmission layer
of sensor networks or the probabilistic nature of their signals [4]. Sensor uncer-
tainty can be classified under two classes, uncertainty due to discrete sampling
(DS) and uncertainty due to Sampling Error (SE). Uncertainty due to DS is
where even if all discrete points are correct, the discrete sensor stream remains
uncertain, because some critical points might be missing due to the limited capa-
bilities of the hardware or the server that hosts the database. On the other side,
uncertainty due to Sampling Error (SE) happens when sensor readings cannot be
replicated, even if the situation is the same, in different times or using different
sensors, the result would be different alongside the sampling errors [6].
The most important factor in sensors’ uncertainty is its size surrounding the
analytical result, to make sure that systems using its reading are making cor-
rect decisions such as predictions of future observations or as basic ingredients
in actuators operations. Quite generally, the smaller the uncertainty, the less
incorrect or very costly are the decisions. For that purpose, there exists many
approaches to tackle and estimate the accuracy of measurements taken by sen-
sors, for both uncertainties classes. First, there exists many types of errors when
it comes to sensors and their measurements:
Systematic errors can happen due multitude of reasons, notably, calibration,
intrusion - where the sensor affects the phenomenon it observes or defects, defined
by average of measured values minus the true value. Whereas, a random error
is defined by the reading minus the average of the precedent readings [2].
3 Complex Space-Time Events Uncertainty
Smart city is by definition a city that employs IoT - sensors and actuators -
into generating a value that improve its key services, to improve the citizen’s
life quality. However, observations are not restricted to natural or industrial
phenomenons, but they are also happening on a very large surface everywhere,
in form of events fired by concepts with a higher level of abstraction, other than
electronic devices. They may be natural, social or urban, also, of quantitative or
828 B. Wadii et al.
qualitative nature, calling the different states of a city in different points of time.
Inspired by The OGC API standing for (Open Geospacial Open Consortium)
API [3], Fig. 1 data model to determine the different concept that surround
an observation of a phenomenon. To begin, a sensor is an entity that observes a
phenomena, which in fact, can be anything that emits readable events happening
in a specific time and in a geospacial feature, designating the state of a thing -
a geospacial positioned object, for example, an Arduino or Raspberry Pi board,
software program that reports train arrivals, crimes occurrences or accidents.
Correspondingly, an actuator receives those events, process and transform them
into a physical energy. Together, each can subscribe to a data stream - sort of
entity that binds related data - linked to one single feature of interest, i.e. an geo-
spacial occurrence, one or many at once and hosts diverse nature of collected
data stream. Furthermore, the event concept is the center of the data model,
as it represents a measurable natural, industrial, urban or social phenomenons,
that can be represented with a quantitative or qualitative reading, and saved to
represent the state of the object experiencing the phenomena, at a given time and
in a geospacial feature. Hence, with this solid rich heterogeneous connectivity,
the system grants an uncountable source of data tailored for well-established
decisions upon usage in analytical operations [1].
Fig. 1. Portunus entity relationship schema
Complex space-time event (CSTE) is a concept that adds an abstract layer

which covers physical readings, e.g, temperature, vehicles speed or air quality,
and also other readings, notably, coming from urban area or services, e.g, train
arrivals, pandemic safety infringement rules and traffic congestion occurrences.
The later events exhibit a measurable behaviour, whom data is already present
in accessible applications, and only need to be exposed publicly in the form of
a complex event as shown in the data model, with unit-less readings. Thus, the
definition of sensor can be extended to these types of data, and therefore com-
plex space-time events are also exposed to the same challenges of uncertainty
- fuzziness, but also, any of the presented patterns and methods to determine
common uncertainties in electronic sensors are present. For instance, readings

coming from non-natural sources such as system applications monitoring urban
services, are necessarily susceptible to disruptions due to network communica-
tion failure or external third party systems involved in the process. Therefore,
CSTE are also target to uncertainties, which in this case, is due to discrete sam-
pling, since the actuators consuming our system’s readings are not continuous.
Moreover, they also cover most of the natural phenomenons events recorded by
sensors, that are susceptible to slow technical degradation which can affect their
clock, their precision, accuracy or increasing their down times, thus, making
CSTE also inclined to uncertainty due sampling errors.
4 Fuzzy Complex Space-Time Events Data Collecting

Framework Portunus
In a previous work [1], Portunus is the name we gave to the system. It repre-
sents the framework that offers the multiple components constituting the system
which manages the data stream between multiple sensors and actuators of dif-
ferent kinds and monitor diverse phenomenons. It functions as a set of four
autonomous microservices communicating asynchronously and independently,
and each have their own set of functional boundaries. In fact, unlike monolith
systems, microservice architecture gives free loose coupling, technology agnostic
development - their implementation don’t have to use the same technology as
long as they provide a consistent interface, and the possibility to scale. Accord-
ingly, having two microservices responsible on receiving incoming sensors read-
ings and distributing this data over actuators respectively, aids to disperse the
heavy load, and make augmenting any of the units separately much easier as
long as the output remain the same, since they are different physical processes
running on top of different node or nodes [7].
Fig. 2. Portunus architecture layers
In Fig. 2, we have dissected its architecture into three layers, each microser-
vice has its own roles in the ecosystem. In details, the data circulation layer holds
830 B. Wadii et al.
the events collection which read observations with the same example’s structure
in Fig. 3 sent by CSTE sensors acknowledged by the system, then saves them
to its own database, and share them with the events distribution microservice.
Moreover, any observation sent is attributed by a data stream Id, which plays the
role of the entity that ties the observations, sensors, actuators, observed phenom-
ena, Correspondingly, the last microservice will take the saved event and share
it with all subscribed actuators. Yet, there will be two use case scenarios, which
the consumers should be aware of, is that they either need to implement a proxy
service to filter and apply corrections to mitigate the uncertainty factor in the
present data, and in a correlative manner increase the accuracy and precision
of the actuators input for quality output, regardless of their nature. For that
reason, Portunus architecture could make use of another microservice in its data
layer, that takes care of applying methods to overcome uncertainties in collected
data.
Fig. 3. Complex Space Time Observation Example
As the system scale, it is important to have a logging service that logs all
the operations that happen within, with an entry point readable and human
usable - in short, an application - to visualize the performance of the different
components, and whether there are any failures or not. Also, the register layer
holds a microservice responsible on archiving all the data that circulate and
have circulated in the data layer. It could play the role of the archive or simply a
database for every record, since most of the events circulating in the data layer
are semi-volatile, i.e. their existence in their database rest upon the need within
the time it is shared among the observers in the data stream. This was the sum-
mary of the microservices roles and features, as for their communication mode,
it is a hybrid of orchestrated and asynchronous choreographic communication,
and uses a message broker to exchange the data. Moreover, it is preferable to
use a cloud infrastructure to manage all the overhead that comes with managing
microservices, notably, their allocated resources, down times, units and the list
goes on.
Modern cloud services are now competing over attracting projects with thou-
sands of products that reduce the whole devops uncountable days of scripting
and monitoring servers in a friendly customisable dashboard and clicks. Amazon
offers many products that are themselves a career to be pursued that start from
mere virtual machines with dynamic resources and pay on the go, databases,
to systems that adds abstract layers to make shipping production applications
faster and easy to maintain. Amazon Elastic Container Service is one of these
products that requires a Dockerfile to prebuild a software with all its applications
peers, run it and with the possibility to control its resource usage plus the num-
ber of containers which acts like sandbox environments to run the application.
Docker is a technology that provides a sort of virtualization that leans on running
an application with a minimal Unix system, then, extended by other software
such as database systems, or other applications depending on its requirements
all through a configuration file to build a snapshot, called an “image”, then as
it runs, its instance becomes a container, running independently from its host,
unless configured otherwise. Although the cost might be higher than renting or
possessing a server, the overall maintenance and management of systems that
rapidly scale, especially in IoT, it becomes wiser to outsource that complexity to
such services, additionally, increasing or decreasing the resources can be achieved
with one button click, instead of having to go through the re-configuration of the
whole system, i.e. more down times to maintenance, loss of data, and therefore
relatively affecting negatively the quality of input provided by Portunus.
5 Conclusion
Finally, Smart cities are urban areas are niches that is rigged by many useful
events occurring in the presence of urban, social and natural phenomenons, on
different geospacial locations and that describe the continuous change of their
states. Sensors or actuators concepts are no longer limited to electronic gadgets,
but, are encompassed with modern applications and systems that can communi-
cate over common protocols to send their observations fitting the CSTE concept.
However, it also means that they are also prone to uncertainty much like their
peers. Therefore, any consuming system - actuator - must take in account this
possibility, and include it in its output. We have proposed Portunus, and we
described how as a Framework, it is able to provide an interface that connects
actuators and sensors to exchange and benefits from the data flow that is created
through data streams, additionally, we’ve shown how important it is to design it
as a microservice based system to increase the scalability factor and load loosen-
ing. Yet, we discussed how as much as it is possible to forward the responsibility
of uncertainty correction to actuators, it could be part of the system as well.
Additionally, there are many methods in the literature, which we consider intro-
ducing to that microservice and implement them in a manner so they can be
customized to fit the exact problematic, depending on the circumstances and the
events in question. Therefore, boosting the credibility of the whole system, since
its core acts around a circuit of data moving between sensors and actuators.
832 B. Wadii et al.
References
1. Basmi, W., Boulmakoul, A., Karim, L., Lbath, A.: Modern approach
to design a distributed and scalable platform architecture for smart
cities complex events data collection. Procedia Comput. Sci. 170, 43–
50 (2020). https://doi.org/10.1016/j.procs.2020.03.008. http://www.sciencedirect.
com/science/article/pii/S1877050920304300
2. Cimbala, J.M.: Errors and Calibration. Penn State University, August 2009
3. Consortium, O.G.: OGC SensorThings API (2017). https://www.opengeospatial.
org/standards/sensorthings. Accessed 17 Dec 019
4. Cuzzocrea, A.: Advances in managing, updating and querying uncertain and
imprecise sensor and stream databases. Inf. Syst. 38(8), 1184–1186 (2013).
https://doi.org/10.1016/j.is.2013.03.004. http://www.sciencedirect.com/science/
article/pii/S0306437913000446
5. Produced jointly with Eurolab, N., Committee, R.A.M.: Measurement uncertainty
arising from sampling A guide to methods and approaches. Eurachem (2019)
6. He, J., Zhang, Y., Huang, G., de Souza, P.: Circe: Correcting imprecise readings and
compressing excrescent points for querying common patterns in uncertain sensor
streams. Inf. Syst. 38(8), 1234–1251 (2013). https://doi.org/10.1016/j.is.2012.01.
003. http://www.sciencedirect.com/science/article/pii/S0306437912000063
7. Newman, S.: Building Microservices Designing Fine-Grained Systems. O’Reilly
Media Inc., Sebastopol (2015)
8. Ren, Q., Li, J., Liu, H.: Energy efficient tracking in uncertain sensor networks.
Ad Hoc Netw. 81, 45–55 (2018). https://doi.org/10.1016/j.adhoc.2018.07.011.
http://www.sciencedirect.com/science/article/pii/S1570870518304657
9. Sholla, S., Naaz, R., Chishti, M.A.: Semantic smart city: context aware application
architecture. In: 2018 Second International Conference on Electronics, Communi-
cation and Aerospace Technology (ICECA), pp. 721–724, March 2018. https://doi.
org/10.1109/ICECA.2018.8474777
10. Wenge, R., Zhang, X., Dave, C., Chao, L., Hao, S.: Smart city architecture: a
technology guide for implementation and design challenges. China Commun. 11(3),
56–69 (2014). https://doi.org/10.1109/CC.2014.6825259
Cloud Based Big Data DNS Analytics at
Turknet
Altan Çakır1 , Yousef Alkhanafseh2 , Esra Karabıyık2(B) , Erhan Kurubaş2 ,

Rabia Burcu Bunyak2 , and Cenk Anıl Bahçevan2
1
Istanbul Technical University, 34467 Sarıyer, Turkey
[email protected]
2
Turknet İletişim Hizmetleri, 34394 Şişli, Turkey
{yousef.alkhanafseh,esra.karabiyik,erhan.kurubas,
burcu.bunyak,cenk.bahcevan}@turknet.net.tr
Abstract. Domain Name System (DNS) is a hierarchical distributed

naming system for computers, services, or any resource connected to the
Internet. A DNS resolves queries for URLs into IP addresses for the
purpose of locating computer services and devices worldwide. As of now,
analytical applications with a vast amount of DNS data are a challenging
problem. Clustering the features of domain traffic from a DNS data has
given necessity to the need for more sophisticated analytics platforms and
tools because of the sensitivity of the data characterization. In this study,
a cloud based big data application, based on Apache Spark, on DNS
data is proposed, as well as a periodic trend pattern based on traffic to
partition numerous domain names and region into separate groups by the
characteristics of their query traffic time series. Preliminary experimental
results on a Turknet DNS data in daily operations are discussed with
business intelligence applications.
Keywords: Domain Name System (DNS) · Big data · Data analysis ·

Apache Spark · Amazon Web Service (AWS) · Amazon Elastic
MapReduce (EMR)
1 Introduction
The Domain Name System (DNS) is part of the core infrastructure of the Inter-
net. Measurement of analytical changes in the DNS traffic over time provides
valuable information about the evolution of business intelligence and predic-
tive maintenance operations for internet companies. We, Turknet, use a com-
plementary approach based on active measurements, which provides a unique,
comprehensive dataset on the evolution of DNS over time. Our cloud based
high-performance infrastructure platforms and integrated developments per-
forms Internet-scale active measurements, currently offline querying over of the
DNS name space on a daily basis. Our infrastructure is designed from the ground
up to enable big data analysis approaches on, e.g., a Apache Spark and Hadoop
cluster. With this novel approach we aim for a customer-oriented DNS-based
measurement and analysis of the Internet traffic at Turknet.
834 A. Çakır et al.
2 Analysis Tools
2.1 Apache Spark
Apache Spark is a parallel processing framework that is designed to perform a

series of tasks at the same time. It has several distinctive features that make
it preferable among other big-data solutions. High processing speed, ability of
recovering data, supporting various software development languages, and work-
ing in-memory computation model are some examples of Spark traits. Due to its
advantages, currently, it is almost used in all data engineering applications such
as network, banking, security, manufacturing, media, etc. Commonly, Apache
Spark structure consists of eight main components, which are Spark core, Spark
SQL, Spark streaming, machine learning libraries, graph processing, and cluster
management [1]. The prime operations of Spark for Big data engineering, includ-
ing elements for memory management, task scheduling, performance improve-
ment, and storage systems, and other, are located in Spark Core design. Fur-
thermore, spark core contains various APIs that determines immutable objects
(Resilient Distributed Datasets (RDDs), Dataframe and DataSets), which can
be considered as Spark’s heart components [2]. Spark SQL is useful with struc-
tured data and it permits querying data via SQL for structured datasets. On the
other hand, analyzing live streams of data can be achieved using spark streaming
including on-they-fly machine learning solutions. Lastly, the spark cluster man-
ager, which could be spark’s own standalone, YARN (Yet-Another-Resource-
Negotiator), or Mesos, specifies the jobs which will be achieved by the spark
cluster. Unlike traditional MapReduce methods, Apache Spark supports both
interactive queries and stream processing which gives it huge processing speed
while treating large set of data. As a consequence, several minutes or maybe
hours will be eliminated from processing data execution time when it is com-
pared to other traditional MapReduce methods in this analysis. Besides, Spark
operations and computations are run in memory not in disk which reduces the
risk of going disk to full usage while dealing with complex applications [3].
2.2 Processing Big Data with Computing Cluster: EMR
Amazon Elastic MapReduce (EMR) is one of the analytic services that Amazon
Web Service (AWS) provides as a platform-as-a-service (PaaS). EMR, basically,
hosts Apache Hadoop framework that is entirely built on instance and storage at
cloud platform, aka Amazon Elastic Compute Cloud (EC2) and Amazon Simple
Storage Service (S3). Thereby, there is no need to deal with the intricacy of
installing Apache Hadoop cluster or paying for it. In general, EMR is used as a
distributed computing environment (cluster) that number of its master and core
nodes can be scaled up or down on demand without the need to establish complex
structure of hardware [4]. EMR Hourly prices can be considered appropriate
when it is compared with its potential force and its prices ranges, such as other
cloud vendors [5]. EMR cluster is able to process several tasks of massive datasets
in parallel within few minutes as well.
Cloud Based Big Data DNS Analytics at Turknet 835
3 Data Analyzing Process

Traditionally, the general data flow diagram in EMR, as shown in Fig. 1, starts
with moving data to an AWS S3 bucket and ends with estimating the cost and
the performance of the process. In between, data collection, aggregation, and
process are accomplished.
Fig. 1. General data flow scheme.
3.1 System Architecture
The Structural Flow of Turknet DNS Data analysis is represented in Fig. 2. This
figure starts from the left side with Turknet DNS servers logs. In this figure, the
customers are represented as client1, client 2, client 3 and client n. Generally,
Turknet company infrastructure contains three different Domain Name Servers
(DNSs), which have the ability to collect consumers information, while they are
browsing various internet sites. Then, by using specific methods, Turknet user’s
data that is located in DNSs can be extracted and saved in log format files. Since
the extracted data from DNSs are considered as high volume (15 TB/month), it
is not plausible to use either internal nor external hard disks to save this massive
volume of data. For example, in order to save 15 TB of data, 4 storage devices
that each of which has the ability to store about 4 TB are required, which
will likely increase the opportunity of losing data. Fortunately, there is cloud
based S3 solution, which helps in overcoming this problem by using its buckets
that can manage and process huge size of data. This method could thereby
eliminate the need of utilizing several data storage devices in order to store large
volume data. However, s3 or similar cloud storage solutions may lead exorbitant
prices, which requires advanced level optimization for the big data operation.
Writing and reading data in s3 cloud storage bucket have extra prices as well.
This problem could be solved by reducing the size of data which is achieved
in Apache Parquet (Apache Spark native) format step. After deploying the log
data in AWS s3 bucket, another required data of multiple information related to
the users, that is taken from Turknet database, should also be uploaded to the
AWS s3 that is just established. Then, the most complex step in this operation
appears as the optimization of analyzing the data.
An Amazon Elastic Map-Reduce (EMR) is used in efficient way to com-
plete this step. EMR, in general, is a cluster that owns one or more master
nodes and multiple core (slave) nodes. It has several built-in programs such as
python programming software language, zeppelin-Apache software, and Apache-

spark engine. Coding part is done on Apache Zeppelin integrated development
environment (IDE). However, analyzing data step consists of reading, filtering,
comparing, and writing these different types of data. The different data is read
by using Apache Spark resilient distributed data-sets (RDD) and DataFrame
immutable object formats. RDD has to be used since the data must be dis-
tributed among the cluster slave nodes while processing it. This contributes to
a serious reduction of the processing time. The second part of this process is
filtering the data. At this part, the useless data-sets such as null, incomplete,
and duplicated ones are completely sanitized by using filter and map functions
in computing layer. This will end up with the filtered data. After that, the
data of DNSs and the data of database are compared with each other by using
Structured Query Language (SQL) Dataframe Spark functions such as join and
groupBy. The last step in analyzing data process is writing the final data. It is
written in Parquet format in partition method by using Spark Dataframe. Par-
quet format is used solely due to its effectiveness in minimizing the size of the
data. Then, parquet files are deployed in another s3 bucket. Finally, the modi-
fied data is ready to be visualized by using the numerous types of visualization
tools. Two different ways are shown in the Fig. 2 below which are Elasticsearch
Logstash and Kibana (ELK), and Plotly methods. In ELK, the parquet data
can be directly used by loading it to elasticsearch. Then, it should be loaded
from elasticsearch to an index in kibana by using Logstash. In Kibana, which is
an open source visualization data, the data can be presented in different chart
types such as time-lion, bar chart, line chart, etc. The other way is to use Plotly.
Using Plotly, only, needs additional step which is converting the data file format
from parquet to CSV format.
Fig. 2. Turknet DNS analytics system architecture.
Several experiences were examined on EMR using three different types of

cluster instances (master and node cores) in order to obtain the most optimal
configuration, and hence, considering it as the typical one in the forward opera-
tions. The tested instances are m5.xlarge, m5.2xlarge, and r5.4xlarge which their
major features of vCPU and RAM can be found from Table 1.
Table 1. Features of three different AWS EMR instances.
Amazon vCPUs RAM Instance Network EBS

node type storage performance bandwidth
(GiB) (Gbps) (Mbps)
m5.xlarge 2 8 G EBS-Only Up to 10 Up to 4,750
m5.2xlarge 8 32 G EBS-Only Up to 10 Up to 4,750
r5.4xlarge 16 128 G EBS-Only Up to 10 4,750
The target of the process is to make the combination of unstructured and

structured data between different data sets and data reduction with increasing
the value. As is pointed before, Turknet has three DNS servers which data can
be collected from, in this procedure, about 300 GB one day data is captured
from DNS servers and analyzed by using Spark application on EMR built-in
Zeppelin. Furthermore, the distributed computing cluster settings of Apache
Spark application should be configured cautiously, depending on the number
of core instances and their virtual CPU cores and their memory, while creat-
ing EMR cluster. For instance, the spark application configurations of an EMR
cluster of one master and 10 core nodes of r5.4xlarge instance, each node has 16
vCPUs and 128 GiB memory, should be modified as in the following: dynamic-
Allocation.enabled - False, executor.cores - 5, executor.memory - 37 GB, execu-
tor.instances - 170, and yarn.executor.memoryOverhead - 5 GB. These values of
spark application have been calculated based on an article published in AWS offi-
cial internet site [6]. The first property which is spark.executor.cores represents
the maximum number of tasks that can be executed in parallel in one executor.
Next, the heap size, amount of allocated memory, of each executor is controlled
by spark.executor.memory. Similarly, spark.executor.instances determines the
number of executors that will be run in a spark job. Ultimately, with chang-
ing spark.yarn.executor.memoryOverhead property the memory requirement to
YARN for each executor can be regulated. As a consequence, the number of par-
allel tasks that will be run at the same time is equal to the result of multiplying
spark.executor.cores with spark.executor.instances.
This operation encourages the Apache spark application to utilize the clus-
ters’ maximum potential ability for its purpose. However, the Apache spark
default configurations maybe used as they are for modest applications. After
adjusting the proper settings of Apache spark, the process can be started. The
first case included 1 + 10 m5.xlarge EMR cluster. In this case, the python script
was run and completed its all tasks within 40 min. Despite its high running time
it is financial appropriate with 0.352 dollar for the whole process. The other
case included the same number of instances but with considering more vCPU
cores and bigger memory size. The elapsed time of this case was approximately
22 min which fulfills a 16 min reduction and a little increase in billing. Finally,
another type of AWS instances was employed which is r5.4xlarge. This type ac
hived astonishing reduction in processing time. The elapsed time of this pro-
cess was approximately 13 min. However, comparing r5.4xlarge with m5.xlarge
leads to save nearly 27 min. It can be said, as a outcome, that the difference
in price (which is 0.2037 per Hour) between m5.xlarge and r5.4xlarge can be
significantly compensated since the running time of r5.4xlarge is less than the
one of m5.xlarge (see Table 2).
Table 2. Three different EMR cluster study cases.
Amazon Cluster Processed Output Running

node type size data size data size time
(GB) (GB) (minutes)
m5.xlarge 40
m5.2xlarge 1 + 10 246 29 22
r5.4xlarge 13
4 Results
The analysis main measurement runs on a cloud-based EMR cluster. Every DNS
log traffic files with CRM and CDR files are orchestrated by a cluster manage-
ment system architecture at AWS. This architecture is responsible for distribut-
ing chunks of work, of 3–4 billion rows including domains each, to a set of
worker nodes. Master node runs custom-built configuration on slave nodes that
performs a pre-defined selection of DNS queries for each domain in a chunk of
work. Queries are matched against a instance running on the worker node. Data
analyzed by workers is sent to a central aggregation point s3 and visualization
for each user for further processing and analysis. As case study, we have analysed
the use of over one week period. We focused on interesting particular domain
categorization with respect to internet traffic. Figure 3 shows the growth in the
fraction of domains consumption per server that use either of these platforms.
Growth is presented as a number of unique household customer relative to the
start of the 7-day period. To perform the analysis, our platform processed over
25 billion query results. The full analysis was performed by 11 r5.4xlarge node
in about 13 min for each day. This can easily be improved by running the anal-
ysis on a larger cluster configuration. This example shows what can be achieved
using our measurement platform and data. The growth in use of DNS servers
illustrates how the Internet usage is evolving from every organisation manag-
ing its own services, including predictive maintenance, to a few large providers
offering these services in future.
The figure (Fig. 3) below presents the hourly changes of active users on the
data collected from different DNS of Turknet. Graph shows the active user
decrease on 4 am. to 6 am. and the increase on 8 pm. to 12 pm. On the other
hand, The heat map of DNS network traffic of Turkey can be clearly shown in
Fig. 4.
Fig. 3. Active unique user numbers between 7 and 13 May 2019 on Turknet network.
Fig. 4. Density map of 30 may 2019 of URL categories distribution.

Lastly, the categories of Uniform Resource Locator (URL) are obtained and
analyzed. Several categories can be handled. However, as it is obvious from
Fig. 5, six categories are neatly considered. Technology/Internet category which
is represented in red, was the highest one among the other categories with nearly
74 thousand of unique usage at 00:00 o’clock.
Fig. 5. Traffic of 30 may 2019 of URL categories distribution over time.
5 Conclusion
We designed a unique active measurement infrastructure for the DNS data with
other various data-sets. Our infrastructure actively measures the total DNS traf-
fic on a daily basis. The resulting output enables reliable DNS-based analysis of
the evolution of the Internet for the first time at Turknet. And not only do we
measure on a large scale, we have also carefully designed for optimal analysis of
the collected data through the Apache Spark cluster optimization in EMR tool-
chain. The simple case study included in this work showcases use of our data-set.
It answers the simple question about the tremendous scale of DNS traffic data
can be tractable and analyzed in an effective implementation of a cloud based
big data platform. This provides a cost effective business intelligence application
for the big data analytic at scale in Turknet.
References
1. Karau, H., Konwinski, A., Wendell, P., Zaharia, M.: Learning Spark, 1st edn.
O Reilly Media Inc., Sebastopol (2015)
2. Cluster Mode Overview. In: Cluster Mode Overview - Spark 2.4.5 Documentation.
https://spark.apache.org/docs/latest/cluster-overview.html. Accessed 15 Mar 2020
3. Ryza, S., May, A., Kestelyn, J., et al.: Apache Spark Resource Management and
YARN App Models. In: Cloudera Engineering Blog (2014). https://clouderatemp.
wpengine.com/blog/2014/05/apache-spark-resource-management-and-yarn-app-
models/. Accessed at 15 Mar 2020
4. Deyhim, P.: Best Practices for Amazon EMR, Whitepaper (2013)

5. Amazon EMR pricing. Amazon. https://aws.amazon.com/emr/pricing/. Accessed
15 Mar 2020
6. Best practices for successfully managing memory for Apache Spark appli-
cations on Amazon EMR — Amazon Web Services. In: Amazon Web
Services. https://aws.amazon.com/blogs/big-data/best-practices-for-successfully-
managing-memory-for-apache-spark-applications-on-amazon-emr/. Accessed 15
Mar 2020
A Combined Bioinspired Algorithm for Big
Data Processing
Elmar Kuliev, Dmitry Zaporozhets, Yury Kravchenko,

and Ilona Kursitys(&)
Southern Federal University, Rostov-on-Don, Russia

[email protected]
Abstract. The paper considers the development of a biologically plausible

algorithm for solving big data processing problems. The intellectual data anal-
ysis is an important part of the big data technologies. In terms of big data
mining, it is reasonable to perform intelligent data analysis to retrieve the
important parameters and find regularities. The authors analyze the classification
task as one of the most important problems in data mining and the popular
solutions for its solving, including bioinspired algorithms as one of the most
effective approaches. The paper suggests a new concept aiming to provide the
population diversity to improve the effectiveness of biologically likely algo-
rithms for the problems of large dimensions. The bioinspired algorithm used in
the paper is represented by an artificial bee colony (ABC) algorithm. The
authors offer a combined algorithm based on the sequential work of evolutionary
adaptation and swarm intelligence. The experimental research was conducted to
prove the effectiveness of the algorithm and estimate its working time. The
results are demonstrated in the paper.
Keywords: Big data Biologically plausible algorithms Bee colony Genetic

algorithm Data mining
1 Introduction
Information processing and data analysis represent an important part of Big Data
technologies. The scientific and engineering fields are considered as the most relevant
for the Big Data areas. Big Data is used in government, business, economics, law, and
other fields. Intelligent analysis is required to retrieve the important parameters of Big
Data. There are a lot of algorithms developed for the mentioned task today. Choosing
the most appropriate algorithm is a tough challenge.
Biologically plausible algorithms compose one of the machine learning branches in
artificial intelligence. They investigate the search space, synthesize the solutions, assess
the quality, which is later used to implement the natural selection. These features allow
them to learn the best areas of the space search. In the Internet space, the volume of the
data rapidly grows, as well as the complexity of the problems needed to be solved for
the data. Hence, the task of adapting the intelligent information technologies based on
the machine learning methods of processing the growing data scopes is very relevant
these days [1, 2].
A Combined Bioinspired Algorithm for Big Data Processing 843
The paper is aimed to develop an approach that can provide the population’s
diversity and increase the effectiveness of biologically plausible algorithms for solving
the tasks of large dimensions. The originality of the paper is assured by the developed
models, architecture, and new combined algorithm, as well as the results of the
experiments presented in the paper.
The remaining parts of the paper are presented in the following way. The next
section is devoted to the literature review of Big Data processing. The third part
presents the proposed models of dividing the data into the parts and their further
sampling. The fourth section is devoted to the proposed bioinspired algorithm, which is
based on the principles of genetic and swarm search. The fifth part presents the results
of the experimental research on comparing the algorithm with other popular algorithms
working with well-known datasets. The last section summaries the paper.
2 Literature Review
There are two types of classification tasks: binary and multiclass classification. In terms
of binary classification, we have only two options. Besides, the data mining tasks can
be solved by “supervised” or “unsupervised training”. The classification algorithms are
considered as “supervised training” because we have information about certain objects.
The objects with the initial information are called “training sampling” [3, 4].
The Support Vector Machine (SVM) is one of the best methods of training for
solving the classification and regression tasks. The main feature of the SVM method is
applying the export of the experimental dataset in the expanded area, where the
hyperplane is built for dividing the classes. This concept is called the “core function”.
The hyperplane divides the classes and pushes them away and pushes them to the
maximum distance from each other. The bigger the distance between the classes, the
less is the average mistake of the classifier. The support vectors are the closest to the
parallel hyperplanes.
This SVM method has a good reputation but using it in the Big Data processing
imposes a lot of expenses. Large computational expenses are caused by the cubic
complexity of the calculations and lead to challenges in using the SVM algorithms for
big data. The current problem is that very few classification methods can be practically
applied today due to the importance of solving the tasks of quadratic programming of
large dimensions. This problem can be solved by different algorithms such as
Sequential Minimal Optimization (SMO), fragment formation, simple SVM algorithm,
etc. The listed algorithms can perform linear calculations and multiple clarifications.
A cascade SVM method is proposed to smooth the problem of large dimensions of
computations. The classifiers are trained on the subsets of the original dataset multiple
times. The support vectors of the classifiers are united to get a new training dataset.
The cascade SVM methods can also be parallelized. For example, Google has
demonstrated a model of parallel computations called MapReduce. This model is
relevant and useful for multi-tasking in the computer clusters. Besides, the Hadoop
MapReduce platform is popular in the field of analysis and processing of big data.
In terms of the Hadoop technology, there is a special file system named HDFS
(Hadoop Distributed File System). It can be represented as an expanded standard file
844 E. Kuliev et al.
system with the table of file descriptors and data area. The HDFS uses a special server,
and the data are distributed among the data servers [5, 6]. Such systems are applied in
creating and implementing the distributed programs performed on the clusters of the
multiple nodes.
Based on the analysis provided above, we can conclude that the SVM algorithm is
quite effective for processing Big Data.
3 Proposed Model
To solve the problems of Big Data optimization, it is impossible to train the classifier
properly. We need to consider the information with the approximately maximal
accuracy as satisfactory. In the training and testing datasets there must be a minimal
number of mistakes. The testing sampling must contain an equal number of mistakes
into training and trained datasets to avoid over-training.
3.1 Multitasking
One of the most optimal strategies of Big Data processing includes three following
steps:
Step 1. Dividing the data into parts
Step 2. Building a model for each part
Step 3. Consolidation of the results
Usually, in Big Data we distinguish several different subsets. For instance, they can
represent the clients with similar behavior, for which we can build a model.
It is reasonable to build several simple models for each part instead of building a
complex model for all the data. This approach is great for increasing the speed and
lowering the memory requirements because we can process less data in one operation.
Besides, processing can be performed in a parallel way, which also provides significant
time savings. The models also can be built by several analyses simultaneously (Fig. 1).
Fig. 1. Dividing the data into several parts and building the models for each part with the
following uniting the results.
One of the advantages of the proposed approach is simple creation and serving the
models. To obtain the results in the shortest time, we can run the model stage-by-stage.
3.2 Sampling from the Frame

In terms of working with big data, we need only a small part of the whole dataset. The
part is called sampling from the general dataset. Properly selected sampling is the basis
for building an effective model. Further analytical processing can be divided into two
parts (Fig. 2):
1. Building the model
2. Applying the model
Building a complex effective model requires investing a lot of resources. The
sampling method allows us to apply the finished model for scattered data hence use
much fewer resources. In such a case, the calculations are reduced to several simple
actions [7–9].
Thus, we can conclude that building a model for a small part of data that is further
applied for the whole dataset can significantly save the calculation time and resources.
Fig. 2. Sampling.
To obtain selection of the data, we can use sampling which allows increasing the
analysis speed leaving the effectiveness at the same level.
4 Combined Algorithm
In terms of the binary classifier, the objective function becomes complex, multi-
extremal and multiparametric. We can use the following algorithms to find the optimal
solutions:
• Genetic Algorithms
• Artificial Bee Colony (ABC) Algorithm
• Particle Swarm Optimization (PSO) Algorithm
The PSO algorithm is a simple evolutionary algorithm for optimization. Its goal is
to find the optimal values of the kernel function and regularizing parameters [10–12].
The authors suggest a bioinspired algorithm based on the Artificial Bee Colony
(ABC) algorithm for solving the Big Data tasks. The approach includes building new
bioinspired architectures, which allow us to solve the problems at an acceptable time

and minimal resources spending. Figure 3 demonstrates the proposed architecture.
Fig. 3. The architecture of the bioinspired search.
In the architecture, swarm intelligence is represented by the ABC algorithm, which

involves determination of the searching area using the controlling parameter a.
The authors propose a combined algorithm which work is based on the sequential
work of the evolutionary adaptation and swarm intelligence (Fig. 4).
Fig. 4. Combined algorithm.

The block of creating the initial population is defined by the number of iterations
and population size. The core includes forming the solution by investigating the source
area applying the adaptive behavior of the bee colony. The obtained solutions are then
transformed into chromosomes (agents) forming the population of the best solutions.
The developed principles of integrating the models of evolutionary adaptation,
genetic search, and swarm intelligence allow us to reduce the shortcomings of the meta-
heuristics used in each model. The special feature of the proposed algorithm is that the
integrated agents are changed in the search process and perform the functions of the
agents from each model.
5 Experimental Research
On the basis of the proposed algorithm, we developed software. The research is

focused on determining the parameters of the classifier, accuracy of classification and
the time spent.
To conduct the experiments, we used the classical Particle Swarm Optimization
(PSO) algorithm, classical Artificial Bee Colony (ABC) algorithm, and the developed
Combined Search (CS) algorithm. The real data were taken from the Statlog project and
the machine learning library UCI. We used two datasets from the medical diagnostics
(WDBC and Heart respectively), two datasets for the loan scoring (Australia and
German respectively), and a dataset for building a predictability model of spam
detection (SPA). We also used two test samples: Test and MOTP12 [13–16].
Table 1 shows the results of searching the optimal parameters of the SVM classifier
using the developed algorithms and the classical PSO algorithm, number of mistakes
made, and execution time of the algorithms. Figure 5, 6 demonstrate the diagrams of
the accuracy and working time.
Table 1. The results of comparing the classical PSO algorithm and the combined algorithms.
Dataset Objects Characteristics Algorithm
PSO ABC Combined search
Time, Accuracy, % Time, Accuracy, % Time, Accuracy, %
sec sec sec
WDBC 569 30 10108 99.12 11235 99.36 14012 99.39
Heart 270 13 6558 93.33 6445 94.15 812 94.51
Australia 690 14 16018 92.9 16000 93.11 17512 94.01
German 1000 24 18198 95.8 19070 95.8 22121 95.98
Spam 4601 57 92645 97.91 93789 97.94 95121 97.96
MOTP12 400 2 15697 95.5 15712 95.5 19510 96.9
Test 300 2 3433 100 3401 100 5901 100
The accuracy for spam classification is 97.6% using the combined algorithm and
97.94% using the ABC algorithm. The classical PSO algorithm gives the result of
97.91% accuracy. The working time of the algorithm is not greater than 2% than the
traditional PSO algorithm.
Fig. 5. Working time of the algorithm.
Fig. 6. The results accuracy.
6 Conclusion
The paper presents the method of solving the important task of Big Data processing.
The authors propose the new approach allowing us to diverse the population and
increase the effectiveness of biologically plausible algorithms. The authors developed
the combined bioinspired algorithm based on the bee colony model. The algorithm uses
the ideas of sequential work of evolutionary adaptation and swarm intelligence.
The experiments were carried out to compare the developed algorithm with the
well-known methods of classification. The algorithm demonstrates high accuracy and
appropriate execution time. The authors achieved the mentioned results of the research
using the parallel processing the searching space by several flows which helps to
prevent the preliminary convergence of the algorithms.
Further research will be focused on the development of new bioinspired models and
algorithms for solving important Big Data tasks.
Acknowledgements. The reported study was funded by RFBR according to the research project
№ 18-29-22019.
References
1. Khan, S., Liu, X., Shakil, K.A., Alam, M.: A survey on scholarly data: from big data
perspective. Inf. Process. Manag. 53(4), 923–944 (2017)
2. Gandomi, A., Haider, M.: Beyond the hype: big data concepts, methods, and analytics. Int.
J. Inf. Manag. 35(2), 137–144 (2015)
3. Singh, S., Chana, I.: A survey on resource scheduling in cloud computing: issues and
challenges. J. Grid Comput. 14(2), 217–264 (2016)
4. Gill, S.S., Buyya, R.: A taxonomy and future directions for sustainable cloud computing:
360 degree view. ACM Comput. Surv. 51(6), 1–37 (2019)
5. Chen, C.P., Zhang, C.Y.: Data-intensive applications, challenges, techniques and technolo-
gies: a survey on big data. Inf. Sci. 275, 314–347 (2014)
6. Wang, J., Wu, Y., Yen, N., Guo, S., Cheng, Z.: Big data analytics for emergency
communication networks: a survey. IEEE Commun. Surv. Tutorials 18(3), 1758–1778
(2016)
7. Ilango, S.S., Vimal, S., Kaliappan, M., Subbulakshmi, P.: Optimization using artificial bee
colony based clustering approach for big data. Clust. Comput. 22, 1–9 (2018)
8. Kune, R., Konugurthi, P.K., Agarwal, A., Chillarige, R.R., Buyya, R.: Genetic algorithm
based data-aware group scheduling for big data clouds. In: Big Data Computing (BDC),
pp. 96–104 (2014)
9. Aslan, S., Karaboga, D.: A genetic Artificial Bee Colony algorithm for signal reconstruction
based big data optimization. Appl. Soft Comput. J. 88, 106053 (2020)
10. Aslan, S.: A comparative study between artificial bee colony (ABC) algorithm and its
variants on big data optimization. Mem. Comput. 12, 129–150 (2020)
11. Ilango, S.S., Vimal, S., Kaliappan, M., Subbulakshmi, P.: Optimization using Artificial Bee
Colony based clustering approach for big data. Clust. Comput. 22, 12169–12177 (2018)
12. Djenouri, Y., Fournier-Viger, P., Lin, J.C.-W., Djenouri, D., Belhadi, A.: GPU-based swarm
intelligence for Association Rule Mining in big databases. Intell. Data Anal. 23(1), 57–76
(2019)
13. Benmounah, Z., Meshoul, S., Batouche, M., Lio, P.: Parallel swarm intelligence strategies
for large-scale clustering based on MapReduce with application to epigenetics of aging.
Appl. Soft Comput. J. 69, 771–783 (2018)
14. Feng Li, D., Zhong, Y., Zhou, J., Yin, J., Zhou, L.J.: Parallel optimization based on artificial
bee colony algorithm. In: 2017 IEEE 2nd International Conference on Big Data Analysis,
pp. 955–959 (2017)
15. Bansal, N., Kumar, S., Tripathi, A.: Application of artificial BEE colony algorithm using
Hadoop. In: Proceedings of the 10th INDIACom; 2016 3rd International Conference on
Computing for Sustainable Global Development, pp. 3615–3619 (2016)
16. Djenouri, Y., Drias, H., Chemchem, A.: A hybrid bees swarm optimization and Tabu search
algorithm for association rule mining. In: 2013 World Congress on Nature and Biologically
Inspired Computing, NaBIC, pp. 120–125 (2013)
Exploration of the Waves of Feminism Using
Sentiment Based Text Mining Techniques
H. Umutcan Ay(&), S. Nazlı Günesen, and Tolga Kaya
Istanbul Technical University, Maslak, 34467 Sarıyer/Istanbul, Turkey

[email protected]
Abstract. The motives and the causes behind the evolvement of the feminist
thought have been in the spotlight for many researchers. This study aims to
explore the evolution of the driving forces of the feminist thought using text
mining and clustering techniques. To do this, first, 443 relevant literary works
published in the progressive time span of three waves of feminism are explored
through bag of words method. Then, to address the wide span of topics within
the third wave, hierarchical clustering is used. Finally, sentiment analysis is
implemented to gain insight on the emotional trends within three successive
waves. Results reveal an increasing emphasis on collectivism and globalization.
Keywords: Text mining Machine learning Feminism Sentiment analysis

Bag of words method Hierarchical clustering
1 Introduction
Feminism emerging from a movement advocating women’s basic rights in male-

dominated society of the early 1800s has now evolved to a broader ideology that
upholds equal rights and legal protection for all gender identities in all areas of life and
even endorses civil rights and environmentalism. Although it has already achieved to
reshape social fabrics of societies considerably almost all over the world, there is still
room for improvement. Therefore, feminism will presumably remain on the agenda for
all human beings in the world for at least a few more decades. Evolvement of feminism
in the world over the years is usually categorized into three waves: the first wave (1800s
to Early 1900s), the second wave (1960s to 1980s), and the third wave (1990s–today).
Within this research, to examine three consequent waves of feminism, machine learning
algorithms will be used alongside the sentiment-based text mining methods to create
unsupervised clusters of different topics. By the usage of unsupervised clustering
methods and also by visualizing the frequent terms within the waves, it is aimed to have
an unbiased determination of different topics and the sentiment analysis indicating
emotions behind the documents, which would create a basis for the future researchers to
pursue their works on these waves.
The rest of the paper is organized as follows: Sect. 2 is devoted to the literature
review. In the third section, the methodology is given. In Sect. 4, data preparation
methods are summarized. In Sect. 5, findings of the models are demonstrated. Finally,
conclusions and suggestions for future research are discussed.
Exploration of the Waves of Feminism 851
2 Literature Review
Men and women in society of the 19th century were assumed biologically created for
predetermined separate roles and duties. Conception of women was evoking role of wife
and mother, physical weakness, emotionality, lower intellectual capacity, intuitiveness,
irrationality, being able to work in manual jobs only, no opportunity to be educated at
school/universities. Right to divorce was far more difficult for women compared to that
of men [9]. The first wave feminist activism arose in 1800s as a movement upholding
abolition of slavery and women’s right to have a voice in society. Towards the late 19th
century, the movement focused on women’s right to vote in particular and ended with
legislation of the 19th amendment granting women voting rights to the US Constitution
in 1919 [8]. Women’s suffrage (right to vote), contract and property rights, right to
divorce, anti-racism (many abolitionists and black right champions spearheaded the
movement) were prominent arguments raised by the first wavers [1].
The society of 1960s was still patriarchal and women were still treated as second
class although they have some legal rights [13]. The second wave unfolded in the
1960s was putting emphasis on the women’s liberation movement for equal legal and
social rights. After the Second World War, it rose in the context of civil rights and anti-
war movements coinciding with awakening of the leftist anti-system minority groups
(68 generation) everywhere in the world. For this reason, mode of the second wave was
more radical [4]. The movement driven by middle-class white women focused on
passing the Equal Rights Amendment (guaranteeing legal gender equality) to the
Constitution. Additionally, equal employment opportunity, equal pay for equal work,
contraception (birth control), racial/sexual discrimination in the workplace, body
positivity, civil rights (anti-war, anti-racism), right of abortion were the leading themes
underlined by the second wavers [8].
The third wave feminism appearing towards the mid of 90s destabilized many
constructs, including gender, body, sexuality, hetero-normativity, and even identity
politics [10]. Moreover, the movement has spent notable effort to address issues
concerning homophobia, social class inequality, sexism, and racism [13]. Additionally,
young feminist movement was emphasizing collective action including gender
minorities and assuming a universal female identity was avoided by the third wave
feminists [1]. Issues that have been emphasized by the third wavers are: LGBTQ rights,
body positivity, fat positivity, sex positivity, brain + beauty (women can have both
brain and beauty), homophobia, social class inequalities, and sexual orientation [10].
3 Methodology
N-gram Tokenization. It is a method independent from language types and identifies

problems morphological issues that lower information retrieval performance [5].
Word Association. It is a method to investigate the relations between text data.
Paradigmatic relations indicate similarity between words after obtaining them from a
bag of cloud, while syntagmatic relations are mostly related with how frequent two
words are combined together [2].
852 H. Umutcan Ay et al.
Hierarchical Cluster Analysis (HCA). HCA clusters individuals based on the dis-
tance between them. Agglomerative and divisive clustering are two distinct types of
HCA. Agglomerative clustering takes a table as an input and converts it into a distance
matrix. It disarranges the distance matrix to make individuals in distance object
accumulated into clusters [3]. Ward’s Clustering Method is the only type of agglom-
erative clustering that works with classical sum-of-squares principle creating groups
that minimize within group dispersion [7]. On the other hand, Canberra distance based
on the following generalized equation (see Eq. 1) is often used for data cumulated
around an origin as it well handles large number of zero entries [3].

Xn1 yi;k yj;k
d ði; jÞ ¼
CAD ð1Þ
k¼0 y þ y
i;k j;k
Sentiment Analysis and Plutchik’s Wheel of Emotions. It is mainly used on texts to

indicate the writer’s emotions or the polarity of their statements depending on the scope
of the analysis. Main idea behind the sentiment analysis is to compare the subject with
the pre-existing dictionary to assign certain emotion ratings onto the text [12]. As a
sentiment dictionary, NRC emotion lexicon was used for the analysis [6]. It assigns 8
different emotions, indicated in a form of wheel of emotions by Plutchik, to its pre-
designated dictionary of 13901 distinct words.
4 Data Preparation
Using the key words of “feminism”, “women rights”, “patriarchy” and “gender
inequality”, 443 different books and articles about feminism were retrieved for analysis.
Categorized by the date of publication, works were then separated into three conse-
quent waves of which the boundaries were set by the previous researches on feminism
where the distribution could be investigated under Table 1 (Table 2).
Table 1. Frequency table for waves

First Second Third
18 35 390
Table 2. Variable types and explanations

Variable Type Explanation
doc_id Character Unique documentation id ranging from 1 to 443
Text Character Text of work subject to investigation
Year Numeric Publication year
Wave Factor Factor of 3
Title Character Main title of the related work
Before going into the research, text of the raw data required refining to eliminate
the noise within the data. During the cleaning process, corpus structure is utilized to
increase the processing speed and efficiency of calculations where following steps were
applied:
1. All characters were shifted to lowercase,
2. Numbers and stop words (e.g., and, but, so, the, very, few) were taken out,
3. All unnecessary punctuations were removed,
4. All words containing non-Latin letters were eliminated to filter names of writers
from different ethnicity.
5. Terms within brackets were eliminated.
6. More than 2 times repeated letters had been decreased to 2 for addressing typos.
7. All extra whitespace was removed to reduce the total space of the document.
As a final step after the cleaning process, terms are stemmed to leave only the roots
of the initial words.
5 Findings: Bag of Words Model and Sentiment Analysis
For the bag of words analysis, cleaned corpuses were turned into term document
matrices indicating only the document IDs and all of the unique words within the texts.
When retrieving those unique words, n-gram tokenization where n = 2 (bi-gram tok-
enization) was utilized as n-gram tokenization outperforms the word-based indexing for
advanced languages like English [5].
First and Second Waves of Feminism
For the first wave, most frequent terms appearing on the left word cloud (in Fig. 1)
indicates that during the initial phases feminism had more intrinsic approach to the
subject. The terms like “womans”, “little girl”, “men women” refers that the most of the
arguments were about position of women within the society. This relation could also be
Fig. 1. Bi-gram word cloud for the first (on left) and second wave (on right)
examined through the word association tests. Within the scope of the texts, the term
“women men” were highly correlated at 0.95 with the usage of “women rights” and
“public opinion” is a high indicator that the main concern was solidifying the place of
women within the society. On the other hand, the discussions within the second wave
(Fig. 1 on right) had mostly been revolved around the separation between “black” and
“white” masses while also emphasizing the equal rights for woman within both the
daily and the professional life as the topics of equal payments, woman suffrage and the
divisions between working class
Third Wave of Feminism. In order to examine the works under the third wave,
frequencies of bi-gram tokenized terms had been determined under the “Term-
Document Structure”, which is a specialized form of a pivot table where terms are
represented in y axis, documents in x axis and the entries represent the frequency of
occurrence. However, to reduce the effect of the most frequent terms, which occurs
almost in every text due to the common topics, the frequencies had been normalized by
giving penalty scores as suggested by the Salton and Buckley by the formula in Eq. 2:
jDj
idfi ¼ log2 ð2Þ
jfdjti 2 d gj
Where ti is the term in term-document matrix, jDj is the total number of documents,
jfdjti 2 dgj is the number of documents including the term ti and idf i is the “inverse
document frequency” [11]. Furthermore, using the re-adjusted frequencies the docu-
ments are clustered according to hierarchical clustering principles by using Ward’s
method by basing the clusters on minimizing the variance within clusters in order to
have more flexible clustering approach. Also as the documents were containing a lot of
zero elements, Canberra distance is utilized (Eq. 1). To have equal amount of docu-
ments in each cluster, cluster number is selected as 3 with the following results.
Fig. 2. Word cloud for cluster 1 (on left) and cluster 2 (on right)
Cluster 1. First thing to notice upon the investigation of the literature within the first
cluster in the Fig. 2 (on left) is that the different ethnicity groups other than “black
people” are now being included within the common discussions in feminist movement
which may be interpreted as upon the third wave the movement had finally reached to
global masses. Also as could be examined, the commonly used term of “African
American” had been introduced to the community.
Cluster 2. The most frequent topics within the literature in the second cluster is
actually about the common rights of the women talking about birth control, common
law, the violence against the women and the presence of women within the labor force
in terms of payment and working rights. Also the globalization of the content could be
examined once again as the regional distribution had grown larger where the Middle
Eastern and Latin American Regions are the most frequent areas of discussion.
Fig. 3. Word cloud for third wave cluster 3
Fig. 4. Radar chart for sentiments of third wave clusters

Cluster 3. The word cloud for the third cluster shown in Fig. 3 is a direct indicator of
the expansion of feminist thought onto the gender minorities. Texts under this cluster
indicate issues with LGBTQI+ community, emphasizing on the recently developed
queer theory. Also as these minorities are included within the core of the feminism, it
could be examined that the movement is yet to be exclusive solely to women but it is
now covering up the other minority groups with unequal conditions.
Using NRC library to attain the emotions dictionary, the radar chart on the clusters
indicates a pattern that most of the literature included within the third wave carries
similar emotional structure as the emotions of “anger”, “trust”, “sadness”, “disgust”
and “fear” being the most dominant ones. Deriving from the Wheel of Emotions by
Plutchik, the emotions stated on the radar in Fig. 4 could be analyzed by first looking at
the high levels of anger and disgust which than creates “contempt” according to him. It
is expected to have “contempt” emotion when talking about the oppression of the
minorities as they may feel humiliated by the society. Furthermore, the “submission”
created by the inclusion of high “fear” and “trust” elements infers that the feminists will
not obey this oppression created by the society and fight for their cause no matter what.
Fig. 5. Radar Charts for “the Vindication of Rights of Women” (on left) (First Wave) and “the
Feminine Mystique” (on right) (Second Wave)
Comparing the clusters of third wave with the sentiment analyses of one of the most
iconic literature of their own era as shown in the Fig. 5, it could be said that the
emotions had changed with the progression of eras. The writers have less “anticipation”
and “joy” on average meaning that the “optimism” within the texts had been reduces
according to Plutchik’s Wheel of Emotions. Also one could infer that the increased
“anger” level within the texts suggests that now the writers expect more immediate
actions and more “aggressive” towards issuing the current subjects of feminism.
6 Conclusion and Recommendations for Future Research
Within the scope of this research, three waves of feminism were analysed by using
machine learning algorithms to provide social sciences with insights about the feminist
movement based on text mining. As a result, it is confirmed that the three waves are
clustered by noticeably distinct topics in accordance with requirements of their eras, as
expected. It is also identified that the movement arising as a basic thought of equality of
men and women, has evolved into an ideology highlighting collectiveness based on a
global perspective. Shaped by their eras’ most heated disputes, the feminism has been
progressively becoming something bigger and broader. However, due to the general
characteristics of the third wave feminism; sample includes much overlapping data as
most frequent words are mostly similar amongst different documents. As a future
research, in order to create more efficient clusters, fuzzy c-means method could be
utilized with increased sample size.
References
1. Cobble, D.S., Gordon, L., Henry, A.: Feminism unfinished: a short, surprising history of
american women’s movements. Juncture IPPR 21(3), 242–243 (2014)
2. Correia, A., Filomena, T., Lobo, V., Teodoro, F., Lobo, V.: Statistical methods for word
association in text mining. In: Recent Studies in Risk Analysis and Statistical Modeling.
Contributions to Statistics, pp. 375–384. Springer, Cham (2018)
3. Desagulier, G.: Visualizing distances in a set of near synonyms: rather, quite, fairly, and
pretty. Corpus Methods for Semantics: Quantitative Studies in Polysemy and Synonymy,
pp. 145–178 (2014)
4. Jacob, K., Licona, A.C.: Writing the waves: a dialogue on the tools, tactics, and tensions of
feminisms and feminist practices over time and place. NWSA J. 17(1), 197–205 (2005)
5. McNamee, P., Mayfield, J.: Character n-gram tokenization for european language text
retrieval. Inf. Retrieval 7(1/2), 73–97 (2004)
6. Mohammad, S.M., Turney, P.D.: Emotions evoked by common words and phrases: Using
Mechanical Turk to create an emotion Lexicon. In: Proceeding of Workshop on
Computational Approaches to Analysis and Generation of Emotion in Text, pp. 26–34
(2010)
7. Murtagh, F., Legendre, P.: Ward’s hierarchical agglomerative clustering method: which
algorithms implement ward’s criterion? J. Classif. 31(3), 274–295 (2014)
8. Rampton, M.: Four Waves of Feminism. Pacific Magazine, Pacific University, Oregon
(2008)
9. Riley, D.: Am I That Name? Feminism and the Category of “Women” in History. University
of Minnesota Press, Third Avenue South, Suite 290 Minneapolis, MN 55401-2520 (1995).
ISBN 0816617309
10. Roseneil, S.: Beyond Citizenship? Feminism and the Transformation of Belonging –
Citizenship, Gender and Diversity. Palgrave Macmillan, New York (2013)
11. Salton, G., Buckley, C.: Term-weighting approaches in automatic text retrieval. Inf. Process.
Manage. 24(5), 513–523 (1988)
12. Silge, J., Robinson, D.: Text Mining with R: A Tidy Approach. O’Reilly Media, Sebastopol
(2017)
13. Walters, M.: Feminism A Very Short Introduction. Oxford University Press, Oxford (2005).
ISBN 9780192805102
SVM-Based Hybrid Robust PIO Fault
Diagnosis for Bearing
Farzin Piltan and Jong-Myon Kim(&)
School of Electrical Engineering, University of Ulsan,

Ulsan 680-749, South Korea
Abstract. Inner, outer, and ball faults are complex non-stationary and non-
linear faults that occurs in rotating machinery such as bearings. Designing an
effective procedure for fault diagnosis (FD) is essential to safe operation of
bearings. To address fault diagnosis issue, a robust, hybrid technique based on
the ARX-Laguerre fuzzy-sliding proportional integral observer (ALFSPIO) for
rolling element bearing (REB) is presented. The main important challenges in the
ARX-Laguerre PI observer are robustness, and estimation accuracy. To address
the robustness issue, sliding observation technique is introduced. Moreover, to
increase the signal estimation accuracy, the fuzzy technique is used in parallel
with ARX-Laguerre sliding PIO. Furthermore, using the ALFSPIO, the residual
energy signals showed more differentiable for fault diagnosis. Beyond the above,
the support vector machine (SVM) is used to fault detection and classification.
The vibration dataset of Case Western Reverse University (CWRU) is used to
validate the effectiveness of the proposed algorithm.
Keywords: Roller bearing element Fault diagnosis ARX-Laguerre

technique PI observer Sliding mode technique Fuzzy algorithm
1 Introduction
Rolling Element Bearings (REBs) are widely used in several industries to reduce
friction. From a dynamic point of view, these components are extremely nonlinear and
complex. Therefore, the complexities of the tasks and the nonlinear parameters in
REBs have led to numerous complications associated with fault detection and classi-
fication. Numerous types of faults have been represented in REBs, which can be
divided into four main groups: inner fault (IF), outer fault (OF), ball fault (BF), and
cage fault [1].
Diverse approaches have been proposed for fault detection and classification in
REBs, including signal-based procedures, data-driven techniques, model-based
approaches, and hybrid techniques [1–5]. Hybrid-based fault detection and classifica-
tion procedures are considered as reliable, stable, and robust because they use various
high-performance techniques. Several techniques have been used in parallel to design
hybrid-based fault detection and classification that are reported in [2, 6]. This research is
divided into three main steps for fault diagnosis in the REBs: a) system modeling, b)
estimation technique based on observation method, and c) decision making. To estimate
SVM-Based Hybrid Robust PIO Fault Diagnosis for Bearing 859
the signal in normal and abnormal conditions based on the observation techniques,
system modeling is the first and significant step. System modeling is categorized as the
following groups: a) mathematical system modeling and b) system identification tech-
nique. Mathematical system modeling is a robust and accurate technique but it has a big
issue in complex systems (e.g., REBs) [6]. To address these issues, system identification
algorithms (e.g., Auto-Regressive with eXogenous inputs (ARX), ARX-Laguerre, and
intelligent-based ARX-Laguerre) have been reported in [6, 7].
To signal estimation based on the observation techniques, various methods have
been proposed, which can be divided into two main groups: linear-based observers and
nonlinear-based observers [3]. However, the linear-based observers such as
proportional-integral observers (PIO) or proportional multiple integral (PMI) observers
are easy to implement but, robustness and accuracy are the main drawbacks [6, 8]. The
nonlinear-based observers such as sliding mode observers (SMO), feedback lin-
earization observers (FLO) are robust, stable, and reliable but they are suffering from
complexity and accuracy [1, 2, 6]. To address the issues of linear-based observers and
nonlinear-based observers, hybrid-based approaches have been described in [2, 6].
In this research, the hybrid technique is used to fault detection and classification in
the REBs. Firstly, the ARX-Laguerre technique is proposed to REB modeling based on
the vibration signals. Beyond the above, the ARX-Laguerre fuzzy-sliding PI observer is
used to accurate signal estimation. After generating the energy residual signal, the
SVM is represented for fault detection and classification. The block diagram of the
proposed technique is represented in Fig. 1. This paper is organized as follows. The
second section outlines the ARX-Laguerre REB modeling based on the vibration
signals. The proposed technique for fault detection and classification is represented in
the third section. In the fourth section, the proposed method for fault detection and
classification in the REB is analyzed. Conclusions are provided in the last section.
2 ARX-Laguerre Rolling Element Bearing Modeling
In complicated systems such as REBs, system modeling is a significant challenge,

especially in uncertain conditions. Hence, to modeling the REB, the ARX-Laguerre
technique is represented. Regarding Fig. 1, this technique has two main steps. In the
first step, the ARX system modeling is defined. Moreover, to improve the robustness
and reliability of the system modeling, the ARX-Laguerre technique is represented in
the second step. The ARX technique for REB modeling based on vibration signals can
be represented by Eq. 1 [6, 7].
XdY XdU
YA ðkÞ ¼ C ðiÞYA ðk iÞ þ
i¼1 Y
i
i¼1
CUi ðiÞUi ðk iÞ; ð1Þ

where YA ðkÞ; Ui ðk Þ; Cx ðiÞ; Cy ðiÞ , and ðdUi ; dY Þ are the REB’s model based on ARX
technique, reference signal for REB’s modeling, REB’s model parameters, and the order
of the REB modeling, respectively. The REB is a nonlinear and complex component and
the ARX technique is a linear algorithm. Therefore, to improve the power of noise
860 F. Piltan and J.-M. Kim
rejection and accuracy, the ARX-Laguerre procedure is represented for REB modeling
in this research. The ARX-Laguerre REB model is represented as follows [6, 7]:
XiY 1 XiU 1
YAL ðk Þ ¼ n¼0
CY ðiÞOn;YAL ðk; cY ÞYAL ðk iÞ þ i
n¼0
CUi ðiÞOn;Ui k; cUi Ui ðk iÞ ;
ð2Þ

Here, YAL ðkÞ; On;YAL ðk; cY Þ and On;Ui k; cUi are the REB’s model based on ARX-
Laguerre technique, the REB’s function of orthonormal, and the REB’s reference
signal orthonormal function, respectively. Therefore, the state-space formulation for
ARX-Laguerre REB’s modeling can be represented as the following equation.

XAL ðk þ 1Þ ¼ ½rX XAL ðkÞ þ rY YAL ðk Þ þ rU Ui ðkÞ
: ð3Þ
YAL ðk þ 1Þ ¼ ðbÞT XAL ðkÞ
Here, XAL ðkÞ and ðrX ; rY ; rU ; bÞ are the state of the model for REB based on
ARX-Laguerre technique and fine-tuning coefficients, respectively.
Fig. 1. Proposed method for fault detection and classification of the REB.
3 Proposed Algorithm for Fault Detection and Classification
The REB modeling based on normal condition of vibration signal is represented in the
previous part. In this section, to improve the estimation accuracy the ARX-Laguerre
fuzzy-sliding PI observer is proposed. This technique is caused to improve the accuracy
of REB’s vibration signal estimation for fault detection and classification. Therefore,
firstly the ARX-Laguerre PI observer is represented as the following equation.
8
>
> ^PIO ðk þ 1Þ ¼ rX X
X ^PIO ðkÞ þ rY Y^PIO ðk Þ þ rU Ui ðk Þ þ u
^ PIO ðk Þ
>
>
>
> þ r1 ePIO
>
<
^
YPIO ðk þ 1Þ ¼ ðbÞT X ^PIO ðk Þ ; ð4Þ
>
>
>
> ePIO ¼ XAL ðk Þ X ^PIO ðk Þ
>
>
>
:
erPIO ¼ Yo ðkÞ Y^PIO ðk Þ
To improve the accuracy uncertainties estimation of the ARX-Laguerre PIO, the

uncertainties can be estimated based on the following equation.
^ PIO ðk þ 1Þ ¼ u
u ^ PIO ðkÞ þ r2 erPIO : ð5Þ
Here, X^PIO ðkÞ; Y^PIO ðkÞ; u

^ PIO ðkÞ; Yo ðkÞ and ðr1 ; r2 Þ are the estimation state of the
model for REB based on ARX-Laguerre PI observer (PIO), the estimation of the REB
signal based on ARX-Laguerre PIO, uncertainties (motor speed variant) estimation
based on the ARX-Laguerre PIO to minimize the error in the normal condition, original
signal of the REB, and the coefficient of the PIO, respectively. Regarding Fig. 1, the
ARX-Laguerre PIO is suffering from robustness. To address this issue the ARX-
Laguerre sliding PIO (SPIO) is represented. In most of the observation/control appli-
cations, the sliding technique is used to increase the rate of the robustness. Various
functions have been used to design sliding algorithm. In this research, the switching
(sign) function is recommended.
8
>
> ^SPIO ðk þ 1Þ ¼ rX X
X ^SPIO ðkÞ þ rY Y^SPIO ðkÞ þ rU Ui ðkÞ þ u ^ SPIO ðk Þ
>
>
>
> þ r1 eSPIO
>
<
Y^SPIO ðk þ 1Þ ¼ ðbÞT X ^SPIO ðk Þ ; ð6Þ
>
>
>
> eSPIO ¼ XAL ðkÞ X ^SPIO ðkÞ
>
>
>
:
erSPIO ¼ Yo ðk Þ Y^SPIO ðkÞ
The sliding technique is applied the ARX-Laguerre PIO to increase the robustness
of the REB signal estimation. So, the following equation is used to estimate the
uncertainties in the REB based on the ARX-Laguerre sliding PIO.
^ SPIO ðk þ 1Þ ¼ u
u ^ SPIO ðk Þ þ r2 erSPIO þ r3 sgn erSPIO : ð7Þ
Here, X^SPIO ðk Þ; Y^SPIO ðkÞ; u

^ SPIO ðk Þ and r3 are the estimation state of the model for
REB based on ARX-Laguerre sliding PIO, the estimation of the REB signal based on
ARX-Laguerre sliding PIO, uncertainties estimation based on the ARX-Laguerre
sliding PIO to minimize the error and increase the robustness in the normal condition,
and the coefficient of the sliding surface, respectively. Apart from improve the
robustness of ARX-Laguerre PIO by ARX-Laguerre sliding PIO, this technique is
suffering from accuracy of signal estimation especially for REB in uncertain condi-
tions. To increase the accuracy in the ARX-Laguerre sliding PIO, the T-S fuzzy logic
algorithm is recommended. The T-S fuzzy algorithm to reduce the estimation error can
be defined based on the following definition.
^ f ðk þ 1Þ ¼ u
if efSPIO ðkÞ is Th then u ^ f ðk Þ þ rf efSPIO : ð8Þ
^ f ðkÞ and rf are the error estimation for REB based on ARX-
Here, efSPIO ðkÞ; Th; u
Laguerre fuzzy-sliding PIO, the threshold value to tuning the estimation accuracy, the
T-S fuzzy estimation behavior, and the coefficient to tuning the fuzzy algorithm,
respectively. Hence, the ARX-Laguerre fuzzy-sliding PIO can be represented by the
following equation.
8
> ^fSPIO ðk þ 1Þ ¼ rX X
X ^fSPIO ðkÞ þ rY Y^fSPIO ðk Þ þ rU Ui ðk Þ þ u
^ fSPIO ðkÞ
>
>
>
>
>
> þ r1 efSPIO
<
Y^fSPIO ðk þ 1Þ ¼ ðbÞT X ^fSPIO ðkÞ ð9Þ
>
>
>
> efSPIO ¼ XAL ðk Þ X ^fSPIO ðkÞ
>
>
>
:
erfSPIO ¼ Yo ðk Þ Y^fSPIO ðk Þ
The T-S fuzzy logic procedure is applied the ARX-Laguerre sliding PIO to increase
the accuracy of the REB signal estimation. Therefore, the following equation is used to
estimate the uncertainties in the REB based on the ARX-Laguerre fuzzy-sliding PIO.
^ fSPIO ðk þ 1Þ ¼ u
u ^ fSPIO ðkÞ þ r2 erfSPIO þ r3 sgn erfSPIO þ r4 u
^ f ðk Þ ð10Þ
^fSPIO ðkÞ; Y^fSPIO ðk Þ; u

Here, X ^ fSPIO ðk Þ and r4 are the estimation state of the model for
REB based on ARX-Laguerre fuzzy-sliding PIO, the estimation of the REB signal
based on ARX-Laguerre fuzzy-sliding PIO, uncertainties estimation based on the
ARX-Laguerre fuzzy-sliding PIO to minimize the error and increase the robustness in
the normal condition, and the coefficient of the T-S fuzzy estimation behavior,
respectively. Regarding Fig. 1, the error of the signal estimation is minimized in the
ARX-Laguerre fuzzy-sliding PIO. So, the residual signal ðrðkÞÞ that can be defined
based on the following equation is minimized in the normal condition.
r ðk Þ ¼ Yo ðk Þ Y^fSPIO ðkÞ: ð11Þ

To increase the fault detection and fault classification performance in the REB, the
feature of
the
energy is extracted from the residual signal in the various conditions. The
energy E of the residual signals can be defined based on the following equation.
Xn
E¼ k¼1
rðkÞ2 : ð12Þ
When the various types of faults issue in REBs, the amplitudes of the energy
signals change drastically. Therefore, the decision-making unit can detect and classify
the faults easily. To detect and classification the faults, a machine learning algorithm
based on the SVM technique is presented in this research. This procedure has a
powerful mathematical background due to the availability of the diverse kinds of
kernels that can be used for training. Therefore, the soft margin SVM is defined by the
following equation [9–13]:

wi £T qðqi Þ þ c wi vi : ð13Þ
where ðqi ; wi Þ; ð£; cÞ; qðqi Þ; and vi are the inputs, outputs, features, and maximum
distance, respectively.
4 Dataset, Results and Discussion
In this research, the Case Western Reserve University (CWRU) bearing dataset [14] is
used to analyze the effectiveness of the SVM-based ARX-Laguerre fuzzy-sliding PI
observer that is shown in Fig. 1. Figure 2 exhibits the data acquisition system.
Fig. 2. The data collection system’s simulator [3, 6].
The vibration sensor collects normal and abnormal conditions that are named by
inner, outer, and ball faults, respectively. The data classification of the CWRU bearing
dataset is represented in [6]. Figure 3 shows the power of proposed method for fault
classification in the REB.
Fig. 3. Energy residual signal for fault classification based on the proposed method.
The experimental results, expressed in terms of the ARX-Laguerre PIO (ALPIO),

ARX-Laguerre sliding PIO (ALSPIO)and SVM-based ARX-Laguerre fuzzy-sliding
PIO (proposed method), are tabulated in Table 1. Furthermore, the confusion matrices
derived for the ARX-Laguerre PIO and SVM-based ARX-Laguerre fuzzy-sliding PIO
(proposed method) are presented in Fig. 4. The results provided in Table 1 and Fig. 4
demonstrate that the proposed method achieve the highest REB fault classification
accuracy with an accuracy about 95.7%.
Table 1. The average of fault detection and classification accuracy when the torque load is 0–3
hp.
Algorithms (Proposed method) ALSPIO ALPIO
Crack diameters (in) 0.007 0.014 0.021 0.007 0.014 0.021 0.007 0.014 0.021
Normal 100% 100% 100% 100% 100% 100% 88% 88% 88%
BF 94% 94.5% 95% 90% 90% 91% 66% 70% 70%
OF 88% 93% 93% 81% 88% 90% 75% 80% 78%
IF 96% 96% 96% 91% 91% 92% 81% 81% 84%
ACA 94.5% 95.9% 96% 90.5% 92.3% 93.3% 78% 80% 80.3%
Fig. 4. Confusion matrices for proposed method, ALSPIO, and ALPIO based on the SVM
method.
5 Conclusions
In this paper, a hybrid technique based on the SVM-based ARX-Laguerre fuzzy-sliding

PI observer is presented for diagnosing REB faults. In the proposed algorithm, the
robust ARX-Laguerre technique is used for REB modeling based on vibration signals.
Next, the ARX-Laguerre PI observer was described, to REB’s signals estimation. After
that, to increase the robustness, the ARX-Laguerre sliding PIO was presented. More-
over, to increase the accuracy, the ARX-Laguerre fuzzy-sliding PIO is proposed.
Additionally, to detect and classification the REB faults, the SVM technique is pre-
sented in this research. The effectiveness of the hybrid observer is validated using a
publicly available vibration dataset of CWRU. Beyond the above, the proposed
algorithm improves the fault classification accuracy from 80% in the ARX-Laguerre
PIO to 95.7%. In our future work, other schemes for combining the control theory and
machine learning approaches will be developed.
Acknowledgements. This research was financially supported by the Ministry of Trade, Industry
& Energy (MOTIE) of the Republic of Korea and Korea Institute for Advancement of Tech-
nology (KIAT) through the Encouragement Program for The Industries of Economic Coopera-
tion Region (P0006123).
References
1. Gao, Z., Cecati, C., Ding, S.X.: A survey of fault diagnosis and fault-tolerant techniques—
Part I: fault diagnosis with model-based and signal-based approaches. IEEE Trans. Ind.
Electron. 62, 3757–3767 (2015)
2. Cecati, C.A.: Survey of fault diagnosis and fault-tolerant techniques—Part II: fault diagnosis
with knowledge-based and hybrid/active approaches. IEEE Trans. Ind. Electron. 62, 3768–
3774 (2015)
3. Piltan, F., Kim, J.M.: Bearing fault diagnosis by a robust higher-order super-twisting sliding
mode observer. Sensors 18, 1128 (2018). https://doi.org/10.3390/s18041128
4. Guo, X., Chen, L., Shen, C.: Hierarchical adaptive deep convolution neural network and its
application to bearing fault diagnosis. Measurement 93, 490–502 (2016)
5. Lou, X., Loparo, K.A.: Bearing fault diagnosis based on wavelet transform and fuzzy
inference. Mech. Syst. Signal Process. 18, 1077–1095 (2004)
6. Piltan, F., Prosvirin, A.E., Jeong, I., Im, K., Kim, J.M.: Rolling-element bearing fault
diagnosis using advanced machine learning-based observer. Appl. Sci. 9(24), 5404 (2019)
7. Bouzrara, K., Garna, T., Ragot, J., Messaoud, H.: Decomposition of an ARX model on
Laguerre orthonormal bases. ISA Trans. 51, 848–860 (2012)
8. Piltan, F., Kim, J.M.: Nonlinear extended-state ARX-laguerre PI observer fault diagnosis of
bearings. Appl. Sci. 9, 888 (2019). https://doi.org/10.3390/app9050888
9. Wu, X., Kumar, V., Ross Quinlan, J., Ghosh, J., Yang, Q., Motoda, H., McLachlan, G.J.,
Ng, A., Liu, B., Yu, P.S.: Top 10 algorithms in data mining. Knowl. Inf. Syst. 14, 1–37
(2008)
10. Prosvirin, A., Kim, J., Kim, J.M.: Efficient rub-impact fault diagnosis scheme based on
hybrid feature extraction and SVM. In: Bhatia, S.K., Tiwari, S., Mishra, K.K., Trivedi, M.C.
(eds.) Advances in Computer Communication and Computational Sciences, pp. 405–415.
Springer, Singapore (2019)
11. Vapnik, V.N.: An overview of statistical learning theory. IEEE Trans. Neural Networks 10,
988–999 (1999)
12. Elangovan, K., Krishnasamy Tamilselvam, Y., Mohan, R., Iwase, M., Takuma, N., Wood,
K.: Fault diagnosis of a reconfigurable crawling-rolling robot based on support vector
machines. Appl. Sci. 7, 1025 (2017)
13. Manevitz, L.M., Yousef, M.: One-Class SVMs for Document Classification (2018)
14. Adams, M.L.: Analysis of rolling element bearing faults in rotating machinery: experiments,
modeling, fault detection and diagnosis. Ph.D. thesis, Case Western Reserve University,
Cleveland, OH, USA (2001)
Modeling Urban Traffic by Means of Traffic
Density Data: Istanbul Case
Tutku Tuncalı Yaman1(&) , Hülya Başeğmez Sezer1 ,

and Emrah Sezer2
1
Department of Management Information Systems, Beykent University,
Istanbul, Turkey
{tutkuyaman,hulyasezer}@beykent.edu.tr
2
Institute of Science and Technology, Informatics Department,
Istanbul University, Istanbul, Turkey
[email protected]
Abstract. The main goal of the proposed research is to perform a predictive

modeling study on Istanbul’s traffic congestion estimation by using traffic
density data. Istanbul Metropolitan Municipality (IMM) shared the data, which
includes traffic density status according to 336 different routes in Istanbul, at the
end of January 2020. Previous studies on the traffic problem in Istanbul have
been limited due to a lack of data. Therefore, it is aimed to perform an initial
study on predictive modeling for Istanbul’s traffic congestion forecast. As a
preliminary result of the analysis, it is seen that the traffic density is low at 93%
accuracy for all locations between 00:00–07:00 am. When the locations are
examined for other hours, it is seen that there was no traffic density at some
locations. In the planned study, intelligent modeling techniques will be per-
formed with identifying out-of-routine situations in traffic flow. Advantages and
disadvantages of predicted models will be discussed according to performance
indicators such as RMSE and MAPE. The superior model will be selected
according to these criteria and it is expected that preferred approach would be a
starting point on future research for predictive forecast studies of Istanbul’s
traffic congestion.
Keywords: Urban traffic Traffic density Traffic forecasting
1 Introduction
The urban freeway is the aorta of urban traffic. However, traffic congestion, which
seriously affects urban traffic efficiency, has become more prominent in recent years.
Istanbul is the most populous city in Turkey and its traffic density is one of the serious
challenges faced by the local population. It is not easy to predict where and when traffic
will be concentrated. According to a recent statistic, the total number of road motor
vehicles in Istanbul is reached 4.187.776 (18% of Turkey) [1]. Detailed figures are
given in Table 1 below.
Table 1. The number of road motor vehicles

Total Car Minibus Bus Small Truck Motorcycle Special Tractor
truck purpose
Share in 18.1% 23.0% 19.8% 19.2% 17.7% 15.9% 10.0% 12.8% 1.3%
Turkey
Share in – 68.7% 2.3% 1.0% 16.1% 3.2% 8.0% 0.2% 0.6%
Istanbul
Istanbul 4187776 2876156 97468 40902 673015 133904 333409 8380 24542
total
As reported in General Directorate of Highways’ (GDH) 2018 Traffic and Trans-

portation Survey, the total road length of Istanbul covers only 5.5% of Turkey while
total vehicle share per km is 10.4% and total passenger share per km is 11%. Detailed
figures are presented in Table 2 below [2].
Table 2. Vehicle-km, Passenger-km and Ton-km values and total road length in 2018
Vehicle-km Passenger-km Ton-km Road length (km)
(1000000) (1000000) (1000000)
High- State High- State High- State High- State Province
way road way road way road way road road
Istanbul 7216.3 4551.5 20291.9 11195.8 1901 5769.9 526 2043 1150
Turkey 23616 89045 66390.0 220255 62118 183705 2159 31021 34153
Share in 30.60% 5.10% 30.60% 5.10% 30.6% 3.10% 24.4% 6.6% 3.40%
total
In the light of the given state of chronic mobility and traffic problems of the city, as
also were stated in [3], and availability of an up-to-date location-based traffic con-
gestion data set, which is shared by the Istanbul Metropolitan Municipality
(IMM) shortly before this study is proposed, the main goal of the research is taken
shape. The hitherto studies which are related to Istanbul’s traffic problem, only cover a
limited perspective since lack of available data (Please refer to [4] and [5]). And none
of them consists of a forecasting model. Thus, we are aimed to conduct an initial study
on predictive modeling for Istanbul’s traffic congestion forecast by pointing out
anomalies in routine road network. IMM’s advanced data set, which will be described
comprehensively in the following chapter, is planned to use. By spending an effort to
elucidate Istanbul’s chronic traffic problem with our model, we are expecting to cast a
light for practitioners in local governments and researchers in subsequent area.
The following sections will briefly introduce the traffic data, the preliminary results
in our study, the various contemporary modeling approaches and anticipated results
based on this traffic data set.
Modeling Urban Traffic by Means of Traffic Density Data 869
2 Traffic Data Set
As a public service, IMM’s Transportation Management Center (TMC) shares both

web and app-based Traffic Density Map (TDM) (see in Fig. 1) for drivers and pas-
sengers by letting them to have a more timely and comfortable journey by coordinating
themselves to alternative routes. TDM is updated with instantaneous traffic density
data, which is transmitted by field traffic measurement devices, meteorological data that
is collected from Automated Road and Weather Observation sensors, and also available
information on road network development work from related departments of munici-
pality. Field traffic measurement devices record bidirectional average speed for high-
ways, which are divided in 400 numbered segments (see in Fig. 3), in 2-min periods.
Fig. 1. Traffic density map [7]
TDM also includes announcements, which cover any car malfunction; car accident
with intervention details and road maintenance & repair work with location and
direction information. Feeding data is obtained by the IMM’s TMC in real-time
monitoring at different time intervals for 24 h.
Fig. 2. Traffic density information in variable message panel [7]

IMM’s TMC is shared a subset of the data used for TDM from IMM’s Open Data
Portal. This portal is launched by IMM’s IT Department and is available online since
January 19, 2020. Measurements made at intervals of 15 min for 258 days were
obtained on 336 different routes. The variables are covered in the dataset are Route-ID
(location), congestion status (low-medium-high), start date, start time, end date, and
end time, respectively [6]. Mentioned data is also used for updating Variable Message
Systems where located on main arteries, by IMM’s TMC. By means of variable
message panels (see in Fig. 2), drivers are informed of traffic accidents, density, and
weather & road condition changes [7].
By request, a finer data is provided from IMM’s TMC for our modeling purposes. It
covers a subset of D.100 state motorway, which only includes segmented bidirectional
traffic density from Altunizade to Avcılar, within the borders of Istanbul. Covered spatial
area is represented by 400 segments and grouped with road segment ID. Along with
segmented-traffic density data, mentioned information that employed in announcements,
was also obtained for anticipated analysis.
Fig. 3. A sample of numerated segments in motorways [7]
3 Preliminary Results
In order to give an idea about current situation of traffic density of such a dense city, a
basic investigation with density data set [6] from IMM’S Open Data Portal is realized
and some representative results are received. According to the outcomes of this
exploratory investigation, traffic density figures in hour intervals, which are calculated
by IMM, are given in Table 3 below.
Table 3. Traffic density in hour intervals

Traffic hour High Medium Low
0 2.7 7.2 90.1
1 1.3 5.5 93.2
2 0.7 5.9 93.5
3 0.6 5.8 93.6
4 0.6 5.8 93.6
5 0.6 5.1 94.3
6 1.2 5.5 93.3
7 6.4 12.3 81.2
8 10.4 16.1 73.5
9 8.3 15.0 76.7
10 7.1 14.4 78.5
11 6.8 14.6 78.6
12 7.8 16.0 76.1
13 8.5 16.1 75.4
14 10.6 18.5 71.0
15 13.1 19.9 67.0
16 15.2 20.6 64.3
17 18.8 20.9 60.3
18 21.6 21.3 57.1
19 17.3 20.0 62.7
20 9.2 14.3 76.5
21 4.9 10.3 84.8
22 4.2 10.0 85.8
23 3.6 9.0 87.4
100
90
80
70
Traffic Density
60
50
40
30
20
10
0
0 25 50 75 100 125 150 175 200 225
Route Id
Fig. 4. Traffic density according to different routes

As illustrated in Fig. 4, different routes have different density shares in total traffic.
First 25 route (coded as Route IDs) which indicate Florya, Sefaköy, Çobançeşme,
Şirinevler, Incirli, Merter, Edirnekapi, Haliç, Okmeydani, Çağlayan, M. Köy, Z. Kuyu,
15 Temmuz Şehitler Köp., Haliç, Cevizlibağ, Merter, Sefaköy, Florya, K.Çekmece,
Istoç, Hasdal, Maslak, FSM Gişeler, Fsm Köp. and Nurtepe directions respectively,
carry nearly 50% of total traffic density.
4 Contemporary Modeling Approaches
Along with the rapid development of the urbanized areas, population growth increased
population density and the total number of car owners, which are increased accord-
ingly; transportation frameworks are confronting many problems on blockage and
reliability of predictive tools. Since transportation time has a crucial role in daily life,
traffic forecast accuracy becomes critical in urban traffic control and navigation.
According to literature, there are a number of important variables exist in traffic
forecast models, such as traffic flow, namely speed, volume, and density [8]. Hence,
there are many difficulties in prediction due to data quality (noise, i.e.), data avail-
ability, and large-scale road infrastructure. Moreover, it is difficult to reveal the
underlying physics of transportation system even if large-scale transportation data are
provided. In the prediction phase, there are two major data-driven approaches: classic
statistical methods (time series analysis, autoregressive integrated moving average
(ARIMA) and its variants) and machine learning (ML) models (k-nearest neighbors
algorithm (KNN), support vector machine (SVM), and neural networks (NN), Random
Forest Regression, Bayesian networks (BNs), Recurrent Neural Network (RNN), Fuzzy
Logic Regression and the hybrid models) [9, 10]. Most of these methods address the
problem as single point forecasting, intended as the ability to predict future values of a
certain physical quantity at a certain location, given only past measurements of the
same quantity at the same location [11]. In literature, most of statistical methods-based
examples were developed for less complex traffic systems. In addition to that, the sizes
of used traffic data sets of these studies were relatively small [12]. At the same time,
most of existing approaches are based only on past and current traffic conditions. Thus,
these limited aspects do not fit well when real-world factors such as traffic accidents
and other unexpected conditions play a part [13].
When we reviewed studies, which aimed to analyze traffic related problems and/or
factors in Istanbul, a number of reports belong to local government professionals were
found. Most of them focus on niche applications and evaluations of smart trans-
portation solutions [14, 15]. Another collaborative study of municipality professionals
and academics in 2016, aimed to predict accurate travel times by picking optimal route.
They analyzed both traffic sensor data and mobile application users’ floating car data
(Probe Data) with an object-oriented approach [16]. A remarkable study that is realized
by independent scholars, use a subset of IMM Traffic Control Center’s traffic speed
data, which was collected via 122 sensors from 1st January to 31st December 2014.
Along with sensor data, weather condition dataset was included in modeling phase.
Since main objective of the stated work was prediction of short-term traffic speed using
KNN and Decision Tree (DTR) regression algorithms, KNN is found as superior
approach. However weather data did not improve their prediction performance [17].
As emphasized previously, most of recent studies in in traffic forecasting modeling
consist of sophisticated algorithms, which cover deep learning models, convolutional
networks, recurrent networks, and generative adversarial networks using multi-source
spatio-temporal data. For further details of these researches, please refer to [9, 12, 18–20].
5 Anticipated Results
In the light of previously mentioned state-of-art approaches and the available data
composition described above, it is aimed to perform an initial predictive modeling
study with a finer data composition comparing preceding studies, for Istanbul’s traffic
congestion forecast by pointing out anomalies in routine road network. In this proposed
research, applicable approaches will be modeled and the pros and cons of predicted
models will be discussed according to well-known performance indicators such as
RMSE and MAPE. The superior model will be selected according to these criteria.
Depending on gathered feedbacks from local government professionals and sug-
gestions of previously published researches, it is expected that the proposed approach
will be shed light on future research for predictive studies of Istanbul’s traffic
congestion.
References
1. TUIK Haber bülteni: Motorlu kara taşıtları. http://www.tuik.gov.tr/PreHaberBultenleri.do?
id=33648. Accessed 30 Jan 2020
2. KGM 2018 Trafik ve ulaşım bilgileri: Otoyollar ve devlet yollarinin trafik dilimlerine göre
yıllık ortalama günlük trafik değerleri ve ulaşım bilgileri. https://www.kgm.gov.tr/
SiteCollectionDocuments/KGMdocuments/Istatistikler/TrafikveUlasimBilgileri/
18TrafikUlasimBilgileri.pdf. Accessed 30 Jan 2020
3. Acar, A., Kocabaş, E., Esen, M.F., Canitez, F.: Green smart technologies for smart cities,
Application of smart city concept to the leading destination: evidence from Istanbul–a case
study. CRC Press, Boca Raton (2019)
4. Dündar, S., Öğüt, K.S.: Determination of basic saturation flow rate in Istanbul. Int. J. Eng.
Technol. 4, 47–52 (2018)
5. Turan, K.Ö.: A solution for traffic jam in Istanbul on east/west direction. Res. Logist. Prod.
5, 311–322 (2015)
6. IBB, Open Data Portal: Değişken mesaj panolarında gösterilen yoğunluk verisi. https://data.
ibb.gov.tr/dataset/degisken-mesaj-panolarinda-gosterilen-yogunluk-verisi. Accessed 23 Jan
2020
7. IBB, Ulaşım Yönetim Merkezi: Trafik bilgilendirme. https://uym.ibb.gov.tr/hizmetler/trafik-
bilgilendirmee. Accessed 25 Apr 2020
8. Yu, H.Y.B., Zhu, Z.: Spatio-temporal graph convolutional networks, A deep learning
framework for traffic forecasting. IJCAI, Stockholm (2018)
9. Smith, B.L., Williams, B.M., Keith Oswald, R.: Comparison of parametric and non-
parametric models for traffic flow forecasting. Transp. Res. Part C 10, 303–321 (2002)
10. Guo, K., Hu, Y., Qian, Z., Liu, H., Zhang, K., Sun, Y., Gao, J., Yin, B.: Optimized graph
convolution recurrent neural network for traffic prediction. IEEE Trans. Intell. Transp. Syst.
1–12 (2020)
11. Lippi, M., Bertini, M., Frasconi, P.: Collective traffic forecasting. In: Machine Learning and
Knowledge Discovery in Databases. In: ECML PKDD 2010. Lecture Notes in Computer
Science, vol. 6322. Springer, Berlin (2010)
12. Cui, Z., Henrickson, K., Ke, R., Wang, Y.: Traffic graph convolutional recurrent neural
network: a deep learning framework for network-scale traffic learning and forecasting. IEEE
Trans. Intell. Transp. Syst. pp. 1–12 (2019)
13. Lin, L., Li, J., Chen, F., Ye, J., Hall, J.: Road traffic speed prediction: a probabilistic model
fusing multi-source data. IEEE Trans. Knowl. Data Eng. 30, 1310–1323 (2017)
14. Dilek, E., Kurt, Y., Öztürk, S., Erşahin, M., Özcan, Ö., Ünal, A.: “İBB yol gösteren” ile
akıllı hareketlilik. In: Istanbul Transport Congress and Exhibition Proceedings, Istanbul
(2017). https://uym.ibb.gov.tr/documents/library/İBB-Yol-Gosteren-Makale.pdf. Accessed
20 Apr 2020
15. Dilek, E., Ayözen, Y.E.: Smart Mobility in Istanbul with “IBB CepTrafik”. In: IEEE/IFIP
Network Operations and Management Symposium Proceedings, pp. 1273–1278, Istanbul
(2016)
16. Yılmaz, Y., Uludağ, S., Dilek, E., Ayizen, Y.E.: A preliminary work on predicting travel
times and optimal routes using Istanbul’s real traffic data. In: Istanbul Transport Congress
and Exhibition Proceedings, Istanbul (2016). https://uym.ibb.gov.tr/documents/file/pdf/
Transist_TKM.pdf. Accessed 20 Apr 2020
17. Gülaçar, H., Yaslan, Y., Oktuğ, S.F.: Short term traffic speed prediction using different
feature sets and sensor clusters. In: IEEE/IFIP Network Operations and Management
Symposium Proceedings, Istanbul, pp. 1265–1268 (2016)
18. Peng, H., Wang, H., Du, B., Bhuiyan, M.Z.A., Ma, H., Liu, J., Wang, L., Yang, Z., Du, L.,
Wang, S., Yu, P.S.: Spatial temporal incidence dynamic graph neural networks for traffic
flow forecasting. Inf. Sci. 521, 277–290 (2020)
19. Djenouri, Y., Belhadi, A., Lin, J.C.-W., Cano, A.: Adapted k-nearest neighbors for detecting
anomalies on spatio–temporal traffic flow. IEEE Access 7, 10015–10027 (2019)
20. Zhang, Y., Wang, S., Chen, B., Cao, J., Huang, Z.: TrafficGAN: network-scale deep traffic
prediction with generative adversarial nets. IEEE Trans. Intell. Transp. Syst. 1–12 (2019)
Realising Newspaper Sales by Using
Statistic Methods
Onur Dogan1(B) and Omer Faruk Gurcan2

1
Department of Industrial Engineering, İzmir Bakırçay University,
35665 Izmir, Turkey
[email protected]
2
Department of Industrial Engineering, İstanbul Technical University,
34367 İstanbul, Turkey
[email protected]
Abstract. Today, some newspapers have started to service online only.

One of the reasons of this situation is competitiveness and increasing
costs in press market. So newspaper sale planning requires adaptive
scheduling and understanding customer behavior well. Estimating the
right number of delivery is crucial. It is succeeded with expert knowl-
edge supported with proper analysis and techniques. This study focuses
on understanding newspaper sales amount by revealing critical variables
and their importance on sales. This is a critical shortcoming especially
for newspaper companies, which always try to reduce the number of
sell outs and increase the number of return in the delivered newspa-
pers offer companies financial benefits. In this study, it was analyzed
that whether different weather conditions (sunny, cloudy or rainy) and
days (weekday or weekend) are effective in daily newspaper sales for
Ankara, İzmir and İstanbul. One-way Anova, Two-way Anova, t-test,
Levene’s, Kolmogorov-Smirnov, and The Kruskal–Wallis tests were used.
Sales data was collected from 71 vendors of a national press firm in three
cities between 12 September–31 December 2018.
Keywords: Newspaper sales prediction · Statistical methods · Anova
1 Introduction
Sales (demand) forecasting is one of the most significant topics for organizations
as it reveals the needed inventory level to satisfy customer demand and prohibit
under-stock and over-stock problems. However, typical forecasting methods do
not commonly satisfy circulation requirements due to many factors.
A newspaper is a special product that has some specific characteristics. It
must be produced daily. No stock available. Delays in delivery cause the infor-
mation to become old. If the information is delayed, the product is not sold
commercially. For this reason, past traditional sales forecast methods are not
applicable to the newspaper industry. Moreover, Turkey’s newspaper reading
rate is low compared to the situation in Europe and America. As an example,
876 O. Dogan and O. F. Gurcan
newspaper readers in Turkey give up to buy a newspaper when the weather rains
or wake up late, the situation is different in developed countries because they are
already subscribers. In summary, the scarcity of readers, the shrinking market
and increasing raw material costs require a specific method to meet the needs
of printed publications, especially the newspaper sector. In this way, it is possi-
ble to prevent waste of various issues such as paper consumption, labor, energy.
Although today’s technology can collect a huge amount of data, failure to use
the collected data correctly and effectively during the decision-making process
is a great loss of information. It is essential to process these data scientifically
using statistical methods. Therefore prediction of the right sales amounts in the
newspaper industry and distributing it to the end sellers is a significant cost
problem.
Newspaper circulation is a daily basis job that needs comprehensive prepara-
tion. It is critical to determine the correct distribution numbers for each vendor.
The transferred too many newspapers will be returned thanks to not be sold.
Each returned newspaper has an extra cost for the company. On the other hand,
when the number of returns is “0” (no sales), it can refer to more sales possi-
bilities, which results in a drop in profit. Because of that, a successful approach
is needed to predict the demands of the newspaper considering some exogenous
factors. Therefore the number of circulation for each vendor should be optimized
and predicted with the least possible error.
In this study, weather conditions and day of the week effect on daily news-
paper sales amount is analyzed with data collected from dealers in 3 cities of
Turkey. Various statistical methods in SPSS are used. In Turkey, there is lim-
ited study in that era. This study shed light on practitioners by comparing two
important variable effect on newspaper sales of different cities.
This study organized as follows: first, related works are given in the literature.
Statistical methods are defined briefly. Details of the problem is given as a case
study and analysis results are presented. Lastly, obtained results are discussed.
2 Related Works
Sales (demand) forecasting is a broad spectrum area in the literature. It has many
applications in finance [9,15,25,26], retail [1,7,10,12], energy [2,13,19], supply
chain [4,5,22] varying from autoregressive integrated moving average (ARIMA),
artificial neural network (ANN), support vector machines (SVM), extreme learn-
ing machines (ELM). Table 1 presents an overview to current works on sales
forecasting.
Because of the special characteristics of newspapers, some studies focused on
determining which exogenous factors have more effect on newspaper sales [18,
21]. Some studies applied different machine learning techniques such as extreme
learning machines (ELM) [6,7,17,23], and deep learning [16,25]. However, these
studies ignored the specific characteristics of the newspaper. Since each method
has some drawbacks, various studies integrated at least two data-based analysis
approaches [7,9,23]. Studies focused on newspapers are few and considered linear
Realising Newspaper Sales by Using Statistic Methods 877
Table 1. An overview to current studies on sales forecasting.
Study Quick fact about study

[6] It expresses that there are many factors affecting sales,
extreme learning machines estimation algorithm is used
[21] Some of the factors affecting sales are listed in this study
[18] Some of the factors affecting sales were investigated using
data mining methods
[16] It realizes the sales forecast in the field of clothing with its
deep learning method
[23] An intelligent sales forecasting system that includes
extreme learning machines has been developed. It can also
be useful if a hybrid system is installed for the future
directions
[17] A strong sales forecast has been made by combining with
the extreme learning machines and ensemble learning
algorithm. Algorithms that give bad results within the
study can be strengthened with ensemble learning
[7] It is a hybrid work with extreme learning machines in the
clothing industry. The study gave good results with limited
data in limited time. It may be needed when creating the
training data set
[8] This is an exemplary study of using clustering methods in
prediction problems. Considering that the firm has 60 end
sellers, clustering methods are needed
[9] A hybrid and intelligent forecasting system has been
developed by combining ARIMA as a classical method and
ANN as a machine learning method
[3] It is a sales forecast study developed to strengthen the
after-sales support of artificial neural networks
[14] It describes the work when the existing system was first
installed in the same company
[11] It is a newspaper distribution study prepared by
considering linear trend and seasonality
[20] It is a study that predicts newspaper sales with ARIMA
method
[24] It is a circulation estimation study performed with
supporter vector regression. To our knowledge, it is the
only study done in this field with basic machine learning
trends [11,14,20]. However, circulation is affected by many exogenous parameters

and cannot be linear.
In conclusion, although data-based methods using historical data are used to
predict the future situation in many areas, no circulation prediction study has
been found in the press (newspaper) domain in this context. Newspaper sales
prediction varies from other types of predictive models because it is in a field
that has special characteristics. In this respect, the study is expected to guide
other researchers by filling a gap in the literature.
3 Statistical Methods
In this study, Levene’s Test, Independent Samples T-Test, Kolmogorov Smirnov

test, One-Way Anova, Two-Way Anova, and Kruskal Wallis test were used. Their
summarizer descriptions are given below:
Levene’s test is an inferential statistic used to assess the equality of vari-
ances for a variable calculated for two or more groups. Some common statistical
procedures assume that variances of the populations from which different sam-
ples are drawn are equal. Levene’s test assesses this assumption. It tests the
null hypothesis that the population variances are equal (called homogeneity of
variance or homoscedasticity).
The Independent Samples t-test compares the means of two indepen-
dent groups in order to determine whether there is statistical evidence that the
associated population means are significantly different.
Kolmogorov-Smirnov test shows that samples are drawn from the same
distribution or not.
The Kruskal–Wallis test by ranks, Kruskal–Wallis H test or one-way
Anova on ranks is a non-parametric method for testing whether samples originate
from the same distribution. It is used for comparing two or more independent
samples of equal or different sample sizes.
One-way analysis of variance (Anova) is a technique that can be used
to compare means of two or more samples (using the F distribution).
The two-way Anova is an extension of the one-way Anova that examines
the influence of two different categorical independent variables on one continuous
dependent variable.
4 Case Study
4.1 Descriptive Information
Data includes newspaper sales amount between 12 September–31 December 2018

from distributors in Ankara, İzmir, and İstanbul. Table 2 shows the details con-
sidering day variable (weekday and weekend) by cities. The average weekday
sales is more than the average weekend sales in Ankara and İstanbul, it is vice
versa in İzmir.
Table 3 shows the average sales amount by weather (sunny, cloudy, and rainy)
and standard deviations of these sales amount by cities. Weather conditions
are grouped into three main categories, sunny, cloudy and rainy and they were
represented as 1, 2 and 3, respectively.
Table 2. Sales amount considering Day variable
City Number of vendors Day Average sales Total sales amount

Ankara 14 Weekday 4 640,39 475 897
Weekend 3 415,81
İzmir 19 Weekday 7 771,9 865 264
Weekend 7 852,63
İstanbul 38 Weekday 17 259,68 1 895 067
Weekend 16 611
Table 3. Sales amount considering weather conditions.
Cities Weather Occurrence Average sales Std. deviation Min. Max.

Ankara 1 36 4 045,78 2 029,567 2 562 9 197
2 66 4 423,90 2 235,12 2 519 9 438
3 9 4 247,56 2 143,66 2 557 8 554
İzmir 1 13 7 318 998,526 6 132 9 135
2 83 7 959,27 1 728,058 5 752 12 093
3 15 7 299,27 1 925,174 5 850 11 978
İstanbul 1 15 16 415,8 3 947,583 12 598 26 362
2 77 17 735,17 4 762,834 10 964 27 644
3 19 14 903,79 3 905,338 11 733 25 563
Table 4. Levene’s and t-tests results considering day variable.
City Variances Levene’s test t-test for equality

for equality of of means
variances
F Sig. t df Sig. Mean Std. error
(2-tailed) difference difference
Ankara Equal 121,085 0 2,863 109 0,005 1 224,58 427,735
Not equal 4,468 81,477 0 1 224,58 274,055
İzmir Equal 62,271 0 −0,23 109 0,824 −80,726 361,325
Not equal −0,34 93,44 0,737 −80,726 239,6
İstanbul Equal 112,046 0 0,668 109 0,506 648,684 971,054
Not equal 1,024 87,79 0,308 648,684 633,213
Table 4 gives the Levene’s and t-test results for the day variable. For Ankara,
according to Levene’s Test or Equality of Variances, Equal variances not assumed
(Sig < 0.05) and t-test show that there is a difference between groups’ means
(weekend and weekday) (Sig < 0.05). According to the Kolmogorov-Smirnov test
result (Sig < 0.05), the distribution of sales amounts by days (weekday-weekend)
is different. For İzmir, according to Levene’s Test or Equality of Variances,
Table 5. One-way Anova results considering weather conditions.
City Sum of squares df Mean square F Sig.

Ankara Between groups 3 357 674,8 2 1 678 837,4 0,374 0,68
Within groups 484 977 230,7 108 4 490 529,9
Total 488 334 905,5 110
İzmir Between groups 8 890 970,1 2 4 445 485,0 1,522 0,223
Within groups 315 355 231,6 108 2 919 955,8
Total 324 246 201,7 110
İstanbul Between groups 129 710 929,3 2 64 855 464,6 3,154 0,04
Within groups 2 220 689 845,0 108 2 919 955,8
Total 2 350 400 774,3 110
Table 6. Two-way Anova results considering weather and day variables.
City Source df Mean square F Sig.

Ankara Intercept 1 668 153 983,193 155,993 ,000
Weather 2 675 692,109 ,158 ,854
Day 1 14 827 959,260 3,462 ,066
Weather * Day 2 663 939,270 ,155 ,857
Error 105 4 283 238,816
Total 111
İzmir Intercept 1 1 382 951 632,345 461,444 ,000
Weather 2 1 943 355,038 ,648 ,525
Day 1 124 212,086 ,041 ,839
Weather * Day 2 333 696,562 ,111 ,895
Error 105 2 997 010,583
Total 111
İstanbul Intercept 1 10 617 117 725,644 512,185 ,000
Weather 2 16 886 583,109 ,815 ,446
Day 1 29 251,994 ,001 ,970
Weather * Day 2 9 274 734,123 ,447 ,640
Error 105 20 729 058,609
Total 111
Equal variances not assumed (Sig < 0.05) and t-test shows that there is not a dif-
ference between groups’ means (weekend and weekday) (Sig > 0.05). According
to the Kolmogorov-Smirnov test (Sig < 0.05), the distribution of sales amounts by
days is different. For İstanbul, according to Levene’s Test or Equality of Variances,
Equal variances not assumed (Sig < 0.05) and t-test shows that there is no differ-
ence between groups’ means (weekend and weekday) (Sig > 0.05). According to
the Kolmogorov-Smirnov test (Sig < 0.05), the distribution of sales amounts by
days is different.
According to One-way Anova given in Table 5, there is no difference between
various weather conditions (sunny, cloudy, rainy) in terms of sales amount for
Ankara and İzmir (Sig. > 0.05). These results are supported by Kruskal Wallis
Test (Sig. > 0.01). For İstanbul, there is a difference between various weather
conditions in terms of sales amount (Sig. < 0.05). This result supported by
Kruskal Wallis Test, weather condition reasons difference in sales amount in 1%
significance level (Sig. < 0.01).
According to two-way Anova shown in Table 6, for Ankara, İzmir and İstanbul
weather conditions and day variables do not affect sales amount (Sig. > 0.05).
There is no evidence of a significant interaction between the two factors (weather
and day) (intercept). There was no statistically significant difference in group
means for various weather conditions and days in terms of sales amount for cities.
When weather conditions (which is categorized in sunny, cloudy and rainy) effect
on newspaper sales amount is analyzed, it is seen that weather condition reason
difference in sales amount in İstanbul but not in Ankara and İzmir. Besides,
the other variable – day (weekday, weekend)-effect was analyzed. Results show
that equal variances not assumed and distribution of sales amounts by day is
different significantly for İstanbul, Ankara, and İzmir. For Ankara, there is a
significant difference between groups’(weekend-weekday) means in terms of sales.
Lastly, when two variable is taken together in two-way Anova, it is obtained that
weather conditions and day do not make a difference.
In this study, two categorical variables’ effect on sales amount taken into
consideration. For future studies, more and not just categorical also continuous
variables can be used. This enables developing more models. In the study, the
analysis period is from September to December 2018. So one-year data will be
more reliable and beneficial. Besides different years data can be analyzed and
compared. Although three major cities in Turkey were analyzed, more cities can
be considered and results can be compared. A monthly sale forecasting study
can be done with proper variables.
References
1. Badorf, F., Hoberg, K.: The impact of daily weather on retail sales: an empirical
study in brick-and-mortar stores. J. Retail. Consum. Serv. 52, 101921 (2020)
2. Beyca, O.F., Ervural, B.C., Tatoglu, E., Ozuyar, P.G., Zaim, S.: Using machine
learning tools for forecasting natural gas consumption in the province of Istanbul.
Energy Econ. 80, 937–949 (2019)
3. Bohanec, M., Borštnar, M.K., Robnik-Šikonja, M.: Explaining machine learning
models in sales predictions. Expert Syst. Appl. 71, 416–428 (2017)
4. Boone, T., Ganeshan, R., Jain, A., Sanders, N.R.: Forecasting sales in the supply
chain: consumer analytics in the big data era. Int. J. Forecast. 35(1), 170–180
(2019)
5. Bruzda, J.: Quantile smoothing in supply chain and logistics forecasting. Int. J.
Prod. Econ. 208, 122–139 (2019)
6. Chen, F., Ou, T.: Sales forecasting system based on gray extreme learning machine
with Taguchi method in retail industry. Expert Syst. Appl. 38(3), 1336–1345 (2011)
7. Choi, T.M., Hui, C.L., Liu, N., Ng, S.F., Yu, Y.: Fast fashion sales forecasting with
limited data and time. Decis. Support Syst. 59, 84–92 (2014)
8. Dai, W., Chuang, Y.Y., Lu, C.J.: A clustering-based sales forecasting scheme using
support vector regression for computer server. Procedia Manuf. 2, 82–86 (2015)
9. Domingos, S.d.O., de Oliveira, J.F., de Mattos Neto, P.S.: An intelligent hybridiza-
tion of ARIMA with machine learning models for time series forecasting. Knowl.
Syst. 175, 72–86 (2019)
10. Floyd, K., Freling, R., Alhoqail, S., Cho, H.Y., Freling, T.: How online product
reviews affect retail sales: a meta-analysis. J. Retail. 90(2), 217–232 (2014)
11. Garcia, A.M., Pak, G., Oduro, F., Thompson, D., Erazo, K., Gilkey Jr., J.: Season-
ality and the newspaper distribution problem: using data visualization to improve
trend line forecasts. Acad. Bus. Res. J. 1, 18–23 (2017)
12. Guidolin, M., Guseo, R., Mortarino, C.: Regular and promotional sales in new
product life cycles: competition and forecasting. Comput. Ind. Eng. 130, 250–257
(2019)
13. Hu, H., Wang, L., Peng, L., Zeng, Y.R.: Effective energy consumption forecasting
using enhanced bagged echo state network. Energy 193, 116778 (2020)
14. İncesu, G., Aşıkgil, B., Tez, M.: Sales forecasting system for newspaper distribution
companies in Turkey. Pakistan J. Stat. Oper. Res. 8(3), 685–699 (2012)
15. Liu, J., Ma, F., Zhang, Y.: Forecasting the chinese stock volatility across global
stock markets. Phys. A 525, 466–477 (2019)
16. Loureiro, A.L., Miguéis, V.L., da Silva, L.F.: Exploring the use of deep neural
networks for sales forecasting in fashion retail. Decis. Support Syst. 114, 81–93
(2018)
17. Lu, C.J., Kao, L.J.: A clustering-based sales forecasting scheme by using extreme
learning machine and ensembling linkage methods with applications to computer
server. Eng. Appl. Artif. Intell. 55, 231–238 (2016)
18. Murray, P.W., Agard, B., Barajas, M.A.: Forecasting supply chain demand by
clustering customers. IFAC PapersOnLine 48(3), 1834–1839 (2015)
19. Nam, K., Hwangbo, S., Yoo, C.: A deep learning-based forecasting model for renew-
able energy scenarios to guide sustainable energy policy: a case study of Korea.
Renew. Sustain. Energy Rev. 122, 109725 (2020)
20. Permatasari, C.I., Sutopo, W., Hisjam, M.: Sales forecasting newspaper with
ARIMA: a case study. In: AIP Conference Proceedings, vol. 1931, p. 030017. AIP
Publishing LLC (2018)
21. Tanizaki, T., Hoshino, T., Shimmura, T., Takenaka, T.: Demand forecasting in
restaurants using machine learning and statistical analysis. Procedia CIRP 79,
679–683 (2019)
22. Wang, C.H., Chen, J.Y.: Demand forecasting and financial estimation considering
the interactive dynamics of semiconductor supply-chain companies. Comput. Ind.
Eng. 138, 106104 (2019)
23. Wong, W., Guo, Z.: A hybrid intelligent model for medium-term sales forecasting
in fashion retail supply chains using extreme learning machine and harmony search
algorithm. Int. J. Prod. Econ. 128(2), 614–624 (2010)
24. Yu, X., Qi, Z., Zhao, Y.: Support vector regression for newspaper/magazine sales
forecasting. Procedia Comput. Sci. 17, 1055–1062 (2013)
25. Zhong, X., Enke, D.: Predicting the daily return direction of the stock market
using hybrid machine learning algorithms. Financ. Innov. 5(1), 4 (2019)
26. Zhou, Z., Gao, M., Liu, Q., Xiao, H.: Forecasting stock price movements with
multiple data sources: evidence from stock market in China. Physica A Stat. Mech.
Appl. 542, 123389 (2019)
Extensions of Fuzzy Sets in Big Data
Applications: A Literature Review
Nurşah Alkan(&) and Cengiz Kahraman

{nalkan,kahramanc}@itu.edu.tr
Abstract. Nowadays, with the increase in technological developments and the

widespread use of the internet, large amounts of data are produced from many
sources, revealing huge and heterogeneous data difficult to process. Therefore,
big data having an enormous volume and high velocity of data with complex
structures have recently drawn substantial interest from not only academics but
also practitioners. While academic researchers focus on understanding the
concept, identifying it, and developing related methodologies, companies focus
on how to transform the potential of this technology into business values and
how they can benefit from this technology. Researchers have proposed new
research paradigms by addressing big data more efficiently to guide both liter-
ature and businesses on these issues. Fuzzy sets have been accepted as a suitable
method to represent and quantify aspects of uncertainty on big data. However,
there are very few systematic research reviews that capture the dynamic nature
of this issue for both academics and businesses who want to research this topic.
Therefore, this study takes into consideration the studies employing fuzzy sets in
big data applications. We aim to present a literature review to lead the researches
on the existing literature and the most recent advances on big data. A large
number of papers employing fuzzy sets in big data applications have been
analyzed with respect to some characteristics such as subject area, published
journal, publication year, source country, and document type.
Keywords: Big data Literature review Data mining Fuzzy sets

Extensions of fuzzy sets
1 Introduction
Data are generated from sensors, devices, video/audio, networks, log files, computing
applications, web, and social media [1]. A large amount of data produced increasingly
from these sources has emerged the concept of big data by creating huge and
heterogeneous data that are difficult to process. Big data is defined as high-variety,
high-volume and/or high-velocity information assets that require cost-effective, inno-
vative forms of information processing, as well as providing enhanced decision making
and process automation [2]. Due to the volumetric size of the data produced by many
devices and systems, the value of the data has been widely understood day by day and
this has become widespread the data analysis applications by revealing the subject of
data processing. In order to obtain meaningful value from big data, the most
Extensions of Fuzzy Sets in Big Data Applications 885
appropriate processing power, analytical capabilities, and information management

skills are needed [3]. Various techniques and technologies that develop to aggregate,
manipulate, analyze, and visualize big data benefit from several fields such as applied
mathematics, economics, computer science, and statistics. Therefore, businesses that
aim to obtain meaningful value from big data should adopt a flexible and multidisci-
plinary approach [4]. However, most techniques and technologies lack meeting the
ideal requirements of each processing step and so it is essential to develop creative and
scalable techniques to improve data management and programming capabilities, to
analyze and understand large-scale datasets with complex structures [5]. Nowadays,
various fuzzy set techniques including extensions of fuzzy sets have become a suitable
methodology and tool for big data analysis. The fuzzy sets introduced by Zadeh [6]
address the lack of information and imprecise information in complex systems, as well
as modeling uncertainties in human decisions and it is applied in various areas such as
control systems, pattern recognition, statistics, and machine learning. Fuzzy sets can be
developed big data techniques as well as mitigate the existing big data challenges.
This study aims to offer a literature review on big data based on fuzzy set tech-
niques. Since it is a new research area, there are still very few studies on this subject.
Therefore, it is important to investigate and summarize for researchers the status of
researches realized on fuzzy set-based big data applications. The originality of this
paper comes from the literature research on fuzzy sets and extensions in big data
applications.
The rest of this paper is organized as follows: the concept of big data is briefly
introduced in Sect. 2. General information about the extensions of fuzzy sets is given in
Sect. 3. A literature review on fuzzy techniques used in the big data field is presented in
Sect. 4. Section 5 includes an analysis of the literature for researchers who want to use
fuzzy techniques in big data problems. Lastly, the obtained results and future research
suggestions have been explained in Sect. 5.
2 Concept of Big Data
There is not a unique definition of big data accepted by researchers, as one of the most
important research topics of the present and future. Depending on what type of soft-
ware tools are widely available and what sizes of datasets are common in a specific
industry, the term big data can change by sector [4]. Laney [7] has defined big data as
volume, variety, and velocity, characterized by 3Vs. However, other characteristics
such as value and veracity, have also frequently been used [5]. A brief explanation of
these characteristics used to define big data is shown in Fig. 1. Many companies are to
adopt data-driven decision making by investing heavily in this technology because big
data doesn’t have a value alone. To enable decision making, organizations need to have
efficient processes that move quickly in high sizes and volumes and turn various data
into meaningful insights [8]. It is of great importance to understand the sources of big
data analysis so that every company can obtain a competitive advantage in its sector by
improving productivity. Big data analytics implements advanced analytic techniques
on huge, diverse data sets with unstructured, semi-structured, and structured data, from
different sources in different sizes.
886 N. Alkan and C. Kahraman
Fig. 1. The 5Vs view at big data
Big data analysis provides businesses, researchers, and analysts with faster and
better decisions-making using data that was previously unattainable or unusable.
Organizations can use advanced analytics techniques such as text analytics, audio
analytics, machine learning, genetic algorithms, predictive analytics, data mining,
neural networks, statistics, and natural language processing to gain new insights from
these data [2, 8]. Advanced tools and technologies that develop Big Data science are
used to manage, analyze, and visualize information from large, diverse, distributed, and
heterogeneous datasets. These tools and technologies such as cloud computing, gran-
ular computing, quantum computing, biological computing systems, business intelli-
gence, Cassandra, Dynamo, Hadoop, MapReduce, SQL, R, and visualization are used
together by supporting big data analysis techniques [4, 9].
Businesses can provide many advantages such as increasing operational efficiency,
developing better customer service, informing strategic direction, identifying new
customers and markets, identifying and developing new products and services by using
big data and analytics. Although big data technologies offer many advantages to its
users, it also reveals difficulties in data capture, storage, search, sharing, analysis, and
visualization [9].
3 Extension of Fuzzy Sets
After fuzzy sets were initially developed by Zadeh [6], fuzzy sets have been expanded
in various forms to better address the lack of information in complex systems and
imprecise information. Type 2 fuzzy sets were introduced by Zadeh [10] as an
extension of ordinary fuzzy sets. As different from ordinary fuzzy sets, intuitionistic
fuzzy sets developed by Atanassov [11] are expressed with membership degree and
non-membership degree of an element. Their sum must be less than or equal to one.
Hesitant fuzzy sets proposed by Torra [12] enable to have different values for the
membership degrees of an element in a set. After intuitionistic type-2 fuzzy sets (IFS2)
were proposed by Atanassov [13], IFS2 has been called as Pythagorean fuzzy sets by
Yager [14]. Pythagorean fuzzy sets are characterized by a membership degree and a
non-membership degree the condition that the sum of their squares has to be equal to
one at most. Q-rung orthopair fuzzy sets introduced by Yager [15] are a general class of
IFSs and PFSs. The sum of qth power of membership degree and a non-membership
degree have to at most equal to one. After that, when q = 3, Senapati and Yager [16]
have expressed as fermatean fuzzy sets (FFSs) to q-rung orthopair fuzzy sets. Neu-
trosophic sets (NSs) developed by Smarandache [17] as an extension of intuitionistic
fuzzy sets are defined with a degree of truthiness, indeterminacy, and falsity of an
element in a set. Spherical fuzzy sets (SFSs) introduced by Kutlu Gündoğdu and
Kahraman [18] are characterized by a membership degree, a non-membership de-gree,
a hesitancy degree which is a generalization of Pythagorean fuzzy sets and neutro-
sophic sets. In SFSs, the square sum of its membership degree, non-membership
degree, and hesitancy degree is equal to or less than one.
4 Literature Review
Various new techniques have been developed to obtain meaningful value from big data
having huge and heterogeneous data, called as data-intensive science. Fuzzy sets
decrease the existing big data challenges by preprocessing data or by reconstructing the
problems thanks to their abilities to represent and quantify aspects of uncertainty.
Although there are many studies on big data in the literature, there are very few studies
using techniques related to fuzzy sets and their extensions in big data. Most of these
studies employ ordinary fuzzy sets in their big data analyses whereas a small portion of
these studies uses extensions of ordinary fuzzy sets. In this paper, due to this lack of
literature, a literature review has been conducted to shed light on academics and
practitioners who want to research in this field. Because of the space constraints, some
of the studies using fuzzy sets techniques for big data problems have been briefly
summarized in the following.
Bi et al. [19] proposed a novel clustering algorithm by integrating the Axiomatic
Fuzzy Set and Subtractive Clustering Method since available churn prediction models
can’t work very well and decision-makers are always faced with uncertain operations
management when dealing with big data in the industry. They indicated that the pro-
posed model fully expressed the significance of fuzzy concept. Zeng et al. [20] utilized
a fuzzy rough set approach for incremental feature selection on hybrid information
systems. Experimental results have shown that the algorithm developed based on fuzzy
rough set increases efficiency. Deng et al. [21] proposed a fused fuzzy deep neural
network that simultaneously extracts information from both fuzzy and neural repre-
sentations. They have confirmed the effectiveness of the model on three practical tasks
containing a high degree of uncertainty in raw data. The fuzzy deep neural network
paradigm has performed significantly better than other non-fuzzy approaches in these
tasks. Xu and Yu [22] introduced a new information fusion method that examines
uncertainty measures by converting the information of each object into a triangular
fuzzy information granule. They indicated that the proposed triangular fuzzy approach
is efficient and effective for information fusion in multi-source datasets by performing
extensive experiments on six datasets. Azar and Hassanien [23] presented a linguistic
hedges neuro-fuzzy classifier for dimensionality reduction, feature selection, and
classification. They showed that the proposed method not only helps to reduce the
dimensionality of large data sets but also can speed up the computation time of a
learning algorithm and simplify the classification tasks. Havens et al. [24] performed a
comparative analysis to demonstrate their effectiveness by using different methods of
fuzzy c-means clustering techniques on big data. Lou et al. [25] proposed a data-driven
approach for customer requirements discernment. They used intuitionistic fuzzy sets to
handle the uncertainty in the semantic expression level. Hosseini et al. [26] developed a
distributed density-based clustering approach that benefits from the hesitant fuzzy
weighted correlation coefficient as its similarity measure. They have obtained that the
proposed approach performed better than MapReduce based approaches in the com-
putational load while showing better sensitivity and validity index compared to
MapReduce based algorithms. Son [27] presented a new distributed picture fuzzy
clustering method based on picture fuzzy sets. As a result of the experimental studies
conducted in various data sets, it has been seen that the clustering quality of the
proposed method obtained better results than other fuzzy methods and algorithms. Ren
et al. [28] developed a consensus model based on hesitant fuzzy linguistic term sets
using a large-scale group decision-making method with social network analysis on
micro-grid planning. Shukla et al. [29] proposed a new method to address the veracity
characteristic of the big data using the concept of the footprint of vagueness in interval
type-2 fuzzy sets. It has been observed that the proposed method reduces the instances
to a manageable extent since it addresses the veracity issue in big data.
5 Analysis of Published Papers
In this section, a comprehensive review of fuzzy set techniques used to solve big data
problems has been presented by using the Scopus database. The literature review has
been realized by considering papers published in international journals indexed in this
database. In the review process, keywords based on “big data” & “fuzzy” have been
analyzed in two ways. Firstly, 182 papers obtained by searching according to “Article
title” in the database have been analyzed. Secondly, 1720 papers obtained by searching
according to “Article title, Abstract, Keywords” have been analyzed. Analyses have
been made based on subject areas, published journals, publication years, source
countries, and document types. The distribution of these studies has been investigated
to identify patterns, trends, and gaps in the literature. The results of the analysis have
been presented through figures.
Firstly, the literature has been analyzed for both search types according to the
publication year as given in Fig. 2. The studies used fuzzy techniques based on big data
have increased together with the increasing importance of big data year by year. There
is a considerable tendency towards fuzzy methods for the solution of big data problems
that are of great importance for businesses in recent years as seen in Fig. 2.
Fig. 2. Number of publications using fuzzy techniques on big data problems according to years.
Figures 3 and 4 are to show the percentage distributions of subject areas that used
fuzzy techniques on big data problems for both search types, respectively. According to
Figs. 3 and 4, computer science is the most implemented subject area with a rate of
43% for both search types.
Fig. 3. Percentage distribution of publications according to subject areas for first search type.
Fig. 4. Percentage distribution of publications according to subject areas for second search type.
Figures 5 and 6 display the studies using fuzzy techniques on big data with respect
to countries for both search types. The countries, China, India, USA, UK, and Spain,
come to the forefront with the number of studies they have conducted for both search
types while Turkey is seen to take place in the last ranks by staying in the background.
Fig. 5. Percentage distribution of publications according to country for first search type.
Fig. 6. Percentage distribution of publications according to country for second search type.
Figure 7 is to indicate the percentages of document types for both search types.
According to the results of the analysis, it is obtained that the document type with the
highest percentage for the first search type is the article and the conference paper for the
second research type.
Fig. 7. Percentage distribution of publications according to document type for both search
types.
The percentage distributions of journals published for both search types have
indicated in Figs. 8 and 9, respectively. IEEE International Conference on Fuzzy
Systems journal publishes most of the fuzzy big data studies as in Fig. 8 while
Advances in Intelligent Systems and Computing journal publishes most of the similar
publications in Fig. 9. These analyzes will be a guide for researchers who want to
publish their work in a journal.
Fig. 8. Percentage distribution of publications according to journals for first search type.
Fig. 9. Percentage distribution of publications according to journals for second search type.
6 Conclusion
Big data analytics handles large amounts of complex data and tries to uncover hidden
patterns, correlations, and other insights. Fuzzy sets theory helps to model vague and
imprecise data in this analysis. This work revealed that most of the fuzzy big data
studies in the literature employ ordinary fuzzy sets. The recent extensions of ordinary
fuzzy sets such as intuitionistic fuzzy sets, hesitant fuzzy sets, Pythagorean fuzzy sets
and neutrosophic sets have been rarely used in fuzzy big data publications. This case
presents an opportunity for researchers to study in this area. Computer sciences,
engineering, and mathematics are the first research area that fuzzy big data analytics are
applied. For further research, we suggest intuitionistic fuzzy big data analytics or
neutrosophic fuzzy big data analytics be employed in the solution of complex data
problems.
References
1. IBM. What is big data analytics? https://www.ibm.com/analytics/hadoop/big-data-analytics.
Accessed 06 Feb 2020
2. Gartner. Big Data. https://www.gartner.com/en/information-technology/glossary/big-data.
Accessed 06 Feb 2020
3. Oztemel, E., Gursev, S.: Literature review of Industry 4.0 and related technologies. J. Intell.
Manuf. 31, 127–182 (2020)
4. Manyika, J., Chui, M., Brown, B., Bughin, J., Dobbs, R., Roxburgh, C., Hung Byers, A.:
Big data: The next frontier for innovation, competition, and productivity. McKinsey Global
Institute (2011)
5. Wang, H., Xu, Z., Pedrycz, W.: An overview on the roles of fuzzy set techniques in big data
processing: trends, challenges and opportunities. Knowl. Based Syst. 118, 15–30 (2017)
7. Laney, D.: 3D data management: controlling data volume, velocity and variety. META
group research note (2001)
8. Gandomi, A., Haider, M.: Beyond the hype: big data concepts, methods, and analytics. Int.
J. Inf. Manage. 35, 137–144 (2015)
9. Philip Chen, C., Zhang, C.: Data-intensive applications, challenges, techniques and
technologies: a survey on Big Data. Inf. Sci. 275, 314–347 (2014)
10. Zadeh, L.: The concept of a linguistic variable and its application. Inf. Sci. 8(3), 199–249
(1975)
11. Atanasov, K.: Intuitionistic fuzzy sets. Fuzzy Sets Syst. 20(1), 87–96 (1986)
13. Atanassov, K.T.: Intuitionistic Fuzzy Sets: Theory and Applications. Springer, Heidelberg
(1999)
14. Yager, R.R.: Pythagorean fuzzy subsets. In: IFSA World Congress and NAFIPS Annual
Meeting (IFSA/NAFIPS), pp. 57–61. IEEE (2013)
(2017)
16. Senapati, T., Yager, R.: Fermatean fuzzy sets. J. Amb. Intell. Hum. Comput. 11, 663–674
(2019)
17. Smarandache, F.: Neutrosophy: neutrosophic probability, set, and logic: analytic synthesis &
synthetic analysis (1998)
18. Kutlu Gundogdu, F., Kahraman, C.: Spherical fuzzy sets and spherical fuzzy TOPSIS
19. Bi, W., Cai, M., Liu, M., Li, G.: A big data clustering algorithm for mitigating the risk of
customer churn. IEEE Trans. Ind. Inf. 12(3), 1270–1281 (2016)
20. Zeng, A., Li, T., Liu, D., Zhang, J., Chen, H.: A fuzzy rough set approach for incremental
feature selection on hybrid information systems. Fuzzy Sets Syst. 258, 39–60 (2015)
21. Deng, Y., Ren, Z., Kong, Y., Bao, F., Dai, Q.: A hierarchical fused fuzzy deep neural
network for data classification. IEEE Trans. Fuzzy Syst. 25(4), 1006–1012 (2017)
22. Xu, W., Yu, J.: A novel approach to information fusion in multi-source datasets: a granular
computing viewpoint. Inf. Sci. 378, 410–423 (2017)
23. Azar, A., Hassanien, A.: Dimensionality reduction of medical big data using neural-fuzzy
classifier. Soft. Comput. 19, 1115–1127 (2015)
24. Havens, T., Bezdek, J., Leckie, C., Hall, L., Palaniswami, M.: Fuzzy c-means algorithms for
very large data. IEEE Trans. Fuzzy Syst. 20(6), 1130–1146 (2012)
25. Lou, S., Feng, Y., Zheng, H., Gao, Y., Tan, J.: Data-driven customer requirements
discernment in the product lifecycle management via intuitionistic fuzzy sets and
electroencephalogram. Journal of Intelligent Manufacturing, in print
26. Hosseini, B., Kiani, K.: A big data driven distributed density based hesitant fuzzy clustering
using apache spark with application to gene expression microarray. Eng. Appl. Artif. Intell.
79, 100–113 (2019)
27. Son, L.: DPFCM: a novel distributed picture fuzzy clustering method on picture fuzzy sets.
Expert Syst. Appl. 42, 51–66 (2015)
28. Ren, R., Tang, M., Liao, H.: Managing minority opinions in micro-grid planning by a social
network analysis-based large scale group decision making method with hesitant fuzzy
linguistic information. Knowl. Based Syst. 189 (2020)
29. Shukla, A., Yadav, M., Kumar, S., Muhuri, P.: Veracity handling and instance reduction in
big data using interval type-2 fuzzy sets. Eng. Appl. Artif. Intell. 88 (2020)
A Big Data Semantic Driven Context
Aware Recommendation Method
Manuel J. Barranco1(B) , Pedro J. Sanchez1 , Jorge Castro2 , and Raciel Yera3

1
University of Jaen, Jaen, Spain
{barranco,pedroj}@ujaen.es
2
University of Granada, Granada, Spain
[email protected]
3
University of Ciego de Avila, Ciego de Avila, Cuba
[email protected]
Abstract. Classical content-based recommender systems (CB) help

users to find preferred items in overloaded search spaces, comparing items
descriptions with user profiles. However, classical CBs do not take into
account that user preferences may change over time influenced by the
user context. This paper propounds to consider context-awareness (CA)
in order to improve the quality of recommendations, using contextual
information obtained from streams of status updates in microblogging
platforms. A novel CA-CB approach is proposed, which provides con-
text awareness recommendations based on topic detection within the
current trend interest in Twitter. Finally, some guidelines for the imple-
mentation, using the Map Reduce paradigm, are given.
Keywords: Content-based · Recommender system · Context-aware

recommendation · User profile contextualization · Map-reduce
paradigm
1 Introduction
In recent years, information in world wide web (www) scenarios have experi-
enced a huge increase, which causes that users need to dedicate great effort to
find relevant information in www search spaces. Recommender systems (RS)
are useful tools to help users in these scenarios, providing successful results in
e-business [1], e-learning [2], e-tourism [3], e-commerce [4], etc.
There are different RS approaches. The most popular are collaborative fil-
tering (CF) [5] and content-based RS (CB) [6]. CF is based on user behavior,
considering that users with similar profiles could like similar items; while CB is
based on item content, recommending items that are similar to such ones that the
user liked in the past. There are other recommendation techniques depending on
the knowledge source [7]: demographic, knowledge-based, community-based, etc.
This work is partially supported by the Spanish Ministry of Economy and Competi-
tiveness through the Spanish National Research Project PGC2018-099402-B-I00.
A Big Data Semantic Driven Context Aware Recommendation Method 895
Moreover, hybrid approaches take advantage of benefits of some techniques to

overcome the drawbacks of other ones. Recently, another important knowledge
source in RSs is the user’s context. In this way, context-aware (CA) recommen-
dation approaches [8] are focused on context information, recommending items
relevant to user needs that change over time.
In this paper, a novel context-aware content-based (CA-CB) recommendation
method is proposed, which improves traditional CBs applying context awareness
based on topic detection within current trend interest. There are previous works
that integrate contextual information to CB recommendations [7,9,10]. Our pro-
posal is to consider current trend interests, coming from microblogging services,
such as Twitter, to build a new CA-CB recommendation method. An impor-
tant step in this method is to remove noise in contextual information, caused
by words that share the same lexical root. In this way, it is necessary to cluster
context, identifying topics to build a contextualized user profile. Moreover, given
that microblogging systems generate data at a high rate, a MapReduce approach
[11] has been considered in our proposal in order to manage big data.
The rest of this paper is structured as follows: Sect. 2 provides a background
of CB, CA and MapReduce, Sect. 3 presents our proposal of CA-CB recommen-
dation method, Sect. 4 introduces some patterns for the implementation, and
finally, Sect. 5 concludes the paper.
2 Preliminaries
In this section, a background of related works is included, about CB, CA and
MapReduce approaches.
2.1 Content-Based Recommender Systems
Regarding the item representation, there are different CB approaches. One of

the most popular is the free-text representation, where each item is described by
means of unstructured data, for example, the content of a web page, a news arti-
cle or a movie synopsis. Commonly, TFIDF technique [12] is applied, converting
the unstructured data in data stemming words [13]. In this way, the quantity of
terms is considerably reduced, unifying words with the same root, for example:
recommend, recommender, recommendation, etc.
Equations 1 and 2 show the calculation of a vector of weights of terms, con-
sidering their importance in the document,
prof iletf
d
idf
= {tft,d ∗ idft s.t. t ∈ d} (1)

|N |
idft = − log (2)
|Nt |
being tft,d the number of occurrences of term t in document d, N the set of all
documents and Nt the set of documents that contain the term t at least once.
896 M. J. Barranco et al.
This set of vectors defines a term space, more specifically, a matrix of weights
of terms (columns) for each document (rows). However, the term space can be
excessively wide and sparse. Latent Semantic Analysis (LSA) [6] is a technique
commonly used to overcome this problem. In LSA, the term-document matrix is
factorized with Singular Value Decomposition (SVD) to reduce it to orthogonal
dimensions and keep the f most relevant singular values.
T F IDF(|D|×|T |) = U(|D|×f ) ∗ s(f ) ∗ V(ft ×|T |) (3)

At this point, we have a reduced feature space that replaces the wide term
space, where each document has a new profile expressed in the new feature space.
Then, user profiles are built through a linear combination of document profiles
of items that users liked in the past [14,15] and finally, the system recommends
those items more similar to the user profile.
2.2 Context-Aware Recommender Systems
User’s context is another source of information which can be considered in rec-

ommendations, in order to provide more suitable results. F. Ricci [16] appointed
that users’ circumstances have an important influence in the users’ behavior
while any decision making activity. Therefore, CA recommendations will be more
accurate and interesting for users, adapted to their context.
Traditional RSs try to approximate a function R applied to a two-dimensional
space R : U ser×Item → Rating in order to make predictions. In contrast, in CA
RSs, the function is applied to a three-dimensional space U ser×Item×Context.
Depending on the moment when context is considered, three classes of CA
approaches can be distinguished [17]: (i) pre-filtering, when the system selects
only information relative to the current context, (ii) post-filtering, if item pre-
dictions are modified regarding the specific context of the users, filtering out
the items which are not according to the context, and (iii) contextual modeling,
when the contextual information is integrated in the recommendation model.
The purpose of the pre-filtering and post-filtering approaches is to reduce
the problem of the CA three-dimensional recommendation function to a two-
dimensional one, to solve it with traditional RS. However, the third CA approach
suggests to integrate the contextual information in the recommendation model.
2.3 MapReduce Paradigm
Scalability is one of the main challenges to deal when facing big data problems.
MapReduce [11] is a paradigm that makes possible to process big data in an
scalable way, becoming one of the most popular paradigms for parallelization in
general purpose applications. There are two main operations: map and reduce:
– Map: this operation takes an input given as a set of <key, value> pairs,
transform it in another set of <key2, value2> pairs, group all pairs that have
the same key and redistributes the work.
– Reduce: this operation combines pairs by key, applies some aggregation func-
tion and produces a smaller pair set.
Apache Spark [18] is one of the most popular MapReduce frameworks, sug-
gested to implement the proposed method. It offers a set of in-memory prim-
itives based on Resilient Distributed Datasets (RDDs), a structure that stores
data in such a way that later computations can be easily parallelized in dis-
tributed machines. RDDs allow to cache or redistribute intermediate results,
which enables the design of data processing pipelines.
3 A Recommendation Method with Context Awareness

Based on Topic Detection in Current Trend Interest
In this section, the proposal of a CA-CB recommendation method is introduced.

It fits into the CA approach of contextual modeling, because it integrates con-
textual information in the model that is used to provide recommendations. The
aim of this method is to build a contextualized user profile, taking into account
user preferences and the current context.
The scheme of the proposal is shown in Fig. 1. It consists of five phases, which
are described in the next subsections.
Fig. 1. General scheme of the proposal.
3.1 Domain Semantic Analysis
The domain D consists of a set of documents that describe items to be rec-

ommended. Initially, the terms that characterize documents are the words that
appear in them. However, this set of terms have to be stemmed using the Porter
Stemmer algorithm [13], in order to unify words that have the same lexical root.
Once terms are stemmed, the TFIDF document profiles, prof ileTd F IDF , are built
according to Eq. 1.
Next, the matrix of TFIDF document profiles is processed by the technique

LSA to reduce its dimensionality. Applying a singular value decomposition, the
initial matrix, T F IDF , consisting of documents, d, described in the term space,
is decomposed in a reduced matrix of documents described in the feature space,
U , a singular value vector, s, and a matrix of terms, t, described in the fea-
ture space, V (see Eq. 3). In this way, we obtain the profile of both terms and
documents in the feature space, which compose the domain semantic model:
prof ileLSA
d = {ut,1 , . . . , ut,f } (4)
prof ileLSA
t = {vt,1 , . . . , vt,f } (5)
3.2 User Profiles Building

The previous phase has built a model consisting of terms and documents profiles.
In addition, user profiles must be available in the same feature space, so that
documents and user profiles are comparable. The system holds a set of ratings,
R = {ru,d }, that shows, for each user, the items in which he/she has expressed
some interest: commenting, voting, buying, etc.:
Ru = {d s.t. ru,d ∈ R} (6)

Given that documents profiles are available in the feature space (see Eq. 4),
we can generate user profiles in the same space, aggregating document profiles
of preferred items:

prof ileLSA
u = prof ileLSA
d ={ prof ileLSA
d,1 , . . . , prof ileLSA
d,f } (7)
d∈Ru d∈Ru d∈Ru
3.3 Context Model Building

In this phase, the context model is built to take into account the current trend
interests in the recommendation process: documents more similar to current
issues will be more relevant. The proposal is to use the status updates of a
popular microblogging service, Twitter, as the source of current trend interests.
From these status updates, the system generates a context model that will be
used to transform the user profile into a contextualized user profile.
Firstly, terms that appear in status updates must be stemmed, and after
that, the system filters out stemmed terms that do not appear in the semantic
model generated in the first phase (see Eq. 5). Next, a clustering of the filtered
stemmed terms takes place, so that, each cluster, ci , determines a context topic,
avoiding topics with very similar meaning. We propose a fuzzy c-means clustering
algorithm [19] that groups the terms using their feature vector, prof ileLSA t ,
calculating distances based on cosine correlation coefficient. After that, context
topic profiles are generated in the feature space, aggregating the profiles of the
terms included in each cluster (see Eq. 8).

prof ileLSA
ci = prof ileLSA
t ={ prof ileLSA
t,1 , . . . , prof ileLSA
t,f } (8)
t∈ci t∈ci t∈ci
3.4 Contextualizing User Profiles
Once we have user profiles and context topic profiles, our aim is to combine
them to provide contextualized and personalized recommendations. Regarding
this purpose, firstly, given a target user, the system selects those topics most
similar to the user’s profile. Our proposal is to calculate the cosine coefficient
between context topics and the user’s profile, selecting the cj that has a greater
coefficient (see Eq. 9).
cj = argmax cosine(prof ileLSA

u , prof ileLSA
ci ) (9)
ci
Next, the user’s profile and the selected context topic’s profile are combined,
providing the user’s contextualized profile. We apply a convex combination (see
Eq. 10): the greater the value of α, the more importance of the user’s profile
over the selected context topic’s profile. In this way, the user’s profile has been
transformed to a contextualized profile, adapted to both user’s preferences and
context.
prof ileLSA LSA

C,u = α ∗ prof ileu + (1 − α) ∗ prof ileLSA
cj (10)
3.5 Prediction
In this phase, for each document that describes an item to be recommended,

we make a prediction of suitability, considering the user’s contextualized profile.
Finally, the system will recommend a list of the most suitable items, described
by the top N documents, sorted by pu,d .

LSA T
pu,d = prof ileLSA
C,u ∗ prof iled (11)
4 MapReduce Implementation
In this section, some guidelines for the implementation with MapReduce are
provided. The phases of the proposed method can be implemented, according to
the MapReduce paradigm, using one or more Map operations and one or more
Reduce operations, that will be distributed and performed in a parallelized way
by the workers, typically, the nodes in a computer cluster (see Sect. 2.3).
Next, we include some MapReduce patterns for implementing the consecutive
phases of the proposed method:
1. Domain semantic analysis.

– Map (stemming):<d, {w}> → <d, {t}>.
– Map: <d, t> → <(d, t), 1>
– Reduce (sum): {<(d, t), 1>} → <(d, t), tf >.
– Map: <(d, t), tf > → <t, 1>
– Reduce (sum): {<t, 1>} → <t, idf >.
For SVD algorithms, see [20,21].
2. Build user profiles.
– Map:
<u, {prof ileLSA
d |d ∈ Ru }> → <(u, x ∈ f eaturespace), {prof ileLSA d,x }>.
– Reduce (sum): {<(u, x), prof ileLSA d,x >} → <(u, x), prof ile LSA
u,x >.
3. Build context model.
– Map (stemming): <d, {w}> → <d, {t}>.
For c-means algorithms see [22].
– Map: <ci , {prof ileLSA
t |t ∈ ci }> → <(ci , x), {prof ileLSA t,x }>.
– Reduce (sum): {<(ci , x), prof ileLSA t,x >} → <(c i , x), prof ileLSA
ci ,x >.
4. Contextualize user profiles.
– Map: <u, (prof ileLSA u , {prof ileLSA
ci })> → <u, (ci , cos(prof ileLSA u ,
prof ileLSA
ci ))>.
– Reduce (argmax): {<u, (ci , cos(prof ileLSA u , prof ileLSAci ))>} → <u, cj >.
5. Prediction:

– Map: <u, prof ileLSA C,u , prof iled
LSA
> → <u, d, prof ileLSA C,u ∗

T
prof ileLSA
d >.

LSA T
– Reduce (top N): {<u, d, prof ileLSA C,u ∗ prof ile d >} → <u, {d1 ,
..., dN }>.
5 Conclusions and Further Work
In this paper, we have studied the integration of contextual information in a CB,

providing more suitable recommendations when considering that global interests
change over time and can affect user preferences. A CA-CB recommendation
method has been proposed, which builds a contextualized user profile, using
the status updates that a microblogging service (Twitter) provides. The context
is made up of a big set of words, which are clustered using a fuzzy c-means
algorithm, in order to obtain a meaningful set of topics. In addition, given the
large size of the data, a MapReduce approach has been considered to achieve
better performance.
Further works will be oriented to apply this method in different scenarios,
using Spark as the MapReduce framework to manage big data.
References
1. Lu, J., Shambour, Q., Xu, Y., Lin, Q., Zhang, G.: A web-based personalized busi-
ness partner recommendation system using fuzzy semantic techniques. Comput.
Intell. 29(1), 37–69 (2013)
2. Yera Toledo, R., Caballero Mota, Y.: An e-learning collaborative filtering approach
to suggest problems to solve in programming online judges. Int. J. Dist. Educ.
Technol. 12(2), 51–65 (2014)
3. Noguera, J., Barranco, M., Segura, R., Martı́nez, L.: A mobile 3D-GIS hybrid
recommender system for tourism. Inf. Sci. 215, 37–52 (2012)
4. Rafailidis, D., Nanopoulos, A.: Modeling users preference dynamics and side infor-
mation in recommender systems. IEEE Trans. Syst. Man Cybern. Syst. 46(6),
782–792 (2016)
5. Koren, Y., Bell, R.: Advances in Collaborative Filtering, pp. 77–118. Springer,
Boston (2015)
6. de Gemmis, M., Lops, P., Musto, C., Narducci, F., Semeraro, G.: Semantics-aware
content-based recommender systems. In: Ricci, F., Rokach, L., Shapira, B. (eds.)
Recommender Systems Handbook, pp. 119–159. Springer, Boston (2015)
7. De Pessemier, T., Courtois, C., Vanhecke, K., Van Damme, K., Martens, L., De
Marez, L.: A user-centric evaluation of context-aware recommendations for a mobile
news service. Multimed. Tools Appl. 75(6), 3323–3351 (2016)
8. Adomavicius, G., Tuzhilin, A.: Context-Aware Recommender Systems, pp. 191–
226. Springer, Boston (2015)
9. Parikh, N., Sundaresan, N.: Buzz-based recommender system. In: Proceedings of
the 18th International Conference on World Wide Web, WWW 2009, pp. 1231–
1232. ACM, New York (2009)
10. Ponzanelli, L.: Holistic recommender systems for software engineering. In: Com-
panion Proceedings of the 36th International Conference on Software Engineering,
ICSE Companion 2014, pp. 686–689. ACM, New York (2014)
11. Dean, J., Ghemawat, S.: MapReduce: simplified data processing on large clusters.
Commun. ACM 51(1), 107–113 (2008)
12. Erra, U., Senatore, S., Minnella, F., Caggianese, G.: Approximate TF-IDF based
on topic extraction from massive message stream using the GPU. Inf. Sci. 292,
143–161 (2015)
13. Porter, M.F.: An algorithm for suffix stripping. Program 14(3), 130–137 (1980)
14. Symeonidis, P., Nanopoulos, A., Manolopoulos, Y.: Feature-weighted user model
for recommender systems. In: Proceedings of the 11th International Conference on
User Modeling, pp. 97–106. Springer-Verlag (2007)
15. Bambini, R., Cremonesi, P., Turrin, R.: A Recommender System for an IPTV Ser-
vice Provider: a Real Large-Scale Production Environment, pp. 299–331. Springer,
Boston (2011)
16. Ricci, F.: Contextualizing recommendations. In: Conjunction with the 6th ACM
Conference on Recommender Systems (RecSys 2012), ACM RecSys Workshop on
Context-Aware Recommender Systems (CARS 2012). ACM (2012)
17. Adomavicius, G., Tuzhilin, A.: Context-aware recommender systems. In: Recom-
mender Systems Handbook, ch. 3, pp. 217–253. Springer, US (2011)
18. Zaharia, M., Xin, R.S., Wendell, P., Das, T., Armbrust, M., Dave, A., Meng, X.,
Rosen, J., Venkataraman, S., Franklin, M.J., et al.: Apache spark: a unified engine
for big data processing. Commun. ACM 59(11), 56–65 (2016)
19. Bezdek, J.C., Ehrlich, R., Full, W.: FCM: The fuzzy c-means clustering algorithm.
Comput. Geosci. 10(2–3), 191–203 (1984)
20. Ulrey, R.R., Maciejewski, A.A., Siegel, H.J.: Parallel algorithms for singular value
decomposition. In: Proceedings of 8th International Parallel Processing Sympo-
sium, pp. 524–533. IEEE (1994)
21. Berry, M., Mezher, D., Philippe, B., Sameh, A.: Parallel computation of the sin-
gular value decomposition. Research Report RR-4694 (2003)
22. Ludwig, S.: MapReduce-based fuzzy c-means clustering algorithm: implementation
and scalability. Int. J. Mach. Learn. Cybern. 6, 04 (2015)
Blended Environment of Naive Bayes
and Support Vector Machine
(SVM) for Designing Simulation Based
E-Learning Respiratory System
Anuradha Verma Babbar(&) and Santosh Kumar Henge(&)
School of Computer Science and Engineering, Lovely Professional University,

Punjab, India
[email protected]
Abstract. Learning process has become smart with use of technology termed as
E-Learning. Simulation based E-Learning revolutionized the education world
where the learners enjoy learning process as a playful-act because of 3D virtual
images and at the same time they have deep understanding of concept because
they experience real life scenarios as they learn by doing. This research has
composed with three major objectives such as: It summarized the analysis of
various data mining algorithms in the current active field of E-Learning by
comparative study of past designed approaches with their advantages and lagging
issues; it has proposed the blended environment of heuristic data mining algo-
rithms such as Naive Bayes with Support Vector Machines which is implicated
with many technical advantages that helped in analyzing the real time learning
predictions; To synchronize the 3D-learning based contents based on the analysis
of learners behavior and their thinking uncertainties. This research is more
helpful for the new and experienced learners for making themselves expertise in
their multi-disciplinary working fields like as bio-medical, bio-engineering, bio-
sciences, engineering-computing and so on. This research is providing the many
technical feasibilities and solutions for the medical and engineering researchers to
improve and share their research experiences for their professional growth.
Keywords: Naive Bayes (NBs) Support Vector Machines (SVM)

Simulation based E-Learning (SbEL) Behavior decision making (BDM)
Behavior uncertainty (BUC)
1 Introduction
The encroachment of internet occupied almost all the domains, no doubt, the field of
education also. Education together with technology or online learning process or
learning through internet is termed as E-Learning [1]. E-Learning has gained popularity
in recent years because of several features like learning away from traditional class-
rooms, no constraints of geographical boundaries, convenient hours of learning,
availability of information with just a click of finger, flexibility to manage with one’s
job or business, no limitations for any sex or age, etc. [2].
904 A. V. Babbar and S. K. Henge
As the technology is updating every day, Learning process also needs to be

updated. The concept of Simulation based E-Learning (SbEL) revolutionized the
market with its very special feature of 3D virtual images where the learners interact
with real life scenarios and gain knowledge through the experience they get by per-
forming the actions in real sense but in virtual atmosphere [3]. The simulation system
used may be games, applications or situations etc. that provide its users the problems of
real world that they may face after entering their careers [4]. Hence SbEL proves to be
useful as it train the users to manage the real life operations well in advance [5].
Despite of enormous benefits of simulation based E-Learning, Many developing
countries still have not widely accepted it. The general challenges like different
behavioral patterns, academic performances, learning styles, adaptability struggle, [6]
high dropout rates [7], lack of collaboration, self-motivation etc. are always there to be
resolved [8]. The solution to these issues can be found with the help of Data Mining
Techniques. The large amount of data available in different fields has always attracted
researchers to mine the data with innovate the concept [3, 9]. Data mining means
extracting knowledge from existing resources or the available database and imple-
menting new ideas for betterment of current trends. In context of Education, Data
mining is known as Knowledge Discovery from Database (KDD) [10].
Large amount of data is available in the educational institutions like students’
attendance, family description, geographical background, academic performances,
learning styles, behavior, collaboration etc. Various data mining algorithms has been
applied on this available raw data, extracting useful information and finding best
technique to predict learners’ acceptance for the course opted thereby enhancing
learning outcomes and efficiency of educational institutions [11]. Data mining tech-
niques used for this knowledge discovery are classification, regression, clustering,
segmentation, association rule mining etc. [12].
The major objective of this paper is to design simulation based 3D learning
environment where the learners can understand the concept by doing practically. To
fulfill this purpose, research mainly focused on the following:
• Summarized the analysis of various data mining algorithms in the current active
field of E-Learning by comparative study of past designed approaches with their
advantages and lagging issues;
• Proposed the blended environment of heuristic data mining algorithms such as
Naive Bayes with Support Vector Machines which is implicated with many tech-
nical advantages that helped in analysing the real time learning predictions;
• To synchronize the 3D-learning based contents based on the analysis of learner’s
behavior and their thinking uncertainties.
The research outcomes produced in this paper is an independent and original piece
of work. We analyzed various data mining algorithms available in existing sources of
research and added to the knowledge that is novel and has not been looked at before.
We duly acknowledge all the sources from which ideas and extracts have been taken.
This paper is organized as follows. In Sect. 1, General introduction of E-Learning, its
characteristics, Simulation based E learning and the challenges that come its way are
briefly described. Also some commonly used data mining techniques are mentioned. In
Sect. 2, past proposed data mining algorithmic methodologies are analyzed and written
Blended Environment of Naive Bayes and Support Vector Machine (SVM) 905
in the tabular form with a brief summary at the end of section. Then in Sect. 3, Training
models through NB and SVM are introduced, and at last, in Sect. 4, conclusion and
suggestions for future research are given.
2 Analyses of Past Proposed Data Mining Algorithmic

Methodologies in E-Learning Development
Table 1. The analysis of past proposed data mining algorithmic methodologies in E-Learning
development
Author and Research implications Techniques used Advantages Lagging issues
citation
P. Kaur, M. Prediction of students Multilayer The teachers can Selected students from
Singh and G. learning outcomes and Perception, predict slow learners High School
Josan [13] identifying slow Naïve Bayes, and modify their participated in this
learners in education SMO, J48 and methods of teaching experiment. Research
sector by comparing REP tree was confined to a
five classification particular section of
algorithms society
R. Sumitha Predicting students Naïve Bayes, Educational Only academic and
and E. future learning MLP, SMO, Institutions modify demographic records
Vinothkumar outcomes by Decision Table, the strategies for of students were
[14] analyzing dataset of REP tree, J48 better management analyzed whereas
senior students to and to remain extracurricular skills
enhance performance competitive were not considered
of educational for this research
institutions
Y. Jiang, B. To propose three layer Naïve Bayes This model helped to This model worked
Qiu, C. Xu knowledge base predict disease more well for diseases with
and C. Li [15] model to diagnose accurately using common symptoms
disease and provide simulative only
clinical environment
recommendation
A. Dutt, M. To review Educational Clustering This paper analyzed Variables like
Ismail and T. data mining algorithms learning styles of Classroom decoration,
Herawan [12] techniques to analyze learners based on exam failure, and
students learning different variables student motivation
outcomes need more research
I. Amra and To set standards for KNN, Naïve Based on prediction Both algorithms can
A. Maghari quality education by Bayes using classification produce different
[16] predicting students algorithms, Ministry results on different
performance at an of Education can datasets
early stage improve level of
performance
(continued)
Author and Research implications Techniques used Advantages Lagging issues
citation
K. Prasada, This paper focus on a J48, Naïve Research showed Social media and
M. Chandra learning model that Bayes and that learner’s internet access
and B. analyze learners’ Random Forest performance can be variables were not
Ramesh [17] behavior by Algorithm improved by included to analyze
identifying rectifying base learners’ behavior
weaknesses variable
E. Amrieh, T. To study relationship Artificial Neural Research proved that Only a few behavioral
Hamtini and I. between learners Network, Naïve academic outcomes patterns were
Aljarah [18] behavior and Bayes, Decision are influenced by considered in this
academic achievement Tree and learners behavior research
Ensemble
Methods
P. Kavipriya To analyze impact of Classification, By analyzing The attributes for
[19] social media on clustering and attendance records, research are collected
behavior and Ensemble assessment scores, from regular students.
academic performance methods assignment It is difficult to apply
submission etc., these for distance
required academic learners
improvements can be
suggested in advance
M. Liyanage, To understand J48, Bayesian Both the learners and Research was confined
K. different learning Network, Naïve instructors can to selected course with
Gunawardena styles using learning Bayes, Random visualize and analyze particular data mining
and M. management system Forest learning styles technique. Result may
Hirakawa [20] vary with other
courses and DM
Techniques
A. Mueen, B. To early predict Decision Tree, Instructors can early Research was carried
Zafar and U. students’ performance Multilayer arrange remedial on particular section of
Manzoor [21] after analyzing Perception, classes for the students with same
academic records Naïve Bayes students who are course instructed by
expected to fail the same lecturer
course
The Table 1 contains outcomes of latest research that is carried after the year 2015.
Researchers have used various data mining techniques like classification, regression,
clustering, association rule mining etc. to predict learners’ behavior and their readiness
to accept E-Learning. They have considered different behavioral patterns, learning
styles, academic and demographic records, understanding level, course opted and
simulative environment etc. for their research. The above table represents the appli-
cation of various Data Mining Algorithms to analyze learners’ behavior and thereby
improving the level of learning outcomes, suggestions for Authorities of educational
institutes, SbEL for better understanding of concept, arrangement of remedial classes
etc. However lagging issues of research are always there.
3 Simulation Based E-Learning Respiratory System by Using

Naïve-Bayes and Support Vector Machine
3.1 Training Modal Through Naïve Bayes
The Naïve Bayes (NB) is an algorithm used for predicting modals with simple and
powerful features. It is a classification technique that works on the principle of con-
ditional probability given by the Bayes’ theorem [22]. NB is a combination of two
words-“naïve” and “Bayes”. In this algorithm, the attributes are assumed to be inde-
pendent. The attributes influence one another but the existence of one is unrelated to the
other i.e. that one attribute itself plays an important role for the classification. For
example, a fruit which is round, orange, pulpy and sliced may be considered as an
orange. Here each attribute separately contribute to the probability that this fruit is an
orange even when these attributes influence one another [23]. The equation given here
is used to calculate posterior probability from P (x|c), P(c) and P(x).
PðxjcÞPðcÞ
PðcjxÞ ¼ ð1Þ
P ð xÞ
PðcjxÞ ¼ Pðx1 jcÞ P ðx2 jcÞ . . .. . . Pðxn jcÞ PðcÞ ð2Þ
P (c|x) is the posterior probability of class (c, target) given predictor (x, attributes).
P (c) is the prior probability of class.
P (x|c) is the likelihood which is the probability of predictor given class.
P (x) is the prior probability of predictor.
The NB classifier helps in making real time predictions as well as multi class
prediction. This model requires less training data so performs better as compared to
logistic regression models. It has some limitations too. Laplace estimation technique is
applied to overcome zero frequency problems. Independent predictors are not always
observed in real life scenarios.
3.2 Training Model Through Support Vector Machine (SVM)

The Support Vector Machine (SVM) is an algorithm used for classification and
regression problems. It helps to handle linear as well as non-linear data by classifying
two distribution classes using hyper plane that best segregates the inputs [24]. To avoid
possibility of wrong classification, hyperplane with maximum margin that creates
largest possible distance from support vectors is selected [25]. Kernel effectively
classifies multi-dimensional data using the feature.
z ¼ x ^ zþy ^ z ð3Þ
The SVM classifiers perform with great efficiency in high dimensional spaces [26].
It is also efficient when number of dimensions is more than number of samples.
Memory efficiency is another advantage of SVM. However SVM classifier has some
limitations too. It takes higher training time for large data sets. Results are not reliable
when target classes overlap. Expensive fivefold cross validation is used to calculate
probability estimates.
4 Conclusion
This paper is an effort to analyze various Data Mining Algorithms in current field of
E-Learning. NB and SVM are blended to design Simulation based E-Learning Res-
piratory System. This is observed that the shortcomings of one algorithm become plus
point of other thereby improving overall performance of the model. SbEL ease learning
process as 3D learning provides the real feel to the learner. It is a boon to many fields
like bio-medical, bio-engineering, bio-sciences, aeronautics etc. as the learners can
interact with 3D virtual atmosphere decreasing risk of committing errors in real life.
The work can be further extended by designing a model that identifies students’
interests, analyzes records, predict their performance and help them to develop their
learning skills by providing suggestions and recommendation for appropriate study
material.
References
1. Abid, A., Kallel, I., Ben Ayed, M.: Teamwork construction in E-Learning system: a
systematic literature review. In: 2016 15th International Conference on Information
Technology Based Higher Education and Training (ITHET), Istanbul, pp. 1–7 (2016)
2. Dunn, T., Kennedy, M.: Technology Enhanced Learning in higher education; motivations,
engagement and academic achievement. Comput. Educ. 137, 104–113 (2019). https://doi.
org/10.1016/j.compedu.2019.04.004
3. Cook, D., Andersen, D., Combes, J., Feldman, D., Sachdeva, A.: The value proposition of
simulation-based education. Surgery 163(4), 944–949 (2018). https://doi.org/10.1016/j.surg.
2017.11.008
4. Hubl, A., Fischer, G.: Simulation based business game for teaching methods in logistics and
production. In: 2017 Winter Simulation Conference (WSC), pp. 4228–4239 (2017)
5. Zulfiqar, S., Zhou, R., Asmi, F., Yasin, A.: Using simulation system for collaborative
learning to enhance learner’s performance. Cogent Educ. 5(1) (2018). https://doi.org/10.
1080/2331186x.2018.1424678
6. Truong, H.: Integrating learning styles and adaptive e-learning system: current develop-
ments, problems and opportunities. Comput. Hum. Behav. 55, 1185–1193 (2016). https://
doi.org/10.1016/j.chb.2015.02.014
7. Alban, M., Mauricio, D.: Predicting university dropout trough data mining: a systematic
literature. Indian J. Sci. Technol. 12(4), 1–12 (2019). https://doi.org/10.17485/ijst/2019/
v12i4/139729
8. Urbina Nájera, A., De la Calleja Mora, J.: Brief review of educational applications using data
mining and machine learning. Revista Electrónica de Investigación Educativa 19(4), 84
(2017). https://doi.org/10.24320/redie.2017.19.4.1305
9. Castro, F., Vellido, A., Nebot, A., Mugica, F.: Applying data mining techniques to e-
learning problems. In: Jain, L., Tedman, D., Tedman, R. (ed.) Evolution of Teaching and
Learning Paradigms in Intelligent Environment. Springer, Berlin (2007)
10. Ionita, I.: Data mining technique for e-learning. J. Appl. Comput. Sci. Math. 10(2), 26–31
(2016). https://doi.org/10.4316/jacsm.201602004
11. Gorade, S., Deo, A., Purohit, P.: A study some data mining classification techniques. Int.
J. Modern Trends Eng. Res. 4(1), 210–215 (2017). https://doi.org/10.21884/ijmter.2017.
4031.zt9tv
12. Dutt, A., Ismail, M., Herawan, T.: A systematic review on educational data mining. IEEE
Access 5, 15991–16005 (2017). https://doi.org/10.1109/access.2017.2654247
13. Kaur, P., Singh, M., Josan, G.: Classification and prediction based data mining algorithms to
predict slow learners in education sector. Procedia Comput. Sci. 57, 500–508 (2015). https://
doi.org/10.1016/j.procs.2015.07.372
14. Sumitha, R., Vinothkumar, E.: Prediction of students outcome using data mining techniques.
Int. J. Sci. Eng. Appl. Sci. 2(6) (2016). Accessed 23 Apr 2020
15. Jiang, Y., Qiu, B., Xu, C., Li, C.: The research of clinical decision support system based on
three-layer knowledge base model. J. Healthc. Eng. 2017, 1–8 (2017). https://doi.org/10.
1155/2017/6535286
16. Amra, I., Maghari, A.: Students performance prediction using KNN and Naive Bayesian. In:
2017 8th International Conference on Information Technology (ICIT) (2017)
17. Prasada, K., Chandra, M., Ramesh, B.: Predicting learning behavior of students using
classification techniques. Int. J. Comput. Appl. 139(7), 15–19 (2016). https://doi.org/10.
5120/ijca2016909188
18. Amrieh, E., Hamtini, T., Aljarah, I.: Mining educational data to predict student’s academic
performance using ensemble methods. Int. J. Database Theory Appl. 9(8), 19–136 (2016).
https://doi.org/10.14257/ijdta.2016.9.8.13
19. Kavipriya, P.: A review on predicting students’ academic performance earlier, using data
mining techniques. Int. J. Adv. Res. Comput. Sci. Softw. Eng. 6(12), 101–105 (2016)
20. Liyanage, M., Gunawardena, K., Hirakawa, M.: Detecting learning styles in learning
management systems using data mining. J. Inf. Process. 24(4), 740–749 (2016). https://doi.
org/10.2197/ipsjjip.24.740
21. Mueen, A., Zafar, B., Manzoor, U.: Modeling and predicting students’ academic
performance using data mining techniques. Int. J. Modern Educ. Comput. Sci. 8(11), 36–
42 (2016). https://doi.org/10.5815/ijmecs.2016.11.05
22. Buenano-Fernandez, D., Lujan-Mora, S.: Exploring approaches to educational data mining
and learning analytics, to measure the level of acquisition of student’s learning outcome. In:
International Conference on Education and New Learning Technologies, pp. 1845–1850
(2016)
23. Catal, C., Sevim, U., Diri, B.: Practical development of an Eclipse-based software fault
prediction tool using Naive Bayes algorithm. Expert Syst. Appl. 38(3), 2347–2353 (2011).
https://doi.org/10.1016/j.eswa.2010.08.022
24. Polat, H., Danaei Mehr, H., Cetin, A.: Diagnosis of chronic kidney disease based on support
vector machine by feature selection methods. J. Med. Syst. 41, 4. https://doi.org/10.1007/
s10916-017-0703-x
25. Osisanwo, F.Y., Akinsola, J.E.T., Awodele, O., Hinmikaiye, J.O., Olakanmi, O., Akinjobi,
J.: Supervised machine learning algorithms: classification and comparison. Int. J. Comput.
Trends Technol. 48(3), 128–138 (2017). https://doi.org/10.14445/22312803/ijctt-v48p126
26. Dewan, M., Wen, D., Lin, F., KinshukBe, D.: Predicting dropout-prone students in e-
learning education system. In: UIC-ATC-ScalCom-CBDCom-IoP 2015, Beijing, China
(2015)
InterCriteria Analysis of Public Health
Data in Bulgaria
Evdokia Sotirova1(&), Valentin Vasilev1, Sotir Sotirov1,

and Hristo Bozov1,2
1
Prof. Dr. Assen Zlatarov University,
Prof. Yakimov, Blvd., 8010 Burgas, Bulgaria
{esotorova,ssotorov}@btu.bg,
2
Oncology Complex Center-Burgas,
86 Demokratsiya Blvd., 8000 Burgas, Bulgaria
Abstract. The assessment of the system of public health has been receiving
significant importance over the past years, because it reflects the trend in opti-
mizing the effectiveness of the health services. The article uses a recently
defined intelligent decision-making method based on an extension of fuzzy logic
called InterCriteria Analysis (ICA). Object of the study of (the) ICA are real data
in the field of public health connected with in-patient, out-patient health care
establishments and their regional distribution in Bulgaria for the years 2010–
2018. Using the ICA approach, we can identify the relationships between and
among indicators of health care facilities and nursing staff, statistical regions and
districts, the doctors in the healthcare facility by medical specialties, and more.
The metrics that have the greatest dependencies and the contrary indicators that
are often independent of one another. This way we can monitor their demeanor
over time.
Keywords: InterCriteria Analysis Intuitionistic fuzzy sets Intuitionistic

fuzzy pairs Intelligent systems Public health
1 Introduction
One of the significant components to determine a country’s quality of life of the people
is the state of health and the associated quality of service. Quality of life assessment has
attracted considerable attention in healthcare systems in recent years. Often, in public
health research, the quality of life idea is analyzed in the context of ‘heath-related
quality of life’.
The analysis of public health data has been increasingly acknowledged as a valid
and appropriate indicator to measure health needs and outcomes. Questions about
perceived physical and mental health and functioning have become an important
component of health surveillance and are generally considered valid indicators of
service needs and intervention outcomes. Self-assessed health status has proved to be a
more powerful predictor of mortality and morbidity than many objective measures of
health. The approach is used when creating a European database on the health status of
InterCriteria Analysis of Public Health Data in Bulgaria 911
individual countries through a unified mechanism for conducting the European Health
Interview Survey [5].
The ICA method is introduced by K. Atanassov, D. Mavrov and V. Atanassova in
[3]. It is based on the extensions of fuzzy sets - intuitionistic fuzzy sets (IFSs, see [2])
and apparatus of index matrices (IMs, see [1]). It is specially introduced to solve
multicriteria problems with intelligent decision-making where some of the criteria have
a higher cost than others. The purpose of the ICA method is to detect the corresponding
high levels of dependency or dependency between these criteria and others. This makes
it possible to offer other criteria that are cheaper, easier, or faster to evaluate or measure
in order to eliminate the disadvantages of future decision-making.
The approach calculates the degree of correlation between all possible pairs of
criteria in the form of intuitionistic fuzzy pairs (IFPs, see [4]) with values in the [0, 1]
interval. This means that it can open new, unknown till now, relationships or rediscover
already established by other methods relationships leading to the generation of new
knowledge, including in the field of intelligent fuzzy systems. The result of applying
the ICA method is to obtain an index matrix that gives the correlations of each pair of
criteria, presented in the form of intuitionistic fuzzy pairs of values. The resulting
dependencies between the criteria are called “positive consonance”, “negative conso-
nance” or “dissonance” [3]. The ICA approach has been applied for analyzing and
intelligent decision making of medical data in [6, 7, 11, 12, 13]. The authors have used
the ICA method for analysis the patients in Burgas and the Burgas region with
malignant neoplasms of the digestive organs for 2014–2018 [9] and with oncological
diseases for 2018 [10], and for assessment the health-related quality of life of the
residents. In this paper, the ICA approach was used for intelligent fuzzy analysis of the
data on health network by type of health establishments in Bulgaria [8].
The paper is organized as follows. Section 2 is the application of the InterCriteria
Analysis approach. The investigated data connected with in-patient, out-patient and
other health establishments, types of beds and their regional distribution in Bulgaria for
2014–2018 are presented. Sections 2.1 and 2.2 show the application of the method
respectively on the number of health establishments and on the number of the beds
health establishments. Sections 3 is a discussion on applying the ICA approach to the
number of health establishments and to the number of beds in the different types of
health establishments for 2010–2018 and comparison with correlation analysis
according to Kendall. Finally, Sect. 4 presents our conclusions and future work. Sec-
tion 5 and Sect. 6 are acknowledgments and references.
2 An Application of the ICA
The InterCriteria Analysis approach is applied to real data connected with in-patient,
out-patient and other health establishments, types of beds and their regional distribution
for 2014–2018. According to the NSI [8]:
• The in-patient health establishments include all hospitals and centres with beds: Multi
profile hospitals (MPHs), Specialized hospitals (SHs), Dermato-venereological cen-
tres (DVCs), Complex oncological centres (COCs), Mental health centres (MHCs);
912 E. Sotirova et al.
• The out-patient health establishments include: Diagnostic and consulting centres

(DCCs), Medical centres (MCs), Dental centres (DCs), Medical-dental centres
(MDCs), Medical-diagnostical and medical-technical laboratories (MLabs).
As of 31.12.2018 in Bulgaria there are 346 in-patient health establishments with
53173 beds, of which 322 hospitals with 50927 beds. Out-patient health establishments
are 2066 with 1253 beds, and other medical and health establishments are 145 with
2086 beds [8].
2.1 An Application of the ICA for the Number of the Health

Establishments in Bulgaria for 2010–2018
An index matrix that contains 10 rows (for type of the health establishments: MPHs,
SHs, DVCs, COCs, MHCs, DCCs, MCs, DCs, MDCs, MLabs) and 9 columns (for
number of the different types of the health establishments (HEs) by years: 2010, 2011,
…., 2018) has been used.
After applying the ICA method, we obtain an index matrix that gives the corre-
lations of each pair of criteria of health establishments presented in the form of intu-
itionistic fuzzy pairs of values (see Table 1). The stronger correlation between criteria
in the IFP is represented by more intense color.
Table 1. Membership part of the IFPs of the relations between the types of health
establishments for 2010–2018 period.
μ MPHs SHs DVCs COCs MHCs DCCs MCs DCs MDCs MLabs
MPHs 1.000 0.083 0.000 0.028 0.028 0.250 0.972 0.611 0.917 0.889
SHs 0.083 1.000 0.778 0.194 0.028 0.583 0.083 0.306 0.111 0.139
DVCs 0.000 0.778 1.000 0.361 0.194 0.639 0.000 0.194 0.056 0.028
COCs 0.028 0.194 0.361 1.000 0.778 0.222 0.000 0.111 0.028 0.028
MHCs 0.028 0.028 0.194 0.778 1.000 0.139 0.000 0.111 0.028 0.028
DCCs 0.250 0.583 0.639 0.222 0.139 1.000 0.222 0.333 0.250 0.222
MCs 0.972 0.083 0.000 0.000 0.000 0.222 1.000 0.639 0.944 0.917
DCs 0.611 0.306 0.194 0.111 0.111 0.333 0.639 1.000 0.611 0.611
MDCs 0.917 0.111 0.056 0.028 0.028 0.250 0.944 0.611 1.000 0.889
MLabs 0.889 0.139 0.028 0.028 0.028 0.222 0.917 0.611 0.889 1.000
2.2 An Application of the ICA for the Number of Beds in the Health
Establishments in Bulgaria for 2010–2018
Data on beds in health establishments refer to the all de facto operating beds as well as
beds temporary out of use for a period of less than six months. The number of beds for
temporary use is excluded [8].
An index matrix that contains 6 rows (beds in the different types of health estab-
lishments: MPHs, SHs, DCCs, MCs, MDCs, Other HEs) and 9 columns (for number of
the different types of the HEs by years: 2010, 2011, …., 2018) will be used.
After applying the ICA method we obtain an index matrix that gives the correla-
tions of each pair of criteria of “beds in the health establishment” presented in the form
of intuitionistic fuzzy pairs of values (see Table 2). The stronger correlation between
criteria in the IFP is represented by more intense color.
Table 2. Membership part of the IFPs of the relations between beds in the health establishments
for 2010– 2018 period.
MPHs, SHs, DCCs, MCs, MDCs, Other HEs,
μ beds beds beds beds beds beds
MPHs, beds 1.000 0.500 0.778 0.917 0.833 0.083
SHs, beds 0.500 1.000 0.278 0.472 0.444 0.472
DCCs, beds 0.778 0.278 1.000 0.806 0.778 0.250
MCs, beds 0.917 0.472 0.806 1.000 0.806 0.111
MDCs, beds 0.833 0.444 0.778 0.806 1.000 0.167
Other HEs, beds 0.083 0.472 0.250 0.111 0.167 1.000
3 Discussions
3.1 After Applying the ICA Approach for the Number of the Health
Establishments for 2010–2018 the Following Conclusions Can Be
Made
After applying the ICA method 1 pair of criteria in strong positive consonance and 5
pairs of criteria in positive consonance were obtained. This means that the number of
types of health establishments has had a very similar tendency to change during the
period 2010–2018. In Table 3 are shown the comparison of the pairs of criteria with
highest value of membership µ from the ICA approach and the correlation coefficients
from correlation analysis according to Kendall.
Table 3. Membership parts of the Intuitionistic fuzzy pairs of the relations between “number of
HEs” according ICA approach and correlation analysis according to Kendall.
ICA Correlation coefficient according
Kendall
1. MCs – MPHs: 〈0,972; 0,000〉 Strong positive MCs – MPHs: 0,990
consonance
2. MCs – MDCs: 〈0,944; 0,028〉 Positive consonance MCs – MDCs: 0,930
3. MPHs – MDCs: 〈0,917; 0,028〉 Positive consonance MPHs – MDCs: 0,914
4. MCs – MLabs: 〈0,917; 0,056〉 Positive consonance MCs – MLabs: 1,000
5. MPHs – MLabs: 〈0,889; 0,056〉 Positive MPHs – MLabs: 0,986
consonance
6. MDCs – MLabs: 〈0,889; 0,056〉 Positive MDCs – MLabs: 0,930
consonance
914 E. Sotirova et al.
3.2 After applying the ICA Approach for the Number of the Beds
in the Different Types of Health Establishments for 2010–2018
the Following Conclusions Can Be Made
After applying the ICA method 1 pair of criteria in positive and 5 pairs in weak positive
consonance were obtained. The other 9 pairs of criteria are in dissonance. For the pairs
of criteria in consonance the number of beds in the different types of health estab-
lishments have had a similar tendency to change during the period 2010–2018. In
Table 4 are shown the comparison of the pairs of criteria with highest value of
membership µ from the ICA approach and correlation analysis according to Kendall.
Table 4. Membership parts of the Intuitionistic fuzzy pairs of the relations between “number of
beds in the HEs” according ICA approach and correlation analysis according to Kendall.
ICA Correlation coefficient according Kendall
1. MPHs, beds – MCs, beds: 〈0,917; 0,083〉 MPHs, beds – MCs, beds: 0,833
Positive Consonance
2. MPHs, beds – MDCs, beds: 〈0,833; 0,139〉 MPHs, beds – MDCs, beds: 0,704
Weak Positive Consonance
3. DCCs, beds – MCs, beds: 〈0,806; 0,194〉 DCCs, beds – MCs, beds: 0,611
4. MCs, beds – MDCs, beds: 〈0,806; 0,167〉 MCs, beds – MDCs, beds: 0,648
5. DCCs, beds – MDCs, beds: 〈0,778; 0,194〉 DCCs, beds – MDCs, beds: 0,592
6. MPHs, beds – DCCs, beds: 〈0,778; 0,222〉 MPHs, beds – DCCs, beds: 0,556
4 Conclusion
The ICA approach was applied for the analysis of the data connected with in-patient,
out-patient health care establishments, their regional distribution as well as on medical
personnel by specialty and categories in Bulgaria. The data for years 2010–2018 were
extracted from the sites, which provide free access to the health statistical information.
With this method, the degrees of correlation between all possible pairs of criteria
are calculated, which means that they can confirm both already known and other
established in the literature dependencies, and to discover completely new, unknown so
far dependencies, and hence make predictions about the sites being surveyed.
In the next our investigation the ICA approach will be applied to data for the
medical personnel by specialty and categories in Bulgaria for the same period.
by DCM # 577/17.08.2018.
Conflict of Interest. The authors declare that there is no conflict of interest regarding the
publication of this paper.
References
1. Atanassov, K.: Index matrices: towards an augmented matrix calculus. Studies in
Computational Intelligence Series, vol. 573. Springer, Cham (2014)
3. Atanassov, K., Mavrov, D., Atanassova, V.: Intercriteria decision making: a new approach
for multicriteria decision making, based on index matrices and intuitionistic fuzzy sets.
Issues in Intuitionistic Fuzzy Sets and Generalized Nets, vol. 11, pp. 1–8 (2014)
4. Atanassov, K., Atanassova, V., Gluhchev, G.: InterCriteria analysis: ideas and problems.
Notes Intuitionistic Fuzzy Sets 21(1), 81–88 (2015)
5. European Health Interview. https://www.nsi.bg/en/content/5630/european-health-interview-
survey. Accessed 01 Apr 2020
6. Ilkova, T., Roeva, O., Vassilev, P., Petrov, M.: InterCriteria analysis in structural and
parameter identification of L-lysine production model. Issues in Intuitionistic Fuzzy Sets and
Generalized Nets, vol. 12, pp. 39–52 (2015/2016)
7. Krumova, S., Todinova, S., Mavrov, D., Marinov, P., Atanassova, V., Atanassov, K.,
Taneva, S.: Intercriteria analysis of calorimetric data of blood serum proteome. Biochimica
et Biophysica Acta General Subjects 1861(2), 409–417 (2017)
8. National Statistical Institute, Section Health establishments. https://www.nsi.bg/en/content/
5574/health-establishments
9. Sotirova, E., Vasilev, V., Bozova, G., Bozov, H., Sotirov, S.: Application of the InterCriteria
analysis method to a dataset of malignant neoplasms of the digestive organs for the Burgas
region for 2014–2018. In: 2019 Big Data, Knowledge and Control Systems Engineering
(BdKCSE), pp. 1–6. IEEE, Sofia (2019)
10. Sotirova, E., Petrova, Y., Bozov, H.: InterCriteria analysis of oncological data of the patients
for the city of Burgas. Notes Intuitionistic Fuzzy Sets 25(2), 96–103 (2019)
11. Todinova, S., Mavrov, D., Krumova, S., Marinov, P., Atanassova, V., Atanassov, K.,
Taneva, S.: Blood plasma thermograms dataset analysis by means of InterCriteria and
correlation analyses for the case of colorectal cancer. Int. J. Bioautomation 20(1), 115–124
(2016)
12. Zaharieva, B., Doukovska, L., Ribagin, S., Michalikova, A., Radeva, I.: Intercriteria analysis
of Behterev’s Kinesitherapy program. Notes Intuitionistic Fuzzy Sets 23(3), 69–80 (2017)
13. Zaharieva, B., Doukovska, L., Ribagin, S., Radeva, I.: InterCriteria approach to Behterev’s
disease analysis. Notes Intuitionistic Fuzzy Sets 23(2), pp. 119–127 (2017)
A Case Study on Vehicle Battery
Manufacturing Using Fuzzy Analysis
of Variance
Abbas Parchami1(&) , Mashaallah Mashinchi1 ,

and Cengiz Kahraman2
1
Department of Statistics, Faculty of Mathematics and Computer,
Shahid Bahonar University of Kerman, Kerman, Iran
{parchami,mashinchi}@uk.ac.ir
2
[email protected]
Abstract. Analysis of variance is an important method in exploratory and

confirmatory data analysis when explanatory variables are discrete and response
variables are continues and independent from each other. The simplest type of
analysis of variance is one-way analysis of variance for comparison among
means of several populations. In this paper, we extend one-way analysis of
variance to a case where observed data are non-symmetric triangular fuzzy
observations rather than real numbers. Meanwhile, a case study on the car
battery length-life is provided on the basis on the proposed method.
Keywords: Fuzzy decision Non-symmetric fuzzy data Arithmetic fuzzy

numbers Analysis of variance
1 Motivation and Literature Review
If the observed data in analysis of variance (ANOVA) are recorded by fuzzy numbers
rather than crisp numbers, we face a quite new and interesting problem which the ordinary
ANOVA method is not appropriate for analysis such vague data. Hence, as the motivation
of this paper, the traditional ANOVA must be generalized in such a way that it can also
analyze in a vague environment as well. Montenegro et al. [5, 6] have presented an exact
one-way ANOVA testing procedure for the case in which the involved fuzzy random
variables are assumed normal as intended by Puri and Ralescu [10]. A one-way ANOVA
study has been developed by Cuevas et al. [1] for the functional data on a given Hilbert
space. An introduction to the asymptotic multi-sample testing of means for simple fuzzy
random variables is also was also sketched by Gil et al. [2], where a bootstrap approach to
the multi-sample test of means for the significance of the difference among population
means on the basis of the evidence supplied by a set of sample fuzzy data is studied. Also
one-way ANOVA with fuzzy data was studied by Wu [11], where the cuts of fuzzy
random variables, optimistic, pessimistic degrees and solving an optimization problem
are used. Lee et al. [4] discussed on the analysis of variance with fuzzy data based on
permutation method as a non-parametric approach.
A Case Study on Vehicle Battery Manufacturing Using Fuzzy Analysis 917
In this paper a simple method for a fuzzy one-way ANOVA, as an extension for
classical ANOVA, is presented where the observations are non-symmetric triangular
fuzzy numbers. The organization of this paper is as follows: In Sect. 2, preliminaries on
fuzzy concepts and some arithmetic operations are stated. In Sect. 3, fuzzy ANOVA is
explained for fuzzy data. Meanwhile in Sect. 3, the decision rule for testing hypothesis
of equality of means of populations data are discussed for FANOVA motel, where the
observations are reported by non-symmetric triangular fuzzy numbers. In Sect. 4 a real
applied example on vehicle battery manufacturing is presented to reveal the ideas of
this paper. Conclusion part is given in the final section.
2 Preliminaries
Let X be a universal set and FðXÞ ¼ fAjA : X ! ½0; 1g. Any A 2 FðXÞ is called a
fuzzy set on X. In particular, let R be the set of all real numbers. We will use
e ja; sla ; sra 2 R; sla ; sra [ 0g, where
FT ðRÞ ¼ f T
8
< ðx a þ sla Þsla if a sla x\a
e ðxÞ ¼ Tða; sla ; sra ÞðxÞ ¼ ða þ sr xÞ sr
T if a x\a þ sra ð1Þ
: a a
0 elsewhere:
Any T e 2 FT ðRÞ is called a triangular fuzzy number (TFN), where a is center point,
sla and sra are the left and right widths of TFN and it may written as Tða; sla ; sra Þ.
e B
Definition 2.1. [9] Let A; e 2 F ðRÞ. Then
Z
1
~ HB
A ~¼ gðaÞ A~ a ðÞB
~ a 2 da 1
2 ð2Þ
0
e and B,
is called distance between A ~ in which for any a 2 ½0; 1
n o1
~ a ðÞB
A ~ a ¼ ½a1 ðaÞ b1 ðaÞ2 þ ½a2 ðaÞ b2 ðaÞ2 2 ð3Þ
measured the distance between A ~ a ¼ ½a1 ðaÞ; a2 ðaÞ and B

~ a ¼ ½b1 ðaÞ; b2 ðaÞ,
a1 ð0Þ, b1 ð0Þ, a2 ð0Þ, b2 ð0Þ are taken as finite real numbers and g is a real valued
R1
non-decreasing function on [0,1] with g(0) = 0 and 0 gðaÞ da ¼ 12 (for instance gðaÞ ¼
mþ1 m
2 a where m ¼ 1; 2; 3; . . .).
Note 2.1. In Definition 2.1, a1 ðaÞ b1 ðaÞ and a2 ðaÞ b2 ðaÞ are the distance between
the left and the right end points of the a-cut of A~ and B,
~ respectively. The value of gðaÞ
2
can be understood as the weight of Aa ðÞB~ ~ a , and the non-decreasing property of
918 A. Parchami et al.
g means that the higher the membership of a-cut, the more important it is in deter-
~ and B.
mining the distance between A ~ This defined operation synthetically reflects the
information on every membership degree. The advantage of this arithmetic operation
on fuzzy numbers is that they can let different a-cuts have different weights.
Remark 2.1. The introduced distance in Definition 2.1 coincides to the absolute
deviation of a and b, when two numbers A ~ and B~ are real numbers a and b,
respectively.
~ ¼ Tða; sl ; sr Þ and B
Theorem 2.1. The distance between TFNs A ~ ¼ Tðb; slb ; srb Þ is
a a
1 h i
A~ HB
~ 2 ¼ ða bÞ2 þ ðsla slb Þ2 þ ðsra srb Þ2
ðm þ 2Þðm þ 3Þ
ð4Þ
ða bÞ r
þ ðsa srb Þ ðsla slb Þ ;
ðm þ 1Þ
where the weighted function gðaÞ ¼ m þ 1 m

2 a for m ¼ 1; 2; 3; . . ..
3 Fuzzy Analysis of Variance (FANOVA)
The classical ANOVA model can be consulted in any book on linear models as well as
well-known references, e.g. [3, 6]. We have been reviewed classical ANOVA model in
our previously works [7–9] with similar notations and we refer the readers of this paper
to them for shortening the length of paper. Suppose that we face with a situation in
which the recorded data/observations are TFNs ~yij ¼ Tðyij ; slyij ; sryij Þ where ~yij is inter-
preted as approximately yij . It must be noted that we are concerned with a classical
ANOVA, where the entire theoretical elements of the model such as random variables,
statistical hypothesis and populations parameter are crisp and hence the model is
considered as Yij ¼ li þ eij , with eij Nð0; r2 Þ in which Yij ’s are ordinary random
variables and the statistical hypotheses are considered as classical ones:
H0: l1 ¼ l2 ¼ . . . ¼ lr
H1: not all li are equal (there is at least one pair with unequal means).
But, just one point that will departed from classical ANOVA assumptions in
classical ANOVA model is that the sampled observations are STFNs rather than being
real numbers and nothing else is altered in the ANOVA model prior to collecting the
data. Regarding to the Definition 2.1, the observed values of the statistics SST, SSTR,
SSE, MSTR, MSE and F can be obtained as follows:
X
r X
ni
2 X
r 2 r X
X ni
2
sf
st ¼ ~yij H ~y:: ; sg
str ¼ f ¼
ni ~yi: H ~y:: and sse ~yij H ~yi:
i¼1 j¼1 i¼1 i¼1 j¼1
P
ni
P
r P
ni
when ~yij ’s are TFNs then ~yi: ¼ n1i ~yij ¼ T yi: ; slyi: ; sryi: ; ~y:: ¼ n1t ~yij ¼
j¼1 i¼1 j¼1
P
ni P
r P
ni

yij yij
sf
g ¼ r1
str e
T y:: ; sly:: ; sry:: ; yi: ¼ j¼1
ni and y:: ¼ i¼1 j¼1
nt . Then, we have mstr g
; m se ¼ nsse
t r
f ¼ nt r sf
and ~f ¼ mstr str
:
f
m se e
r1 sse
The decision rule: Let ~f be the observed value of the test statistic and F1a;r1;nt r be
the a th quantile of the fisher distribution with r 1 and nt r degrees of freedom. At
the given significance level a, we accept the null hypothesis H0 if ~f F1a;r1;nt r ;
otherwise we accept the alternative hypothesis H1.
In testing ANOVA based on fuzzy numbers the p-value can be calculated by p-

value ¼ P F [ ~f in which ~f is the observed value of the test statistic on the basis of
fuzzy observations.
Remark 3.1. Considering Remark 2.1, when the observed data are precise numbers yij ,
that is they are indicator functions Ifyij g for i ¼ 1; . . .; r and j ¼ 1; . . .; ni , then all the
introduced extended statistics in above for fuzzy ANOVA coincide to statistics of
classical ANOVA.
Theorem 3.1. In ANOVA model, suppose the observed data are ~yij ¼ Tðyij ; slyij ; sryij Þ
2 FT ðRÞ, i ¼ 1; 2; . . .; r; j ¼ 1; 2; . . .; ni , then observed values of sf
st, sg f are
str and sse
as following real values:
1 h i
sf
st ¼ ssty þ sstsry þ sstsly ; ð5Þ
ðm þ 2Þðm þ 3Þ
1 h i
sg
str ¼ sstry þ sstrsry þ sstrsly ð6Þ
ð m þ 2Þ ð m þ 3Þ
1 h i
f ¼ ssey þ
sse ssesry þ ssesly ; ð7Þ
ð m þ 1Þ ð m þ 3Þ
where ssty ; sstry and ssey are the crisp values of sst, sstr and sse for the centre points of
Tðyij ; slyij ; sryij Þ, and sstsly ; sstrsly and ssesly the crisp values of sst, sstr and sse for the left
widths of Tðyij ; slyij ; sryij Þ, and sstsry ; sstrsry and ssesry the crisp values of sst, sstr and sse for
the right widths of Tðyij ; slyij ; sryij Þ, respectively.
Result 3.1. Under the same assumption of Theorem 3.1, the observed values of the
g m
mean squares mstr, g se and the test statistic ~f are respectively as follows:
sg
str 1 h i
g ¼
mstr ¼ mstry þ mstrsly þ mstrsry ð8Þ
r1 ðm þ 2Þðm þ 3Þ
f
sse 1 h i
g
m se ¼ ¼ msey þ msesly þ msesry ð9Þ
nt r ðm þ 2Þðm þ 3Þ
and
h i
g
mstr mstry þ ðm þ 2Þ1ðm þ 3Þ mstrsly þ mstrsry
~f ¼ ¼ h i : ð10Þ
g
m se msey þ ðm þ 2Þ2ðm þ 3Þ msesly þ msesry
4 Computational Results of FANOVA on the Average Length

of Car Battery Life
Under ideal conditions, the average length of car battery life manufactured by an
Iranian factory is estimated to be about 3 to 4 years. In addition to the quality of battery
building, factors such as the amount of battery usage, maintenance, ambient temper-
ature, vibration, audio system usage, charging and discharging can also affect the
length life of the batteries used.
Consider this fact that the length of car battery life does not end at a moment and its
useful life/capability will down slowly over the time. Therefore, the length of car
battery life can be recorded by a non-precise/fuzzy number.
On the other hand, on most car batteries, an visual marker is Embedded to measure
the electrolyte concentration of the battery (which is in fact a simple
hydrometer/acidometer) and shows the battery status to the user by displaying one of
the three following colors:
• the green color means a healthy battery
• the black (or red) color means the battery is lower than the standard and needs to be
charged, but the car is still able to start, and
• the white color means the battery is completely discharged and the car can not start.
Therefore, in this research, asymmetric triangular numbers are used to record the
data on the battery lifetime.
Three assembly lines at an Iranian factory are simultaneously producing car bat-
teries with a similar brand. Recently, a claim has been made that the length of battery
life is different in these three assembly lines. Therefore, in order to check the inde-
pendency of the produced batteries lifetime from the production line number, the
factory wish to test whether the batteries lifetime from the different production line
number are the same or not. In other words, we are going to test “H0 : l1 ¼ l2 ¼ l3 ”,
versus “H1 not all li ’s are equal, for i = 1,2,3”. To this lifetime test, 29 triangular fuzzy
batteries lifetimes are selected as the experimental units in FANOVA test with sample
sizes n1 ¼ 10, n1 ¼ 8 and n1 ¼ 11, respectively (see Table 1) (Fig. 1).
Fig. 1. Triangular fuzzy observed data for the batteries lifetime in assembly lines 1, 2 and 3
(from top to bellow).
Considering Theorem 3.1 and Result 3.1, one can obtain the observed values of
FANOVA statistics which are reported in Table 2, based on the given non-symmetric
TFNs. For instance, the total sum of squares is calculated for m ¼ 1 by Theorem 3.1 in
below
" #
X
3 X
ni
1 3 X
X ni X3 X ni
r l
sf
st ¼ 2
ðyij y:: Þ þ 2
ðs sry:: Þ þ ðsyij sly:: Þ2
i¼1 j¼1
3 4 i¼1 j¼1 yij i¼1 j¼1
h i 1 h
i
¼ ð1:60 2:37Þ2 þ þ ð0:56 0:80Þ2 þ þ ð0:36 0:80Þ2 þ . . . ¼ 8:68:
34
And, also by Eq. (10) from Result 3.1 the observed value of FANOVA test statistic
is ~f ¼ 17:09 for m ¼ 1 (see Table 2).
Table 1. Fuzzy observed data for the length of car battery life in three factory assembly lines.
Observation # Line 1 Line 2 Line 3
1 T(1.60, 0.79, 0.56) T(2.06, 0.86, 0.80) T(2.13, 0.54, 1.17)
2 T(2.73, 1.21, 1.25) T(3.30, 0.87, 0.36) T(2.76, 0.53, 1.07)
3 T(2.91, 0.78, 0.42) T(3.28, 1.11, 0.50) T(2.14, 1.08, 0.47)
4 T(1.90, 0.78, 1.15) T(3.15, 1.04, 1.31) T(2.16, 0.72, 1.20)
5 T(1.75, 0.66, 0.56) T(2.80, 1.57, 0.81) T(2.26, 1.09, 0.90)
6 T(1.55, 0.84, 0.62) T(3.39, 0.64, 0.77) T(2.54, 0.68, 1.27)
7 T(1.77, 0.86, 0.91) T(2.92, 1.41, 0.85) T(2.42, 1.04, 0.44)
8 T(2.46, 0.94, 0.53) T(3.20, 0.51, 0.72) T(1.95, 0.95, 0.57)
9 T(1.53, 1.07, 0.83) T(2.77, 0.75, 1.09)
10 T(1.51, 0.51, 1.10) T(2.13, 0.68, 0.59)
11 T(1.74, 1.26, 0.36)
Table 2. Details of ANOVA for the length of car battery.

Source of variation e
ss fs
Degrees of freedom m ef
Between treatments 4.93 2 2.46
Within treatments (error) 3.75 26 0.14 17.09
Total 8.68 28
By comparing the computed FANOVA test statistic, one can accept the alternative
hypothesis H1 at significance level 0.05. The critical value of ANOVA test is
F1a; r1; nt r ¼ F0:95;2;26 ¼ 3:37 and also the computed p-value ¼ 1:82 105
strongly shows accuracy of H1 . Therefore, we conclude that there is a relation between
produced batteries lifetime and the production line number, based on the recorded
fuzzy data in Table 1.
5 Conclusion and Future Research Works
In applied sciences such as economics, agriculture and social sciences, it may be

confront with vague/fuzzy concepts, such as the threshold of patient tolerance, the
degree of utility of life and the monthly income of a Taxi driver. In such situations, the
classical ANOVA can not solve the vague test and it need to generalize based on fuzzy
data. The proposed fuzzy ANOVA (FANOVA) model is a generalized version of the
classical ANOVA using non-symmetric triangular fuzzy observations, such that when
all observations are real numbers, FANOVA reduces to ANOVA because vagueness of
the fuzzy statistics are removed and what remain is only the center point of them.
FANOVA is easy to be implemented as it is almost developed similar to classical
ANOVA for its components, and it is easy to be used by professional clients who are
familiar with ANOVA. The proposed approach in this paper can be extending for other
experimental designs such as the latin-square design in a fuzzy environment for future.
References
1. Cuevas, A., Febrero, M., Fraiman, R.: An ANOVA test for functional data. Comput. Statist.
Data Anal. 47, 111–122 (2004)
2. Gil, M.A., Montenegro, M., Gonzalez-Rodriguez, G., Colubi, A., Casals, M.R.: Bootstrap
approach to the multi-sample test of means with imprecise data. Comput. Statist. Data Anal.
51(1), 141–162 (2006)
3. Hocking, R.R.: Methods and Applications of Linear Models: Regression and the Analysis of
Variance. Wiley, New York (1996)
4. Lee, W.J., Jung, H.Y., Yoon, J.H., Choi, S.H.: Analysis of variance for fuzzy data based on
permutation method. Int. J. Fuzzy Logic Intell. Syst. 17, 43–50 (2017)
5. Montenegro, M., Colubi, A., Casals, M.R., Gil, M.A.: Asymptotic and bootstrap techniques
for testing the expected value of a fuzzy random variable. Metrika 59, 31–49 (2004)
6. Montenegro, M., Gonzalez-Rodriguez, G., Gil, M.A., Colubi, A., Casals, M.R.: Introduction
to ANOVA with fuzzy random variables. In: Lopez-Diaz, M., Gil, M.A., Grzegorzewski, P.,
Hryniewicz, O., Lawry, J. (eds.) Soft Methodology and Random Information Systems,
pp. 487–494. Springer, Berlin (2004)
7. Nourbakhsh, M.R., Mashinchi, M., Parchami, A.: Analysis of variance based on fuzzy
observations. Int. J. Syst. Sci. 44, 714–726 (2013)
8. Parchami, A., Ivani, R., Mashinchi, M., Kaya, İ.: An implication of Fuzzy ANOVA: metal
uptake and transport by corn grown on a contaminated soil. Chemom. Intell. Lab. Syst. 164,
56–63 (2017)
9. Parchami, A., Nourbakhsh, M.R., Mashinchi, M.: Analysis of variance in uncertain
environments. Complex Intell. Syst. 3, 189–196 (2017)
10. Puri, M.L., Relescu, D.A.: Fuzzy random variables. J. Math. Anal. Appl. 114, 409–422
(1986)
11. Wu, H.C.: Analysis of variance for fuzzy data. Int. J. Syst. Sci. 38, 235–246 (2007)
New Suggestion for Fuzzy Random
Variable and Its Statistical Simulation
Abbas Parchami(B)
Department of Statistics, Faculty of Mathematics and Computer, Shahid Bahonar

University of Kerman, Kerman, Iran
[email protected]
Abstract. A new kind of fuzzy random variable, named the “piece-

wise linear fuzzy random variable”, is introduced in this paper based on
the nested random cuts. Illustrative examples are presented for better
understanding of the simulation method from the piecewise linear fuzzy
random variable.
Keywords: Fuzzy random variable · Piecewise linear fuzzy number ·

Empirical cumulative distribution function
1 Introduction and Preliminaries

In this paper, it is assumed that the reader is familiar with the preliminary
concepts of the fuzzy sets theory. In this section we introduce several definitions
and symbols on fuzzy numbers which are needed throughout the paper.
Definition 1. Fuzzy subset Ã from real line IR, is said to be a fuzzy number
and defined by the membership function Ã : X → [0, 1], if:
1. Ã(x) is piecewise continuous,

2. there exists exactly one x0 ∈ IR with Ã(x0 ) = 1,
3. the α-cut Ãα = {x ∈ IR; Ã(x) ≥ α} is a convex set for each α ∈ (0, 1].
The set of all fuzzy numbers on real line are denoted by F (IR).
Several different approaches have been proposed for modeling fuzzy random
variables till now. Regarding to their modelling idea, we would like to briefly
classify the available definitions of fuzzy random variables in three classes as
follows:
1. Fuzzy random variable based on random cuts, e.g. see [1,3,4,13–18],

2. Fuzzy random variable with fuzzy-valued parameter, e.g. see [11,12],
3. Fuzzy observation of a precise random variable, e.g. see [5–7].
In order to shortening the length of the article, we will explain in detail the
advantages, disadvantages and criticisms of each category elsewhere. But in this
paper, we are going to present a different approach, comparing to the three
Fuzzy Random Variable 925
mentioned methods, that can have many applications in fuzzy random variables
simulation.
The organization of this paper is as follows. A new type of fuzzy random
variable called a piecewise linear fuzzy random variable, is defined in Sect. 2.
Then, a better formula for fuzzy random variable is presented in Sect. 3 based on
the continuous empirical cumulative distribution function. Moreover, numerical
examples are provided in Sect. 3 to simulate the proposed fuzzy random variable
in Sects. 2 and 3. Finally, conclusions are presented with concerning possible
developments to the presented work.
2 Piecewise Linear Fuzzy Random Variable

In this section, we define a new type of fuzzy random variable called a piecewise
linear fuzzy random variable, which will be a base for the definition of LR fuzzy
random variable in the next section. A piecewise linear fuzzy random variable,
is a polygon arising from the straight conjunction of several random points, can
be obtained by simulation steps as presented in the following algorithm.
Algorithm 1. Simulation of “the piecewise linear fuzzy random variable”

X̃knot.n ≈ CX ∼ fC , SX l
∼ fSXl , SX ∼ fS r
r
X
knot.n
l r
with core CX , support [SX , SX ] and the number of knots knot.n has three steps:
Step 1 (core and support generation): Simulate the core of piecewise linear
fuzzy random variable X̃knot.n from distribution fC , and simulate its left and
right spreads from distributions fSXl and fSX r , respectively. Therefore, in the
first step, the observed core and the observed support of piecewise linear fuzzy
random variable X̃knot.n are Cx and [Cx − Sxl , Cx + Sxr ], respectively.
Step 2 (knots generation): Each left and right sides of the piecewise linear
fuzzy random variable X̃knot.n must have knot.n random points. The coordinates

of knots in the left side for i = 1, 2, . . . , knot.n are Cx − l(knot.n+1−i) , knot.n+1
i
in which l(i) ’s are the ordered simulated data from the truncated distribution
iid
l on the interval 0, Sx ; i.e., l1 , . . . , lknot.n ∼ fS l ;T run[0,S l ] .
l
fSX
X x
i
The coordinates of knots in the right side are Cx + r(i) , knot.n+1 for i =
1, 2, . . . , knot.n, where r(i) ’s are the ordered simulated data from the truncated
iid
r on the interval [0, S ]; i.e., r1 , . . . , rknot.n ∼ f r
r
distribution fSX x r ].
SX ;T run[0,Sx
Step 3 (knots conjunction): After sequentially joining the generated knots in

steps 1 and 2 by several straight lines, the membership function of the observed
“piecewise linear fuzzy random variable” x̃knot.n is formed and its simulation is
completed.
926 A. Parchami
Note 1. The simulated “piecewise linear fuzzy random variable” in Algorithm

1 is on the basis of three crisp random variables (one real-valued and two non-
negative-valued). Therefore, the name of all three crisp random variables (the
first one for a random core, the second for the left random spread and third for
the right random spread) is used in the name of “piecewise linear fuzzy random
variable”. For more details, see Example 1.
Example 1. To simulate a fuzzy number from the piecewise linear fuzzy random
variable
X̃2 ≈ CX ∼ N (5, 2), SX
l
∼ E(3), SX
r
∼ U (0, 1) 2 ,
so-called piecewise linear fuzzy random variable Normal-Exponential-Uniform
with respective parameters 5, 2, 3, 0 and 1, we simulate the core and spreads of
x̃2 in Step 1 from CX ∼ N (5, 2), SXl
∼ E(3) and SXr
∼ U (0, 1) as follows:
core(x̃2 ) = Cx = 1.717,
supp(x̃2 ) = [Cx − Sxl , Cx + Sxr ]
= [1.717 − 0.057, 1.717 + 0.186]
= [1.660, 1.903].
In Step 2, we first simulate knot.n
=
2 random data from the truncated dis-
tribution E(3) on the interval 0, Sxl = [0, 0.057] with the probability density
function (pdf)
l ;T run[0,S l ] = fE(3);T run[0,0.057]
fSX x
fE(3) I(0 < x < 0.057)

=
FE(3) (0.057) − FE(3) (0)
3e3x
= , 0 < x < 0.057.
0.95
Our simulation result are l2 = 0.017 and l1 = 0.028. Therefore l(1) = 0.017,
l(2) = 0.028, so the coordinates of simulated knots for the left side are

i i
Cx − l(knot.n+1−i) , = 1.717 − l(3−i) , , i = 1, 2
knot.n + 1 3
and thus two knots/points (1.689, 0.333) and (1.700, 0.667) have been simulated
for the left side of piecewise linear fuzzy random variable.
For determination of two knots for the right side of fuzzy random variable,
we first generate two observation r1 = r(1) = 0.052 and r2 = r(2) = 0.156 from
the following pdf
r ;T run[0,S r ] = fU (0,2);T run[0,0.186]
fSX x
fU (0,2)
= I(0 < x < 0.186)
FU (0,2) (0.186) − FU (0,2) (0)
1
= I(0 < x < 0.186)
2 × 0.093
= fU (0,0.186) .
Therefore, considering Algorithm 1 and

i i
Cx + r(i) , = 1.717 + r(i) , , i = 1, 2,
knot.n + 1 3
the simulated knots for the right side are (1.873, 0.333) and (1.769, 0.667).
Then, in Step 3, we connect the simulated knots in steps 1 and 2 by a set of
straight lines to show the membership function of x̃2 as in Fig. 1.
Fig. 1. The membership function of the observed piecewise linear fuzzy random vari-
able x̃2 with 2 knots in Example 1.
3 A Better Formula for Piecewise Linear Fuzzy Random

Variable Based on the Continuous ecd Function
First, we recall the definitions of “the empirical cumulative distribution function”

and “the continuous empirical cumulative distribution function”.
Definition 2. The empirical cumulative distribution function (ecdf ) of random
sample data x1 , . . . , xn is defined as
Fx1 ,...,xn (x) = Fn (x)

1
n
#(xi ≤ x)
= I(xi ≤ x) =
n i=1 n
928 A. Parchami
⎧
⎪
⎪ 0 x < x(1)
⎨i
n x(i) ≤ x < x(i+1) ,
= (1)
⎪
⎪ i = 1, 2, . . . , n − 1
⎩
1 x(n) ≤ x
Definition 3. A continuous empirical cumulative distribution function based on

random sample data x1 , . . . , xn is obtained by sequentially connecting the points
x(i) , ni , for i = 0, 1, . . . , n, where x(i) is i-th ordered data (for i = 1, . . . , n) and
x(0) is the smallest point of X domain. Therefore, the continuous ecdf based on
the random sample data x1 , . . . , xn is defined by
⎧
⎪ 0 x < x(0)
⎪
⎨i + x−x(i)
x ≤ x < x(i+1) ,
Fxc1 ,...,xn (x) = n n(x(i+1) −x(i) ) (i) (2)
⎪
⎪ i = 0, 2, . . . , n − 1
⎩
1 x(n) ≤ x
Corollary 1. Regarding to Algorithm 1 and Definition 3, the membership func-

tion of the piecewise linear fuzzy random variable

X̃knot.n ≈ CX ∼ fC , SXl
∼ fSXl , SX ∼ fS r
r
X
knot.n
can be rewritten as follows:

⎧
⎪
⎪ 0 x ≤ Cx − Sxl
⎪
⎪
⎨ 1 − Fl1 ,...,lknot.n (Cx − x) Cx − Sxl < x < Cx
c
X̃knot.n (x) = 1 x = Cx (3)

⎪
⎪
⎪
⎪ 1 − F c
(x − Cx ) C r
x < x < Cx + Sx
⎩ r 1 ,...,r knot.n
0 x ≥ Cx + Sxr
where F c is the continuous ecdf.
Example 2. Considering Corollary 1, we simulate the membership function of

the piecewise linear fuzzy random variable

X̃knot.n ≈ CX ∼ N (3, 1), SX
l
∼ E(3), SX
r
∼ U (0, 2) knot.n
with different number of knots. These membership functions are shown in Fig. 2
for knot.n = 2, 4, 6, 12, 20, 50, 200, 1000. Intuitively, Fig. 2 inspires a kind of con-
vergency in the limiting case when knot.n −→ ∞ which can be discussed more
in future works from the both theoretical and applied point of views.
Fig. 2. The membership function of simulated piecewise linear fuzzy random variable
in Example 2 with respect to increasing the number of knots.
4 Conclusions Future Scientific Works
A logical algorithm for simulating of fuzzy random variable, named “piecewise

linear fuzzy random variable”, is proposed at the first of this paper. The pre-
ciseness of the “piecewise linear fuzzy random variable” can be controlled with
the number of random knots by the user. We may use the “piecewise linear
fuzzy random variable” in many fuzzy statistics topics (such as fuzzy regression,
testing fuzzy hypothesis by fuzzy data, fuzzy quality control and etc.) in future
scientific works.
Finally, R packages F uzzyN umbers, Calculator.LR.F N s and Sim.P LF N
are proposed to the interested readers from [2,9] and [10], which are powerful
tools for statistical simulation based on the fuzzy numbers and the “piecewise
linear fuzzy random variable” [8].
930 A. Parchami
References
1. Ferraro, M.B.: A linear regression model for LR fuzzy random variables: properties
and inferential procedures. Ph.D. thesis (2008)
2. Gagolewski, M., Caha, J.: FuzzyNumbers: Tools to deal with fuzzy num-
bers in R, R package version 0.4-1 (2015). https://cran.r-project.org/web/
packages=FuzzyNumbers
3. Katagiri, H., Kato, K., Uno, T.: Possibility/necessity-based probabilistic expecta-
tion models for linear programming problems with discrete fuzzy random variables.
Symmetry 9, 1–34 (2017)
4. Malakoori, M., Moghadam, M.B.: Simulation of LR fuzzy random variables with
normal distribution. Middle-East J. Sci. Res. 15, 768–779 (2013)
5. Nourbakhsh, M.R., Mashinchi, M., Parchami, A.: Analysis of variance based on
fuzzy observations. Int. J. Syst. Sci. 44, 714–726 (2013)
6. Parchami, A., Sadeghpour-Gildeh, B., Nourbakhsh, M., Mashinchi, M.: A new
generation of process capability indices based on fuzzy measurements. J. Appl.
Stat. 41, 1122–1136 (2014)
7. Parchami, A., Nourbakhsh, M.R., Mashinchi, M.: Analysis of variance in uncertain
environments. Complex Intell. Syst. 3, 189–196 (2017)
8. Parchami, A.: Calculator for fuzzy numbers. Complex Intell. Syst. 5(3), 331–342
(2019)
9. Parchami, A.: Calculator.LR.FNs: calculator for LR fuzzy numbers. R package
version 1.1 (2016). https://CRAN.R-project.org/package=Calculator.LR.FNs
10. Parchami, A.: Sim.PLFN: simulation of Piecewise Linear Fuzzy Numbers. R pack-
age version 1.0 (2017). https://CRAN.R-project.org/package=Sim.PLFN
11. Puri, M.L., Ralescu, D.A.: The concept of normality for fuzzy random variables.
Ann. Probab. 13, 1373–1379 (1985)
12. Puri, M.L., Ralescu, D.A.: Fuzzy random variables. J. Math. Anal. Appl. 114,
409–422 (1986)
13. Wu, C.W.: Assessing process capability based on bayesian approach with subsam-
ples. Eur. J. Oper. Res. 184, 207–228 (2008)
14. Wu, C.W.: Decision-making in testing process performance with fuzzy data. Eur.
J. Oper. Res. 193, 499–509 (2009)
15. Wu, C.W., Liao, M.Y.: Estimating and testing process yield with imprecise data.
Expert Syst. Appl. 36, 11006–11012 (2009)
16. Wu, H.C.: Statistical hypotheses testing for fuzzy data. Inf. Sci. 175, 30–56 (2005)
17. Wu, H.C.: The central limit theorems for fuzzy random variables. Inf. Sci. 120,
239–256 (1999)
18. Wu, H.C.: The laws of large numbers for fuzzy random variables. Fuzzy Sets Syst.
116, 245–262 (2000)
Networks
Fuzzy Centrality Analysis for Smart City
Trajectories
Lamia Karim1,2 , Azedine Boulmakoul2(B) , Ghyzlane Cherradi2 ,

and Ahmed Lbath3
1
LISA Laboratory ENSA Berrechid, Hassan 1st University, Settat, Morocco
[email protected]
2
Casablanca, Morocco
3
LIG/MRIM, CNRS, University Grenoble Alpes, Saint-Martin-d’Hres, France
[email protected]
Abstract. Space entities in smart city are directly concerned with the
detection, location and identification of people to ensure safety and com-
fort requirements. A variety of sensors, actuators and appropriate analyt-
ical methods and architecture are proposed to process large data volume
from sensor networks and to discover knowledge patterns to act accord-
ingly. In this work, a solution based on fuzzy centrality to monitor and
control the strategy of analyzing persons movement of people in intelli-
gent spatial entities based on trajectories gathered from outdoor spaces,
is proposed. We used structural analysis techniques and fuzzy graphs
to determine the fuzzy centrality of trajectories. The fuzzy centrality
developed in this paper is based on the lattice of fuzzy transitive rela-
tions max-min or max- The modeling of the outdoor trajectories is also
an undeniable contribution of this work. Our meta-model integrates the
OGC CityGML to represent and enable the exchange of geoinformations
necessary for the construction and operation of outdoor navigation sys-
tems to produce person’s trajectories. We also describe the framework
based on microservices to carry out this work.
Keywords: Trajectories meta-model · Outdoor trajectories · Path

algebra · Fuzzy centrality · OGC CityGML · Location based services
1 Introduction
Cities should handle around different concepts of urban planning to deal with
the problems they are facing today and tomorrow [11]. The smart city concept
involves the integration of information and communication technologies to pro-
vide various services. The modeling of trajectories captures the movement of
a mobile object evolving in space-time over a certain period [2–4,9,10]. Nowa-
days, mobile phones leave positioning logs, which specify their location, at all
times, they are connected to a mobile network. Thus, there is an opportunity to
934 L. Karim et al.
discover, from these trajectories, spatio-temporal patterns which convey useful

knowledge. This article presents a solution based on fuzzy centrality to monitor
and control the strategy of analyzing persons movement of people in intelligent
spatial entities based on trajectories gathered from outdoor zones. We use struc-
tural analysis techniques and fuzzy graphs to determine the fuzzy centrality of
trajectories. The fuzzy centrality developed in this paper is based on the lattice of
fuzzy transitive relations. The modeling of the outdoor trajectories is also briefly
described this work. We also describe the architecture based on microservices to
implement requirements developed in this work.
Structure of the Paper. In Sect. 2 we introduce foundations on path algebra
and fuzzy centrality. We draw fuzzy centrality analysis of outdoor trajectories
and the process calculus in Sect. 3. In Sect. 4, we briefly describe our smart
trajectories meta-model in interaction with OGC CityGML, and we propose a
simple deployment scenario of our approach by an application to trajectories;
while some conclusions are drawn in Sect. 5.
2 Path Algebra and Fuzzy Centrality
In this section, we recall necessary elements for understanding this work.
2.1 Preliminaries
1. Defintion 1. A semiring (path algebra) is a structure (E, ⊕, ⊗) which satisfies
the following properties:
(i) ⊕ has neutral element ε
(ii) ⊗ has a neutral element e
(iii) ε is absorbing for ⊗, that is to say: ∀a ∈ E : a ⊗ ε = ε ⊗ a.
A right semiring(resp.left) is a right pre-semiring(resp.left) satisfying prop-

erty (i) and properties (i) and (ii) below.

(i) ⊗ has e as a right neutral element (a ⊗ e = a, ∀a)(resp.left: e ⊗ a = a, ∀a)

(ii) ε is a right absorbing element (a ⊗ ε = ε, ∀a)
A semiring in which the operation ⊗ is commutative is said to be commutative.
2. Defintion 2. We call dioid a set (E, ⊕, ⊗) endowed with two internal laws
⊕ and ⊗ satisfying the following properties:
(i) (E, ⊕) is a commutative monoid with neutral element ε;
(ii) (E, ⊗) is a monoid with neutral element e;
(iii) The canonical preorder relation relative to ⊕; (defined as a ≤ b ⇔ ∃c :
b = a ⊕ c) is an order relation i.e. satisfies: a ≤ b and b ≤ a ⇒ a = b;
(iv) ε is absorbing for ⊗ i.e. ∀a ∈ E : a ⊗ ε = ε ⊗ a = ε;
(v) ⊗ is right and left distributive with respect to ⊕;
Fuzzy Centrality Analysis for Smart City Trajectories 935
The following semering

was proposed
for the Widest path problem solv-
ing [1,6] (R + ∪+∞, = max, = min, ε = 0, e = +∞);
3. Defintion 3. Let A ∈ Mn (E) be a square n × n matrix with entries in E. We

define the graph G(A) associated with A as follows [6]:
– The set of vertices of G(A) is 1, 2, ...n the set of indices of the rows (or
columns) of A;
– The set of arcs of G(A) is the set of oredred pairs (i, j) corresponding to
the terms aij of A distinct from ε.
We observe that the graph G(A) thus defined is a valued graph: each arc(i,j)
is endowed with the value ai j ∈ (aij = ε) of the corresponding term of
the matrix A. The matrix A can be considered as the generalized adjacency
matrix. Let A ∈ Mn (E) be a square n × n matrix with elements in a semiring
(E, ⊕, ⊗). For any K ∈ N, denote by AK the K th power of A, i.e. A⊗A⊗...⊗A
(k times) and define the matrices A(k) by:
k
A(k) = I ⊕ A⎡⊗A ⊕
2
⎤ ... ⊕ A
e ··· ε
⎢ .. . . .. ⎥
Where I = ⎣ . . . ⎦ is the identity matrix of Mn (E)
ε ··· e
4. Defintion 4 (([0, 1], ⊕q , p ) path algebra). A valuation set Y is the unit set
[0, 1] on which two families of internal lows ⊕q and q are thus defined [6,12]:
∀a ∈ [0, 1], ∀b ∈ [0, 1]

(i) a ⊕q b = M in[1, (aq + bq )1/q ] where q ≥ 1
(ii) a q b = 1 − M in1, [(1 − a)q + (1 − b)p ]1/q where p ≥ 1.
It is easy to show that:
A. a ⊕1 b = M in[1, a + b]
B. a⊕∞ = M ax[a, b] often noted a ∨ b
C. a 1 b = M ax[0, a + b − 1] which is nothing less than a b
D. a ∞ b = M in[a, b] often noted a ∧ b.
2.2 Fuzzy Centrality
We briefly define in this section the most used and well known fuzzy centralities
[5,7,8]
1. Defintion 5 (Fuzzy centrality). A Fuzzy centrality C defined on a fuzzy graph

= (V, E) is a function C : (G,
G
v) → R+ . The centrality function is defined
over all vertices V (G) of a given graph G.
2. Defintion 6 (Fuzzy degree centrality). For any node v ∈ V (G)

of a network

G, let Γ(v) be the set of neighbors of node Γ(v) : (v) = {w ∈ V |μE (v, w) > 0.
936 L. Karim et al.
The fuzzy degree centrality FDC of a node v ∈ V is defined as: F DC(G, v) =

|Γ(v)|. Where |.| is the fuzzy cardinality.

let
3. Defintion 7 (Fuzzy betweeness centrality). For any pair (v, w) ∈ E(G),
σ(v, w) be the total number of different fuzzy shortest paths between v and
w that pass through x ∈ V . The fuzzy betweenness centrality FBC of a node

v) =
v ∈ V is defined as: F BC(G, x,w∈V σv (x, w)/σ(x, w).
3 Fuzzy Centrality Analysis of Outdoor Trajectories
A trajectory can be represented in several ways [2]. Let Ω a set of trajectories

defined on time interval period [ts , tE ] which describes a set of attributes and
denoted by:
Ω[tS ,tE ] = {T1 , T2 , ..., Ti , .....Tn }, |Ω[tS ,tE ] | = n; tS : starting time, tE : ending
time.
We define a relevant place as a point-of-interest (POI): a location where
goods and services are provided, geometrically described using a point (lon, lat),
and semantically enriched with at least an interest category. A point of interest
(lon, lat, interest) is generally any specific location that a trajectory can visit
and that can be recommended when answering a query. The type of trajectory
analysis that is used to extract visited POIs from raw trajectory data has been
discussed in the literature as stop zone, and we follow definitions given in [3] for
a stop: a temporary suspension of the travel for some reason. In this work, we
are particularly interested in those stops that take place at a predefined POI.
Let Πc set of point-of-interest Πc = {P1 , ..., Pi , ..., Pm };
∧ : Ω[tS ,tE ] × Π;
∧ij : Ti Pj ; where is a spatial relationship that mean StopAt

⎧ ⎫
⎪
⎪ P1 . . . Pj . . . Pm ⎪
⎪
⎪
⎪ .. ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ t 1 . ⎪
⎪
⎪
⎨ .. .. ⎪
⎬
. .
⎪
⎪ ⎪
⎪
⎪ ti . . . . . . wij
⎪ ⎪
⎪
⎪
⎪ . ⎪
⎪
⎪.
⎪ . ⎪
⎪
⎪
⎩ ⎪
⎭
tn
where: wij represents the stay time at Pj .
(w −m )+∂
w¯ij = ij 2∂jj j ; wij ∈ [0, 1]: represents the reduced form of wij . mj is the
Pj column average and ∂j is the deviation.
We denote ∧ ¯ the transformation form of ∧. = ∧¯t ⊗ ∧
¯ and σij =
⊕p t
k=1...n ∧ (i, k) ⊗q ∧(k, j); σij corresponds to capacity between nodes i and j.
¯ ¯
4 Smart Trajectories Meta-Model and OGC CityGML

In this paragraph, we briefly present trajectories’ elements modeling. The class
diagram given below (see Fig. 1) summarizes this modeling. A trajectory is an
ordered associative container of events that could be fuzzy. An event is char-
acterized by its spatio-temporal attributes and by its measurement relative to
a given physical observable. Various kinds of trajectories can be derived from
the model (see further work in [2–4,9,10]). The trajectory model is also easy to
integrate within the OGC CityGML meta-model [11] (see Fig. 2).
Fig. 1. Smart trajectories meta-model and OGC CityGML
Fig. 2. UML diagram of the transportation model in CityGML
4.1 Application to Trajectories

The geographic context given in Fig. 3, the points of interest concerned cor-
respond to some centers of economic interest in the city of Casablanca.
938 L. Karim et al.
They are described as follows: TW (Twin Center), ZR (Zara apparel retailer),

MD (MD Mc Donald’s), MC (Massimo Dutti). As an illustration of our approach,
we consider four trajectories associated with four different people making tours
in the commercial area mentioned above. We quantify in minutes the staying
time made by a person when visiting a point of interest (see Table. 1). Figure 4
describes the process of obtaining the matrix defined on Πc × Πc (see Sect. 3)
by means ([0,1], max, min) path algebra. The fuzzy degree centrality used for
this purpose is given by the formula below and summarized in Fig. 5.

f dc poi = μΣ poi , poj (1)
poj ∈Γ(poi )
Fig. 3. A map view of the considered points of interest
Table 1. Staying time of points of interest
Ω TW ZR MD MC
T1 30 90 10 0
T2 80 30 0 100
T3 0 100 100 20
T4 100 0 0 120
T5 120 60 50 0

Fig. 4. process computing

poi f dc poi
TW 0.7
ZR 0.3
MD 0.3
MC 0.9
Fig. 5. Fuzzy degree centrality by Fig. 6. Fuzzy degree centrality distribution

point of interest
The distribution of the fuzzy centrality given in Fig. 5, shows that the point
of interest MC is leader, followed by TW. ZR and MD follow after and define the
weak centralities. Admittedly, the differences in centralities are not significant,
however they demonstrate the importance of the rank of each point of interest.
The objective here is to deploy the calculation process by a simple example. In
the future, the tests will be established on real scenarios (Fig. 6).
4.2 System Architecture
In this section, the general architecture of the proposed system is presented. The
platform has a microservice architecture, as shown in Fig. 7. We implement the
microservices with reactive principles (Responsive, Resilient, Elastic and Mes-
sage Driven) to offer great possibilities whereby each component is individually
developed, released, deployed, scaled, updated and removed.
Fig. 7. View of the system architecture
Data are received in push and/or pull, and almost every data stream can
be considered as an IoT data source. They may come from IoT Brokers, social
media, streams, etc. The system is composed with small microservices such as:
a spatial database engine, traffic stream trajectory, path algebra builder, fuzzy
centrality processing, etc. Those micoservices use asynchronous messaging to
communicate with other similar services.
940 L. Karim et al.
5 Conclusion
The trajectories cover several areas. The spatio-temporal routes of navigation
of a mobile object in an urban environment constitute the most popular tra-
jectories in the literature. These trajectories relate to transport vehicles or per-
sons. Consequently, there is an opportunity to discover, from these trajectories,
spatio-temporal patterns which convey useful knowledge. In this paper, we have
proposed an analytical solution based on fuzzy centrality to monitor and discover
behavior models according to persons movements evolving in a geographic space,
from collected semantic trajectories. We used fuzzy structural analysis practices
to determine the fuzzy centrality of trajectories. Our trajectories meta-model
can easily be integrated within OGC CityGML to represent and to allow spatial
informations exchange necessary to produce the trajectories of persons. We also
presented an ecosystem of microservices to implement requirements introduced
in the system specifications. The implementation of the tests of our prototype
is planned in a future work.
References
1. Baras, J.S., Theodorakopoulos, G.: Path problems in networks. Synth. Lect. Com-
mun. Netw. 3(1), 1–77 (2010)
2. Boulmakoul, A., Karim, L., Lbath, A.: Moving object trajectories meta-model and
spatio-temporal queries. Int. J. Database Manag. Syst. 4(2), 35–54 (2012)
3. Boulmakoul, A., Karim, L., Elbouziri, A., Lbath, A.: A system architecture for
heteroge-neous moving-object trajectory metamodel using generic sensors: tracking
airport security case study. IEEE Syst. J. 9(1), 283–291 (2013)
4. Boulmakoul, A., Karim, L., Lbath, A.: T-Warehousing for hazardous materials
transportation. Ingénierie des Systèmes d’Information 21(1), 39–59 (2016)
5. Fan, T.F., Liau, C.J., Lin, T.Y.: A theoretical investigation of regular equivalences
for fuzzy graphs. Int. J. Approximate Reasoning 49(3), 678–688 (2008)
6. Gondran, M., Minoux, M.: Semirings: New Models and Algorithms. Springer Pub-
lishing Company Incorporated, Operations Research/Computer Science Interfaces
Series (2008)
7. Hu, R., Zhang, G.: Structural holes in directed fuzzy social networks. J. Appl.
Math. 2014, 1–8 (2014)
8. Hu, R.J., Li, Q., Zhang, G.Y., Ma, W.C.: Centrality measures in directed fuzzy
social networks. Fuzzy Inf. Eng. 7(1), 115–128 (2015)
9. Karim, L., Boulmakoul, A., Lbath, A.: Real time analytics of urban congestion
trajectories on hadoop-mongoDB cloud ecosystem. In: Second International Con-
ference on Internet of things, Data and Cloud Computing, pp. 1–11 (2017)
10. Maguerra, S., Boulmakoul, A., Karim, L., Badir, H.: A distributed execution
pipeline for clustering trajectories based on a fuzzy similarity relation. Algorithms
2019, 12–29 (2019)
11. OGC City Geography Markup Language (CityGML) Encoding Standard. http://
www.opengis.net/spec/citygml/2.0
12. Yager, R.R.: On a general class of fuzzy connectives. Fuzzy Sets Syst. 4, 235–242
(1980)
Improving Customer Experience
for an Internet Service Provider: A Neural
Networks Approach
Özge H. Namlı1,2, Seda Yanık2(&), Faranak Nouri2,

N. Serap Şengör2, Yusuf Mertkan Koyuncu3, and İrem Küçükali3
1
Turkish-German University, Beykoz, Istanbul, Turkey
2
Istanbul Technical University, Macka, Istanbul, Turkey
[email protected]
3
Turkcell İletişim Hizmetleri AŞ, Maltepe, Istanbul, Turkey
Abstract. Today one of the challenges of companies is to decrease call center

costs while improving the customer experience. In this study, we make pre-
diction and proactively take action in order to solve customer problems before
they reach the customer call center. We use machine learning techniques and
train models with a dataset of an internet service provider’s several different
systems.
We first use supervised techniques to classify the customers having slow
internet connection problems and normal internet connection. We apply two
classification approaches, multi perceptron neural networks and radial basis
neural networks. Then, we cluster the same dataset using unsupervised tech-
niques, namely Kohonnen’s neural networks and Adaptive Resonance Theory
neural networks. We evaluate the classification and clustering results using
measures such as recall, accuracy and Davies-Bouldin index, respectively.
Keywords: Call center problem prediction Classification Clustering

Artificial neural networks
1 Introduction
As a result of the rise of the customer needs in today’s work, call centers is used as an
important tool to increase the customer satisfaction. However, the service given by the
call centers also incurs significant costs for the companies. To gain competitive
advantage, companies look for innovative approaches and new technologies to
decrease the call center costs while retaining, even improving, the customer experience.
Internet providers are depending mostly on the call center services to solve the cus-
tomer problems. While system variables, and many customer variables can be observed,
sophisticated prediction approaches can help to identify problems and proactively solve
these problems before customers access the call center with the problems if not recognize
the problems at all. Such a proactive approach would help to improve the customer
experience as well as decrease the load, hence costs of the call center services.
There exists a vast literature related to customer experience measurement using
machine learning in the setting of the telecommunication sector. Anchuen et al. [2]
942 Ö. H. Namlı et al.
develop neural networks using quality of service levels and customer evaluations data
to predict customer experience. Charonyktakis et al. [3] compare neural networks,
support vector machines and decision trees used for customer experience prediction.
Jung et al. [5] develop a SCOR model for customer experience prediction using
decision trees. Diaz-Aviles et al. [4] develop a random forest model to predict the
customer experience provision result in call center calls. Amour et al. [1] compare
naïve bayes, decision trees, random forest, NN and ANFIS to classify the factors
affecting the customer experience.
In this study, we focus on a specific customer experience problem that results with a
call center call of connection speed complaint. We aim to develop classification and
clustering methods to predict the connection speed problem using system variables and
identify the factors and their provisions resulting in a call related to connection speed.
Using this approach, similar provisions can be detected and proactively speed con-
nection problems can be solved. We use multi-layer perceptron (MLP) neural network
(NN) and radial basis function (RBF) network approaches for the classification prob-
lem and self-organizing map (SOM) and adaptive resonance theory (ART) approaches
for the clustering problem. Then we compare the prediction performance of these
methods.
The study is organized as follows: Sect. 2 presents the procedures of the machine
learning methods used, Sect. 3 describes the application, Sect. 4 summarizes the results
obtained, finally Sect. 5 presents the conclusion and the future studies.
2 Methodology
Classification is a supervised learning method, which focuses on the prediction of

known labels. MLP and RBF networks are used for classification in this study.
Multi-Layer Perceptron (MLP) is the most widely used feed-forward neural net-
works for classification problems where non-linear transformation is required [6]. The
back propagation learning algorithm is initiated by assigning all weights (Wn) ran-
domly. The input data and the added bias term are multiplied by the weights of the first
hidden layer and the calculated value is passed into the activation function. A sigmoid
function is commonly used as the activation function. The output of the sigmoid
functions is the output of the hidden layer. This process is repeated for all layers,
including the output layer. As a result of the process, output values (estimate values) for
the network were calculated. The instantaneous value of the error is calculated as in
Eq. (1):
1 X l 2
E ðlÞ ¼ ydj ylj ð1Þ
2 j2C
For optimizing the weight of the network, local gradients are calculated. Each
neuron calculates its output and local gradient. Weight correction for all layers is done
as the last step as in Eq. (2) where g and d denotes learning rate and local gradient,
respectively. The momentum term (a) is included in the equation to escape the local
Improving Customer Experience for an Internet Service Provider 943
minimum. To complete the learning, weights are updated until the termination criteria
are met.
Wji ðk þ 1Þ ¼ Wji ðk Þ þ g:dj :yi þ aðWji ðkÞ Wji ðk 1ÞÞ ð2Þ
The radial basis function (RBF) network, which has two layers except the input layer,
is another widely-used feed-forward network. The hidden units in RBF networks
commonly use a bell-shaped Gaussian activation function [6] and the output unit
activations are calculated by the sum of the weighted hidden unit activations. The
computational steps of the algorithm are as given in [6].
The error instantaneous cost function is calculated as in Eq. 2. While, gw ; gc ; gr
show the learning rate, the weights are updated as in Eq. (3–5) with momentum term as
in Eq. 2.
wji ðk þ 1Þ ¼ wji ðkÞ þ gw ej /i ð3Þ

X
m xr cir ðkÞ
cir ðk þ 1Þ ¼ cir ðkÞ þ gc ew /
j¼1 j ji i
ð4Þ
r2i
!
Xm kX c i k 2
ri ðk þ 1Þ ¼ ri ðkÞ þ gr ew /
j¼1 j ij i
ð5Þ
r3i ðk Þ
Clustering is an unsupervised learning method in which the set of objects are assigned
into subsets (clusters) in such a way that points in the same subset are more similar to
each other than to those in other subsets. SOM and ART networks are used for the
clustering problem in this study.
The Self-Organizing Map (SOM) is used both to cluster data and to transform an
incoming signal pattern of arbitrary dimension into a one or two dimensional discrete
map. The SOM algorithm can be summarized as follows [7]:
First, the weights of the neurons that are placed at the chosen lattice are initialized.
Chosen a random input x from data set, the similarity between weights of the nodes
of the neurons and the input is determined.
In SOM, each weight vector has a location in the gird chosen, it also has neigh-
boring neurons that are close to it. The weight of the winning neuron w based on
similarity matching and the weights of its neighboring neurons hj;iðxÞ are updated to
become more alike to the randomly selected sample.
Each neighboring node’s weights are updated. The closer a node to the winning
neuron, the more its weight is altered. The weight vector wj is updated by Eq. 6.
wj ðn þ 1Þ ¼ wj ðnÞ þ gðnÞhj;iðxÞ ðnÞðxðnÞ wj ðnÞÞ ð6Þ
At the end of each iteration, the winning neuron matrix is compared to the previous
matrix to determine when the properties do not change, if the winning neuron matrix is
the same in two successive iterations, it is continued up to 500 * n iterations to better
represent the winning data by the neurons.
Adaptive resonance theory uses the unsupervised learning technique [5]. The ART
algorithm checks whether an input matches to one of the previously stored clusters. If
so, the entry is added to the most matched cluster, otherwise, a new cluster is created
adaptively. ART1 procedure which uses binary vectors has the following components
and steps: (i) Input unit (F1 layer), combines the signal from the input with the signal of
the F2 layer. The F1 layer is attached to the F2 layer with bottom-up (Wji ) weights and
the F2 layer is attached to the F1b layer via the top-down ðZij ) weights. (ii) Competitive
Unit (F2 layer) is the attentional subsystem’s output layer. The unit with the largest net
entry will be selected to learn the entry order and all other cluster units will be enabled
by setting to 0. (iii) Reset Mechanism is based on the similarity between the top-down
weight and the input vector. If the degree of this similarity is less than a pre-specified
vigilance parameter, the cluster will not be allowed to learn the pattern.
There is no stopping criterion for the algorithm, and it is continued by creating a
new cluster whenever necessary with each incoming data.
3 Application
The application of the clustering and classification approaches presented in Sect. 2 is

conducted for one of the largest internet service provider in Turkey. The aim is to use
features from different databases of the company and predict the connection speed
problems proactively. The flow of the application is presented in Fig. 1.
Evaluation of
Classification
-MLP
Results
-RBF
Classification
Data - Correctly Classified Rate
- Precision
Pre-Processing
Data Collection - Recall
& - F-measure
- ROC Curve
Coding
Clustering
Clustering - DB Index
-ART - Dunn Index
-SOM - Silhouette Coefficient
- V-Measure
Fig. 1. The flow of the application study.
In the data collection module, a database of the customer set whose data will be
collected and a database of specified variables are used. Accordingly, the data col-
lection module connects to various systems such as CRM, XDG, HDM etc. using
different protocols, then draws the required data and finally records them in a new
database. The customer set is specified according to the connection infrastructure and
the location of the customers manually. A time is set for the data collection process.
In the data pre-processing module, the variables in the data set is encoded because
most of the data are text variables. In order to use such variables more effectively for
machine learning, data are coded as 0–1 binary variables as follows: The value cate-
gories of each text variable are assigned numbers from 1 to n, and then these numbers
are converted into numbers written in mod 2. The variables which are real numbers are
also written in mod 2 and represented as 0–1 values. The following modules of the
study, classification, clustering and the evaluation of the results are presented in Sect. 4.
MLP and RBFN algorithms are applied on the data in order to detect the speed problem
of the customers. The connection speed of the customers is estimated with three labels
as “uncertain”, “normal” and “slow”. The adjusted parameter values and the results of
the performance criteria of the classification algorithms are presented in Table 1. When
we examine the results of RBFN and MLP applications in our standard dataset, we
observe that the prediction success is good. However, when we examine the ROC
curves of these algorithms from Fig. 2, we observe that all the samples in the test set
are assigned to only two classes and the “slow” label is not predicted at all because its
ratio is very low in the training data set. In order to overcome this problem, we
duplicate the samples, with slow label, and create an equal frequency training set.
When we examine the applications of the classification algorithms on the “Duplicated
Dataset”, we can observe in Fig. 3 and Table 1 that all classes are taken into consid-
eration and the results are obtained with high accuracy prediction success.
Table 1. MLP and RBFN algorithms results.

MLP
# of Neurons Learning Rate Sigmoid Coefficient for Termination
Function Sigmoid Protocol
Hl. 1: # of Hl. 1: 0.1 Hl. 1: Logistic Hl. 1: 1 Mean
Input + 1 Hl. 2: 0.2 Hl. 2: Logistic Hl. 2: 2 Error < Epsilon
Hl. 2: # of Output l.: 0.3 Output l.: Output l.: 3
Input + 2 Logistic
Output l.: 3
Alfa # of Iteration # of Run W Epsilon
0.2 100 20 [−0.25, 0.25] 0.1
Dataset Accuracy Precision Recall F-Measure Itr. Number
S Dataset 0.9616 0.9325 0.9616 0.9460 31.25
D Dataset 0.9226 0.9732 0.9226 0.9429 6.35
DTDRD Dataset 0.912 0.9831 0.9120 0.9416 2
FADRD Dataset 0.8925 0.9717 0.8925 0.9250 1
RBFN
# of Neurons Alfa # of Sigma Termination Protocol
Runs
89 0.7 20 [0.5, 1.5] mean error < epsilon
Learning Rate Epsilon C W # of Iteration
[0.20, 0.80] 0.1 [0.1, 0.9] [0.01, 0.25] 100
Dataset Accuracy Precision Recall F-Measure Itr. Number
S Dataset 0.8274 0.8090 0.8274 0.8164 15.85
D Dataset 0.8728 0.9293 0.8728 0.8935 18.75
DTDRD Dataset 0.3234 0.3987 0.3234 0.3474 91.1
FADRD Dataset 0.3439 0.5111 0.3439 0.3858 100
(S Dataset: Standard Dataset, D Dataset: Duplicated Dataset, DTDRD Dataset: Decision Tree - Dimension
Reduction - Duplicated Dataset, FADRD Dataset: Principal Component Analysis - Dimension Reduction -
Duplicated Dataset)
Fig. 2. ROC curves for standard dataset.
Fig. 3. ROC curves for duplicated dataset.
In order to increase the success and effectiveness of the applied classification

algorithms, dimension reduction is made using Decision Tree Algorithm and Principal
Component Analysis on the data set. Application results are presented in Table 1 and
Figs. 4 and 5 for dataset named “DTDRD Dataset” and “FADRD Dataset”.
When all the results are examined, it is seen that the dimension reduction do not
yield curative results for classification success. It is seen that the best and most effective
results are obtained with “D Dataset” for both MLP and RBFN algorithms. Although
the results are close, MLP algorithm is more successful and faster than RBFN.
SOM and ART algorithms are applied to data for the detection of the speed
problem of the customers. The adjusted parameter values and the results of the per-
formance criteria of the clustering algorithms are presented in Table 2.
Fig. 4. ROC curves for dimension reduction dataset by decision tree.

Fig. 5. ROC curves for dimension reduction dataset by factor analysis.
We use four performance measure indices to evaluate clustering. Davies-Bouldin

index is defined as a ratio between the cluster scatter and the cluster’s separation and a
lower value close to zero will mean that the clustering is better, the V-Measure is
defined as the harmonic mean of homogeneity and completeness and 1 stands for
perfectly homogeneous or completeness. The silhouette value is a measure of how
similar an object is to its own cluster, high value close to 1 indicates that the object is
well matched to its own cluster. The Rand Index computes a similarity measure
between two clusters, value 1 indicates the identical clustering.
According to Fig. 6 and Table 2 by applying ART and SOM on the “DTDR” data,
the performance measures are better than the other datasets for both algorithms.
Table 2. SOM and ART algorithms results.

SOM
# of # of # of W Neighborhood Match Learning Effective Width
Neurons Iteration Run Function Similarity Rate Function
100 20 100 [0,1] Gaussian Euclidean Exponential Exponential
distance decreasing decreasing
Dataset Davies V_Measure V_Measure Adjusted Rand Silhouette # of Cluster
Bouldin (Beta = 0.6) (Beta = 1.6)
Homogeneity Completeness
S Dataset 2.149 0.167 0.112 0.0057 0.093 99
D Dataset 1.993 0.215 0.146 0.014 0.077 94
DTDRD 0.830 0.126 0.085 0.013 0.805 87
Dataset
FADRD 1.584 0.217 0.148 0.017 0.128 83
Dataset
ART (Vigilance = 0.6)
Dataset Davies V_Measure V_Measure Adjusted Rand Silhouette # of Cluster
Bouldin (Beta = 0.6) (Beta = 1.6)
Homogeneity Completeness
S Dataset 1.040 0.071 0.046 0.00016 −0.096 3394
D Dataset 0.931 0.296 0.206 0.006 0.376 3962
DTDRD 0.842 0.369 0.274 0.0533 0.646 193
Dataset
Fig. 6. ART and SOM different performance measures results for different dataset.
5 Conclusion
In this study, we aim to predict the connection speed problems of internet users. We
apply two classification and two clustering algorithms on the dataset obtained from an
internet service provider. For the classification approaches, the sparsity of the dataset is
found to be important for the prediction performance. MLP is found to be superior than
RBFN. Data duplication, a remedy for data sparsity, improves the prediction perfor-
mance of both MLP and RBFN. With both SOM and ART clustering algorithms, the
best clustering performance is obtained when dimension reduction is applied to the
duplicated dataset. However, the superiority between SOM and ART can’t be specified
because it changes with respect to different performance evaluation criteria.
In future research, we will make a comparison whether the generated clusters match
the labels used in the classification. And we will use clustering prior to classification as
a hybrid algorithm in order to improve the prediction accuracy.
References
1. Amour, L., Souihi, S., Hoceini, S., Mellouk, S.: A hierarchical classification model of QoE
influence factors. In: 13th WWIC, Malaga, Spain (2015)
2. Anchuen, P., Uthansakul, P.M.: QOE model in cellular networks based on QOS
measurements Using NN Approach. In: 13th ECTI-CON, Chiang Mai, Thailand (2016)
3. Charonyktakis, P., Plakia, M., Tsamardinos, I., Papadopouli, M.: On user-centric modular
QoE prediction for VoIP based on machine-learning algorithms. IEEE Trans. Mob. Comput.
15(6), 1443–1456 (2015)
4. Diaz-Aviles, E., Pinelli, F., Lynch, K., Nabi, Z., Gkoufas, Y., Bouillet, E., Salzwedel, J.:
Towards real-time customer experience prediction for telecommunication operators. In: IEEE
International Conference on Big Data, pp. 1063–1072 (2015)
5. Grossberg, S.: Adaptive resonance theory. Encyclopedia of cognitive science (2006)
6. Haykin, S.: Neural networks: a comprehensive foundation. Prentice Hall PTR, Upper Saddle
River (1994)
7. Kohonen, T.: The self-organizing map. Proc. IEEE 78(9), 1464–1480 (1990)
Segmentation of Larynx Histopathology
Images via Convolutional Neural
Networks
Ahmet Haşim Yurttakal1(B) and Hasan Erbay2

1
Computer Technologies Department, Technical Sciences Vocational School,
Bozok University, Yozgat, Turkey
[email protected]
2
Computer Engineering Department, Engineering Faculty,
University of Turkish Aeronautical Association, Ankara, Turkey
Abstract. The larynx is a part of the throat, between the base of the
tongue and the trachea. Larynx cancer is a disease in which malign cells
form in the tissues of the larynx. It amounts to about 3% of newly diag-
nosed cancers and has a poor prognosis. The histopathologic analyses
of larynx cancer have been performed manually by pathologists. It is
time-consuming and correct diagnosis highly depends on pathologists’
experience. However, computerized methods are crucial for early detec-
tion and monitoring treatment progress in medicine, also free of human
error.
In this study, the Convolutional Neural Network-based U-Net model
developed for the automatic segmentation of larynx histopathology
images. The dataset comprises 55 laryngeal cancer cases. There was a
total of 224 P63 stained images of different grades. Among them, 87 were
Grade I, 73 Grade II, and 64 Grade III cases. According to the simulation
results, the model can quickly and accurately differentiate cell structures
on tissue and allows advanced image analysis operations. Moreover, it is
suitable for use in a laboratory environment. It also helps pathologists
in the decision-making process.
Keywords: Segmentation · Larynx · Histopathology · Convolutional

Neural Networks
1 Introduction
Cancer is the common word for all malignant tumors, and there are thousands of
cancer types differentiated according to the pathological differences in appear-
ances of tumors. Moreover, their clinical analyses are also different. Cancer clas-
sifications in pathology mean to collect those tumors with similar pathological
features under the same type and distinguish those with different features. An
effective classification should be easy to implement, highly reproducible among
pathologists, and easily understood by clinicians. Furthermore, it is expected to
950 A. H. Yurttakal and H. Erbay
facilitate diagnosis, help to predict patients’ clinical courses, and guide treat-
ment. Larynx cancer is a disease in which malign cells form in the tissues of the
larynx [9]. The prevalence of cancer at the time of diagnosis is the most cru-
cial factor and not only determines the prognosis but also helps in choosing the
most appropriate treatment based on the experience obtained previous treat-
ment results of patients in a similar stage. One of the methods frequently used
at this stage is pathological staging. One of the methods frequently used at this
stage is histopathological staging. It includes all data obtained on the spread of
the disease after surgery, which is applied as an initial treatment or performed
within 4 months after diagnosis unless there is any systemic or radiation therapy
[4]. To relate biologically relevant features to clinical variables, it is essential to
apart the nuclei from tissue images. However, the segmentation of the nuclei is
a time-consuming process. Moreover, tissue type, staining differences, and cell
type give different visual features [10].
On the other hand, in recent years, there have been very successful results
in image processing and machine learning-based studies in the classification and
segmentation of digital pathology images. Lung cancer [19], brain cancer [2],
cardiac cancer [8], liver cancer [13] are some of the successful studies in the
classification of histopathological images. Prostate cancer [15], skin cancer [16],
gastric cancer [18], breast cancer [14] are some of the successful studies in the
segmentation of histopathological images.
In this study, initially, we employed Otsu segmentation for generating ground-
truth. Then, we proposed a fully automatic larynx histopathology segmentation
system using U-Net based Convolutional Neural Network (CNN). The model
might assist pathologists in the decision-making process for the correct staging
of laryngeal cancer tumors and the association of biological features with clinical
data.
The rest of the paper is organized as follows. In Sect. 2, the dataset is detailed,
and the methods used in the study discussed. Section 3 describes simulation results
as well as the effectiveness of the proposed model via some statistical metrics.
Finally, Sect. 4 concludes the paper with a summary and the paper’s contribution.
2 Materials and Methods
2.1 Dataset
Medisp Histology Image Collection Library (HICL) dataset was used in the
study. The original data set includes 55 cases of laryngeal cancer and contains 450
P63 painted images with a different magnification of x20 and x40. Among them,
175 images are of grade I cases, 147 grade II cases, and 128 grade III cases. The
light microscopy imaging system, comprised of a LEICA DM 2500 microscope
connected to a LEICA DFC 420C camera (Leica Microsystems GmbH), was used
for digitization. Raw clinical materials collected at Archives of Patras University
Hospital in Greece [11].
Experiments were carried out on 224 images, 87 of them were Grade I, 73 Grade
II, and 64 Grade III. Figures 1 and 2 shows some sample images from dataset.
Segmentation of Larynx Histopathology Images 951
Fig. 1. Grade I Fig. 2. Grade II
2.2 Groundtruth Generating

Groundtruth mask of each image was generated to compare the segmentation
success of the model. In this step, Gaussian Blur was applied to images, then
each segmented with Otsu [12], one of the well-known thresholding algorithm.
Then, closing morphological procedure was performed with a 5 × 5 elliptical disk.
2.3 Segmentation Model

U-Net architecture is composed of 31 layers at which 23 are convolution, 4 down-
sampling, and 4 upsampling layers. The downsampling process is carried out
using a 2-step (2-max max pooling) operation and the upsampling process using
the 2 × 2 up convolution step. ReLU activation function is applied after 3 × 3
convolutional layer at each stage, and dropout applied after both the convolu-
tion layers. It consists of 16 filters in the first layer and the number of channels
increased by twice in each downsampling layer, reaching 256 filters in the deepest
layer. The total number of trainable parameters is 1962625.
3 Simulation Results
Application implemented in Python environment. Images in the dataset are
1728 × 1296 in size and 24 in bit depth. Initially, all images were resized to
256 × 256 in dimensions. After the ground-truth operation, the dataset was
divided into two parts, 70% for training, 30% for testing. 20% of the train-
ing set is reserved for validation. Then U-Net model was applied to segment the
laryngeal images.
The performance of the model was measured in terms of accuracy, mean
intersection-over-union (mean-IOU), and mean squared error(MSE). Figure 3
shows accuracy graph over 50 epoch. IOU is the area of overlap between the
predicted segmentation and ground-truth. It is divided into the area of unity
between the predicted segmentation and ground-truth. It is often used in seg-
mentation processes. At the end of the training, the mean-IOU, accuracy, and
MSE values of the model are 0.7624, 0.9341, 0.0474, respectively. Figure 4 shows
segmentation result.
Fig. 3. Model accuracy curve
Fig. 4. Original image and its segmentations

Segmentation is one of the most important tasks used in histopathological image
analysis. It can make some structures like Nuclei more obvious for the next clas-
sification. Many studies have been conducted in the literature on segmentation
and classification of histopathological images [7]. On the other hand, state-of-
the-art deep learning approaches give successful results in complex segmentation
problems [5]. However, studies on larynx cancer are limited. Some studies are
given below.
Atlı and Gedik [1] proposed a U-Net based VESUNET model for automatic
classification of retinal vessels. Dong et al. [3] proposed a fully automated method
on BRATS 2015 datasets containing 220 high-grade brain tumors and 54 low-
grade tumor cases for brain tumor segmentation developed using U-Net-based
deep convolution networks. Tong et al. [17] an improved lung nodule segmenta-
tion algorithm based on the U-Net network. Lin et al. [6] compared the success
of the U-net based approach with the K-means based approach for skin lesion
segmentation.
In this study, 224 larynx cancer histopathological images, which were auto-
matically ground-truth with Otsu, were segmented with U-Net. The advantage
Segmentation of Larynx Histopathology Images 953
of the study is that the proposed method is completely automatic and it can
divide various histopathological images into sections. Even in a relatively small
dataset, it achieved high accuracy in classifying three different grade levels. It
also requires less preprocessing. The simulation results are promising. Therefore,
it can help doctors make decisions as a second reader.
In the next study, ground-truth images will be created with a specialist doc-
tor. Then, it will be classified by Residual U-Net and classified into different
grades with transfer learning models.
References
1. Atli, İ., Gedık, O.S.: Vesunetdeep: a fully convolutional deep learning architecture
for automated vessel segmentation. In: 2019 27th Signal Processing and Commu-
nications Applications Conference (SIU), pp. 1–4. IEEE (2019)
2. Barker, J., Hoogi, A., Depeursinge, A., Rubin, D.L.: Automated classification of
brain tumor type in whole-slide digital pathology images using local representative
tiles. Med. Image Anal. 30, 60–71 (2016)
3. Dong, H., Yang, G., Liu, F., Mo, Y., Guo, Y.: Automatic brain tumor detection and
segmentation using u-net based fully convolutional networks. In: Annual Confer-
ence on Medical Image Understanding and Analysis, pp. 506–517. Springer (2017)
4. Edge, S.B., Byrd, D.R., Carducci, M.A., Compton, C.C., Fritz, A., Greene, F.,
et al.: AJCC Cancer Staging Manual, vol. 7. Springer, New York (2010)
5. Garcia-Garcia, A., Orts-Escolano, S., Oprea, S., Villena-Martinez, V., Garcia-
Rodriguez, J.: A review on deep learning techniques applied to semantic segmen-
tation. arXiv preprint arXiv:1704.06857 (2017)
6. Lin, B.S., Michael, K., Kalra, S., Tizhoosh, H.R.: Skin lesion segmentation: U-nets
versus clustering. In: 2017 IEEE Symposium Series on Computational Intelligence
(SSCI), pp. 1–7. IEEE (2017)
7. de Matos, J., Britto Jr, A.d.S., Oliveira, L.E., Koerich, A.L.: Histopathologic image
processing: a review. arXiv preprint arXiv:1904.07900 (2019)
8. Mazo, C., Alegre, E., Trujillo, M.: Classification of cardiovascular tissues using
LBP based descriptors and a cascade SVM. Comput. Methods Programs Biomed.
147, 1–10 (2017)
9. Napier, S.: Laryngeal carcinoma. In: Histopathology Reporting, pp. 177–183.
Springer (2020)
10. Naylor, P., Laé, M., Reyal, F., Walter, T.: Nuclei segmentation in histopathology
images using deep neural networks. In: 2017 IEEE 14th International Symposium
on Biomedical Imaging (ISBI 2017), pp. 933–936. IEEE (2017)
11. Ninos, K., Kostopoulos, S., Kalatzis, I., Sidiropoulos, K., Ravazoula, P., Sakel-
laropoulos, G., Panayiotakis, G., Economou, G., Cavouras, D.: Microscopy image
analysis of p63 immunohistochemically stained laryngeal cancer lesions for predict-
ing patient 5-year survival. Eur. Arch. Otorhinolaryngol. 273(1), 159–168 (2016)
12. Otsu, N.: A threshold selection method from gray-level histograms. IEEE Trans.
Syst. Man Cybern. 9(1), 62–66 (1979)
13. Pang, W., Jiang, H., Li, S.: Sparse contribution feature selection and classifiers
optimized by concave-convex variation for HCC image recognition. BioMed Res.
Int. 2017 (2017)
14. Peikari, M., Salama, S., Nofech-Mozes, S., Martel, A.L.: Automatic cellularity
assessment from post-treated breast surgical specimens. Cytometry Part A 91(11),
1078–1087 (2017)
15. Salman, S., Ma, Z., Mohanty, S., Bhele, S., Chu, Y.T., Knudsen, B., Gertych, A.: A
machine learning approach to identify prostate cancer areas in complex histological
images. In: Information Technologies in Biomedicine, vol. 3, pp. 295–306. Springer
(2014)
16. Saraswat, M., Arya, K.: Supervised leukocyte segmentation in tissue images using
multi-objective optimization technique. Eng. Appl. Artif. Intell. 31, 44–52 (2014)
17. Tong, G., Li, Y., Chen, H., Zhang, Q., Jiang, H.: Improved u-net network for
pulmonary nodules segmentation. Optik 174, 460–469 (2018)
18. Yu, F., Ip, H.H.: Semantic content analysis and annotation of histological images.
Comput. Biol. Med. 38(6), 635–649 (2008)
19. Zhang, R., Shen, J., Wei, F., Li, X., Sangaiah, A.K.: Medical image classifica-
tion based on multi-scale non-negative sparse coding. Artif. Intell. Med. 83, 44–51
(2017)
Glioma Brain Tumor Grade Classification
from MRI Using Convolutional Neural
Networks Designed by Modified FA
Timea Bezdan , Miodrag Zivkovic , Eva Tuba , Ivana Strumberger ,

Nebojsa Bacanin(B) , and Milan Tuba
Singidunum University, Danijelova 32, 11000 Belgrade, Serbia

{tbezdan,mzivkovic,istrumberger,nbacanin}@singidunum.ac.rs,
{etuba,tuba}@ieee.org
Abstract. Gliomas represent the most common form of brain tumors.

The most often used technique, to establish the diagnosis, is based on
magnetic resonance imaging. To establish the diagnosis in the early stage
is sometimes very difficult even for a specialist with much experience,
thus an efficient and reliable system is needed that helps the special-
ist in the interpretation. The convolutional neural network has excel-
lent achievement in image classification; though, adjusting the values of
hyperparameters is a very time-consuming process. In this paper, we
propose to adjust the hyperparameters of convolutional neural networks
by a modified firefly algorithm and apply it to glioma grade classifica-
tion. We evaluated the proposed approach on magnetic resonance images
from more data collections. The typical brain images are obtained from
the IXI dataset. The glioma brain tumor images are used from the can-
cer imaging archive. The obtained results confirm superiority related to
other techniques in the same research area.
Keywords: Convolutional neural network · Medical image

classification · Metaheuristics · Firefly algorithm · Computer-aided
diagnosis
1 Introduction
The most frequent brain tumors are gliomas. Tumors can be benign or malignant.
Benign tumors are non-cancerous and will not spread to other organs. On the
other hand, malignant tumors are cancerous, and they can spread to other tissues
and organs in the body. Depending on cell activity and potential growth, gliomas
can be categorized into four different grades (from grade I to grade IV), from
benign to the most malignant. Grade I glioma brain tumors are benign, and the
cells are growing slowly. Grade II tumors have abnormal cells and can develop
into more aggressive into grade III or grade IV. Grade III are malignant gliomas;
the cells are more aggressive and grow faster. Grade IV represents the most
aggressive malignant glioma brain tumor, where the cells grow fast and spread
out to other parts in the brain. The most often used technique for establishing
tumor diagnosis is based on magnetic resonance imaging (MRI). The clinical
practice depends on accurate tumor classification. Thus it is important to develop
a glioma classification system, which can support specialists in the MRI-based
diagnosis. The objective of this paper is to develop a method for the glioma brain
tumor grade classification, the grades on the scale from I to IV in MR images
by using convolutional neural network evolved by swarm intelligence approach.
Convolutional neural networks (CNNs) [7] has three essential types of layers,
convolution, pooling, and dense layers (also called fully-connected layers). In the
convolution layer, the kernel (filter) convolves the input image and extract low-
level features, in the architecture, every subsequent convolution layer extracts
higher and higher-level features from the raw pixels of the input images with-
out image pre-processing. After the convolution process, the activation func-
tions transform the convolved feature into non-linear output. The next layer is
the pooling layer, which is used for dimensionality reduction, and it results in
faster computation. The structure ends with flattening dense layers, and the
final dense layer classifies the input image. Oftentimes the model learns the spe-
cific data (image) and fails in new data classification; this is a high-variance
problem (overfitting). The utilization of the regularization technique prevents
the issue of high-variance; one strong regularization method is dropout [11]. The
dropout indicates to dropping out the neurons from the dense layer. One of the
most challenging tasks in CNN design is to find the fittest values for various
hyperparameters, which is considered as a typical NP-hard class optimization
problem. Metaheuristic algorithms are utilized extensively in complex optimiza-
tion problems. They do not always secure to obtain the optimal global solution,
yet they will find an approximate solution in a reasonable time. The two main
stages of the algorithm are exploration (diversification) and exploitation (inten-
sification). The exploration stage investigates the search space on a global scale;
on the other hand, the exploitation stage is seeking nearby for the fittest solu-
tions. Commonly, the metaheuristic algorithms are subdivided into two broad
categories, swarm intelligence algorithms and evolutionary algorithms.
The rest of the paper is organized as follows: Sect. 2 presents the proposed
method, in Sect. 3 we exhibit the simulation results, and Sect. 4 gives the con-
clusion of this research and future work.
2 Methodology
This section details the proposed methodology. First, we give an overview of the
original firefly algorithm [18], and then we describe the modified version.
2.1 Original Firefly Algorithm
In 2008, Yang proposed the firefly algorithm (FA). FA has plenty succeeded
implementation in various real-life problems, image processing [14,17], convolu-
tional neural network design [13], radio frequency identification network planning
Glioma Brain Tumor Grade Classification 957
[3], portfolio optimization [4], and different hybridized versions are also devel-
oped, such as [12,15,16].
The algorithm is inspired by the fireflies’ flashing light activity. The FA fol-
lows three rules:
1. Each firefly can be attracted by any other firefly since all of them are unisex;
2. The attractiveness of firefly is determined by the light level, by decreasing the
distance, the attractiveness of fireflies’ increasing and they become brighter;
3. The fitness function is determined by the fireflies’ brightness.
Initially, the algorithm generates random solutions (fireflies) by the following

formula:
xi,j = lbj + rand · (ubj − lbj ) (1)
where the jth element of ith individual (firefly) is denoted by xi,j ; lbj and ubj
are the lower and upper bound, respectively. And rand is a random number
drawn from the uniform distribution.
The Cartesian distance measures the distance between the two individuals:

d

ri,j = ||xi − xj || = (xi,k − xj,k ) (2)
k=1
where the distance between the ith and jth firefly is denoted by ri,j , the problem
dimension is denoted by d.
The light intensity is approximated by the following formula:
I0
I(r) = (3)
1 + γr2
where I(r) indicates to the light intensity, while I0 is the light intensity of the
source, the distance is denoted by r. The light absorption coefficient is denoted
by γ.
The attractiveness between two fireflies is calculated as:
β0
β(r) = (4)
1 + γr2
where the attractiveness is denoted by β and β0 represents the attractiveness at

r = 0.
The solution will update the position in search space by utilizing the following
formula:
2
xt+1
i = xti + β0 r−γri,j (xj − xi ) + α(rand − 0.5) (5)
where the new position of the ith firefly is denoted by xt+1 i ; the current position
at iteration t is denoted by xti ; the solution moves toward the brighter solution xj ;
the randomization control parameter is drawn from the Gaussian distribution,

and it is denoted by α; rand is a random number between 0 and 1.
In summary, the algorithm is controlled by three parameters, light absorption
coefficient (γ), attractiveness (β), and the randomization parameter (α).
2.2 Modified Firefly Algorithm

In this paper, we propose a modified version of the FA algorithm. We introduce
an additional operation in the search process that allows us to find the global
optimum in less iteration and avoiding to get trapped in the local minima.
A random variable will determine whether the firefly movement Eq. (5) will
be utilized for position update or the current solution will move toward the
current fittest solution by using the following formula:
xt+1
i = xt∗,j + λ(xt∗,j − xti,j ) (6)
where the updated solution is denoted by xt+1 i,j ; the jth element of the current
best solution at iteration t is denoted by xt∗,j ; the jth element of the current
solution at iteration t is denoted by xti,j . λ is a control parameter and its value
is reduces over the course of the iteration according to Eq. (7).
λ = a(2r − 1) (7)
where r is a random number from the uniform distribution and a is calculated
as follows:

t
a=2 1− (8)
maxIter
where the current iteration is denoted by t and maxIter refers to the maximum
number of iterations.
If the light intensity of jth solution is less then the light intensity of the ith
solution, the firefly will be discarded and new random solution will be generated
instead by employing Eq. (1). The pseudocode of the proposed modified FA
algorithm is depicted in Algorithm 1.
3 Experimental Setup and Results

3.1 Dataset Details
In this paper, we used the same dataset likewise in the first case study in [2].
The normal, healthy brain MR images are collected from the IXI dataset [1]; the
dataset contains almost 600 images. MR images with a glioma brain tumor are
obtained from three datasets from the cancer imaging archive (TCIA) [5]. The
first dataset, the repository of molecular brain neoplasia data (REMBRANDT)
dataset [9] comprises MR images from 130 patients glioma brain tumor grade
on a scale of I to IV. The second data collections are from the cancer genome
Algorithm 1. Pseudocode of the modified firefly algorithm

Initialize the population of N solutions (fireflies)
Evaluate the light intensity
Set the value of γ (the light absorption coefficient); iteration counter (t) to 0; define
the maximum iteration M axIter
while t < M axIter do
Update the value of λ
for i = 1 to N do
for j = 1 to i do
if p > 0 then
if Ij > Ii then
Move the current solution xi toward xj by using Eq. (5)
else
Generate new random solution by using Eq.(1)
end if
else
Update the position toward the best solution by utilizing Eq. (6)
end if
Evaluate the fitness of the new solution
end for
end for
Sort the solution according to their fitness and save the current best solution
end while
Return the best solution
atlas glioblastoma multiforme (TCGA-GBM) [10]; this collection contains MR

images from 262 patients. The third data collection contains MR images from
199 patients, and it is obtained from the cancer genome atlas low-grade glioma
(TCGA-LGG) [8]. In this study, we only used T1 axial post gadolinium contrast
agent images.
The pixel values of the images are between 0 and 255; we normalized the
values in the range from 0 to 1. In order to make the dataset more extensive, we
employed the data augmentation technique. Besides increasing the training data
size, data augmentation reduces the high-variance problem. In our proposal,
we have created new images by mirroring and rotating the original images in
different angels. After the augmentation, the total number of MR images is
8000, with various glioma brain tumor grade and 8000 normal brain MR images.
We split the dataset into 80% training set and 20% test set.
3.2 Experimental Setup
The initial population in the modified firefly algorithm generates 50 random CNN
architectures and evaluates the fitness function, which is the classification error
rate for each structure. The algorithm needs to optimize the hyperparameters
of the networks through the progression of the iteration and output the best
CNN structure. In this study, we optimize the following hyperparameters in the
structures: number of convolutional layers, number of filters, the size of each

filter, pooling layer, number of the dense layers and the number of neurons in
each layer, type of the activation function, type of the optimizer, learning rate
and the dropout rate. In order to reduce the space of feasible solutions and
computation time, the values of each hyperparameter are defined between the
lower and upper bound.
The number of convolutional layers (CL) is between 2 and 5. Similarly to the
set of convolution layers, the number of pooling layers (P L) is between 2 and 5.
We did not optimize the type of pooling layer; we only applied the max-pooling.
The size of the max-polling layer is 2 × 2. The set of the dense layer (DL) has
three possible values 1, 2 or 3. The filter size ranges from 2 to 7, and the number
of filters (F N ) is between 16 and 128. The number of hidden units (neurons) in
the dense layer is between 128 and 512. The dropout rate is defined in the range
from 0.1 to 0.5. The set of the activation functions (AF ) contains ReLU, Leaky
ReLU (LReLU), ELU, and SELU. As the loss function, the categorical cross
entropy is selected. To optimize the loss function, we define a set of six different
optimizers, SGD optimizer, Adam, Adamax, N adam, Adagard and Adadelta.
The set of learning rate (LR) contains the following values:{0.01, 0.001, 0.0001}.
We use zero-padding around the input in order to keep the output size the
same. The configuration of the CNN structure hyperparameters is similar to [2],
to make a better and fair comparative analysis. The population size is set to 50.
The maximum number of iteration is set to 15. The three control parameters
that control the algorithm are the randomization parameter, and its value is set
to 0.5; the light absorption coefficient is set to 1, and the attractiveness at a
distance zero is set to 0.2.
3.3 Experimental Results
We run the algorithm in 15 iterations, and the best CNN structure is generated
after 12 iterations. The accuracy of the consecutive iterations of the proposed
algorithm is visualized in Fig. 1.
Fig. 1. Average accuracy through the course of iterations
The best CNN architecture resulted in 92.6% accuracy on the test set. The
structure has four convolution layers and four max-pooling layers, and at the
end of the structure, one dense layer, the second dense layer, is the classification
layer. The filter size is 5 × 5 in the first two convolutional layerers and 3 × 3 in
the third and fourth layer. The number of filters is 16 in the first convolution
layer, 32 in the second layer, in the third layer 96, and in the fourth layer 128.
The number of the hidden units is 256 in the fully-connected layer. As transfer
function, the ReLU is selected by the algorithm, and Adam optimizer is utilized,
with the learning rate of 0.001. The dropout rate is 0.3. The proposed approach is
compared to two similar methods on the same dataset. The results of CNN+GA
and SVM+RFE [6] are taken from [2]. The comparison is presented in Table 1,
and its visual representation is in Fig. 2.
Table 1. Comparison of the proposed method
Method Accuracy(%)
SVM+RFE [6] 62.5
CNN+GA [2] 90.9
CNN+mFA 92.6
Fig. 2. Comparative analysis
4 Conclusion
In this paper, we propose a novel approach, CNN-based approach, to classify
the grades of the glioma brain tumors from MR images, and we propose a meta-
heuristic approach to design the convolutional neural network automatically. To
select the best hyperparameters for the CNN structure, we employ the modified
firefly algorithm. The proposed approach is evaluated on MR images from more
data collections, the normal brain images are used from the IXI dataset, and the
glioma brain tumor images are used from the cancer imaging archive (TCIA).
The proposed method is compared to similar techniques, and the results show
the superiority of the proposed method, and it is a strong candidate for glioma
brain tumor grade classification from MR images. In future research, we are going
to apply other basic metaheuristics, as well as improved, modified metaheuristic
algorithms to the same problem dataset.
Acknowledgment. The paper is supported by the Ministry of Education, Science

and Technological Development of Republic of Serbia, Grant No. III-44006.
References
1. IXI dataset (2015). http://brain-development.org/ixi-dataset/
2. Anaraki, A.K., Ayati, M., Kazemi, F.: Magnetic resonance imaging-based brain
tumor grades classification and grading via convolutional neural networks and
genetic algorithms. Biocybern. Biomed. Eng. 39(1), 63–74 (2019). https://doi.
org/10.1016/j.bbe.2018.10.004. http://www.sciencedirect.com/science/article/pii/
S0208521618300676
3. Bacanin, N., Tuba, M., Jovanovic, R.: Hierarchical multiobjective RFID network
planning using firefly algorithm. In: 2015 International Conference on Information
and Communication Technology Research (ICTRC), pp. 282–285, May 2015
4. Bacanin, N., Tuba, M.: Firefly algorithm for cardinality constrained mean-variance
portfolio optimization problem with entropy diversity constraint. Sci. World J.
2014, 1–16 (2014)
5. Clark, K.W., Vendt, B.A., Smith, K.E., Freymann, J.B., Kirby, J.S., Koppel, P.,
Moore, S.M., Phillips, S.R., Maffitt, D.R., Pringle, M., Tarbox, L., Prior, F.W.: The
cancer imaging archive (TCIA): maintaining and operating a public information
repository. J. Digit. Imaging 26, 1045–1057 (2013)
6. Kalbkhani, H., Shayesteh, M.G., Zali-Vargahan, B.: Robust algorithm for brain
magnetic resonance image (MRI) classification based on garch variances series.
Biomed. Signal Process. Control 8(6), 909–919 (2013). https://doi.org/10.1016/j.
bspc.2013.09.001
7. Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to
document recognition. Proc. IEEE 86(11), 2278–2324 (1998). https://doi.org/10.
1109/5.726791
8. Pedano, N., Flanders, A., Scarpace, L., Mikkelsen, T., Eschbacher, J., Hermes,
B., Ostrom, Q.: Radiology data from the cancer genome atlas low grade glioma
[TCGA-LGG] collection. The Cancer Imaging Archive (2016)
9. Scarpace, L., Flanders, A.E., Jain, R., Mikkelsen, T., Andrews, D.W.: Data from
rembrandt. Cancer Imaging Archive 10, K9 (2015)
10. Scarpace, L., Mikkelsen, L., Cha, T., Rao, S., Tekchandani, S., Gutman, S., Pierce,
D.: Radiology data from the cancer genome atlas glioblastoma multiforme [TCGA-
GBM] collection. Cancer Imaging Archive 11(4) (2016)
11. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.:
Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn.
Res. 15, 1929–1958 (2014)
12. Strumberger, I., Bacanin, N., Tuba, M.: Enhanced firefly algorithm for constrained
numerical optimization. In: 2017 IEEE Congress on Evolutionary Computation
(CEC), pp. 2120–2127. IEEE (2017)
13. Strumberger, I., Tuba, E., Bacanin, N., Zivkovic, M., Beko, M., Tuba, M.: Design-
ing convolutional neural network architecture by the firefly algorithm. In: Pro-
ceedings of the 2019 International Young Engineers Forum (YEF-ECE), Costa da
Caparica, Portugal, pp. 59–65 (2019)
14. Tuba, M., Bacanin, N., Alihodzic, A.: Multilevel image thresholding by fire-
works algorithm. In: 2015 25th International Conference Radioelektronika
(RADIOELEKTRONIKA), pp. 326–330, April 2015. https://doi.org/10.1109/
RADIOELEK.2015.7129057
15. Tuba, M., Bacanin, N.: Artificial bee colony algorithm hybridized with firefly algo-
rithm for cardinality constrained mean-variance portfolio selection problem. Appl.
Math. Inf. Sci. 8(6), 2831 (2014)
16. Tuba, M., Bacanin, N.: Improved seeker optimization algorithm hybridized with
firefly algorithm for constrained optimization problems. Neurocomputing 143,
197–207 (2014). https://doi.org/10.1016/j.neucom.2014.06.006
17. Tuba, M., Bacanin, N.: JPEG quantization tables selection by the firefly algo-
rithm. In: 2014 International Conference on Multimedia Computing and Systems
(ICMCS), pp. 153–158. IEEE (2014)
18. Yang, X.S.: Firefly algorithms for multimodal optimization. In: Watanabe, O.,
Zeugmann, T. (eds.) Stochastic Algorithms: Foundations and Applications, pp.
Dynamic System Control Using Z-number
Based Fuzzy Neural Network
Rahib H. Abiyev(&)
Department of Computer Engineering, AAIRC, Near East University,

Mersin-10, North Cyprus, Turkey
[email protected]
Abstract. Z-number based fuzzy neural network (ZFNN) for dynamic system
control is proposed. Using the interpolative reasoning mechanism, the structure
and design algorithms of the ZFNN are presented. The gradient algorithm is
designed for adjusting the parameters of the ZFNN. The proposed ZFNN
structure is tested for control of a dynamic system and the performance of the
ZFNN based controller is compared with the performances of the other con-
trollers. The results of the simulation show the efficiency of using the designed
system in control of dynamic systems.
Keywords: Z-number Dynamic system control Fuzzy neural network
1 Introduction
Input and output variables of fuzzy rule base systems are basically described by fuzzy
linguistic values [1, 2]. The reliabilities of these fuzzy values are important for the
accurate construction of the rules and therefore for modelling of the control systems.
Zadeh proposed Z-number in order to handle the uncertainty of information using its
reliability degree [3]. Z-number uses restriction and reliability information for each
fuzzy variable in the rule base. Using this information Z-number more adequately
describes the imperfect knowledge in the rule base.
In the literature, a set of research studies related to the development of the Z-number
based systems have been published [3–7]. The references [6–8] applied Z-number for
solving the multi-criteria decision making problems. In these papers, Z-number is
transformed into crisp numbers in order to solve the problems. In [9] Z-number is used
to solve AHP problem using the approach proposed in [8]. These research papers
present the advantages of the presented methodology. But, the transforming of
Z-numbers in real numbers may cause information’s loss that may lead to decreasing the
performance of the designed fuzzy system. The studies [10–13] used Z-numbers for the
designing of decision making systems.
The fuzzy reasoning under Z information is an important problem. The inference
mechanisms of the existing fuzzy system are based on the compositional rule of
inference, interpolation, the concept of distance measure, analogy and similarity. Koczy
and Hirota, based on distance information, purposed interpolative reasoning for the
sparse fuzzy rules [14, 15] and used for the control problem in [16]. [17] presents fuzzy
Dynamic System Control Using Z-number Based Fuzzy Neural Network 965
control of dynamic plants using Z-number and interpolative reasoning. The references
[18, 19] present the development of a Z-number based controller for omnidirectional
robots. [20] uses Z-number to estimate the security risk level of food.
The design of fuzzy rules is a very important step in designing of the fuzzy system.
One of an effective approach for the designing of the fuzzy system is the use of artificial
neural networks (NNs) [21–24]. This paper presents a combination of Z-number based
fuzzy sets and neural networks for the construction of ZFNN based control system.
Contributions can be summarized as follows: the structure of the ZFNN is proposed;
the design algorithms of the ZFNN are presented using interpolative reasoning; the
learning algorithm of the ZFNN is developed.
The remaining part of the paper is organized as follows. Section 2 gives fuzzy rule
interpolation. Section 3 presents the structure and design algorithms of the ZFNN.
Section 4 gives simulation example and the conclusions are presented in Sect. 5.
2 Fuzzy Rule Interpolation
In this paper, the fuzzy interpolative reasoning proposed in [14, 15] is used for the
design of the ZFNN. This reasoning mechanism is used for sparse rule base and it was
applicable for the continuous, normal and convex fuzzy sets, with bounded support.
Let’s consider SISO type fuzzy rules. Assume that the value of input variable X is
A*. Let’s calculate the corresponding output Y of the fuzzy rule. Assume that input A*
value is between fuzzy sets A1 and A2 (see Fig. 1). Let’s compute the value of the
corresponding output variable Y of a fuzzy system using the rules that include A1 and
A2. That is, let’s determine B* in If X = A* Then Y = B*?
Fig. 1. The membership functions used in fuzzy rules.
The problem is the finding of the output fuzzy sets B* of the rule. Let’s use a-cut for
the given fuzzy rules. a-cuts of A1 and A2 fuzzy sets are denoted as inf{A1a}, sup{A1a}
and inf{A2a}, sup{A2a} which are called infimum and supremum of A1 and A2,
respectively (Fig. 1). Based on conditions A1 < A* < A2 and B1 < B2 and using linear
interpolation, Koczy and Hirota [14, 15] have shown that
966 R. H. Abiyev
d ðA ; A1 Þ d ðB ; B1 Þ
¼ ð1Þ
d ðA ; A2 Þ d ðB; B2 Þ
Here d(*) is distance between two fuzzy sets. Euclidian distance, Hamming dis-
tance, Kaufman-Gupta distance, Hausdorff distance, measures [17] can be used to find
d(*). In the paper, the Euclidian distance is used.
Koczy and Hirota [14, 15] used a-cut to calculate the lower dL and upper dU
distances between the two fuzzy sets. By this way the fuzzy distance between two
fuzzy sets is estimated

dL Aa ; Aa1 ¼ d inf Aa ; inf Aa1 ¼ inf Aa inf Aa1
a a a a a
dU A ; A2 ¼ d sup A ; sup A2 ¼ sup A sup Aa2
ð2Þ
dL Ba ; Ba1 ¼ d inf Ba ; inf Ba1 ¼ inf Ba inf Ba1

dU Ba ; Ba2 ¼ d sup Ba ; sup Ba2 ¼ sup Ba sup Ba2
Using distance measure, interpolative reasoning was proposed in [14, 15]. In the
next section the design of ZFNN is presented.
3 Z-number Based Fuzzy Neural Network
The design of the Z-number based fuzzy neural network (ZFNN) consists of the devel-
opment of the fuzzy IF-THEN rules in the network structure using the training capability
of neural networks. We use Mamdani type IF-THEN rules in rule base [18–20]. Let’s
consider Z-rules and its reasoning mechanism. Multi-input single-output (MISO) fuzzy Z
rules are presented as follows.

If x1 is A1j ; R1j and . . . xi is Aij ; Rij and. . . xm is Amj ; Rmj Then y is Bj ; Rj
ð3Þ
where j = 1,..,n, i = 1,…,m, n is the number of rules and m is the number of input
signals. Aij and Bj are restrictions, Rij and Rj are reliabilities defined for input and output
fuzzy sets. xi and y are the system’s input and output signals, correspondingly.
In the paper, using interpolative reasoning we are presenting the design of the
ZFNN and its learning algorithm. The proposed ZFNN structure is given in Fig. 2.
The ZFNN includes five layers. The first layer is designated for distributing of input
signals xi. The second layer is the rule layer, that is represented by R1, R2,…,Rn. The
weights of connection between first and second layers are represented
by cij and rcij.
The a cut will be denoted as caij ¼ cLa
ij ; cij
Ua
; rcaij ¼ rcLa
ij ; rcij
Ua
. In the first layer
using a-cut the distances between the input signals xi and fuzzy values cij are
calculated.
Fig. 2. The structure of the ZFNN.
2
dc caij ; xai ¼ caij xai ð4Þ
At first, the formula (9) is used to find the distance dcai between constraint C and
input X, and then the same formula is adapted to find the distance dria between relia-
bilities rxi and rcij. Then the sum of lower and upper distances are determined.
Xm Xm
Dcaj ¼ DcLa
j ; Dcj
Ua
ij ;
¼ ð i dcLa i ij Þ;
dcUa ð5Þ
Xm Xm
Drja ¼ ðDrjLa ; DrjUa Þ ¼ i
drijLa ; i
drijUa ; ð6Þ
The total distance will be determined by summing the distances computed for
constraint and reliability parameters.
j ¼ Dcj þ Drj ; Dj ¼ Dcj þ Drj

DLa ð7Þ
La La Ua Ua Ua
The layer three is used to compute the inverse of output signals of the second layer.
1 1
j ¼
QLa Ua ; Qj ¼ DLa ð8Þ
Ua
Dj j
The next fourth layer computes the weighted output signals of the third layer
Xn Xn Xn
Ya ¼ j¼1
Qaj waj ; RaY ¼ j¼1
Qaj rwaj ; Sa ¼ j¼1
Qaj ð9Þ
a La Ua
where Y a ¼ Y La ; Y Ua ; RaY ¼ RLa
Y ; RY , S ¼ S ; S
Ua
. The last forth layer com-
putes the output of the network
968 R. H. Abiyev

a Y a RaY Ya RaY
ðU ; RaU Þ ¼ ; ¼ ð Pn a ; P n aÞ ð10Þ
Sa Sa j¼1 Qj j¼1 Qj

Here U a ¼ ðU La ; U Ua Þ and RaU ¼ RLa U ; RU
Ua
are outputs of the network obtained
for constraint and reliability parameters.
Using formulas (4–10), the ZFNN output is computed. We are using triangular type
fuzzy sets. Using a = 0 and a = 1 levels we can obtain left Ul, middle Um and right Ur
values of the triangle. Left (Ul, RUl) and right (Ur, RUr) values are related to a = 0 level,
middle (Um, RUm) - to a = 1 level. The formula U = ((Ul + 4 * Um + Ur)/6) * ((RUl
+ 4 * RUm + RUr)/6) is used to find the crisp output value. The design of the ZFNN
includes the finding of appropriate correct values of the parameters of the antecedent
(cij and rcij) and consequent (wj and rwj) parts. In this paper, the gradient descent
algorithm is developed for updating the parameters of the ZFNN. The update of the
parameters has been carried out using errors calculated on the output of the network.
1 XO
E¼ ðUkd Uk Þ2 ð11Þ
2 k¼1
Here Uk is current and Ukd is desired output signals of the network (k = 1,..,n), O is
number of network outputs. The parameters wj, rwj, cij and rcij in the consequent and
premise parts of the ZFNN are adjusted as
@Ea
waj ðt þ 1Þ ¼ waj ðtÞ þ c þ b wa
ðt þ 1 Þ wa
ð t Þ ð12Þ
@waj j j
@Era
rwaj ðt þ 1Þ ¼ rwaj ðtÞ þ c a þ b rwaj ðt þ 1Þ rwaj ðtÞ ð13Þ
@rwj
@Ea
caij ðt þ 1Þ ¼ caij ðtÞ þ c a þ b caij ðt þ 1Þ caij ðtÞ ð14Þ
@cij
@Era
rcaij ðt þ 1Þ ¼ rcaij ðtÞ þ c þ b rc a
ðt þ 1 Þ rc a
ð t Þ ð15Þ
@rcaij ij ij
where c and b are learning rate and momentum respectively. Using (12–15) the update
of the parameters have been performed.
4 Simulation Studies
Z-number based fuzzy neural network presented above is used for the dynamic system
control. The architecture of the ZFNN based control system is presented in Fig. 3. Here
using set-point signal g(k) and output of the plant y(k) the value of error e(k) is
determined. ThisP error signal is used to find the values of the change of error e’(k) and
the sum of error e(k). These signals are inputs for the ZFNN controller. The output of
ZFNN is control signal given to the plant. The error, change of error and sum of error
are used to determine the corresponding value of control signal.
Fig. 3. The architecture of the ZFNN based control system.
ZFNN is characterized by weight coefficients. These coefficients are parameters of

the If-Then rules (3). Determination of proper values of these parameters is important
issue and they have affect on the accuracy on designed ZFNN based control system.
The proposed ZFNN is applied for controlling the following dynamic plant.
yðk 1Þyðk 2Þðyðk 1Þ þ 2:5Þ

yð k Þ ¼ þ u(k) ð16Þ
1 þ y ð k 1Þ 2 þ y ð k 2Þ 2
where y(k − 1), y(k − 2) are one-step and two-step delayed plant’s outputs, u(k) is
control signal.
The excitation signal used for the design of the ZFNN is given as:
8
>
> 10; 0 k\50
<
15; 50 k\100
GðkÞ ¼ ð17Þ
>
> 10; 100 k\150
:
15; 150 k\200
The problem is the determination of unknown parameters of the ZFNN which are
the coefficients of the second layer (cij and rcij, i = 1,…,m; j = 1,…,n)- antecedent parts
of the rule (1) and the coefficients of the fourth layer (wj, rwj, j = 1,…,n)- consequent
part of the rule (1). The gradient descent learning algorithm described above section is
used for updating the values of parameters. The obtained parameters are used for the
construction of the premise and consequent parts of the fuzzy rules (3).
During learning the values of the ZFNN coefficients are initialized randomly. The
learning is carried out for 500 and 1000 epochs. The simulations were performed using
a different number of rules. At first, the ZFNN based control system is trained using 16
neurons and 500 learning epochs. In simulation root mean square error (RMSE) is used
for estimation the performance of the ZFNN. During learning, the RMSE value was
obtained as 0.014621. After learning, the ZFNN system is tested using the plant model
(16) and the reference signal (17). RMSE value was obtained as 0.014609. After
training the ZFNN is used for the control of the dynamic plant. Figure 4 depicts the
response characteristics of the ZFNN based control system. Table 1 shows simulation
results the ZFNN based control system for the plant (16) using a different number of
970 R. H. Abiyev
rules and also different epoch number. For comparison, the simulations results of the
ZFNN are compared with simulations results of the other models. Table 2 shows the
results of simulations of different controllers. As shown from the table, the values of
RMSE for the ZFNN based control system is less than of other control systems.
Fig. 4. Time response characteristic of the ZFNN based control system for different set-point
signals. The dashed line is a set-point signal, the solid line is a plant output.
Table 1. Simulation Results. Table 2. Comparative Results.

Number Epochs Training Test error Test Models Number RMSE
of error (Mean) (Mean) error of rules Mean Best
rules (Best)
Type-1 FNS [2] 3 0.2896 0.2214
3 500 0.050629 0.052624 0.037616
9 0.1886 0.1476
1000 0.049491 0.048032 0.017384
FWNN [21] 3 0.35159 0.3181
8 500 0.029387 0.031182 0.010965
Type-2 TSK FNS [2] 3 0.1802 0.0967
1000 0.017744 0.014037 0.004278
9 0.1469 0.0725
16 500 0.014621 0.014609 0.002654
ZFNN 3 0.052624 0.0376
1000 0.002654 0.002654 0.001845
8 0.031182 0.01097
5 Conclusions
The paper proposes the ZFNN system for dynamic plant control. Interpolative rea-
soning is implemented for designing the ZFNN based control system. The structure and
operating algorithms of the ZFNN are presented. Using the proposed structure and
gradient algorithm the design of the ZFNN is performed. The ZFNN is used for the
designing of the control system. The obtained simulation results show the suitability of
using ZFNN in dynamic plants’ control. Future research will be focusing on the
investigation of ZFNN structure using the compositional rule of inference and
implementation of the ZFNN for solving different engineering problems.
References
1. Zadeh, L.A.: The concept of linguistic variable and its application to approximate reasoning-
I. Inf. Sci. 8(3), 199–249 (1975)
2. Abiyev, R.H., Kaynak, O.: Type-2 fuzzy neural structure for identification and control of
time-varying plants. IEEE Trans. Ind. Elect. 57(12), 4147–4159 (2010)
3. Zadeh, L.A.: A note on a Z-number. Inf. Sci. 181, 2923–2932 (2011)
4. Aliev, R.A., Huseynov, O.H., Aliyev, R.R., Alizadeh A.V.: The arithmetic on Z-numbers.
Theory and Application. World Scientific Publishing Co. (2015). 316 p.
5. Aliev, R.A., Zeinalova, L.M.: Decision making under Z-information. In: Guo, P., Pedrycz,
W. (Eds.) Human-Centric Decision-Making Models for Social Sciences (Studies in
Computational Intelligence), pp. 233–252. Springer, Cham (2014)
6. Kang, B., Wei, D., Li, Y., Deng, Y.: Decision making using z-numbers under uncertain
environment. J. Comput. Inf. Syst. 8(7), 2807–2814 (2012)
7. Kang, B., Wei, D., Li, Y., Deng, Y.: A method of converting Z-number to classical fuzzy
number. J. Comput. Inf. Syst. 9(3), 703–709 (2012)
8. Xiao, Z.-Q.: Application of Z-numbers in multi-criteria decision making. In: Proceedings of
Informative and International Conference on Cybernetics for Computational Social Systems
(ICCSS), pp. 91–95 (2014)
9. Azadeh, A., Saberi, M., Atashbar, N.Z., Chang, E., Pazhoheshfar, P.: Z-AHP: a Z-number
extension of fuzzy analytical hierarchy process. In: Proceedings of International Conference
on Digital Ecosystems and Technologies (DEST), pp. 141–147 (2013)
10. Zamri, N., Ahmad, F., Rose, A.N.M., Makhtar, M.: A fuzzy TOPSIS with Z-numbers
approach for evaluation on accident at the construction site. Adv. Intell. Syst. Comput. 549,
41–50 (2017)
11. Aliev, R.A., Pedrycz, W., Alizadeh, A.V., Huseynov, O.H.: Fuzzy optimality based decision
making under imperfect information without utility. Fuzzy Optim. Decis. Making 12(4),
357–372 (2013)
12. Wang, J.Q., Cao, Y.X., Zhang, H.Y.: Multi-criteria decision-making method based on
distance measure and choquet integral for linguistic Z-numbers. Cogn. Comput. 9(6), 827–
8426 (2017)
13. Yager, R.R.: On Z-valuations using Zadeh’s Z-numbers. Int. J. Intell. Syst. 27, 259–278
(2012)
14. Koczy, L.T., Hirota, K.: Approximate reasoning by linear rule interpolation and general
approximation. Int. J. Approximate Reasoning 9(3), 197–225 (1993)
15. Koczy, L.T., Hirota, K.: Interpolative reasoning with insufficient evidence in sparse fuzzy
rule bases. Inf. Sci. 71(1), 169–201 (1993)
16. Kovács, S., Kóczy, L.T.: Application of interpolation-based fuzzy logic reasoning in
behaviour-based control structures. In: Proceedings of the FUZZIEEE, IEEE International
Conference on Fuzzy Systems, Budapest, Hungary, vol. 6, pp. 25–29 (2004)
17. Abiyev, R.H.: Z-number based fuzzy inference system for dynamic plant control. advanced
in fuzzy systems (2016)
18. Abiyev, R.H., Akkaya, N., Gunsel, I.: Control of Omnidirectional robot using Z-number-
based fuzzy system. IEEE Trans. Syst. Man Cybern. Syst. 49(1), 238–252 (2019)
19. Abiyev, R.H., Gunsel, I., Akkaya, N.: Z-number based fuzzy system for control of
omnidirectional robot. In: ICSCCW 2019, Advances in Intelligent Systems and Computing,
vol. 1095. Springer, Cham (2019)
20. Abiyev, R.H., Uyar, K., Ilhan, U., Imanov, E., Abiyeva, E.: Estimation of food security risk
level using Z-number-based fuzzy system. J. Food Quality 2018, 1–9 (2018)
972 R. H. Abiyev
21. Abiyev, R.H.: Controller based on fuzzy wavelet neural network for control of technological
processes. In: Proceedings of IEEE International Conference on Computational Intelligence
for Measurement Systems and Applications, Sicily, Italy, pp. 215–219 (2005)
22. Abiyev, R.H.: Fuzzy wavelet neural network for prediction of electricity consumption.
AIEDAM Artif. Intell. Eng. Design Anal. Manuf. 23(2), 109–118 (2009)
23. Abiyev, R.H., Kaynak, O., Kayacan, E.: A type-2 fuzzy wavelet neural network for system
identification and control. J. Franklin Inst. Eng. Appl. Math. 350(7), 1658–1685 (2013)
24. Abiyev, R.H., Abizade, S.: Diagnosing Parkinson’s diseases using fuzzy neural system.
Comput. Math. Methods Med. 2016(3), 1–9 (2016)
Comprehensive Research of the Commodity
Market: Theoretical, Methodological
and Modern Approaches Using Neural
Networks with Fuzzy Logic
Mikayilova Rena Nuru Kizi(&)
Digital Economy and ICT, Azerbaijan State Economic University,

AZ1147 Baku, Azerbaijan
[email protected]
Abstract. In market conditions, entering a new foreign market is always an

urgent problem for all types of commodity market units, which puts forward
special requirements for the effectiveness of the economic management system,
and also requires solving marketing research tasks, differentiating and seg-
menting the commodity market. The article gives a general methodology for
marketing segmentation and positioning in the product market on the basis of
which you can really assess the condition of the product market, the potential of
the enterprise and the formation of its future activities. The issues of marketing
segmentation and forecasting of various units of the commodity market using
neural and neuro-fuzzy adaptive systems with the aim of exploring opportunities
to enter a new foreign market are considered.
Keywords: Product market marketing segmentation Consumer market

Agricultural market Industrial market Neural network systems Fuzzy neural
network Fuzzy logic Adaptive networks Anfis
1 Introduction
In modern conditions, the commodity market takes on such an economic form that no
developed economy of any country can do without [1]. The category of the commodity
market and its structure includes the key positions of economies: the commodity market
of producers of non-food products; product market for food producers; consumer market.
Under market conditions, the formation of a new, more efficient economic man-
agement system that is able to model the main components of the commodity market
on the basis of research and forecasting of the consumer demand being a factor shaping
the efficiency of market subsystems with the aim of entering new foreign markets is of
particular importance.
Entering a new foreign market has always remained an urgent problem. In this
regard, the urgent task is to study the issues of marketing research, differentiation and
segmentation of the product market, in order to determine the possibility of entering
foreign markets, taking into account the real state of the product market and the
potentials of enterprises.
974 M. R. N. Kizi
The use of segmentation and positioning methods in the domestic product market
contributes to the successful entry into a new foreign market [2, 3]. This is evidenced
by studies conducted by Kaplin and Karachev [1]. Marketing research of product
market segmentation and product positioning makes it possible to decide on access to
the foreign market. The phases of the methodology for selecting a foreign market and
modeling economic indicators of export are constructed as follows:
1. an assessment of the capacity of the commodity market, within which the organi-
zation of the sale of goods is supposed.
2. diagnostics of the structure of the target product market.
3. assessment of entry and exit barriers within product markets.
Under the capacity of the consumer market, as was noted, the volume of sales of a
certain product in this market of a country over a certain period of time is understood.
Assessment of the dynamics of the market capacity in general and of individual
types of goods is the main one for diagnosing their capabilities in the product market,
developing strategies and marketing policies. There are many options for calculating
the real market capacity. The simplest way is to determine it on the basis of the balance
method using statistical information about the proposal:
O ¼ F ðP; E; I Þ ð1Þ
where O is the real market capacity; P is the volume of production of this product in the
region; E is the export of goods outside the country; I is the import of goods into the
country.
In order to diagnose the market environment in the republic, Table 1 shows the
ratios of the above indicators (see Formula 1) from 2013 to 2017 in accordance with
individual units of the commodity market of the manufacturing industry.
World experience shows that firms should have monthly data on the capacity of the
consumer market and the proportion of firms on it, which makes it possible to quickly
respond to market shifts in the product market. However, there is no single approach to
such studies.
In this regard, the article gives a general methodology for marketing segmentation
and positioning in the product market on the basis of which it is possible to assess the
real state of the product market, the potential of the enterprise and the formation of its
further activities in order to reach acceptable segments of foreign product markets (see
Fig. 1).
The developed methodology consists of the following stages:
1. Collection, filtering and normalization of retrospective economic indicators affect-
ing the assessment of the capacity of the commodity market,
2. Segmentation by level of development of commodity market units using neural
networks,
3. Determination of the most profitable segments of the product market with the aim of
further forecasting their development,
Forecasting the further development of the most profitable segments of the com-
modity market using a neuro-fuzzy adaptive forecasting system.
Comprehensive Research of the Commodity Market 975
Table 1. Economic indicators of the commodity market of Azerbaijan Republic for 2013–2017
(in million AZN) [4].
Year Export Import Production Real capacity
(E) (I) volume (O)
(P)
Food products 2013 518.24 849.6 2286.4 2700.6
2014 518.56 785.84 2422 2756.1
2015 687.61 1001.55 2307.6 2593
2016 812.94 2022.49 2964.7 3676.2
2017 1022.38 2095.08 2999.8 3630.8
Total 3559.73 6754.56 12980.5 15356.7
Drinks and Tobacco 2013 33.36 320.16 231.7 580.2
2014 35.68 375.44 244.8 669.5
2015 38.72 396 224.9 569.6
2016 63.92 357 265.4 378.8
2017 63.24 394.06 302.6 497.2
Total 234.92 1842.66 1269.4 2704.3
Industrial goods 2013 212 1546.96 1441.8 3110.1
2014 160.72 1422.8 1463.5 3023.1
2015 197.78 2598.53 1375.3 3555.8
2016 465.29 3169.14 1582.2 3172.7
2017 549.78 2863.65 2280.8 3641.9
Total 1585.57 11601.08 8143.6 16503.6
Chemical products 2013 140.32 747.2 200.3 958.9
2014 184.88 700 247.2 891.1
2015 209.22 922.13 353.2 1001.3
2016 270.64 1522.35 374.2 1110.5
2017 303.79 1768 513.9 1375.2
Total 1108.85 5659.68 1688.8 5337
Machinery and 2013 119.44 3338.64 262.7 4272.5
equipment 2014 152.08 2763.84 360.9 3728.5
2015 97.79 4045.25 304.4 3773.3
2016 26.69 4769.01 300 3002
2017 32.81 5065.66 280.6 3200.4
Total 428.81 19982.4 1508.6 17976.7
The developed methodology consists of the following stages:

1. Collection, filtering and normalization of retrospective economic indicators affect-
ing the assessment of the capacity of the commodity market,
2. Segmentation by level of development of commodity market units using neural
networks,
3. Determination of the most profitable segments of the product market with the aim of
further forecasting their development,
976 M. R. N. Kizi
Fig. 1. Stages and methods used for marketing segmentation and positioning in the product
market.
2 Program-Theoretical Implementation of Segmentation

and Positioning Methods in the Commodity Market
Let’s consider each of these stages separately:
2.1 Collection, Filtering and Normalization of Retrospective Economic

Indicators Affecting the Assessment of the Capacity
of the Commodity Market
In order to normalize retrospective economic indicators, the economic indicators shown
in Table 1 were used. Here, the following economic indicators were used as input
values, respectively: x1 – export, x2 – import, x3 – production volume, y1 – real capacity
of commodity market. Normalization of these indicators is carried out using the
Formula 2:
xik xmini
~xik ¼ ð2Þ
xmaxi xmini
where xik - i-th input value of the k-th example of the initial sample in traditional units
of measurement adopted in the problem being solved.
2.2 Segmentation by Level of Development of Commodity Market Units

Using Neural Networks
The segmentation problem in terms of cluster analysis is formulated as a set of T
commodity market objects (food products, drinks and tobacco, industrial goods,
chemical products, machinery and equipment), which are characterized by certain
levels of development of certain values of the same set of classification features
(quantitative and quality indicators).
It is required to break the set of objects into disjoint subsets that are similar to each
other in a number of ways. In this regard, the task of marketing segmentation of the
commodity market is solved using neural networks (or ANN) [5, 6].
The main essence of a neural network is to estimate the difference - d (distance) of
the weight vector w of each neuron in the network from the input template vector xp
and search for a neuron in which the weight coefficient (w) has a minimum distance
from x. For each j-th neuron and i-th factor of the p-th input data template, the
minimum distance is usually calculated by the Formula 3:
X
m 2
dj ¼ wij xpi ; ð3Þ
i¼1
where j = 1,…,m, m is the total number of neurons.

The neuron with the smallest difference d is considered the winner and further
adjustments to its weights are carried out according to the following Formula 4:
j
j
wnew ¼ wold ð1 aÞ þ axp ; ð4Þ
where a: [0..1] is the learning coefficient.

Software implementation of marketing segmentation of the commodity market was
carried out using a neural network with a Kohonen layer based on the MATLAB
system application package Neural Network Toolbox [7, 8].
For input data x ¼ x1 ; x2 ; . . .; x20 (20 is the number of data templates) (see
Table 1), 3 clustering centers were selected for all 5 product markets. Kohonen coef-
ficient was taken 0.9999. With a learning speed of 0.01, the maximum number of
learning epochs was found to be 9, after which the number of objects assigned to each
centroid did not change. The training lasted for 19.6166 s. The total distance of objects
from centroids was 2.4870. The results of the neural network on the distribution of
objects according to the corresponding centroids - clusters – is the following: 1st
segment - machinery and equipment, 2nd segment - drinks and tobacco and chemical
products, 3rd segment - food products and industrial goods.
In Fig. 2 in the form of dynamic graphs, these are displayed, respectively: on the
first graph (a), all objects and corresponding centers according to the selected criteria
(for 2017, x5 is export, x10 is import and x15 is the production volume, respectively),
the second graph (b) shows the values of individual criteria for different centroids, the
third graph (c) shows the cluster weights (the number of objects assigned to each
centroid), the fourth graph (d) shows the dynamics of the change in the total distance of
individual objects from the corresponding centroids.
Fig. 2. Neural network segmentation of product markets.

978 M. R. N. Kizi
2.3 Identification of the Most Profitable Segments of the Product Market

with the Aim of Further Forecasting Their Development
The results of the 2nd stage for marketing segmentation and positioning in the product
market show that segment 1 has large imports, low exports and low production
volumes.
The 2nd segment has low indicators for all criteria. The most attractive segment is
the 3rd segment, where all indicators are at an average level. And if you predict further
dynamics, you can determine how appropriate is for enterprises of food product and
industrial goods markets to develop their production in order to enter the foreign
market.
2.4 Predicting the Further Development of the Most Profitable Segments

of the Commodity Market Using a Neuro-Fuzzy Adaptive
Forecasting System
A fuzzy neural network was chosen as a method for predicting the commodity market of
the manufacturing industry [5, 6, 11, 12]. Studies have shown that fuzzy logic is the most
popular method that combines the human style of thinking of fuzzy systems [5, 9, 10]
through the use of fuzzy sets and a linguistic model consisting of a set of fuzzy IF-THEN
rules. On the other hand, having a large amount of information in hand, one should
analyze their retrospectiveness and further predict their dynamics of change based on
learning algorithms based on neural networks. Using Matlab software and mathematical
tools [10] in particular anfis, a model of a fuzzy-neural forecasting system was developed.
As input parameters, we used the total values of economic indicators for 2003–2017, such
as I, P, O, E (see Table 2).
Table 2. The total values of economic indicators for 2003–2017.

x1- Import x2– Production x3–Real capacity y1 - Export
(I) volume (O) (E)
(P)
Food products 6754.56 12980.5 15356.7 3559.73
Drinks and Tobacco 1842.66 1269.4 2704.3 234.92
Industrial goods 11601.08 8143.6 16503.6 1585.57
Chemical products 5659.68 1688.8 5337 1108.85
Machinery and 19982.4 1508.6 17976.7 428.81
equipment
As the output parameter of the model, export indicators of units of the commodity
market of the manufacturing industry were selected. Using anfis, the structure of which
is shown in Fig. 3 (Anfis model Structure), 27 rules of inference were formulated (see
Fig. 3). Sugeno was chosen as the method of implication. For each input parameter,
three linguistic terms with a Gaussian membership function were defined. Below is a
view of the response surface of the input and output indicators (see Fig. 4).
Fig. 3. Anfis Model Structure and 27 rules of inference.
Fig. 4. A view of the response surface of the input and output indicators.
3 Conclusion
Our model of an adaptive system for forecasting the dynamics of growth in the
commodity market of the manufacturing industry proves that with an increase in
production turnover and a decrease in imports, it is possible to increase the export of
the commodity market.
This method allows, as a toolbox, to calculate and predict the necessary ratio of
economic indicators in order to determine the export potential of the commodity
market.
The studies and the results obtained confirm the correctness and feasibility of using
neural and neuro-fuzzy systems to solve the problems of segmentation and forecasting
the performance of the commodity market.
But in continuation of the problem under study, it is necessary to conduct a
comprehensive study based on the proposed method in order to identify and determine
the degree of influence of other important factors on the process of increasing the
export potential of the commodity market.
980 M. R. N. Kizi
References
1. Kaplina, O.V., Karachev, I.A.: Ekonomika i upravleniye narodnim khozyastvom
[Bывeдeниe тoвapa нa нoвый зapyбeжный pынoк]. Vestnik Finanovoqo Universiteta 1,
33–46 (2016)
2. Abryutina, M.S.: Ekonomicheskiy analiz tovarnogo rynka i torgovoy deyatel’nosti
[Экoнoмичecкий aнaлиз тoвapнoгo pынкa и тopгoвoй дeятeльнocти]. Delo i Servis,
Moscow (2010)
3. Kotler, F.: Osnovi Marketinqa [Ocнoвы мapкeтингa]. Vilyams, Moscow (2015)
4. Sənayesi, A.: Statistik məcmuə. AR DSK, Bakı (2018)
5. Aliev, R.A., Fazlollahi, B., Aliev, R.R.: Soft computing and its applications in business and
economics. Springer-Verlag, Berlin (2004)
6. Jang, J.S.R., Sun, C.T., Mizutani, E.: Neuro-Fuzzy and Soft Computing, pp. 335–368.
Prentice Hall, Upper Saddle River (1997). ISBN 0-13-261066-3
8. Dyakonov, V.: Matematicheskiye pakety rasshireniya MATLAB [Maтeмaтичecкиe пaкeты
pacшиpeния MATLAB. Cпeциaльный cпpaвoчник], SPb, Piter (2001)
9. Zadeh, L.A.: Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets Syst. 1, 3–28 (1978)
10. Dubois D., Prade H.: Fuzzy Sets and Systems. Academic Press, NewYork (1980)
11. Alekperov, R.B., Ibrahimova, K.A.: Neural network modeling and estimation of the
effectiveness of the financing policy impact on the socio-economic development of the
socio-educational system. In: 13th International Conference on Theory and Application of
Fuzzy Systems and Soft Computing, ICAFS-2018, Advances in Intelligent Systems and
12. Balashirin, A.R., Tarlan, I.I.: Application of neural networks for segmentation of catering
services market within the overall system of consumer market on the model of restaurant
business with the aim to advance the efficiency of marketing policy. In: Aliev, R.A.,
Kacprzyk, J., Pedrycz, W., Jamshidi, M., Sadikoglu, F.M. (eds.) ICAFS 2018. AISC, vol.
896, pp. 905–913. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-04164-9_120
The Maximum Lexicographic Contraflow
Finding in a Fuzzy Dynamic Network
Evgeniya Gerasimenko(&) and Vladimir Kureichik
Southern Federal University, Taganrog, Russia

{egerasimenko,vkur}@sfedu.ru
Abstract. Currently, the tasks of emergency evacuation based on the flow

theory have been studying extensively due to the increasing number of emer-
gency situations, natural disasters, fires, terrorist attacks and other phenomena.
However, the main issue of the evacuation simulation is the necessity to tackle
dynamic environmental factors, such as road capacities and evacuation time
parameters, as well as the fuzzy nature of the evacuation. Traffic delays have
become ubiquitous during evacuation, whereas evacuation and rescue flows
move in opposite directions. To address the aforementioned issues a method for
the partial contraflow evacuation is examined to increase the total outbound
capacity and reverse the arcs of the modelled network up to the required
capacity. The main contribution to the paper is the proposal for the fuzzy
lexicographic maximum dynamic evacuation according to the partial lane
reversal strategy using multiple attribute group decision making method based
on TOPSIS in intuitionistic fuzzy setting for ordering the terminals during
evacuation. The main goal of the proposed algorithm is to transport the maxi-
mum number of aggrieved from origins to destinations that have different levels
of priority due to the level of reachability, capacity of destination nodes, reli-
ability (security) and total expenses. A case study is conducted to illustrate the
proposed algorithm.
Keywords: Lexicographic flow Contraflow design Fuzzy network
1 Introduction
Man-maned and natural disasters became ubiquitous and threaten the life, health and
behavior of huge groups of people. The safety pattern of evacuation is the challenging
issue in the modern evacuation modelling. Such a way of transportation assumes
transferring the aggrieved to the reliable shelters. However, the concept of reliability or
safety in the scope of evacuation is a fuzzy concept that should be evaluated by
different linguistic criteria with a high degree of uncertainty. To address the issues of
uncertainty in the instantaneous dynamic evacuation networks the tools of fuzzy logic
are successfully used. Fuzzy logic and its modern extensions [1] are powerful tools for
considering vague nature of the parameters of the modelled network. Despite the
extensive research in the fuzzy flows [2, 3], a comprehensive approach to the fuzzy
evacuation modelling is still lacking [4].
982 E. Gerasimenko and V. Kureichik
The key concept of the evacuation process is the examined transportation network.
There are various evacuation strategies based on network flow algorithms [2, 3].
Authors H. W. Hamacher, S. A. Tjandra provide a state-of-the-art study of the
macroscopic and microscopic evacuation flow models and relevant network flow
algorithms [5].
The comprehensive survey on the flow tasks based on the concept of contraflow
was conducted in [6]. Lane reversal technique is an advanced method of increasing the
transported flow value by enhancing the outbound capacity of the network. When the
traffic is reversed, it enables releasing the loaded road segments and redistributing the
movement towards the empty roads eliminating traffic congestions. The concept of
lane-based evacuation routing was proposed in [7]. Dhamala and Pyakurel [8] incor-
porated contraflow techniques into conventional flow models to increase the total
outbound capacity and introduce the concept of partial contraflow to reverse the arcs of
the modelled network up to the required capacity.
Methods of the multiple group decision making are of the crucial importance in the
process of complex decision making. Single experts are not able to comprehend each
feature in the simulated problem. In such conditions it is necessary to apply expertise
and skills of the qualified group of experts. However, the problem of determining the
weights of experts raises. Therefore, in the complex tasks of decision making experts
should have an opportunity to estimate various attribute features not evenly but
assigning the different weight coefficient to each attribute. If the decision maker
endeavors to incorporate all weighted assessments of decision variables into the
solution, algorithm tends to become time-consuming. Consequently, methods of
ranking the alternatives based on the technique for order of preference by similarity to
ideal solution (TOPSIS) [9] allows overcoming time-consuming calculations and
simultaneously introduce weights of experts on different attributes. Incorporating
multiple decision making in the evacuation modelling in intuitionistic fuzzy conditions
[1] leads to the correct choice of the alternatives and assumes rather high level of
hesitation and uncertainty than fuzzy numbers provide.
The main contribution to this study is proposal for the fuzzy maximum lexico-
graphic dynamic flow algorithm according to the partial lane reversal strategy using
multiple attribute group decision making method based on TOPSIS in intuitionistic
fuzzy setting to order the terminals during evacuation.
The rest of the paper is organized as follows. In Sect. 2, we observe preliminaries
and problem statement of the maximum lexicographic dynamic evacuation problem
with the partial lane reversal in fuzzy conditions based on TOPSIS. Proposed algorithm
underlying this paper is given in the Sect. 3. Section 4 observes a case study of the
proposed method. Finally, Sect. 5 concludes the paper.
2 Preliminaries and Problem Statement
Definition. Let X be a nonempty set. An intuitionistic fuzzy set A in X is defined as:

A ¼ hx; lA ð xÞ; mA ð xÞ : x 2 X i, where the functions lA ð xÞ; mA ð xÞ : X ! ½0; 1 are the
degrees of membership and nonmembership of the element x 2 X, such as for each
The Maximum Lexicographic Contraflow Finding in a Fuzzy Dynamic Network 983
x 2 X: 0 lA ð xÞ þ mA ð xÞ 1: Furthermore, we have pA ð xÞ ¼ 1 lA ð xÞ mA ð xÞ called

the intuitionistic fuzzy set index or hesitation margin of x in N. It represents the degree
of indeterminacy or hesitation of x 2 A. For each x 2 X : 0 pA ð xÞ 1. The pair
a ¼ ðlA ð xÞ; mA ð xÞÞ is called an intuitionistic fuzzy value (IFV). The main operations
with intuitionistic fuzzy values used in the algorithm, are performed in [9].
Network flow model (1)–(5) under consideration for the lexicographic maximum

dynamic evacuation modelling is a multiterminal fuzzy network G ~ ¼ X; A ~ , where X
is a crisp set of nodes, A ~ – a fuzzy set of arcs with the ranked set of sources
s1 s2 . . .sm s and sinks d1 d2 . . .dm d; where s1 ; d1 have the highest priority
and sm ; dm – the lowest one. Each arc has a time-depended assigned fuzzy arc capacity
~uij ðhÞ and traversal time sji ðh). The set of time horizon is denoted by T = {0,1,…,T}.
The fuzzy flow ~nij ðhÞ conveying along the arc (i, j) should satisfy flow conservation-
constraints (2–4) and capacity constraint (5). The objective is to find the maximum
dynamic lexicographic flow ~nid ðhÞ from the sources to the sinks within a given time
horizon and priority order (1). Sources and sinks are categorized due to the experts’
assessments via multiple attribute intuitionistic fuzzy group decision making algorithm
based on TOPSIS.
X
T X
Maximize ~nid ðhÞ; 8T ¼ 0; . . .; T; ð1Þ
h¼0 xi 2D
Xp X X
Subject to : ~n ðhÞ ~
n h s ji ð h Þ ¼ 0; xi 6¼
h¼0 x 2Cðx Þ ij
j i x 2Cðx Þ ij
j i ð2Þ
s; d; h 2 T;
Xp X X XT X
~n ðhÞ ~nji h sij ðhÞ ¼ ~
n ðhÞ; xi ¼ s;
h¼0 x 2Cðx Þ ij
j i xj 2C 1
ðx Þ
i h¼0 i2R ir
ð3Þ
Xp X X
~nij ðhÞ ~
nij h sij ðhÞ ¼
h¼0 xj 2Cðxi Þ xj 2C1 ðxi Þ
XT X ð4Þ
~
n ij ð h Þ ; xi ¼ d;
h¼0 i2R

0 ~nij ðhÞ ~uij ðhÞ; 8 xi ; xj 2 A;
~ h 2 T; h ¼ 0; . . .; T sij ðhÞ: ð5Þ
3 Lexicographic Maximum Dynamic Contraflow Evacuation

in a Fuzzy Network Based on TOPSIS
In this section, we propose an algorithm for the maximum lexicographic dynamic

evacuation modelling in a fuzzy network with partial lane reversal. Modified TOPSIS
method in intuitionistic fuzzy setting with all interim calculations was prosed in [9].
Step 1. Determine the priority order of terminals according to the TOPSIS method. (See
[9] for detailed description).
1:1. Evaluate the alternatives according to the attributes. Construct decision matrices
Dk ¼ ðakij Þmn .
1:2. Compose the positive ideal decision matrix D þ ¼ ðaijþ Þmn and the negative
ideal decision matrices Dd ¼ ðadij Þmn and Du ¼ ðauij Þmn .

1:3. Compose the collective decision matrix D ¼ aij mn according to the values of
closeness coefficients applying intuitionistic fuzzy weighted averaging operator.
0
1:4. Determine the weighted decision matrix D0 ¼ ðaij Þmn .
1:5. Calculate the intuitionistic fuzzy positive ideal evaluation value A+ and negative
ideal evaluation value A−. Based on the determined values, calculate the distance
þ
d1::n and d1::n of each alternative’s collective evaluation value to A þ and A .
1:6. Calculate each alternative’s closeness coefficient ci .
1:7. Rank the alternatives’ closeness coefficients [9].
Step 2. Constructing the time-expanded modified network for the partial contraflow
reversal evacuation [9].
The time-expanded modified network for the lane reversal is such a fuzzy network that
replaces initial fuzzy arc capacities ~uði; j; h; #Þ and ~uðj; i; h; #Þ by the new modified arc
capacity ~ul ði; j; h; #Þ = ~uði; j; h; #Þ þ ~uðj; i; h; #Þ with the initial time parameter
sl ði; j; h; #Þ ¼ sði; j; h; #Þ. ~uði; j; h; #Þ stands for the capacity of the arc connecting
node-time pair ði; hÞ with ðj; #Þ, where the node-time pair ði; hÞ is a node i at the time
period h in the dynamic network.
Step 3. Finding augmenting trails in the residual network G ~ lr

e
Find the augmenting trails that seeks for the augmenting shortest path ~ ple from the
0 0
artificial source s to the artificial sink d according to the priorities in the residual
network G ~ lr . If ~nl ði; j; h; #Þ\~ul ði; j; h; #Þ in G~ lr , then ~ulr ði; j; h; #Þ ¼ ~
ul ði; j; h; #Þ
e e
~nl ði; j; h; #Þ. If ~nl ði; j; h; #Þ [ ~0, then ~ulr ðj; i; #; hÞ ¼ ~ nl ði; j; h; #Þ.
Step 4. Partial reversal of the arcs

4:1 For each time period h 2 T reverse the arc connecting ðj; #Þ and ði; hÞ up to the
capacity ~nl ði; j; h; #Þ ~uði; j; h; #Þ if ~nl ði; j; h; #Þ [ ~ uði; j; h; #Þ, replace
~uði; j; h; #Þ by ~0 if the arc ði; hÞ; ðj; #Þ 62 A:
e
4:2 For each time period h 2 T and each arc ði; hÞ; ðj; #Þ 2 A ~ if the arc ði; hÞ,ðj; #Þ is
reversed, the saved arc capacity ~sði; j; h; #Þ ¼ ~ u ði; j; h; #Þ ~
lr
nðj; i; #; hÞ and
~ ~
~sðj; i; #; hÞ ¼ 0. ~sði; j; h; #Þ ¼ ~uði; j; h; #Þ nði; j; h; #Þ if both node-time pairs
ði; hÞ, ðj; #Þ and ðj; #Þ,ði; hÞ are not reversed.
4 Case Study
In this section, we model the lexicographic maximum partial contraflow evacuation

scenario in a dynamic fuzzy network (see Fig. 1). The problem is in conveying the
maximum amount of aggrieved from the dangerous areas s1 ; s2 to the shelters d1 ; d2 ; d3
within the time horizon T = 4. Three decision makers Ci (i = 1,2,…,3) are selected to
evaluate the evacuation priority of terminals d1 ; d2 ; d3 in the form of intuitionistic fuzzy
values towards four attributes: the level of reachability (B1 ), capacity of destination
nodes (B2 ), reliability (security) (B3 ), and total expenses (B4 ), which is shown in
Table 1. The attribute weight vector is W = (0,23, 0.3, 0,22, 0.25).
Following the steps 1.1–1.7 of the algorithm calculate intuitionistic fuzzy negative
ideal (Tables 2, 3) and positive ideal (Table 4) decision matrices. Intuitionistic fuzzy
collective and weighted decision matrices are performed in Tables 5, 6. According to
the step 1.5. the distances of alternatives’ evaluation values to the values A þ and A
are d1þ ¼ 3.112, d2þ ¼ 3.011, d3þ ¼ 2.954,d1 ¼ 0.888, d2 ¼ 0.99, d3 ¼ 1.046. The
relative closeness coefficients: c1 ¼ 0.222, c2 ¼ 0:247, c3 ¼ 0:62 (step 1.6).
Fig. 1. Real evacuation network. Fig. 2. Graph image of the real network.
Table 1. Intuitionistic fuzzy decision matrix of the DMs.

B1 B2 B3 B4
C1
d1 (0.6, 0.2) (0.2, 0.6) (0.4, 0.3) (0.7, 0.1)
d2 (0.5, 0.4) (0.3, 0.5) (0.6, 0.3) (0.8, 0.1)
d3 (0.5, 0.3) (0.7, 0.2) (0.4, 0.3) (0.7, 0.2)
C2
d1 (0.7, 0.2) (0.4, 0.5) (0.5, 0.3) (0.8, 0.2)
d2 (0.6, 0.3) (0.2, 0.5) (0.7, 0.2) (0.9, 0.1)
d (0.5, 0.2) (0.6, 0.1) (0.4, 0.2) (0.7, 0.3)
C3
d1 (0.7, 0.3) (0.3, 0.4) (0.5, 0.4) (0.7, 0.2)
d2 (0.5, 0.3) (0.4, 0.4) (0.4, 0.2) (0.7, 0.2)
d3 (0.4, 0.3) (0.6, 0.3) (0.5, 0.2) (0.8, 0.2)
Table 2. Intuitionistic fuzzy negative ideal Table 3. Intuitionistic fuzzy negative

decision matrix Du : ideal decision matrix Dd :
B1 B2 B3 B4 B1 B2 B3 B4
d1 (0.7, 0.2) (0.4, 0.5) (0.5, 0.3) (0.8, 0.2) d1 (0.6, 0.2) (0.2, 0.6) (0.4, 0.3) (0.7, 0.2)
d2 (0.6, 0.3) (0.4, 0.4) (0.7, 0.2) (0.9, 0.1) d2 (0.5, 0.4) (0.2, 0.5) (0.4, 0.2) (0.7, 0.2)
d3 (0.4, 0.3) (0.7, 0.2) (0.5, 0.2) (0.8, 0.2) d3 (0.4, 0.3) (0.6, 0.3) (0.4, 0.3) (0.7, 0.3)
Table 4. Intuitionistic fuzzy positive ideal decision matrix D þ :

B1 B2 B3 B4
d1 (0.666, 0.232) (0.302, 0.497) (0.465, 0.334) (0.734, 0.162)
d2 (0.532, 0.334) (0.302, 0.468) (0.58, 0.232) (0.815, 0.129)
d3 (0.465, 0.266) (0.633, 0.185) (0.432, 0.232) (0.734, 0.232)
Table 5. Intuitionistic fuzzy collective Table 6. Intuitionistic fuzzy weighted deci-

decision matrix D. sion matrix D0 .
B1 B2 B3 B4 B1 B2 B3 B4
d1 (0.675, (0.312, (0.474, (0.74, d1 (0.228, (0.106, (0.132, (0.286,
0.243) 0.554) 0.338) 0.156) 0.722) 0.838) 0.788) 0.628)
d2 (0.534, (0.302, (0.648, (0.854, d2 (0.161, (0.102, (0.205, (0.382,
0.329) 0.469) 0.251) 0.1) 0.774) 0.797) 0.738) 0.562)
d3 (0.5, (0.639, (0.437, (0.736, d3 (0.155, (0.263, (0.119, (0.283,
0.243) 0.176) 0.226) 0.228) 0.722) 0.594) 0.721) 0.691)
Rank alternatives’ relative coefficients to get the priority d3 d2 d1 : The process

of conversion of the real evacuation network to the graph depiction is performed in
Fig. 2. Initial values of time-dependent arc capacities and traversal time parameters are
performed in Table 7.
Time-expanded modified network (Fig. 3) for the partial contraflow reversal is
constructed based on Fig. 2 and the step 2 of the algorithm by summing the capacities
of opposite arcs.
Following the step 2 of the algorithm, search the augmenting trails in the residual
network in Fig. 3 and submit the maximum possible flow value to the sinks with the
calculated priority d3 d2 d1 , as shown in Fig. 4. Fuzzy maximum dynamic flow
value in the given dynamic evacuation network with partial lane reversal method is f 65
units.
Fig. 3. The time-expanded network.
Table 7. Transit arc capacities ~uij ðhÞ at time periods h.

T Arc capacities, traversal times
(s1,1) (s2,2) (2,s2) (1,2) (2,1) (1,3) (3,1)
(2,d1) (d1,2) (2,d2) (3,d2) (d2,3) (3,d3) (d3,3)
20; 1 ~
0 f 4; 2 ~ 5; 2 f 11; 2 f12; 2 f
10; 2 ~
8, 2~
7; 2 ~8; 2 f 14; 1 ~
9, 3 ~
8, 3 f
11; 2 f10; 2
f 1 5;
1 22; ~ 2 4, ~ 2 13; f 2 12;f 2 e 9; 2 f 2 ~
10; 9, 1 f 1 14;
10; f 1 ~
9; 3 ~
8, 3 f f 1
13; 1 12,
2 f
20; 2 ~
6; 1 ~
5; 1 f
10; 1 f
15; 1 f 3 10; 3 f
11; f 10; 1 f 15; 1 ~
12; 1 f 9; 3 ~ 3
8, f
13; 1 f12, 1
3 f
19; 1 ~
5; 2 ~
5; 2 f
14; 1 10; 1 f
f 2 f
12; 10; 2 f
10; 1 f
12, 1 15; 1 f
f 10; 3 f 11; 1 f
11, 3 f 10; 1
f 1
4 20; ~
6, 1 ~
7; 1 15; 1 ~
f f 2 f
9; 1 13; 10; 2 f
10; 3 f
12; 3 f f 2 f
15; 2 10; f 2 10,
11; 2 11; f 2
Calculate the values of saved dynamic arc capacities according

the partial lane
reversal. The dynamic arcs that are reversed completely: 22 ! s02 ; ð33 ! 11 Þ,
ð23 ! s22 Þ, ðd14 ! 23 Þ: Partially reversed arcs: ð23 ! 12 Þ up to the capacity f 12,
4
d3 ! 33 up to the capacity ~6. 3. Saved arc capacity of ~
1 units: s01 ! 11 , s12 ! 23 ;
of ~6 units: ð22 ! d 3 Þ; of ~4 units: ð23 ! s1 Þ.
2 2
Fig. 4. The time-expanded network with the flow distribution.
5 Conclusion and Future Study
The paper illustrates the approach to the lexicographic maximum dynamic partial
contraflow evacuation in the fuzzy dynamic network with the priority list of terminal
nodes via the technique based on order of preference by similarity to ideal solution.
Elaborate scenarios of the real world evacuation that have been complicating every
year require expertise of various experts so that the final solution would be reliable and
flexible to changes in environment. Proposed method is based on implementing the
scheme of the multiple attribute group decision making in fuzzy setting during the
evacuation modelling. In particular, the extra level of the experts’ hesitation can be
introduced in the model based on the intuitionistic fuzzy setting that allows defining the
hesitation margin. Hazardous and dynamic pattern of evacuation assumes considering
time-varying nature of the network parameters and lexicographic order of the terminals:
destinations are ranked by experts during the evacuation process. In the proposed
method, evaluation of terminal nodes is conducted by the multiple attribute intuition-
istic fuzzy group decision making algorithm based on TOPSIS. Evacuation process
utilizes the concept of partial contraflow to reverse the arcs of the modelled network up
to the required capacity for smooth traffic management and decreasing the level of
congestions. A case study of the maximum lexicographic dynamic evacuation in fuzzy
conditions with partial lane reversal is conducted. The simulation of the TOPSIS
method in intuitionistic fuzzy setting is provided to rank the terminals. Microscopic
models of the evacuation in fuzzy conditions will be proposed as a part of the future
research.
Acknowledgments. This work has been supported by the Russian Foundation for Basic
Research, Project № 19-01-00559 a.
References
1. Mohanty, R.K., Tripathy, B.K.: Intuitionistic hesitant fuzzy soft set and its application in
decision making. In: Advances in Intelligent Systems and Computing, vol. 517. Springer,
Singapore (2017)
2. Bozhenyuk, A., Gerasimenko, E., Kacprzyk, J., Rozenberg, I.: Flow tasks solving in dynamic
networks with fuzzy lower, upper flow bounds and transmission costs. In: Studies in
Fuzziness and Soft Computing, vol. 346, pp. 77–159. Springer, Heidelberg (2017)
3. Bozhenyuk, A., Belyakov, S., Gerasimenko, E., Savelyeva, M.: Fuzzy optimal allocation of
service centers for sustainable transportation networks service. In: Kahraman, C., Sari, İ.
(Eds.) Intelligence Systems in Environmental Management: Theory and Applications.
Intelligent Systems Reference Library, vol. 113. Springer, Cham (2017)
4. Gerasimenko, E., Rozenberg, I.: Earliest arrival dynamic flow model for emergency
evacuation in fuzzy conditions. In: IOP Conference Series: Materials Science and
Engineering, vol. 734, pp. 1–6 (2020)
5. Hamacher, H.W, Tjandra, S.A.: Mathematical modeling of evacuation problems: a state of the
art. In: Pedestrain and Evacuation Dynamics, pp. 227–266 (2002)
6. Rebennack, S., Arulselvan, A., Elefteriadou, L., Pardalos, P.M.: Complexity analysis of
maximum flow problem with arc reversal. J. Comb. Optim. 29, 200–216 (2010)
7. Cova, T., Johnson, J.P.: A network flow model for lane-based evacuation routing.
Transp. Res. Part Policy Pract. 37, 579–604 (2003)
8. Pyakurel, U., Nath, H.N., Dempe, S., Dhamala, T.N.: Efficient dynamic flow algorithms for
evacuation planning problems with partial lane reversal. Mathematics 7(10), 1–29 (2019)
9. Yang, W., Chen, Z., Zhang, F.: New group decision making method in intuitionistic fuzzy
setting based on TOPSIS. Technol. Econ. Dev. Econ. 23(3), 441–461 (2017)
An Enhanced Particle Swarm Optimization
with Levy Flight for RBF Neural Network
in Typical Karst Area, South China
Zhangjun Cao1, Dong Wang1(&), and Lachun Wang2

1
Key Laboratory of Surficial Geochemistry, Ministry of Education,
Department of Hydrosciences, School of Earth Sciences and Engineering,
State Key Laboratory of Pollution Control and Resource Reuse,
Nanjing University, Nanjing, People’s Republic of China
[email protected]
2
School of Geographic and Oceanographic Sciences,
Nanjing University, Nanjing, China
Abstract. This paper applied an enhanced particle swarm optimization

(PSO) technique with levy flight algorithm for training the radial basis function
(RBF) neural network to forecast the data of runoff in the Houzhaihe River
basin, a typical karst area in Guizhou Province, southwest China. The karst
aquifer system is a highly nonlinear and complex system due to its unique
aqueous medium, the complexity of its hydrogeological conditions makes the
traditional hydrological model research results unsatisfactory, and the estab-
lishment of a physical distributed model based on hydrological mechanism
requires a large number of hydrogeological parameters, which are often
unavailable in karst areas. Radial Basis Function (RBF) neural network has been
widely used in various fields because of its simple structure, high-speed cal-
culation and ability to approximate any nonlinear function. Based on the RBF
neural network, this paper established a time series prediction model for typical
karst regions. In order to improve the performance of RBF network model, we
applied an enhanced particle swarm optimization with levy flight in this
research. The results show the proposed enhanced RBF model performs much
better than the one without improvement by the levy flight.
Keywords: Karst area Radial basis function (RBF) neural network Particle
swarm optimization Levy flight
1 Introduction
There are two types of models for runoff forecasting: physically based and data-driven
[1]. Physically based models aim to reproduce hydrological processes and the basin
behavior of the hydrological processes is expressed by mathematical equations.
However, the study area is a typical Karst region and the physically based approach
needs detailed information about the study area, which, on account of complex geo-
logic condition, is typically not available [2]. In the karst area, runoff generation differs
vastly depending on preferential flow through fissures and fractures and the karst
An Enhanced Particle Swarm Optimization with Levy Flight 991
aquifer is highly anisotropic. Thus, hydrological models, based on physical informa-

tion, homogeneous media and methods of wave theory, do not perform well enough in
karst areas as in non-karst areas [3].
Artificial neural networks (ANNs) is branch of artificial intelligence. ANN is
basically data driven and treated as black box models. These models are capable of
accommodating the non-linear relationship between rainfall and runoff [4]. Radial basis
function (RBF) neural networks, which have received considerable applications in
plenty of areas, such as approximating functions and, time series prediction and
recognition, is a widely used ANN model. The RBF neural network is preferred to the
other ANNs in this research because of its faster convergence speed and more precise
approximation with only one hidden layer [5]. The particle swarm optimization
(PSO) algorithm is kind of evolutionary techniques which has been adopted to improve
the performance of conventional neural networks [6]. The PSO algorithm has enhanced
capabilities of solving complex problems, high convergence speed, and good gener-
alization capabilities for a wide variety of problems. However, the standard PSO neural
network (SPSONN) could be easily to become trapped in local minima and hence may
fail to converge to the global optimum [7]. In order to overcome the deficiency and
improve the performance of PSO algorithm, Levy flight algorithm had been introduced
which is a class of random walk, generalized Brownian motion to include non-
Gaussian randomly distributed step sizes for the distance moved [8]. Levy flight, with
small steps, enables the optimization algorithm to search around the current solutions
and meanwhile, large jumps, which are occasionally produced by Levy flight, could
also do help to prevent the optimization algorithm from getting stuck in the local
minimum. Many natural and artificial facts can be depicted by Levy flight, such as fluid
dynamics, earthquake analysis, the diffusion of fluorescent molecules, cooling behavior
and noise etc. [9]. Researchers also use the Levy flight to improve the performance of
other evolutionary techniques such as Cockoo search algorithm [10], dragon fly
algorithm [11] and so on.
An enhanced PSO algorithm with Levy flight has been implemented to train the
RBF neural network and the improved Levy flight PSORBF have been employed to
predict runoff of the Houzhaihe River basin, GuiZhou Province, southwest China in
this paper. The results obtained from the Levy-PSORBF for runoff forecasting would
be compared with that from the standard PSORBF(SPSO-RBF) model. All the models
were developed using Matlab. The second part would introduce the methodology while
the third part indicated the optimization procedure, then the improved model was used
in typical karst area in China compared with the model without Levy flight. We draw
the conclusion and would have some discussion and suggestion in the end.
2 Methodology
2.1 Radical Basis Function Neural Network
The RBF neural network was applied to predict real time runoff. Compared to other neural
networks, RBF networks have faster training and more precise approximation with simple
feed-forward architecture, which has three layers including a single hidden layer of units.
992 Z. Cao et al.
The input layer consists of n units which represent the elements of the input vector
x. The activation of a neuron in the output layer is determined by computing the
weighted sum of outputs of hidden layer. RBFNN can be expressed in the form:
XM
yj ð x Þ ¼ i¼1
Oji zi ð xÞ; j ¼ 1; 2; . . .; L ð1Þ
Where Oji is synaptic weight between the ith neuron of the hidden layer and the jth
neuron of the output layer, M, L—respectively denotes the number of the hidden
neurons and the number of the output neurons, zi ð xÞ—the ith output of the hidden layer
of the RBFNN. As the hidden layer is composed of the hidden center vectors and the
kernel functions, which are generally Gaussian kernel functions, zi ð xÞ can be written as:
!
jjx ci jj2
zi ð xÞ ¼ K ðjjx ci jj; ri Þ ¼ exp ð2Þ
2r2i
Where || . . . || represents Euclidean norm; ci is i-th hidden center vector, ri is the

controlling parameter of the kernel function. The j-th output yj ð xÞ of the RBFNN can be
expressed as:
!
XM jjx ci jj2
yj ð xÞ ¼ Oji exp ð3Þ
i¼1 2r2i
The training parameters in the RBF network are much fewer than those in other
feed forward neural networks, which are centers (ci ), widths (r2i ), and weights (Oji )
between the hidden layer and the output layer.
2.2 Particle Swarm Optimization (PSO) Algorithm

The RBF model with gradient descent algorithm training the network may suffer from
convergence problems so that the RBF model spent much more time in the training while
came to less accurate. The standard particle swarm optimization (PSO) was applied
instead of gradient descent algorithm to improve the performance of the network.
In the standard PSO, each particle is attracted toward the position of the current
global best and its own best position in history right, while at the same time it has a
tendency to move randomly. For each particle with position and velocity vector, they
can be updated in iteration as:

Vik þ 1 ¼ x Vik þ c1 r1 Pki Xik þ c2 r2 Pkg Xik ð4Þ
Xik þ 1 ¼ Xik þ Vik þ 1 ð5Þ
Where Pki is the local best position of the particle i and Pkg is the global best position
ever found so far,r1 and r2 are two random vectors drawn from a uniform distribution;
parameters c1 and c2 are often called the learning parameters or acceleration constants;
parameters x is named inertia weight.
2.3 Levy Flight Algorithm

PSO has the problems of being trapped in the local optimum. Levy flights method is
used for solving these problems and enable PSO generate more efficient results. In this
method, it is ensured by Levy flight that PSO, which was unable to perform global
search well, would perform global search more effectively and reduce the possibility of
being trapped in local optimum by increasing the diversity of the particle. The new
position of the particle is calculated as:
Xik þ 1 ¼ Xik þ a LevyðbÞi ¼ 1; 2; . . .; n ð6Þ
Where a is the step size which should be related to the scales of the problem of
interest. In the proposed method a is random number for all dimensions of particles.
means entry-wise multiplications. LevyðkÞ is the generation of Levy step that can be
summarized as (Hakli and Uguz, 2014):
l
LevyðbÞ ¼ ð7Þ
jtj1=b
where l and l are drawn from normal distributions which is:

l N 0; r2l t N 0; r2t ð8Þ
and:
8 91=b
<Tð1 þ bÞsin pb2
=
rl ¼ h i ðb1Þ rt ¼ 1 ð9Þ
: T ð1 þ bÞ b2 2 ;
2
Where T is a standard Gamma function. Usually b is defined as 1.5.
3 RBF Model Improved by Levy-PSO Algorithm
Dealing with the drawbacks of RBFNN and increasing the performance of the pre-
diction model, a Levy flight PSO algorithm was adopted to optimize the parameters of
the RBFNN model. The optimization procedure can be described as follows:
1. Specify the lower and upper values of position and velocity. Set the value of
maximum iteration kmax.
2. Confirm the architecture of the RBF model due to the case you are working on. The
size of swarm is determined by the structure of the RBFNN model. Each particle
represents the ci , ri and Oji of the RBF model.
3. Initialize swarm of a population. Individuals are randomly generated and each of
them includes two parts: position and velocity (we can treat the position as the value
of the parameters).
994 Z. Cao et al.
4. Decode these particles into the RBF neural network architecture.

5. Calculate the fitness value for each particle of the population using the fitness function.
Then, initialize the individual best position Pi and the global best position Pg .
6. Compare the result of individual current fitness with its experienced best position, if
the current one is better, set current position as the best position. Then compare
individual current fitness with the fitness of the global best position, if current
fitness is better, set current position to be the global best position instead of the
former position.
7. Update the velocity and the position for the whole particle swarms by Levy flight
according to Eq. (4) and (6).
8. If the training error is reduced to an acceptable value or the iteration reaches the
maximum number, terminate the iteration and obtain the best ci , ri and Oji from the
global best solution. Otherwise, go back to step 5.
4 Application Results and Discussion
The study area is Houzhaihe River basin, a typical Karst region in PuDing in the mid-
west of Guizhou Province, southwest China. The area of the watershed is 80.65 km2.
The average rainfall per year is 1314.6 mm. Most of the rain falls between the months
of May and October, accounting for approximately 85% of the annual rainfall. The
average yearly rate of evaporation is 920 mm.
The study employs the monthly runoff data collected from Laoheitan station in the
basin. The measured monthly runoff data is from the January 1990 to December 2006.
The data from 1990 to 2005 were used to train the models in the study while the
monthly runoff data of 2006 was the validation set.
In this paper, the structure of RBF neural network was confirmed like this: one
input layer with two neurons whilst a hidden layer with 100 neurons and output layer
with one neuron. Generally, the more neurons made the higher accuracy, but the more
neurons also took much more time in training process and the more neurons made the
harder to converge, in the study the number of hidden layer neurons was determined as
100. The input neuron represented the runoff data of two continuous months whilst the
output node denoted the runoff data one month behind.
Mean absolute error (MAE), root mean square error (RMSE), mean absolute per-
centage error (MAPE) and the coefficient of efficiency (dc) were used to evaluate the
models’ performance.
All the monthly runoff data of networks were scaled in the interval [−1, 1] by the
equation:

x xmin
xnorm ¼ 2 1 ð10Þ
xmax xmin
When initializing the particle position, we set its gauge between (−1,1), the max-
imum velocity Vmax = 0.3, the minimum velocity Vmin = − 0.30, the population size
n = 20, and the maximum number of iterations K = 200, inertia weight = 0.5, c1 =
c2 = 2. For the Levy flight algorithm, we set the parameter a = 0.5, b = 1.5.
Table 1. Model prediction results. unit: cubic meters per second per month.
2006 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Real 2.55 2.81 3.42 5.99 5.33 11.68 11.27 8.53 9.22 20.45 7.97 5.07
SPSO 2.68 2.89 3.78 7.44 4.91 15.22 11.02 6.75 9.74 26.91 0.69 2.24
LPSO 3.85 3.10 3.27 5.06 5.39 9.43 10.98 9.44 9.15 16.46 11.64 6.51
Table 2. Model evaluation index.

Model MAE RMSE MAPE Dc
SPSO-RBF 2.09 3.18 0.24 0.56
LPSO-RBF 1.28 1.83 0.174 0.86
Fig. 1. Model prediction results. unit: cubic meters per second per month.
Table 1 and Fig. 1 show the measured data and predicted results of both SPSO-
RBF model and Levy-PSORBF model. Table 2 indicates the performance of the Levy-
PSO-RBF model and SPSO-RBF model from the perspective of evaluation index given
in this paper. Figure 2 and Fig. 3 are scatter plots of the predicted results of the two
models.
From Table 2, we can tell the MAE of the SPSO-RBF model is 2.09 meanwhile the
MAE of the Levy-PSO-RBF model is 1.28, and the performance of Levy-PSO-RBF
model improved by 38.8% compared with the SPSO-RBF model; the RMSE of the
SPSO-RBF model is 3.18 meanwhile that of the Levy-PSO-RBF model is 1.83,
996 Z. Cao et al.
the performance of Levy-PSO-RBF model improved by 42.5% compared with the

SPSO-RBF model; the MAPE of the SPSO-RBF model is 24% meanwhile that of the
Levy-PSO-RBF model is 17.4%, Levy-PSO-RBF model performed 27.5% better than
SPSO-RBF model; the Dc of the SPSO-RBF model is 0.56 meanwhile that of the
Levy-PSO-RBF model is 0.86, Levy-PSO-RBF model performed 53.6% better than
SPSO-RBF model. Figure 2 and Fig. 3 are scatter plots of the predicted results of the
two models. According to Fig. 2 and Fig. 3, it can be concluded that the slope of the
fitting curve by the Levy-PSO-RBF model is 0.76 while the slope of the fitting curve by
the SPSO-RBF model is 1.346 which meant the predicted results of Levy-PSO-RBF
more accurate. From all the perspective of evaluation index applied in the paper, we
can find out that the Levy-PSO-RBF model outperformed the SPSO-RBF model at
least 27.5%.
Fig. 2. Scatter diagram of SPSO-RBF model.
Fig. 3. Scatter diagram of Levy-PSO-RBF model.

5 Conclusion
In this study, a coupling intelligence algorithm has been applied to replace the physical
model to predict the data of runoff int typical karst area. PSO algorithm was used taking
the place of gradient descent algorithm to train the RBF neural network model. In order
to reduce the possibility of PSO being trapped in the local optimal solution and increase
the diversity of the population, Levy flight strategy was introduced to the PSO, and the
Levy-PSO-RBF model was finally created. The enhanced model was applied to predict
the runoff in typical karst area in China and compared with the SPSO-RBF model
without Levy flight. According to the results, we could draw the conclusion: The
introduction of Levy flight increased the accuracy of the PSO-RBF model, in other
word, the Levy-PSO-RBF model outperformed the SPSO-RBF model; the Levy-PSO-
RBF model provided an alternative option for application in karst areas.
In further research, we will focus on how to combine the artificial intelligence with
the physical model to make the new artificial intelligence models closer to the objective
physical process.
References
1. Nguyen, P.K.-T., Chua, L.H.-C., Son, L.H.: Flood forecasting in large rivers with data-
driven models. Nat. Hazards 71(1), 767–784 (2013)
2. Hu, C.H., Hao, Y.H., Yeh, T., Pang, B., Wu, Z.: Simulation of spring flows from a karst
aquifer with an artificial neural network. Hydrol. Process. 22(5), 596–604 (2008)
3. Fleury, P., Plagnes, V., Bakalowicz, M.: Modelling of the functioning of karst aquifers with a
reservoir model: application to Fontaine de Vaucluse (South of France). J. Hydrol. 345(1–2),
38–49 (2007)
4. Chen, C., Chen, B.P., Chou, F.N., Yang, C.: Development and application of a decision
group Back-Propagation Neural Network for flood forecasting. J. Hydrol. 385(1–4), 173–
182 (2010)
5. Wang, Y.M., Chang, J.X., Huang, Q.A.: Simulation with RBF neural network model for
reservoir operation rules. Water Resour. Manage 24(11), 2597–2610 (2010)
6. Asadnia, M., Chua, L., Qin, X.S., Talei, A.: Improved particle swarm optimization-based
artificial neural network for rainfall-runoff modeling. J. Hydrol. Eng. 19(7), 1320–1329
(2014)
7. Razfar, M.R., Asadnia, M., Haghshenas, M., Farahnakian, M.: Optimum surface roughness
prediction in face milling X20Cr13 using particle swarm optimization algorithm. Proc. Inst.
Mech. Eng. Part B-J. Eng. Manuf. 224(B11), 1645–1653 (2010)
8. Haklı, H., Uğuz, H.: A novel particle swarm optimization algorithm with Levy flight. Appl.
Soft Comput. 23, 333–345 (2014)
9. Chen, Y.: Research and simulation on Levy Flight model for DTN. In: 3rd International
Congress on Image and Signal Processing, pp. 4421–4423. Yantai, China (2010)
10. Gandomi, A.H., Yang, X.S., Alavi, A.H.: Cuckoo Search Algorithm: a metaheuristic
approach to solve structural optimization problems. Eng. Comput. 29, 17–35 (2013)
11. Mirjalili, S.A.: Dragonfly algorithm: a new meta-heuristic optimization technique for solving
single-objective, discrete, and multi-objective problems. Neural Comput. Appl. 27, 1053–
1073 (2016)
A Method for Forecasting the Demand
for Pharmaceutical Products in a Distributed
Pharmacy Network Based on an Integrated
Approach Using Fuzzy Logic
and Neural Networks
Alekperov Ramiz Balashirin Oglu1(&)

and Isgandarova Ilhama Tarlan Kizi2(&)
1
Information Technology and Programming, Azerimed LLC,
Baku AZ1147, Azerbaijan
[email protected]
2
Department of Computer Engineering, Odlar Yurdu University,
Baku AZ1072, Azerbaijan
[email protected]
Abstract. This article discusses the use of fuzzy logic and a neural network to
predict the demand for pharmaceutical products in a distributed network, in
conditions of insufficient information, a large assortment and the influence of
risk factors. A comprehensive approach to solving forecasting problems is
proposed using: the theory of fuzzy logic - when forecasting emerging and
unmet needs and a neural network - if there is a lot of retrospective information
about the actual sale of drugs and drugs. Using this approach to solve the
problems of forecasting demand allows you to get statistics and experience. The
general algorithm, mathematical interpretation and examples of forecasting the
demand for pharmaceutical products in the face of uncertainty of information are
given, and the general structure of the system for forecasting the demand for
drugs is described.
Keywords: Pharmaceutical market Fuzzy logic Neural networks ABC

analysis XYZ analysis Fuzzy classification Demand forecasting
Uncertainty
1 Introduction
The pharmaceutical market is an important area of the country’s economy, which must be
given special attention due to the fact that it is one of the necessary factors for the timely
provision of human health. Today, there are a large number of pharmaceutical products
(medicines and equipment), which are mass-market goods by terms, the use of which can
be divided into durable goods (for example, sanitation, hygiene, medical devices) and
short-term use (medicines, medicine plant materials, medical cosmetics) [1].
The main task of studying the situation in the sale of medicines is to predict and
determine to what extent their state of sale corresponds to demand, how to change these
A Method for Forecasting the Demand for Pharmaceutical Products 999
indicators in the near future, and what methods should be used for forecasting in order
to satisfy the failure-free provision of the population and pharmacy network with
medicines and drugs.
As you know, the demand is divided into realized (satisfied - the actual sale of
medicines with sufficient and constant availability in the pharmacy network), unsat-
isfied (medicines that are supplied to the pharmacy network in insufficient quantities or
unevenly) and the emerging demand - clearly expressed demand for medicines, which,
as a rule, include new and little-known types of pharmaceutical products.
Currently, there are a large number of forecasting methods [2]: - formal methods,
expert forecasting, adaptive methods, etc. Studies have shown that at present there is no
universal forecasting method that allows obtaining a forecast system for drug demand
with sufficient accuracy for practical purposes.
In each case, it is necessary to choose the appropriate forecasting methods, focusing
on the availability of reliable information. On the other hand, with insufficient infor-
mation about demand satisfaction or with a small amount of retrospective, factual data,
for example, in cases of emerging or unsatisfied demand, managers usually use their
experience and intuition, and use unclearly formulated requirements, which leads to the
use of an approach using soft computing technology [3].
We propose a comprehensive approach to solving the problem of forecasting using
[3, 4]: the theory of fuzzy logic in predicting emerging and unmet demand and the
neural network in the presence of a large amount of retrospective information about the
actual sale of drugs.
The general algorithm for predicting the demand for drugs is given below.
Step 1. Collection (taking into account the seasonality factor), filtering information
on the sale of drugs and normalizing statistics. Formation of interval dynamics
between sales.
Step 2. Classification of sales taking into account the facts of profitability and the
coefficient of variation, which allows to assess the stability of sales.
Step 3. If the drug belongs to a class for which there is not enough sales infor-
mation, go to step 5.
Step 4. The use of a neural network to predict the time series of drug sales. To take
into account the seasonality factor, choose the appropriate neural network archi-
tecture for predicting the insurance stock for drugs over a certain period of time.
Step 5. Applying fuzzy logic to determine the desired best quantity of demand for
unmet and emerging demand. The base of fuzzy rules will be formed according to
the experience, intuition of managers and taking into account the characteristics of
medicines.
Step 6. According to the results of forecasting and fuzzy logic, to form a demand for
medicines for a certain period.
Step 7. Calculate the total demand for medicines, taking into account the current
balance of the drug in the warehouses of the pharmacy network.
1000 A. R. B. Oglu and I. I. T. Kizi
2 Classification of Drugs and Preparations According

to Various Criteria and the Choice of Method for Its
Implementation
The main purpose of the classification of drugs and drugs is to understand how
profitable and stable they are in demand and to determine the unprofitability of the
goods sold and to give recommendations on choosing the forecasting method in step 2
of the algorithm for forecasting the demand for drugs that is described above.
Typically, these tasks in practice are solved using special methods - ABC (ranks
products by contribution to total sales, which allows you to determine the importance
of products) and XYZ (how easy it is to sell products and how predictable they are)
analyzes [3, 4]. Both methods of analysis are based on the Pareto principle and allow
you to break down the weight of an assortment of goods into the following general
classes:
1. ABC analysis: a). A - the most important (20% of the assortment = 80% of sales);
in). B - intermediate (medium importance (30% of the assortment = 15% of sales);
c). C - less important (50% of the assortment = 5% of sales),
2. XYZ analysis: a). X - steady demand and high forecast accuracy, so their sales
volume is easy to predict (0 10%); in). Y is volatile demand (in particular, due to
seasonality and stocks), but a forecast is possible. For more reliable results, an
additional analysis can be carried out (10–25%); from). Z random demand, forecast
is impossible, since there are no patterns in consumption (from 25%).
Combining both methods, you can get 9 classes of goods on the basis of which you
can conduct a cross-analysis (see Table 1) and give recommendations on choosing the
method of forecasting them.
Table 1 shows how important it is to correctly identify the goods belonging to the
listed classes. Based on the clear interval values of the boundaries of ABC and XYZ
classes based on Pareto principles, when moving between boundaries, you can
incorrectly determine, for example, which class the product belongs to, for example,
AX or AY, BX or BY, where it is recommended to predict and determine the insurance
for classes AY and BY stock, and for classes AX and BX to conduct a forecast minus
excess insurance stock.
In this regard, in this paper, a method for the classification of drugs and drugs using
fuzzy logic is proposed. The foundations of the theory of fuzzy logic were developed
by L. Zadeh [5]. In the works of R.A. Aliyev [6] and other scientific researchers list the
following basic algorithms for fuzzy inference:
1. Mamdani algorithm,
2. Tsukamoto algorithm,
3. Larsen algorithm,
4. Sugeno algorithm,
5. Aliyev algorithms (ALI-1, ALI-2, ALI-3, ALI-4).
Suppose that there is

1. T {t1, t2,…, tM}, where ti € T is the set of medicines and preparations (marketable
products) for which classification is carried out;
2. set A - profitability-ordered goods (sales share), calculated as a percentage of total
income, sorted in descending order and summed up by an cumulative total, which
consists of fuzzy sets, for ABC classification, respectively, A1 (A), A2 (B), A3 (C);
3. the set Z - ordered by the value of the coefficient of variation (cv) of sales, which
shows the percentage deviation of sales from the average and consists of fuzzy sets,
for the XYZ classification, respectively, Z1 (X), Z2 (Y), Z3 (Z) and the set G {g1,
g2,…, g9}, where gi € G is the set of combined classes described in Table 1.
We assume that the lower the coefficient of variation, the greater the sales stability.
The algorithm for ABC and XYZ analyzes is described in detail in papers, for
example [5].
Assume that N rules are used to solve the problem:
R1 : if x1 is A1 and x2 is Z1 then g is G1 ;
R2 : if x1 is A2 and x2 is Z2 then g is G2 ;
; ð1Þ
...
Rn : if x1 is An and x2 is Zn then g is Gn ;
where i € 1… N, where N is the total number of fuzzy inference rules; x and y are the
names of the input variables; g is the name of the output variable; A1, A2… An; B1,
B2..Bn; G1, G2… Gn are some given membership functions, and a clear knowledge of
g (0) must be determined on the basis of the information given and the clear values of
x1 (0), x2 (0).
The maximum number of rules depends on the number of terms of fuzzy linguistic
variables that correspond to fuzzy sets A and Z.
We present the Sugeno algorithm for the general case where the number of rules is
not known in advance.
1. We determine the degrees of truth for the premises of each rule (Rn, n € 1..N):
R1 : fA1 ðx1 ð0ÞÞ; Z1 ðx2 ð0ÞÞg;

R2 : fA2 ðx1 ð0ÞÞ; Z2 ðx2 ð0ÞÞg;
ð2Þ
...
Rn : fAn ðx1 ð0ÞÞ; Zn ðx2 ð0ÞÞg:
2. Applying the operation of the fuzzy logical minimum (in the formulas through “˄”),
we find the “cut-off” levels for the premises of each of Rn, where n € 1… N.
R1 : a1 ¼ A1 ðx1 ð0ÞÞ ^ Z1 ðx2 ð0ÞÞ;

R2 : a2 ¼ A2 ðx1 ð0ÞÞ ^ Z2 ðx2 ð0ÞÞ;
ð3Þ
...
Rn : an ¼ An ðx1 ð0ÞÞ ^ Zn ðx2 ð0ÞÞ
Table 1. Properties of fuzzy classes and recommendations for choosing a method for
forecasting sales.
Class Class Class properties Recommendations for choosing a
number name forecasting method
1 CZ Low importance of goods, Insufficient information about the
irregular consumption, low sale - to forecast based on the
predictability experience and intuition of
managers (fuzzy logical
conclusion)
2 CY Low importance of goods, The constant amount of orders is to
seasonal fluctuations, medium predict, based on the neural
degree of predictability network, with a delay of 3 months
(fuzzy inference), safety stock with
the possibility of the company
3 CX Low importance of goods, stable A large amount of retrospective
consumption, high predictability static data - to forecast safety stock
based on a neural network
4 BZ Medium importance of goods, Insufficiency of information about
irregular consumption, low the sale - to predict, based on the
predictability experience and intuition of
managers (fuzzy inference), safety
stock
5 BY The average degree of Neural Network Needs to Predict
importance of goods, seasonal Safety Stock
fluctuations, the average degree
of predictability
6 BX Medium importance of goods, A large amount of retrospective
stable consumption, high static data - it is necessary to
predictability forecast demand based on a neural
network and minus excess safety
stock
7 AZ Very important goods, irregular Insufficiency of information about
consumption, low predictability the sale to predict, based on the
experience and intuition of
managers (fuzzy inference), safety
stock
8 AY Very important products, Neural Network Needs to Predict
seasonal variations, medium Safety Stock
predictability
9 AX Very important products, stable A large amount of retrospective
consumption, high predictability static data - it is necessary to
forecast demand based on a neural
network and minus excess safety
stock
3. We determine the individual outputs for each of Ri, where i € 1..N:
R1 : g1 ¼ a1 x1 ð0Þ þ b1 x1 ð0Þ;

R2 : g2 ¼ a2 x1 ð0Þ þ b2 x2 ð0Þ;
ð4Þ
...
Rn : gn ¼ anx1 ð0Þ þ bnx2 ð0Þ
4. Determine the clear value of the output variable:

XN ai g
g0 ¼ i
ð5Þ
i¼1 ai
Based on this algorithm, Sugeno developed a fuzzy inference system for the
classification of medicines and drugs, which includes two inputs: share of income and
coefficient variation of sales determining the level of sales stability, as well as one
output - classification groups. The following is a description of a fuzzy inference
system for classifying drugs and drugs.
2.1 Definition of Input Variables

As a linguistic variable for the input variable share of income (x1), three sets of fuzzy
variables were chosen (see Fig. 1) having the following nonlinear forms of membership
function [5, 6]:
1. high profitability - zmf (Z-shaped membership function) with parameters [0.7685
0.8241];
2. average profitability - miidle gbellmf (Generalized bell-shaped membership func-
tion) with parameters [0.0467 5.46 0.8443];
3. low profitability - smf (S-shaped membership function) parameters [0.8717 0.9114].
For the input variable x2 (sales stability - determined by the values of the coefficient
of variation of sales), three sets of fuzzy variables were selected (see Fig. 2) having the
following nonlinear forms of membership function:
1. High sales stability (corresponding to the concept - the coefficient of variation is
low) - zmf c parameters [0.09656 0.167];
2. average sales stability (corresponds to the concept - the coefficient of variation is
average) - gbellmf with parameters [0.082 8.17 0.218];
3. low sales stability (consistent with the concept - the coefficient of variation is high)
sigmf (Sigmoidal membership function) with parameters [150 0.2976].
Note that the calculated values of these input parameters are in the range x €
[0..100]. Normalization of these indicators is carried out with the separation of the
values of the input parameters to their maximum value.
h ig h m i d d le lo w
1
0.8
Degree of membership
0.6
0.4
0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

s h a r e o f in c o m e
Fig. 1. Membership functions of the input variable x1 (share of income)
h ig h m i d d le lo w
1
0.8
Degree of membership
0.6
0.4
0.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

c o e f - f ic ie n t v a r ia t io n o f s a le s
Fig. 2. Membership functions of the variable x2 (coefficient variation of sale (sales stability))
2.2 Output Variable

The output variable Qroups has 9 linear expressions corresponding to discrete class
values shown in Table 1, respectively: Output = 9 - group ‘AX’, Output = 8 - group
‘AY’, Output = 7 - group ‘AZ’; Output = 6 - group ‘BX’, Output = 5 - group ‘BY’,
Output = 4 - group ‘BZ’, Output = 3 - group ‘CX’, Output = 2 - group ‘CY’, Out-
put = 1 - group ‘CZ’.
2.3 Rules
Nine rules were created for the connection of input and output variables, formed on the
basis of Table 2. These rules provide the choice of one of the output expressions,
depending on the values of the input variables. A weighted average (whatever) was
used as a defuzzification method [3]. The resulting approximated surface of the
inference system is shown in Fig. 3.
Table 2. Inference rules.

Share of income Sales stability
High Middle Low
High AX AY AZ
Middle BX BY BZ
Low CX CY CZ
Fig. 3. Approximate surface
2.4 Analysis of the Results

Evaluation of the effectiveness of the developed inference algorithm for the classifi-
cation of drugs and drugs was carried out using the approximate data shown in
Tables 3.
Table 3. The share of sales of goods and the coefficient variation of sales.
Drugs Share of income Coefficient variation of sales
Income Share Sales
March April May June July
Drug1 284 18.66% 0.75 80 68 75 76 73 0.05
Drug2 154 10.12% 0.85 36 37 28 40 18 0.25
Drug3 28 1.84% 0.99 26 25 30 18 23 0.16
Drug4 16 1.05% 1 20 10 15 7 21 0.37
Drug5 70 4.60% 0.97 20 7 30 21 28 0.38
Drug6 360 23.65% 0.57 27 32 12 10 10 0.52
Drug7 110 7.23% 0.93 15 27 32 18 10 0.39
Drug8 500 32.85% 0.33 18 19 22 25 21 0.12
The results of a fuzzy logical conclusion are given, where the corresponding classes
for each drug are determined: Drug1 corresponds to class AX, Drug2 corresponds to
class.
BY, Drug3 corresponds to class CY, Drug4 corresponds to class CZ, Drug5 cor-
responds to class CZ, Drug6 corresponds to the class AZ, Drug7 corresponds to the
class CZ, Drug8 corresponds to the class AX.
An analysis of the results shows the high efficiency of the proposed inference
algorithm for the classification of drugs. After determining the class in accordance with
the proposed recommendations, you can choose a method for forecasting demand for
each product individually:
1. In conditions of sales stability, a large amount of retrospective static data and
uncertainty, it is mainly proposed to use a neural network solution as a forecasting
method [3, 4, 7, 8]. The forecasting problem is posed as the problem of approxi-
mating a number of dynamics, i.e. It is proposed to use the Generalized Regression
Neural Network (GRNN) [7] to construct a function by a finite set of points, by the
number of sales of medicines taken by months, and for ease of calculation and the
reliability of forecasting sales for goods from classes such as AX, BX, CX.
2. For goods from the class AY, BY, CY which are characterized by quarterly, steady
changes repeating from month to month in levels, i.e. seasonal variations, it is
proposed to use a neural network with a delay at the entrance.
3. The lack of information and the small amount of statistics on the sale determines the
use of an expert system for making decisions on determining the demand for goods
from the class AZ, BZ, CZ. To this end, to solve this subproblem, we built an expert
system based on the Mamdani fuzzy logic inference algorithm in the Fuzzy Logic
Toolbox of the MATLAB computing environment [7].
3 Conclusion
Using this approach to solve the problem of demand forecasting allows taking into
account both statistical data and the experience and intuition of the managerial staff of
the pharmacy network at various stages of demand forecasting and is flexible in terms
of the availability of sufficient information about demand.
Since, with an increase in the sale of medicines and drugs using this approach,
demand moves from the class of dissatisfaction to realized demand.
Due to the fact that drugs have their own specific features, we are conducting
further research to improve the expert system in order to determine the desired best
amount of demand for unmet and emerging demand.
References
1. Osobennosti sprosa na farmatsevticheskom rynke [Ocoбeннocти cпpoca нa фap-
мaцeвтичecкoм pынкe]. http://www.alppp.ru/law/informacija-i-informatizacija/44/statja–
osobennosti-sprosa-na-farmacevticheskom-rynke.html. Accessed on 26 Mar 2020
2. Tikhonov, E.Y.: Prognozirovaniye v usloviyakh rynka [Пpoгнoзиpoвaниe в ycлoвияx
pынкa], Nevinnomyssk (2006)
3. Zadeh, L.A.: Fuzzy Logic, Neural Networks, and Soft Computing. Commun. ACM 37, 77–84
(1994)
4. Aliev, R.A., Bijan, F., Aliev, R.R.: Soft Computing and its Applications in Business and
Economics. Springer, Heidelberg (2004)
5. How to Use ABC Analysis for Inventory Management (and the Added Value of XYZ Analysis).
https://www.eazystock.com/blog/how-to-use-abc-analysis-for-inventory-management. Acces-
sed on 26 March 2020
6. Takagi, T., Sugeno, M.: Fuzzy identification of systems and its applications to modeling and
control. IEEE Trans. Syst. Man Cybern. 15, 116–132 (1985)
8. Balashirin, A.R., Tarlan, I.I.: Application of neural networks for segmentation of catering
services market within the overall system of consumer market on the model of restaurant
business with the aim to advance the efficiency of marketing policy. In: Aliev, R., Kacprzyk,
J., Pedrycz, W., Jamshidi, M., Sadikoglu, F. (eds.) 13th International Conference on Theory
and Application of Fuzzy Systems and Soft Computing, ICAFS-2018, Advances in Intelligent
Systems and Computing, vol. 896, pp. 905–913. Springer, Cham (2018)
Marketing Strategy Selection Using Fuzzy
Analytic Network Process
Serhan Hamal , Bahar Sennaroglu(&) ,

and Mahmure Övül Arıoğlu
Marmara University, Kadikoy, 34722 Istanbul, Turkey

[email protected]
Abstract. This study presents a multiple criteria decision making problem for
selecting best marketing strategy among three alternatives according to five
criteria, which are managerial capabilities, customer linking capabilities, market
innovation capabilities, human resource assets, and reputational assets. The
marketing strategies alternatives are product proliferation, product innovation,
and manufacturing cost reduction. The objective is to select the best marketing
strategy for the Istanbul store of a multinational furniture retailer. The alterna-
tives according to the respective criteria are evaluated based on the linguistic
statements of the marketing expert of the company. The fuzzy analytic network
process is used for the problem at hand due to human subjective decision
making and interdependence among criteria. Chang’s extent analysis method is
employed in computing the pairwise comparison matrices. According to the
results, among the marketing strategies alternatives, manufacturing cost reduc-
tion is selected as the best strategy.
Keywords: Marketing strategy selection Fuzzy analytic network process

Chang’s extent analysis
1 Introduction
Fuzzy-set theory is designed to model the vagueness or imprecision of human cognitive

processes [1]. Strategy selection involves human subjective decision making process
and is a multiple criteria decision making (MCDM) problem. Therefore, fuzzy MCDM
methods are appropriate decision tools for strategy selection problems. Hence, this
study uses a fuzzy MCDM method for marketing strategy selection of the Istanbul store
of a multinational furniture retailer. The objective is to select the best marketing
strategy to be used in the marketing planning process of the company.
In the literature, there are studies using fuzzy-set theory on strategy selection in
various areas. Lin et al. [2] implement fuzzy analytic network process (ANP) for
selecting the best marketing strategy for a private hotel. Gürbüz et al. [3] apply a hybrid
method using fuzzy analytic hierarchy process (AHP) and fuzzy metric distance
methods to marketing strategy selection. Cevik Onar et al. [4] develop an MCDM model
determining the weights of the factors by interval type-2 fuzzy AHP and selecting the
best strategy for strategic decisions by hesitant fuzzy technique for order preference by
similarity to ideal solution (TOPSIS). Ge et al. [5] propose a methodology based on an
Marketing Strategy Selection Using Fuzzy Analytic Network Process 1009
integrated logarithmic fuzzy preference programming to obtain the unique normalized

optimal priority vector for fuzzy pairwise comparison matrices of AHP for the optimum
maintenance strategies selection problem. Zhang [6] uses fuzzy Delphi method and
ANP to evaluate regional low-carbon tourism strategies. Baykasoğlu and Gölcük [7]
propose an interval type-2 fuzzy MCDM model by integrating TOPSIS method and
decision making trial and evaluation laboratory method and implement it in a strengths,
weaknesses, opportunities, and threats based strategy selection problem. Adabi et al. [8]
design a fuzzy negotiation strategy selection system for conducting negotiator agents in
how to select a suitable negotiation strategy according to the critical condition of the
cloud trading market. Seiti et al. [9] propose a model based on fuzzy axiomatic design
and implement it in a maintenance strategy selection problem. Güler and Büyüközkan
[10] find criteria weights by fuzzy AHP and select the most appropriate digital trans-
formation strategy by fuzzy axiomatic design. Kuzu [11] applies fuzzy AHP to the
strategy selection problem of a state university in Turkey. Jain et al. [12] implement an
integrated methodology of fuzzy AHP and fuzzy TOPSIS to select green marketing
strategy for a manufacturing firm. Edjossan-Sossou et al. [13] suggest a hybrid fuzzy
MCDM approach using fuzzy AHP, fuzzy weighted average method and fuzzy PRO-
METHEE for identifying the most sustainable risk management strategy under uncer-
tainty. Even though there are strategy selection studies in the literature, only a few
studies are on marketing strategy selection.
This study presents an application of fuzzy ANP to the marketing strategy selection
problem with the objective of selecting the best marketing strategy for the Istanbul
store of a multinational furniture retailer. Three alternatives were identified among
many marketing strategies defined by Kotler [14] by the marketing expert of the
company by taking into consideration the target market and market position of the
company. In order to evaluate alternatives, five criteria were determined from among
the marketing resources proposed by Hooley et al. [15], which are those resources that
create competitive advantage when a company possesses and deploys them. The
contribution of this paper is that the criteria used in evaluating the marketing strategies
alternatives belong to marketing resources [15], hence organizational factors creating
competitive advantage are considered, and also their interrelations are taken into
account. Due to imprecise nature of linguistic statements in the evaluations, fuzzy ANP
is used as the appropriate fuzzy MCDM method.
In the remainder of the paper, firstly fuzzy analytic network process is briefly
described, then application of fuzzy ANP for marketing strategy selection is presented,
and finally conclusion is given.
2 Fuzzy Analytic Network Process
In this study, fuzzy ANP was used for the strategy selection problem due to human
subjective decision making and interdependence among criteria.
The steps of the ANP method can be summarized as follows [16]:
Step 1. Build the network structure of the decision problem for ANP.
Step 2. Compare the criteria in the whole system to form the supermatrix.
1010 S. Hamal et al.
Step 3. Derive the weighted supermatrix by transforming all column sums to unity
exactly.
Step 4. Raise the weighted supermatrix to its limiting power to get the global
priority weights.
The criteria and the alternatives according to the criteria were evaluated based on
the linguistic statements of the marketing expert of the company. These linguistic
statements were recorded in the form of triangular fuzzy numbers in the pairwise
comparison matrices. Chang’s extent analysis method was employed in computing the
fuzzy comparison matrices to form ANP supermatrix.
Chang’s extent analysis method utilizes the concept of extent analysis combined
with degree of possibility to calculate weights from fuzzy comparison matrices [17].
The steps of the Chang’s extent analysis method is as follows [18–20]:
Each object is taken (X ¼ fx1 ; x2 ; . . .; xn g is an object set) and for each goal
(G ¼ fg1 ; g2 ; . . .; gm g is a goal set) extent analysis is performed. Therefore, m extent
analysis values for each object is obtained as Mg1i ; Mg2i ; . . .; Mgmi ; i ¼ 1; 2; . . .; n where all
the Mgji ðj ¼ 1; 2; . . .; mÞ are triangular fuzzy numbers.
Step 1. The value of fuzzy synthetic extent with respect to the ith object is obtained.
Xm hXn Xm i1
Si ¼ Mj
j¼1 gi i¼1
Mj
j¼1 gi
ð1Þ
where
Xm Xm Xm Xm
Mj ¼
j¼1 gi
l;
j¼1 j
m;
j¼1 j
u
j¼1 j
ð2Þ
Xn Xm Xn Xn Xn
i¼1
Mj ¼
j¼1 gi
l;
i¼1 i
m;
i¼1 i
u
i¼1 i
ð3Þ
hXn Xm i1
1 1 1
Mj
j¼1 gi
¼ Pn ; Pn ; Pn ð4Þ
i¼1
i¼1 ui i¼1 mi i¼1 li
Step 2. The degree of possibility of M2 ¼ ðl2 ; m2 ; u2 Þ M1 ¼ ðl1 ; m1 ; u1 Þ is

defined.
8
<1 m2 m1
VðM2 M1 Þ ¼ 0 l 1 u2 ð5Þ
: l1 u2
otherwise
ðm2 u2 Þðm1 l1 Þ
Step 3. The weight vector is given by
0
0 0 0
T
W ¼ d ðA1 Þ; d ðA2 Þ; . . .; d ðAn Þ ð6Þ
by defining the degree possibility for a convex fuzzy number to be greater than
k convex fuzzy numbers Mi ði ¼ 1; 2; . . .; kÞ as
V ðM M1 ; M2 ; . . .; Mk Þ ¼ minV ðM Mi Þ; i ¼ 1; 2; . . .; k ð7Þ
and assuming
0
d ðAi Þ ¼ minV ðSi Sk Þ for k ¼ 1; 2; . . .; n; k 6¼ i ð8Þ
where Ai ði ¼ 1; 2; . . .; nÞ are n elements.

Step 4. The normalized weight vectors are obtained via normalization.
W ¼ ðd ðA1 Þ; d ðA2 Þ; . . .; dðAn ÞÞT ð9Þ
where W is a nonfuzzy number.
3 Application of Fuzzy ANP for Marketing Strategy Selection
The network structure of the marketing strategy selection problem is shown in Fig. 1.
Arrows show the interdependencies and the supermatrix is formed according to this
network structure through pairwise comparison matrices. There are three marketing
strategies which are product proliferation, product innovation, and manufacturing cost
reduction. The marketing strategies are evaluated according to five criteria, which are
managerial capabilities, customer-linking capabilities, market innovation capabilities,
human resource assets, and reputational assets [15].
Managerial Market
Capabilities Customer-Linking Innovation
Capabilities Capabilities
Reputational
Human Assets
Resource Assets
Product Product
Manufacturing
Proliferation Innovation
Cost Reduction
Fig. 1. ANP network structure of marketing strategy selection problem.

A questionnaire was prepared according to the network structure and applied to the
marketing expert of the company. The linguistic statements of the expert for pairwise
comparisons were obtained using nine-point linguistic scale and represented with the
triangular fuzzy numbers (Table 1) in the pairwise comparison matrices. A triangular
fuzzy number is defined as ðl; m; uÞ and satisfies l m u, where l; mandu represent
respectively the smallest possible value, the most promising value, and the largest
possible value that describe a fuzzy event [19].
Chang’s extent analysis method was employed in computing the fuzzy comparison
matrices to derive the local weights with nonfuzzy numbers in order to form ANP
supermatrix based on the network structure and local weights. Then the weighted
supermatrix was formed and raised to its limiting power to get the global priority
weights. The obtained limiting supermatrix is given in Table 2. Manufacturing cost
reduction has the highest global priority weight of 0.262 and the normalized weight of
0.525 (Table 3), and hence among the marketing strategies alternatives, manufacturing
cost reduction is selected as the best strategy.
Table 1. Linguistic scale and triangular fuzzy numbers.

Equally important (1, 1, 1) (1, 1, 1)
Intermediate (1, 2, 3) (1/3, 1/2, 1)
Weakly important (2, 3, 4) (1/4, 1/3, 1/2)
Intermediate (3, 4, 5) (1/5, 1/4, 1/3)
Essentially important (4, 5, 6) (1/6, 1/5, 1/4)
Intermediate (5, 6, 7) (1/7, 1/6, 1/5)
Very strongly important (6, 7, 8) (1/8, 1/7, 1/6)
Intermediate (7, 8, 9) (1/9, 1/8, 1/7)
Absolutely important (9, 9, 9) (1/9, 1/9, 1/9)
Table 2. ANP limiting supermatrix.

Managerial Customer- Market Human Reputational Product Product Manufacturing
capabilities linking ınnovation resource assets proliferation ınnovation cost reduction
Capabilities Capabilities Assets
Managerial 0.028 0.028 0.028 0.028 0.028 0.028 0.028 0.028
capabilities
Customer- 0.189 0.189 0.189 0.189 0.189 0.189 0.189 0.189
linking
capabilities
Market 0.157 0.157 0.157 0.157 0.157 0.157 0.157 0.157
ınnovation
capabilities
Human 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
resource assets
Reputational 0.127 0.127 0.127 0.127 0.127 0.127 0.127 0.127
assets
Product 0.172 0.172 0.172 0.172 0.172 0.172 0.172 0.172
proliferation
Product 0.065 0.065 0.065 0.065 0.065 0.065 0.065 0.065
ınnovation
Manufacturing 0.262 0.262 0.262 0.262 0.262 0.262 0.262 0.262
cost reduction
Table 3. Ranking and priority weights of alternatives.

Ranking of alternatives Global priority weight Normalized weight
1. Manufacturing cost reduction 0.262 0.525
2. Product proliferation 0.172 0.345
3. Product ınnovation 0.065 0.130
4 Conclusion
The ANP overcomes the restrictions of the hierarchical structure of independent cri-
teria. It is an effective tool in decision making by incorporating interdependence
between criteria.
In this study, the fuzzy ANP was used for the marketing strategy selection problem
due to human subjective decision making and interdependence among criteria. The
fuzzy triangular numbers and Chang’s extent analysis method was used in forming
ANP supermatrix.
This application of MCDM in the real world involved a single decision maker who
was the marketing expert of a multinational furniture retailer. The best marketing
strategy was selected for the Istanbul store of the company. The marketing strategy
alternatives were identified based on relevant literature and opinion of the expert who
took into consideration the target market and market position of the company. The
evaluation criteria were determined based on marketing resources that create, maintain
or contribute competitive advantage in the market.
As a further study, other methods to defuzzify weights, such as the modification to
normalization in Chang’s method or the centroid defuzzification, can be used and new
results can be compared by the results of this study. In addition, as a future research,
interval-valued intuitionistic fuzzy (IVIF) numbers can be used to represent the lin-
guistic statements and IVIF-ANP can be applied to the marketing strategy selection
problem.
References
1. Lootsma, F.A.: Fuzzy logic for planning and decision making. APOP, vol. 8. Springer,
Boston, MA (2013)
2. Lin, C.-T., Lee, C., Wu, C.-S.: Optimizing a marketing expert decision process for the
private hotel. Expert Syst. Appl. 36(3), 5613–5619 (2009)
3. Gürbüz, T., Albayrak, Y.E., Alaybeyoğlu, E.: Criteria weighting and 4P’s planning in
marketing using a fuzzy metric distance and AHP hybrid method. Int. J. Comput. Intell. Syst.
7(sup1), 94–104 (2014)
4. Cevik Onar, S., Oztaysi, B., Kahraman, C.: Strategic decision selection using hesitant fuzzy
TOPSIS and interval type-2 fuzzy AHP: a case study. Int. J. Comput. Intell. Syst. 7(5),
1002–1021 (2014)
5. Ge, Y., Xiao, M., Yang, Z., Zhang, L., Hu, Z., Feng, D.: An integrated logarithmic fuzzy
preference programming based methodology for optimum maintenance strategies selection.
Appl. Soft Comput. 60, 591–601 (2017)
6. Zhang, J.: Evaluating regional low-carbon tourism strategies using the fuzzy Delphi-analytic
network process approach. J. Clean. Prod. 141, 409–419 (2017)
7. Baykasoğlu, A., Gölcük, İ.: Development of an interval type-2 fuzzy sets based hierarchical
MADM model by combining DEMATEL and TOPSIS. Expert Syst. Appl. 70, 37–51 (2017)
8. Adabi, S., Mosadeghi, M., Yazdani, S.: A real-world inspired multi-strategy based
negotiating system for cloud service market. J. Cloud Comput. 7(1), 1–40 (2018). https://doi.
org/10.1186/s13677-018-0116-5
9. Seiti, H., Hafezalkotob, A., Fattahi, R.: Extending a pessimistic–optimistic fuzzy information
axiom based approach considering acceptable risk: application in the selection of
maintenance strategy. Appl. Soft Comput. 67, 895–909 (2018)
10. Güler, M., Büyüközkan, G.: Analysis of digital transformation strategies with an integrated
Fuzzy AHP-axiomatic design methodology. IFAC-PapersOnLine 52(13), 1186–1191 (2019)
11. Kuzu, Ö.H.: Strategy selection in the universities via fuzzy AHP method: a case study. Int.
J. High. Educ. 9(2), 107–117 (2020)
12. Jain, A., Darbari, J. D., Kaul, A., Jha, P.C.: Selection of a green marketing strategy using
MCDM under fuzzy environment. In: Soft Computing for Problem Solving, pp. 499–512.
13. Edjossan-Sossou, A.M., Galvez, D., Deck, O., Al Heib, M., Verdel, T., Dupont, L., Chery,
O., Camargo, M., Morel, L.: Sustainable risk management strategy selection using a fuzzy
multi-criteria decision approach. Int. J. Disaster Risk Reduct. 45(101474), 1–15 (2020)
14. Kotler, P.: Marketing Management. The Millennium Edition. Prentice Hall, Upper Saddle
River (2000)
15. Hooley, G.J., Greenley, G.E., Cadogan, J.W., Fahy, J.: The performance impact of
marketing resources. J. Bus. Res. 58, 18–27 (2005)
16. Tzeng, G.-H., Huang, J.-J.: Multiple Attribute Decision Making Methods and Applications.
CRC Press, Boca Raton, FL (2011)
17. Ahmed, F., Kilic, K.: Comparison of fuzzy extent analysis technique and its extensions with
original eigen vector approach. In: Proceedings of the 18th International Conference on
Enterprise Information Systems, vol. 2, pp. 174–179. SCITEPRESS (2016)
18. Chang, D.Y.: Applications of the extent analysis method on fuzzy AHP. Eur. J. Oper. Res.
95(3), 649–655 (1996)
19. Wei, J. Y., Sun, A.F., Wang, C.H.: The application of fuzzy-ANP in the selection of supplier
in supply chain management. In: 2010 International Conference on Logistics Systems and
Intelligent Management, vol. 3, pp. 1357–1360. IEEE (2010)
20. Sennaroglu, B., Akıcı, O.: Supplier selection using fuzzy analytic network process. In:
International Conference on Intelligent and Fuzzy Systems, pp. 829–834. Springer, Cham
(2019)
A Deep Learning Model for Skin Lesion
Analysis Using Gaussian Adversarial
Networks
Fatih Ergin(B) and Ismail Burak Parlak(B)
Department of Computer Engineering, Galatasaray University,

Ciragan Cad. No: 36, Ortakoy, 34349 Istanbul, Turkey
Abstract. Computer assisted radiology becomes an interdisciplinary

domain between mathematics, medicine and engineering. Tumor detec-
tion, analysis, classification are main problems in digital radiology for
diagnosis and follow-up. A physician or an oncologist involves in the care
of patients by regarding detailed reports of carcinoma in situ that analyze
the pathology of suspicious lesions. Deep learning applied to several fields
in medicine is considered as an intervention for oncology. Even if the final
treatment of the lesion is decided by the oncologists or the surgeons in
a case of resection, image based analysis of lesions (benign or malign)
promises automated decision making for radiology. Skin lesion detection
and classification are current challenges in medical image analysis. Der-
matologic image processing benefits from the evaluation scores of neural
nets. Gaussian Adversarial Networks (GAN) bring a new architecture in
machine learning by adding generator and discriminator steps in data
analysis. In this article, GAN architecture has been implemented on two
dimensional skin lesion images. After the preprocessing, colored images
have been trained in GAN. The experiment setup has been enriched by
adding incremental noise on tumor images before GAN training. The
evaluation has been tested through accuracy, sensitivity, specificity, Dice
coefficient and Jaccard coefficient parameters. In conclusion, test results
showed that GAN architecture provides a robust approach in skin lesion
analysis.
Keywords: Deep neural networks · Image processing · Histology ·

Tumor detection · Image segmentation
1 Introduction
During the last few years, deep neural networks have gained considerable atten-
tion in several problems of computer vision. The new network hierarchies present
complex transfer modalities to deal with adaptive learning tasks. Deep neural
This research has been financially supported by Galatasaray University Research Fund
Project 19.401.001.
1016 F. Ergin and I. B. Parlak
networks (DNN) allow machines to learn hybrid data structures of mathematical

models. Image semantics would be resolved using relevant models. The learning
rate is measured in DNN by the achievement of comprehensive data analysis.
Generative adversarial networks (GANs) deal with a new hierarchy in intelligent
systems. The network scheme conforms to better training performance with less
annotations. GAN achievement is derived through a competitive learning where
the model consists of different stacks. The processing in layers is characterized
with multiple levels of abstraction from high-dimensional input data.
In medicine, lesion detection becomes more efficient with new models based
on deep learning networks from histological to radiological acquisitions. Recent
studies [1,2] reveal that detection performance of deep networks has even
matched or exceeded human-level performance in several tasks such as diabetic
retinopathy and tumor detection. Early detection of cancer is considered as one
the most complex and hard problems in radiology. The follow-ups and repeated
cases are also challenges for the correct decision making. Over the last decade,
the progress and the integration of DNN enable rapid diagnosis of patients at
these risk groups. Even if the final medical decision must be taken with a spe-
cialist, DNNs might reduce the time for the diagnostic errors and workload of
physicians. Therefore, DNN performance is not compared with physicians in our
study. The evaluation of DNN based segmentation is performed through Ground
Truth; a mask which identifies the whole area or volume in target images.
The common goal of deep learning techniques is to recursively learn compu-
tational model parameters using a training data set to gradually improve the
model in performing the desired purpose. Using many previously unseen data,
models can also perform the same task accurately once a computer is trained
for a specific task. The strong generalization ability of deep learning now distin-
guish it from the other techniques of machine learning. The detection and eval-
uation criteria result that the use of multilayered hierarchy of GAN shows valid
scores. The variation of inter observer, the inhomogeneity in image scale encom-
pass the complexity of automatic lesion detection. In skin lesion segmentation,
International Skin Imaging Collaboration (ISIC) focuses on the analysis and the
improvement of big datasets. Annotated image corpora is considered as a chal-
lenge in deep learning for detection and classification. Even if recent studies in
data challenges promised valid results for clinical applications, the performance
evaluation shows that training datasets might cause variation in skin lesion detec-
tion. In order to promote automatic analysis in this field, GAN technique that
represents promising scores is preferred. Although lots of work has been pro-
posed, there is still a margin of performance improvement for both skin lesion
segmentation and classification. The International Skin Imaging Collaboration
(ISIC) is a cooperation focusing on the automatic analysis of skin lesion, and has
continuously expanded its datasets since 2016. In ISIC 2017, annotated datasets
for three processing tasks related to skin lesion images, including lesion segmen-
tation, dermoscopic feature extraction and lesion classification, were released for
researchers to promote the accuracy of automatic melanoma detection methods.
In this paper, we provide a new application area of deep neural networks in
skin lesion analysis. We note that dermoscopic feature extraction is relatively
Deep Learning Model for Skin Lesion Analysis Using GANs 1017
a new problem in deep learning to address the detection and the classifica-
tion of lesions. The following section presents the automatic diagnosis of skin
lesions from dermoscopic images and related studies of DNN for medicine. The
third section shows the methodology of GAN, our corresponding formulation
through computational parameters and the statistical evaluation. Our detection
results are given through statistical parameters in the fourth section. Finally,
the assessment of GAN in skin lesion detection is concluded through the current
state-of-the-art and prospective improvements.
2 Related Works
Recent studies of DNN have increased the impact of Boltzmann Machines and
Convolutional Neural Networks (CNNs) in the applications of object detection
and classification. The automatic extraction of complex features is considered
as the main advantage of CNNs. Moreover, researchers in radiology are focused
on deep network layers through CNN characteristics to bypass the challenges
in feature extraction. However, the noisy nature of medical image acquisitions
provides additional issues. In order to handle the noise, GAN based models
become more popular in big medical datasets. The data augmentation serves as
a robust alternative to minimize segmentation errors in lossy images. Goodfellow
et al. [4] explored a special case of generative models by adding a multilayer
perceptron for the random noise against the competitive multilayer perceptron
in decision making. In the review paper of Kazeminia et al. [5], the assesment of
GANs is summarized through several studies.
The advantages are presented as scarcity of annotated data and rich feature
extraction. On the other hand, the disadvantages are given as unstable training
and uninterpretability. Xue et al. [6] developed a new neural network; SegAN
with two sub steps; the segmentor and the critic with a multi-scale L1 loss func-
tion. Frid-Adar et al. [7] proposed a DNN architecture using GAN where liver
images were analyzed by adding synthetic datasets. This step of data augmen-
tation permitted the synthesis of high quality liver lesion region of interests
and improved the performance of image classification in CNN application. Iqbal
et al. [8] tested the performance of a comparable network architecture (SkipNet,
IntNet, SENet) using multimodal voxels in brain tumor images.
3 Methodology
Our study is composed of three step; preprocessing, implementation of GAN
and evaluation. During the preprocessing, image normalization procedures have
been applied; the image resizing and the conversion of file formats. Moreover,
data size has been augmented by creating additional images files in different
noise levels. The implementation of GAN is based on the adversial architecture.
The learning generative models has been leaded by Goodfellow et al. [4]. The
fundamental hierarchy is based on two steps; discrimination and generation. The
first objective is to determine whether a data sample is arising from a false or
true distribution in the measurement. The second objective competes with the
previous step by perturbing with noisy false distribution. The challenge between
two steps causes to give high probability value to the most likely samples to
be a real data. On the other hand, low probability values are identified as false
samples.
In our study, we have divided these objectives through a general deep learning
architecture. A generative adversarial network has been implemented with two
sub networks denoted as discriminator and generator. The generator network has
produced false skin lesion data using random noise. The discriminator network
has determined if a candidate skin lesion image belongs to the real dataset.
Figure 1 illustrates the GAN scheme applied into our study. The mechanism is
based on the approach of Xue et al. [6]; the SegAN hierarchy where two levels
segmentor and critic are competing. The optimization function is determined
using L1 loss function.
Fig. 1. The flowchart of Gaussian adversial network layers.
Before the implementation GAN network, we have increased the number of

experiments by giving additive noise into initial dataset. Five different noise
levels have been tested using Gaussian noise given as it follows through the
initial image Ii ;
If inal = Ii + In (1)
If inal , In are the noisy image and the additive noise, respectively.
The Gaussian noise is defined using the following equation;

1 (z−μ)2
N (Z) = √ e 2σ2 (2)
σ 2π
N (Z) denotes the noise distribution in single channel image. Our images have
been encoded in RGB channels. Therefore, the noise has been applied for
all channels. μ and σ represents the mean value and the standard deviation,
respectively.
All images in this study have been extracted from ISBI Challenge 2017;
Melanoma Detection dataset [3]. The dataset has been divided into three levels;
train, validation and test datasets. During our data design, the training dataset
has consisted of 2000 dermoscopic JPEG images. The related image masks has
been stored in PNG format.
The dataset includes various type of lesions namely malignant melanoma,
nevus and seborrhoeic keratosis. All JPEG images have been converted to PNG
format to be compatible with our framework by watching out the additive noise
during this process. This step is required to use our GAN model. The initial
images [3] were of various dimensions. Moreover, all images have been resized
into same dimension scale to reduce the memory consumption and increase the
accuracy as a preprocessing stage. Also, RGB images have been transformed to
grayscale images to prevent memory limitations.
Our network has been implemented using Python language. NumPy, SciPy
package have been used in data structures and predefined functions. The open
source library TensorFlow has been integrated to our model where data shaped
as tensors, which is a generalization of N-dimensional arrays, can be fed into
the computational graph for the network scheme. Keras has been choosen for
neural networks API where all required parts; layers, objectives, activation func-
tions and optimizers are defined. The components include several parameters in
network optimization through convolutions, pooling, dropout and batch normal-
ization. Pytorch has been set for the machine learning framework to compute
the weight matrices and biases during a session of GAN model. Training a model
becomes the implementation of an optimization function where running the opti-
mizer is performed as a batch of data inside a session. The segmentation step
given in Fig. 1 shows the computation for the training process where Convolu-
tion operations denoted as Conv are propagated to the computation of Rectified
Linear Unit (ReLU) function. All models were trained and tested using Tesla
K80 GPU which has 12 GB of video memory on a Ubuntu 18.04 during this
study.
During tumor detection, our results have been evaluated through several
metrics to ensure a valid performance of the study The experiments have been
reviewed to provide statistics in different Gaussian noise levels. Table 1 shows
Dice index, Jaccard index, Accuracy, Sensitivity and Specificity scores with dif-
ferent additive noise levels. Several evaluation metrics have been used to deter-
mine the quality of the models. Sensitivity, specificity, Jaccard and Dice simi-
larity coefficients have been used to compare the target segmentation mask and
the predicted segmentation mask.
In statistical image analysis, segmentation, classification and detection per-

formance deal with the accuracy as a common metric. It figures out the percent-
age of the predicted pixelwise area in test tumors. The following equations are
defined with respect to false positives (Fp ), true positives (Tp ), false negatives
(Fn ) and true negatives (Tn ),
Tp + Tn
Accuracy = (3)
T p + T n + Fp + Fn
Sensitivity represents the performance of avoiding false negatives in tumor
like pixels. It is characterized as the number of true positives (Tp ) over the
number of true positives plus the number of false positives (Fp ) as follows;
TP
P recision = (4)
T p + Fn
Specificity represents the performance of avoiding false positives in tumor
like pixels. It is set as the number of true positives (Tp ) over the number of true
positives plus the number of false negatives (Fn ) as follows;
Tn
Recall = (5)
T n + Fp
Sensitivity and Specificity are used criteria to infer how useful and valid a
technique is. Specificity gives the fraction of detected relevant pixels compared
to the computed areas. Sensitivity computes the fraction of relevant segmented
areas compared to the total true positives. Beyond these standards, Jaccard
Similarity and Dice Coefficient are given as follows;
2 ∗ Tp
Dice = (6)
(Tp + Fp ) + (Tp + Fn )
Segmented ∩ GT
JaccSim(Segmented, GT ) = (7)
Segmented ∪ GT
where Segmented and GT represents the detection and ground truth, respectively.
4 Results
The GAN network has been trained for 200 epochs. The batch size was 8 with an
adaptive learning rate. Adam Optimizer [9] has started with 2.0 ∗ 10−4 and has
decayed with 0.5 every 25 epochs. We have noticed that model learning limits
have been reached. Therefore, we have selected 200 as the number of epochs.
In a nutshell, GAN introduces fake skin lesions during the generator level
and the discriminator takes a decision after the training if test image is a lesion.
We note that the loss value in generation is lowered in order to outperform
the segmentation during the discrimination level. Table 1 shows the correspond-
ing results through statistical metrics. Figure 2 presents the original skin lesion
image, ground truth (GT) and segmented area.
Fig. 2. Lesion detection compared to ground truth.
Table 1. Evaluation of detection through different metrics.
Noise(%) Dice Jaccard Accuracy Sensitivity Specificity

0 0.8203 0.7076 0.9264 0.9178 0.9312
10 0.8002 0.6845 0.9192 0.9167 0.9291
20 0.8055 0.6894 0.9248 0.903 0.9294
30 0.8112 0.6958 0.9226 0.9376 0.9263
40 0.811 0.6969 0.9227 0.9081 0.9285
50 0.7864 0.6625 0.9134 0.9282 0.9208
In the evaluation of skin lesion images, GAN approach had high accuracy,
sensitivity and specificity of detection with respect to ground truth even in high
level of additive noise. On the other hand, Dice index and Jaccard similarity
values have decreased during the detection when additive noise has added onto
images.
5 Conclusion
Lesion detection becomes more efficient with new models based on deep learning
networks from histological to radiological acquisitions. Medical image analysis
is a cutting edge in image processing especially for cancer studies. However,
medical imaging and tumor detection require specific care due to detection area
and patient survey. In this paper, medical images have been analyzed through
image intrinsic parameters to segment tumor areas. Gaussian Adversial Networks
have been tested on different images with additive noise levels. Detection results
are given on different statistical scores. Output scores indicate that GAN would
be a reliable alternative in tumor based medical imaging studies. Our GAN
approach provided high accuracy, sensitivity and specificity of detection with
respect to ground truth even in high level of additive noise. However, Dice index
and Jaccard similarity values have decreased during the detection when additive
noise has added onto images. We conclude that further studies are necessary to
reveal the feasibility of applying this algorithm. It is crucial to assess in clinical
settings whether GAN applications could lead to improved care and outcomes in
the diagnosis and follow up of skin tumors. In prospective studies, other medical
modalities would be mapped with the similar techniques to expand the utility
of GAN.
References
1. Gulshan, V., Peng, L., Coram, M., Stumpe, M.C., Wu, D., Narayanaswamy, A.:
Development and validation of a deep learning algorithm for detection of diabetic
retinopathy in retinal fundus photographs. Jama 316(22), 2402–2410 (2016)
2. Isin, A., Direkoglu, C., Şah, M.: Review of MRI-based brain tumor image segmen-
tation using deep learning methods. Procedia Comput. Sci. 102, 317–324 (2016)
3. Codella, N., Rotemberg, V., Tschandl, P., Celebi, M., Dusza, S., Gutman, D., Helba,
B., Kalloo, A., Liopyris, K., Marchetti, M.: Skin lesion analysis toward melanoma
detection 2018: a challenge hosted by the international skin imaging collaboration
(ISIC). arXiv:1902.03368 (2019)
4. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S.,
Courville, A., Bengio, Y.: Generative adversarial nets. In: Advances in neural infor-
mation processing systems, pp. 2672–2680 (2014)
5. Kazeminia, S., Baur, C., Kuijper, A., van Ginneken, B., Navab, N., Albarqouni, S.,
Mukhopadhyay, A.: GANs for medical image analysis. arXiv:1809.06222 (2018)
6. Xue, Y., Xu, T., Zhang, H., Long, L.R., Huang, X.: Segan - adversarial network
with multi-scale L1 loss for medical image segmentation. Neuroinformatics 16(3–4),
383–392 (2018)
7. Frid-Adar, M., Diamant, I., Klang, E., Amitai, M., Goldberger, J., Greenspan, H.:
GAN-based synthetic medical image augmentation for increased CNN performance
in liver lesion classification. Neurocomputing 321, 321–331 (2018)
8. Iqbal, S., Ghani, M.U., Saba, T., Rehman, A.: Brain tumor segmentation in multi-
spectral MRI using convolutional neural networks (CNN). Microscopy Res. Tech.
81(4), 419–427 (2018)
9. Kingma, D.P., Ba, J.: Adam: a method for stochastic optimization. In: International
Conference on Learning Representations (ICLR) (2014)
Learning Channel-Wise Ordered
Aggregations in Deep Neural Networks
Iris Dominguez-Catena(B) , Daniel Paternain , and Mikel Galar
Institute of Smart Cities, Public University of Navarre, Pamplona, Spain

{iris.dominguez,mikel.galar,daniel.paternain}@unavarra.es
Abstract. One of the most common techniques for approaching image

classification problems are Deep Neural Networks. These systems are
capable of classifying images with different levels of detail at different
levels of detail, with an accuracy that sometimes can surpass even man-
ual classification by humans. Most common architectures for Deep Neu-
ral Networks are based on convolutional layers, which perform at the
same time a convolution on each input channel and a linear aggregation
on the convoluted channels. In this work, we develop a new method for
augmenting the information of a layer inside a Deep Neural Network
using channel-wise ordered aggregations. We develop a new layer that
can be placed at different points inside a Deep Neural Network. This
layer takes the feature maps of the previous layer and adds new feature
maps by applying several channel-wise ordered aggregations based on
learned weighting vectors. We perform several experiments introducing
this layer in a VGG neural network and study the impact of the new
layer, obtaining better accuracy scores over a sample dataset based on
ImageNet. We also study the convergence and evolution of the weight-
ing vectors of the new layers over the learning process, which gives a
better understanding of the way the system is exploiting the additional
information to gain new knowledge.
Keywords: Neural nets · RNN · Deep Learning · Ordered

aggregations
1 Introduction
Image Classification can be considered one of the most common problems to be
solved with Deep Learning and Convolutional Neural Networks (CNNs). CNNs
are neural networks that allow us to work on data with spatial structures, such as
images, where the meaning of a pixel depends strongly on its neighborhood. This
is achieved in CNNs by applying convolutional operations that impose local con-
nectivity constraints on the network weights. Most CNN architectures alternate
these layers that exploit local information with pooling layers that down-sample
This work was partially supported by the Public University of Navarre under the
projects PJUPNA13 and PJUPNA1926.
1024 I. Dominguez-Catena et al.
the feature maps, making the network resistant to local translations. This archi-
tecture means that, for most CNNs, only local information is exploited in early
layers of the network, and only by down-sampling will we combine this informa-
tion into a more global perspective in later layers.
Our main hypothesis is that some of this global information could be recov-
ered via ordered weighted aggregations using global metrics, and used on the
convolutional layers of the network to gain new knowledge. Thus, our main
objective is to implement a new layer that augments the feature maps of a CNN
network, generating information with channel-wise ordered weighted aggrega-
tions, and study the impact on the system. We are interested in both the impact
on the final performance of the network and the convergence of the weighting
vectors of the new aggregations.
The rest of this work is as follows. In Sect. 2 we will overview the literature
related to our work. Then, in Sect. 3 we will develop our proposed methodology.
In Sect. 4 we will present a case of study, with a specific implementation of the
method and the results we have found. Finally, in Sect. 5 we will conclude this
work and propose some future work lines.
2 Related Work
In the literature we can find several approaches for applying ordered weighted
aggregations, usually Ordered Weighting Averaging operators (OWA for short)
and other fuzzy measures to CNNs [1,3–5,7,9,10,13].
The most common approach is to use fuzzy measures to aggregate the result
of an ensemble of classifiers, in the image classification case usually CNNs [1,5,9,
10]. In these systems, we have several independent classifiers, and the aggregation
only operates on the results of the classifiers.
Another approach is to use Fuzzy Measure-based operators in the pooling
layers of CNN classifiers [3,4], replacing the common aggregations used for pool-
ing, maximum and average.
A third interesting approach are Linear Order Statistic Neurons [13], where
the neuron at the core of neural networks is redesigned based on OWA operators.
Finally, the work presented here is mainly inspired by previous attempts
to employ OWA operators to summarize the information on a CNN [7]. The
authors proposed the creation a “Fuzzy Layer”, designed to get the information
in a certain point of the network and replace it with the result of applying six
predefined OWA operators (max, min, soft-max, soft-min, average and a random
operator) channel-wise, sorting the channels by entropy. Our proposal works in a
similar fashion, but with several key differences. The first one is that we will try
to augment the information of the network, instead of replacing it. The second
one is that we will not use predefined operators, but we will allow the network to
learn the weights of the aggregations. Finally, we will not use OWA operators, as
the common constraints make the learning progress more difficult, and instead,
we will use a generalized version that does not constraint the weights.
Learning Channel-Wise Ordered Aggregations in Deep Neural Networks 1025
3 Proposal
To develop our proposal, we will first present the general layer structure in
Sect. 3.1. The, in Sect. 3.2 we will present the metric used for sorting image
channels. Finally, we will consider the way of aggregating channels in Sect. 3.3.
3.1 Layer Structure

Our proposed augmented layer structure will be a combination of ordering and
aggregation, used to augment the information in the network. We should note
that by combining ordering and aggregation, we are creating an approximation
of a channel-wise OWA operator [14], with some of the restrictions that make
them averaging operators lifted. For these aggregations to be real OWA, we
n
would require that wi ∈ [0, 1] for every i = 1, . . . , n and i=1 = 1. In our
experience these constraints impair learning, so we avoid this constraint for the
practical implementation of OWAs.
The layer will be placed in a CNN between two layers, and will take as input
the image of resolution I × J with Cin channels, each channel being a feature
map corresponding to a convolutional filter of the previous layer. As output,
the layer will generate a new image of I × J resolution, but with additional
channels making a total of Cout number of channels, with Cout = Cin + Cf eat .
To generate the new output, we will take the input channels of an image, sort
them according to a channel-wise metric (explained in Sect. 3.2), and aggregate
them with Cf eat channel-wise weighted aggregations (explained in Sect. 3.3).
Finally, we will concatenate the new Cf eat generated channels with the original
Cin channels to preserve the original information of the network.
3.2 Channel-Wise Ordering Metrics

To sort the channels we define an ordering metric, a function that will take as
input a single channel image X of size I × J, and output a single m(X) measure
value corresponding to that value.
For this work, we have to chosen to use as the ordering metric the Total
Variation [8]. This metric can be defined as:
I
J
T Vv (X) = |xi,j − xi−1,j | (1)
i=2 j=1
I
J
T Vh (X) = |xi,j − xi,j−1 | (2)
i=1 j=2
T V (X) = T Vv (X) + T Vh (X) (3)

This metric measures the absolute differences between each pixel and its
horizontal and vertical neighbors. This will result in a higher T V for images
with a lot of crisp borders or peak values, and lower T V for flatter images with
constant pixel values.
3.3 Channel-Wise Weighted Aggregation
For the aggregation, we use a simple weighted aggregation, with a single weight
per channel. We will perform several of these aggregations, thus taking an input
image X of Cin channels and a resolution of I rows J, and producing a new image
of Cf eat channels with the same resolution of I ×J, generating each channel with
an independent weighted aggregation of the input image.
For these aggregations, we impose the restriction that all the weights Wi > 0
for i ∈ (1, Cin ). This restriction is implemented by using a ReLU over the raw
weights, where:
ReLU (X) = max(x, 0) (4)
The matrix of weights (Cf eat × Cin ) corresponding to these aggregations
(where one row corresponds to one aggregation) are initialized consisting of
random numbers following a normal distribution N (0, 1). Since we work with
positive weights, we obtain the absolute values of the generated random numbers.
After that, these weights are considered in the same as the rest of the param-
eters in the network, in such a way that they are learned by back-propagation.
4 Case of Study
As a test for the proposed architecture, we implement the layer inside a VGG13
network [11]. This is a well known CNN architecture with 10 convolutional
blocks, each one composed of a convolutional layer, a batch normalization layer
and a ReLU activation layer. We test our new proposed layer by inserting it on
the points marked as P1 and P2 , one in the early layers of the network and one
in the later layers. The layer structure is presented in the Table 1.
For evaluating our proposal, we consider the Imagenette dataset, a subset of
ImageNet [2]. This dataset includes 10 classes from the original ImageNet, each
one with around 1,300 training images and 50 test images, for a total of 12,894
training images and 500 test images. These images are in color, with 3 RGB
channels, and a variable image size. We will resize all of them to a 256 × 256
resolution upon loading. The implementation of this experiment is done using
PyTorch 1.3.1 and Fastai 1.0.58.
We use as a reference an unmodified version of the network. For both the
reference and the test configurations, we run 10 repetitions, each one training
the network from scratch for 10 epochs, following the 1cycle policy [12], with
a maximum learning rate of 6 × 10−5 . We then get the last accuracy score
of each run, and compute both the average, and the non-parametric Mann-
Whitney U test [6] comparing each configuration with the reference network.
For the statistical test, the null hypothesis is that the configuration has the
same performance as the reference, and we will consider the threshold of 0.05
for the resulting p-value.
The accuracy results of the reference and our two test configurations, one
with insertion point in P1 and 64 learned feature maps, and the other with
insertion point in P2 and 16 feature maps, are shown in the Table 2. We choose
Table 1. Network architecture.
Name Kernel size Stride Output size

input data - - 256 × 256 × 3
conv1 1 3×3 1 256 × 256 × 64
conv1 2 3×3 1 256 × 256 × 64
maxpool 2×2 2 128 × 128 × 64
conv2 1 3×3 1 128 × 128 × 128
P1 - - 128 × 128 × (128 + Cf eat )
conv2 2 3×3 1 128 × 128 × 128
maxpool 2×2 2 64 × 64 × 128
conv3 1 3×3 1 64 × 64 × 256
conv3 2 3×3 1 64 × 64 × 256
maxpool 2×2 2 32 × 32 × 256
conv4 1 3×3 1 32 × 32 × 512
conv4 2 3×3 1 32 × 32 × 512
maxpool 2×2 2 16 × 16 × 512
conv5 1 3×3 1 16 × 16 × 512
P2 - - 16 × 16 × (512 + Cf eat )
conv5 2 3×3 1 16 × 16 × 512
maxpool 2×2 2 8 × 8 × 512
Linear - - 256
Linear - - 256
Linear - - 10
this learned feature map numbers to keep a proportion of one generated feature
map for every 8 original layers at that point. We can observe an improvement
of 1% with both configurations, with a more marked improvement in the second
configuration (on a higher insertion point and using more learned features). Both
configurations have statistically significant results, with a p − value ≤ 0.05.
We present some of the weighting matrices in the Fig. 1. The first column
corresponds to a run of the first configuration (with insertion point in the new
layer in P1 and 64 learned feature maps), while the second column to a run of
Table 2. Accuracy scores.
Insertion point Cf eat Accuracy p − value

Reference - 89.45 ± 0.52 -
P1 16 90.48 ± 0.55 0.0079
P2 64 90.78 ± 0.51 0.0016
(a) Insertion point P1 , Cf eat = 16 (b) Insertion point P2 , Cf eat = 64
Fig. 1. Normalized weighting vector evolution.
the second configuration (with insertion point in P2 and 16 feature maps). On

both columns, from top to bottom, the weighting matrices are plotted on each
epoch of the training.
We can observe a clear and fast convergence in both cases, where after the
third epoch the matrices remain stable. On both configurations we can observe
very clear patterns, similar to soft-max and soft-min operators. These patterns
seem to repeat for each aggregation, converging to just two main models, one for
the min and one for the max. Both configurations evolve with similar speed and
to similar patterns, although with a bit more noise on the configuration with
insertion point in P1 . This supports the theory that the network is extracting
knowledge from these aggregations, and the fact that the models are smooth
proves that the channel metric that we are using is not generating a fixed order
(which would defeat the purpose of using ordered aggregations).
5 Conclusion
We have presented and studied the augmentation of a CNN with information
derived from the application of ordered weighted aggregations to its feature
maps. The study shows that the systems is able to extract knowledge from the
new layer, obtaining an improvement in accuracy and showing a clear conver-
gence of the weighting matrices corresponding to the new layer.
Further research is needed to stabilize these results, studying the impact of
the layer on different points of the network, different network architectures, and
under different circumstances.
We also believe that the general architecture proposed for feature map aug-
mentation could be extended with different techniques, using different fuzzy
measures and other methods of generating the new channels.
References
1. Anderson, D.T., Scott, G.J., Islam, M.A., Murray, B., Marcum, R.: Fuzzy choquet
integration of deep convolutional neural networks for remote sensing. In: Compu-
tational Intelligence for Pattern Recognition, pp. 1–28. Springer, Cham (2018)
2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: a Large-Scale
Hierarchical Image Database. In: CVPR 2009 (2009)
3. Dias, C.A., Bueno, J.C.S., Borges, E.N., Botelho, S.S.C., Dimuro, G.P., Lucca, G.,
Fernandéz, J., Bustince, H., Drews Junior, P.L.J.: Using the choquet integral in
the pooling layer in deep learning networks. In: Barreto, G.A., Coelho, R. (eds.)
Fuzzy Information Processing, pp. 144–154. Springer, Cham (2018)
4. Dias, C.A., Bueno, J.C.S., Borges, E.N., Lucca, G., Santos, H., Dimuro, G.P.,
Bustince, H., Junior, P.L.J.D., Botelho, S.S.C., Palmeira, E.: Simulating the
behaviour of choquet-like (pre) aggregation functions for image resizing in the
pooling layer of deep learning networks. In: International Fuzzy Systems Associa-
tion World Congress, pp. 224–236. Springer, Cham (2019)
5. Du, X., Zare, A.: Multiple instance choquet integral classifier fusion and regression
for remote sensing applications. IEEE Trans. Geosci. Remote Sens. 57(5), 2741–
2753 (2019)
6. Mann, H.B., Whitney, D.R.: On a test of whether one of two random variables
is stochastically larger than the other. In: The Annals of Mathematical Statistics,
pp. 50–60 (1947)
7. Price, S.R., Price, S.R., Anderson, D.T.: Introducing fuzzy layers for deep learning.
In: 2019 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), pp. 1–6
(2019)
8. Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal
algorithms. Physica D 60(1), 259–268 (1992)
9. Scott, G.J., Hagan, K.C., Marcum, R.A., Hurt, J.A., Anderson, D.T., Davis, C.H.:
Enhanced fusion of deep neural networks for classification of benchmark high-
resolution image data sets. IEEE Geosci. Remote Sens. Lett. 15(9), 1451–1455
(2018)
10. Scott, G.J., Marcum, R.A., Davis, C.H., Nivin, T.W.: Fusion of deep convolu-
tional neural networks for land cover classification of high-resolution imagery. IEEE
Geosci. Remote Sens. Lett. 14(9), 1638–1642 (2017)
11. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale
image recognition. arXiv 1409.1556 (2014)
12. Smith, L.N.: A disciplined approach to neural network hyper-parameters: part 1 –
learning rate, batch size, momentum, and weight decay. arXiv 1803.09820 (2018)
13. Veal, C., Yang, A., Hurt, A., Islam, M.A., Anderson, D.T., Scott, G., Keller, J.M.,
Havens, T.C., Tang, B.: Linear order statistic neuron. In: 2019 IEEE International
Conference on Fuzzy Systems (FUZZ-IEEE), pp. 1–6 (2019)
14. Yager, R.R.: On ordered weighted averaging aggregation operators in multicriteria
decisionmaking. IEEE Trans. Syst. Man Cybern. 18(1), 183–190 (1988)
Design of Nano-scale Synaptic Neural Network
Using RRAM
Ali Mohamed1(&) and Osama Rayis2

1
University of Garden City, Khartoum, Sudan
[email protected]
2
Sudan University of Science and Technology, Khartoum, Sudan
[email protected]
Abstract. An electronic design of a leaky integrate-and-fire (LIF) artificial

neuron built with one Vanadium Dioxide Resistive Random Access Memory
switch (RRAM VO2 switch), which samples the functionalities of biological
neurons, is presented. Built with (RRAM VO2 switch) neurons, a two layer
spiking neural network is modeled in the LTSPICE simulator. A one transistor
one resistor (1T1R) design which integrates CMOS and RRAM in a structure
serving as a synaptic neural connection is proposed.
Keywords: RRAM Artificial neurons Biosimilar 1T1R Brain computing
1 Introduction
Improving integrated circuits is going in two main directions. The first direction is
replacing the conventional electronics by new Nanoelectronic devices like Molecular
electronics, Nanotubes, and Nanowires. This direction does not manifest its compe-
tence; as well it require novel infrastructure. The second direction is improving the
current electronic devices, put them in Nano-scale condition, and enhance their func-
tionalities by integrating with new Nanoelectronic devices. This work corresponds to
the second direction by integrating CMOS with RRAM. During past twenty years,
intensive research efforts in neuromorphic engineering domain guided to design Nano-
scale neuromorphic systems based on CMOS technology. Yet using mere CMOS
technology in designing dense synaptic connections does not meet the design needs
with regard to size, responsiveness, memory, and switching properties. This restriction
drove to search for components with properties of fast switching; low power con-
sumption, nonvolatile memory, and voltage-current features imitate neural networks.
These components like resistive random access memory (RRAM) and phase change
memory (PCM) can be integrated with CMOS to give similar neuron functions.
The work is written in the following layout: the introduction is presented in Sect. 1,
artificial neuron built with RRAM in Sect. 2, vision of integrating CMOS and RRAM
in Sect. 3, and the conclusion is written in Sect. 4.
1032 A. Mohamed and O. Rayis
2 Artificial Neuron Built with RRAM

2.1 RRAM Neuron Model
The RRAM (simulated with 50 50 nm bi-stable two-electrode VO2 switch)-neuron
model is designed based on the LIF model, which can produce biosimilar spikes [1, 2].
The function of RRAM is based on phase transition of the metal-insulator in VO2 films,
which occurs around temperature Tth = 340 K. [3], also in electric switching process
the electric field effect on charge carriers is considered [4, 5]. The I – V curve of
RRAM is compatible with practical I – V features (Fig. 1), got in this work, on RRAM
with a channel size of 20.5–30 nm, film thickness of *50 nm, and a series resistance
of 250 W.
Fig. 1. Practical I–V features of RRAM
Figure 2 indicates the dependence of RRAM current Isw on the supplied voltage
Vsw, RRAM passes from a high to a low-resistance state and vice versa when threshold
voltage Vth = 5.6 V and hold voltage Vh = 2.2 V are reached. The high and low-
resistance branches are approximated with resistance values Roff *14 kW and Ron
*300 W, depending on the I-V characteristics and voltage Vsw. To perform SPICE
simulations on RRAM artificial neuron, a controlled voltage- switch is used with
parameters (Roff, Ron, Vth, and Vh). The electronic design of RRAM neuron is shown in
Fig. 2. The neuron design has n inputs, and one output Vout. Resistances R1w . . .Rnw
represent synaptic weights between neurons. The i-th input signal to a neuron is
inversely proportional to the resistance. The input spikes are added and charge the
capacitance Csum. The capacitance Csum is discharged through the resistance Rin and
affects the neuron current. The supply voltage Vdd is adjusted less than the switching
voltage Vth, so that RRAM stays off when no input signal. RRAM is off when its
voltage is less than threshold voltage; (Vsw < Vth) and so the neuron is inactive, and
when RRAM is on (Vsw > Vth); the neuron is active. This is done by applying negative
Vdd and positive spikes input.
The neuron will be active when the capacitor charges and reaches a threshold
voltage Vc-th. When RRAM is switched on its resistance fall to Ron which leads to
discharging capacitance Cc through resistances Rin and Rout, and due to that a positive
spike is generated at Vout. By connecting the outputs of some neurons to the inputs of
other ones, SNNs with excitatory coupling can be obtained. Rs is adjusted so that
Design of Nano-scale Synaptic Neural Network Using RRAM 1033
Fig. 2. Electronic design of RRAM neuron.
RRAM switches off after the capacitance Cc discharges and so the neuron circuit
generates a spike through RRAM.
Figure (3a) shows the electronic circuit of RRAM neuron. The pulses voltage
generator is applied to the input of the neuron, and the output is connected to the input
stage of the next neuron. Resistance and capacitance values: R1w = 500 W, R2w = 1 kW,
Rs = 700 W, Rin = 1 kW, Rout = 10 kW, Csum = 1 nF, Cc = 10 nF, and Vdd = –5.75 V,
with applied positive pulse of amplitude 2 V and width of 0.3 µs. Figure (3b) describes
the oscillograms of the input Vin and output Vout, as well as the voltages Vc of
capacitance Csum, which indicates the spikes drive.
Fig. 3. (a) electronic design of RRAM neuron activated by a voltage generator, and
(b) oscillograms of voltages Vin, Vout, and Vc showing the spikes’ drive.
To initiate the output spike, the threshold voltage needed at Csum is Vc_th *0.33 V
(dashed strait line in Fig. (3b)). When RRAM switches on, the capacitance Cc begins
discharging; as a result a spike with voltage amplitude of *3.2 V initiates. After
switching on RRAM, Vc falls to negative values according to active recharging of Csum
resulted of the negative voltage on Cc and Vdd. When (Vsw > Vh), the spike width is
*170 ns, which is function of the discharging time of Cc,. The cantilever tip of the
pulse appears when RRAM goes off. The initiating pulse width can be far smaller than
output spike width. Figure (4a) shows that the delay between input and output leading
edges of pulses is inversely proportional to the amplitude of the input pulse, called
spike latency. The latency between the input and output signals are 140 ns for pulse
amplitude of 2 V, and are 440 ns for 1 V. However, the output pulse will not start if the
input pulse is less than Vc_th. If the input pulse is appropriately long, various spikes can
be at the output. Figure (4b) shows the reaction of RRAM neuron to a pulse with
amplitude 1 V and width 3.6 µs, which forms five spikes at the output. The refractory
period (Tr) (delay between the spikes) which depends on the pulse amplitude, is
approximately 630 ns and is specified by time of charging capacitor Csum to voltage
Vc_th. For instance, the period Tr is 300 ns for pulse amplitude of 2 V, as well the
refractory period Ts is a bit rising Fig. (4b), resulted of the little rise in Vc_th from spike
to spike, since there is no enough time for capacitance Cc to charge to its full values.
The rise in the period between the spikes under constant exposure is similar to bio-
logical neurons (spike frequency adaptation).
Fig. 4. (a) Oscillograms of Vin, Vc, and Vout, applied to RRAM-artificial neuron (b) one long
pulse (c) two pulses separated by small delay time
If subsequent spikes delay is less than the refractory period, the artificial neuron
generates only one spike. Figure (4c) displays two input pulses with amplitude of 1 V and
a delay of 300 ns, and the artificial neuron generate a spike only for the first input pulse.
To perform the Winner Take All rule (WTA), inhibitory connections do great task in
the SNN output layer. Such connections allow just one neuron, which is associated with a
specific class, to be activated, so the first neuron-generated spike deactivates all other
linked neurons using the inhibitory connections. Figure (5a) shows inhibitory connec-
tions of two neurons interconnected via capacitances Cinh = 10 nF. The values of
capacitance and resistance correspond to the single artificial neuron design shown in
Fig. (3a), except Rin = Rout = 200 X. The capacitance Cc voltage of an inactive neuron
drops as a result to the existence of Cinh capacities, during activation of one of the neurons
and the discharging of its capacitance Cc, so, the first (in time) activated neuron will inhibit
all other neurons linked to it by inhibitory connections. It is proper to monitor the current
Isw and voltage Vsw on two switches, in order to trace the activation of neurons in this
circuit (Fig. (5b)). The delay between the two applied pulses Vin_1 and Vin_2 is 2 ls.
When the first pulse Vin_1 strikes at the first switch (Fig. (5b)), the switch turns on, the
current Isw_1 rises sharply, and the on mode continues for *4.2 ls. Switching on occurs
because the voltage Vsw_1 reaches the threshold value Vth (Fig. (5b)).
Fig. 5. (a) Design of two oscillators with inhibitory connections. (b) Oscillograms of the input
signals Vin_1, Vin_2, switch voltages Vsw_1, Vsw_2 and switch currents Isw_1, Isw_2, and
(c) oscillograms of input signals Vin_1, Vin_2 and voltages of capacitors Vc_1, Vc_2, when two
voltage pulses separated with a delay of 2 µs are applied.
The voltage Vsw_1 falls quickly after turning on the first switch, leading to a
decrease in voltage Vsw_2 on the second switch, as the signal propagates through the
capacitors Cinh. The second pulse striking the input of the second neuron (Vin_2) will
not activate it, because the voltage Vsw_2 is less than the threshold value (Vsw_2 < Vth).
If an excitation pulse is applied to the second neuron after deactivation of the first
neuron, then the second neuron will be activated.
The Isw current pulses must be converted to output voltage pulse to determine the
activity of output layer neurons. The voltages Vc_1 and Vc_2 can be used as activity
markers and their drive are shown in Fig. (5c)
2.2 Spiking Neural Network Architecture

A simple model of SNN can be presented by a two-layer network (Fig. (6a)), where
information is supplied to the input (first layer), and one of the neurons related to a
certain class of information is activated at the output (second layer) [6, 7]. The con-
nection strength between any two neurons is set by changeable synaptic weights. When
signals are applied to the first layer neurons, they are activated and release an excitation
signal to the second layer. The neuron of the first layer, which is activated first, sends
an inhibiting signal to the rest output neurons to inhibit them from activation. So the
WTA rule is executed, when data is classified by determining the only active neuron in
the output layer. The activation speed of the output layer neurons is directly propor-
tional to the input signals and synaptic weights between the specific output neuron and
each input layer neuron.
Figure (6b) presents a coupling diagram of the input and output layers neurons of a
pulsed neural network. A signal from the generator is applied to each neuron in the
input layer. The generator is connected to the neurons of the input layer by an exci-
tation connection. All neurons in output layer are connected to inhibiting connections.
This is done by connecting the neurons of output layer to the inhibiting bus using the
Fig. 6. (a) Design of neural network which consists of two layers and (b) performance of input
and output layers neurons
capacitance Cinh = 10 nF. The other components, in Fig. (6b), have the following
estimations: Rw = 500 W, Rs = 700 W, Rin_i = 1 kW, Rin_j = 200 W, Rout_i = 10 kW,
Rout_j = 200 W, Csum = 1 nF, and Cc = 10 nF. The supply voltage for all neurons is
Vdd = –5.75 V.
The network training process depend on STDP technique [8, 9]. The technique is
an execution of the Hebbian learning rule and causes a change in synaptic weight
which is a function of the delay Δtin-out between pre and post-synaptic spikes. Since
synaptic weight is inversely proportional to resistance Rw_i, j, the function is inverted in
relation to the axes of an ordinate. Resistance falls, if the spike of post-synaptic (from
output layer neurons) arrives with a delay [0, 0.5] µs after the spike pre-synaptic (from
input layer neurons). Otherwise, the resistance Rw_i, j increases as shown in Fig. 7.
Fig. 7. The value of resistance between the input and output neuron ΔRw_i, j changes depending
on the delay between pre and post-synaptic spikes Δtin out.
3 Vision of Integrating CMOS and RRAM
The major challenge consists of executing the synaptic plasticity designs considered at
the basis of biological learning such as Spike-timing-dependent plasticity (STDP) and
spike-rate-dependence plasticity (SRDP) at RRAM-CMOS integrated device level. To
perform synaptic plasticity hardware design, the integration of RRAM devices and
CMOS serving as both cell selectors and current limiters has been widely used leading
to the design of hybrid synaptic structures such as the one-transistor/one-resistor
(1T1R) structure and the four-transistors/one-resistor (4T1R) structure. Figure 8 shows
a circuit schematic where a hybrid structure based on serial connection of a Ti/HfO2/

TiN RRAM cell and a metal oxide semiconductor transistor; referred to as 1T1R
structure, works as electronic synapse connecting a Pre to a Post synaptic with an
integrate-and-fire (I&F) design. This building block is designed to achieve STDP rule
which is considered one of the key techniques organizing learning in breasted organ-
isms. According to STDP rule, synaptic weight can vary as a function of relative time
delay Dt between spikes released by Pre and Post synaptics. If the Pre spike overrides
the Post spike, Dt is positive resulting in an increase of synaptic weight. Otherwise, if
Pre spike generation delayed a little after the Post spike, Dt has negative value leading
to a decrease of synaptic weight. The STDP performance in 1T1R synapse is achieved
as follows: when the Pre synaptic transmits a 10-ms-width voltage pulse at MOS gate, a
current proportional to RRAM conductance flows across the synapse since its TE is
biased by a continuous voltage with low amplitude used for communication phase.
This current thus enters Post synaptic where it is integrated, causing an increase of Post
membrane/internal potential Vint. As this integral signal reaches a threshold Vth, the
Post sends both a forward spike toward next neuron layer and a suitably-designed spike
including a 1-ms- pulse width with positive amplitude followed, after 9 ms, by a 1-ms-
width negative pulse, which is backward delivered at TE to activate a synaptic weight
update according to STDP rule. Here, as the Pre spike overrides the Post spike
(0 < Dt < 10 ms), Pre voltage interferes only with the positive pulse of the Post spike,
thus causing a set transition within RRAM. Otherwise, if the Pre spike delayed after the
Post spike (−10 ms < Dt < 0), interference occurs between Pre spike and the negative
pulse in the Post spike, thus activating a reset transition in RRAM.
Fig. 8. A 1T1R RRAM structure serving as synaptic connection between a Pre circuit and a Post
circuit with I&F architecture.
4 Conclusion and Recommendations
A new model of an LIF neuron based on one RRAM is presented. During the simu-
lation, the RRAM artificial neuron model demonstrates biosimilar properties. A two-
layer SNN was designed. The coupling between the neurons of the input and output
layers was implemented using excitatory connections, and, inside the output layer, the
coupling used inhibitory connections. A 1T1R RRAM structure serving as synaptic

connection between a Pre circuit and a Post circuit with I&F architecture is proposed.
The following recommendations are essential for future work: Focus research
efforts on new Nano-scale devices like non-volatile resistive random access memory
(RRAM), and explore its unique properties and capabilities to simulate synaptic neu-
rons. RRAM is compatible to be integrated with conventional devices used in common
computation industry like CMOS, so it is worth to search rising functionalities of
conventional devices by merging with new Nano-scale devices. Support the orientation
of merging non volatile memory with logic unit, which will open the door for new
generation of very high speed, low power consumption, and light computers which use
new computation techniques derived from human brain computing, and to overcome
the so-called von Neumann bottleneck of conventional hardware, novel non-von
Neumann computing paradigms have been intensively explored with a view of
bringing data processing closer to where data are stored. In this wide range, neuro-
morphic computing has emerged as one of the most promising approaches since it aims
at improving dramatically computation mainly in terms of energy efficiency taking
inspiration from how the human brain processes information via biological neural
networks.
References
1. Pergament, A., Velichko, A., Belyaev, M., Putrolaynen, V.: Electrical switching and
oscillations in vanadium dioxide. Phys. B Condens. Matter. 536, 239–248 (2018)
2. Belyaev, M.A., Boriskov, P.P., Velichko, A.A., Pergament, A.L., Putrolainen, V.V.,
Ryabokon’, D.V., Stefanovich, G.B., Sysun, V.I., Khanin, S.D.: Switching channel
development dynamics in planar structures on the basis of vanadium dioxide. Phys. Solid
State 60, 447–456 (2018)
3. Karda, K., Mouli, C., Ramanathan, S., Alam, M.A.: A self-consistent, semiclassical
electrothermal modeling framework for Mott devices. IEEE Trans. Electron Devices 65,
1672–1678 (2018)
4. Yi, W., Tsang, K.K., Lam, S.K., Bai, X., Crowell, J.A., Flores, E.A.: Biological plausibility
and stochasticity in scalable VO2 active memristor neurons. Nat. Commun. 9, 4661 (2018)
5. Tavanaei, A., Ghodrati, M., Kheradpisheh, S.R., Masquelier, T., Maida, A.: Deep learning in
spiking neural networks. Neural Netw. 111, 47–63 (2019)
6. Jeong, H., Shi, L.: Memristor devices for neural networks. J. Phys. D Appl. Phys. 52, 023003
(2019)
7. Kwon, M.-W., Baek, M.-H., Hwang, S., Kim, S., Park, B.-G.: Spiking neural networks with
unsupervised learning based on STDP using resistive synaptic devices and analog CMOS
neuron circuit. J. Nanosci. Nanotechnol. 18, 6588–6592 (2018)
8. Zhou, E., Fang, L., Yang, B.: Memristive spiking neural networks trained with unsupervised
STDP. Electronics 7, 396 (2018)
9. Yousefzadeh, A., Stromatias, E., Soto, M., Serrano-Gotarredona, T., Linares-Barranco, B.: On
practical issues for stochastic STDP hardware with 1-bit synaptic weights. Front. Mol.
Neurosci. 12, 665 (2018)
Neural Network-Based Control Framework
for SISO Uncertain System: Passive Fault
Tolerant Approach
Sejal Raval1, Himanshukumar R. Patel2(&) , and Vipul A. Shah2

1
Government Polytechnic, Ambawadi, Ahmedabad 380015, Gujarat, India
2
Dharmsinh Desai University, Nadiad 387001, Gujarat, India
[email protected],
[email protected]
Abstract. This brief deals with fault-tolerant control framework designed for
an uncertain SISO level control system using a neural network. The level control
system is common and very important in modern chemical processes, food
processing, refinery or engineering industries. They are widely used to process
or store some raw material for further processing or packaging of final product.
In the show of all level control methods, actuator and sensor play an important
part. Numerous difficulties rising in SISO level control system operations are
linked to actuator or system component faults. In various cases, the exactness of
the apparatuses and maneuvers used to observer and regulate the level control
scheme is extremely reliant on the dynamic performance of the system
manipulator (actuator), sensor, and system component (leak). In this paper, the
passive fault-tolerant control approach proposed using neural networks in which
residue signal features and control/manipulated variable signals are used for
training and thereafter predicting he control signal for actuator/sensor/system
component fault tolerance. The article proposed novel structure for fault-tolerant
control design with neural network for all possible faults in SISO level control
process with parameter uncertainty.
Keywords: Actuator fault System component fault Sensor fault SISO

uncertain system Passive fault tolerant control Neural network
1 Introduction
Fault Tolerant Control (FTC) of industrial process is a scientific discipline aimed at the
maintaining the acceptable control performance and the system stability under faulty
situation [1]. The main task of the FTC is to avoid that modest faults develop into
serious disaster and henceforth the system accessibility increased and minimize the
danger circumstances of safety vulnerability [1–3]. FTC has been the theme of
demanding investigation for the last thirty years [4, 5]. This investigation determination
has flourished in various real-world industrial applications [6, 7]. There are basic two
structured approach for FTC active and passive [8]. Passive FTC has designed based on
predetermined conditions and magnitude of the system faults and designed robust
controller for the system, other side in active FTC approach Fault Detection and
Diagnosis (FDD) is key component required, FDD has three important function
1040 S. Raval et al.
initialize with detection of the faults in process, then isolation and finally there iden-
tification [8, 9]. The attractiveness of passive FTC algorithms originates as of their
usual capability to deliberate all kind of possible faults at design level. The another
significant motive is that the working ideologies remain understandable plus compar-
atively informal to describe to beginners, which looks to be a critical part throughout
execution of a novel control system in engineering industries.
An important part of passive FTC strategies is a thorough knowledge of the system
and all kind of possible faults into the system. There are many controllers (i.e. H∞,
Model Predictive Control, Sliding mode control, PID controller, artificial intelligence
base control etc.) are used to design passive FTC algorithm [10–12]. In recent [13] the
author used hybrid control scheme to design passive FTC algorithm, the author used
conventional PID controller with fuzzy logic based regulator for SISO and MIMO level
control process with all possible faults. As pointed out in [14, 15], passive FTC
algorithms has stretched a great level of development, particularly for linear control
systems. To date, many methods of passive FTC for SISO/MIMO level control system
have been derived, e.g. PID + type-1 fuzzy logic control, PID + type-2 fuzzy logic
control, fractional order PID + type-1 fuzzy logic control and fractional order PID +
type-2 fuzzy logic control [16]. Certainly, neural networks are the furthermost standard
models used. They are actual flexible, common and freely used in belongings when
there is no precise mathematical model of the process [14]. That is the motive why
recurrent neural networks immerging to use in designing the one-step ahead
prognosticator.
The paper is systematized as follows. After Introduction, in Sect. 2, contains a
description of the SISO uncertain level control process for which the controller scheme
was designed. Then, in Sect. 3, the Passive FTC design is proposed using neural
network. The last section contains conclusions.
2 Model Description of the SISO Uncertain Level Control

System
2.1 SISO Uncertain Level Control System

The presented process in figure is the single-tank level control structure and contains of
single water chamber, single sump chamber, pneumatic control valve and one electrical
pump deposited in Fig. 1. The level control system was realized in Matlab.
Table 1. Specifications of process variables.

Variable Specification Range
CV1 Control valve/Actuator 0–100%
fi/fo Inlet/Outlet flow rate 0–5 m3/h
fsys Leak flow rate 0–1 m3/h
L Water level in tank 0–0.5 m
Neural Network-Based Control Framework for SISO Uncertain System 1041
Table 2. Specifications of faulty scenarios.

Fault Description Type Nature Size/Unit
fsys Leak flow rate Additive Intermittent 0–1 m3/h
fa Actuator fault in CV1 Multiplicative Intermittent 0–20%
While the explanation of the process variables are presented in Table 1.

Acronyms LS, LC, PM, CV1, CV, and V1 stand for level sensor, level controller, pump,
actuator/final control element, manual operated valve for leak, and outlet flow control
valve respectively. The specification of faults is accessible in Table 2. The system
parameter uncertainty pu consider in the SISO level control system in terms of math-
ematical modeling inaccuracy, the variation of time constant s parameter from 0 to
±10% and system gain k variation of 0 to ±10%.
Fig. 1. Schematic diagram of the single-tank level process [8].
2.2 SISO Level Control System Mathematical Modeling

The process input is (inlet flow rate to tank fi using CV1) and the output is (tank height).
The process model of the single-tank level system given by mass balance and Ber-
noulli’s law yields:
dh
A ¼ ðfi f0 Þ ð1Þ
dt
where,
dh
dt Rate of change of liquid height in tank,
A Cross section are of tank,
fi Inlet flow rate of tank,
f0 Outlet flow rate of tank.
From the process reaction curve method obtain the model of SISO level control
process as given following [5, 8],
5
G P ðsÞ ¼ ð2Þ
100s þ 1
3 Proposed Framework for Passive FTC
The proposed methodology for the Passive Fault Tolerant Control (PFTC) involves of
four stages explicitly data generation layer, the pre-processing layer, the training layer,
and control output prediction layer. The proposed methodology is shown in Fig. 2 and
explained in the subsections in detail.
Fig. 2. Proposed framework/methodology for Passive FTC.
3.1 Data Generation Layer

The residue (r) signal verses manipulated variable/control valve opening signal data for
this work is assimilated in the data generation layer.
For data generation, the contextual information, such as residue signal, manipulated
variable/control valve opening signal, and so forth; faulty and healthy mathematical
model of the SISO uncertain level control system have been used. For data generation
purpose in faulty SISO level control model consider intermittent actuator (fa) and system
component (leak) (fsys) fault and system parameter uncertainty (pu) ware consider. The
mathematical model of SISO level control system with fault and without fault is taken
from [8]. The conceptual diagram of the data generation system is shown in Fig. 3.
Fig. 3. Data generation system for SISO uncertain level control system model.
For the data generation purpose fault magnitude is consider (±0% to 20%) and
system uncertainty consider (±0% to 10%). From the data generation layer total 10000
data ware generated for different residue signals and appropriate different 10000 vales
generated for manipulated variable/control output signals in different 10 simulation.
3.2 Pre-processing Layer

In this layer, data has been preprocessed in order to make it smooth for additional
processing. Different smoothing filter can be used for this purpose, such as moving
average, loess, lowess, Rloess, RLowess, Savitsky–Golay, and so forth. In this work,
we have used the moving average method which is a very major data smoothing filter
used by many researchers for data smoothing. Equation (3) is the mathematical rep-
resentation of moving average filter.
1 XM1
y½ i ¼ X ði þ jÞ ð3Þ
M J¼0
In this calculation, x [] is the input, y [] is the output, and M is the figure of points
used in the moving average. In the pre-processing layer, first, we have calculated the
statistical moments and concatenated with original data. The dataset comprises one
parameter as input; namely residue (r) signal (P1), and one parameter as output;
explicitly control output/manipulated variable signal (M1). The statistical moments,
explicitly mean, variance, skewness, and kurtosis can be calculated using Eqs. (4)–(7).
1 Xn
l¼ x
i¼1 i
ð4Þ
n
1 Xn
r¼ i¼1 i
ð x lÞ 2 ð5Þ
n
3
1 X n xj l
S¼ ð6Þ
n i¼1 r
4
1 X n xj l
K¼ ð7Þ
n i¼1 r
Where l, r, S, K symbolize mean, variance, skewness, and kurtosis, corre-

spondingly. Pi denotes values of residue signal (P1) arithmetic mean of the data (P2),
standard deviation of the data (P3), skewness of the data (P4), and kurtosis (P5).
Values, i = 1, 2, 3 to 5. For trial and test purpose, we have normalized the data by using
Eq. (8).
x xmin
xnew ¼ ð8Þ
xmax xmin
Where xnew signify the output normalized value, x indicate the current value, xmin
signifies the minimum value in the set and xmax indicates maximum value [17].
3.3 Training Layer

The ANN model applied in the suggested work is the Feed Forward Back-Propagation
Neural Network (FFBPNN) as shown in Figs. 4 with original, normalized data, and
data with statistical moments. Inputs to the neural network are normalized data and
statistical moments data. We have used one parameter as inputs namely residue signal
(P1), and, arithmetic mean of the data (P2), standard deviation of the data (P3),
skewness of the data (P4), and kurtosis (P5) and one output control valve
opening/manipulated variable (M1) as shown in the following figures. It usually has
three layers, namely the input layer, hidden layer, and output layer.
Fig. 4. Structure of NN model for five inputs.
The detailed mathematical formulations of the artificial neural network are given
below which have been taken from Reference [17]. In order to calculate the hidden
layer value, the following Eq. (9) can be used.
Xn 1
#j ¼ 1 þ exp 1 i¼1
x i wij ð9Þ
where # j represents node j in the hidden layer, xj represents node I the input layer, and
the weight between nodes are represented by wij. The output layer node value can be
calculated by (10).
Xj 1
y ¼ 1 þ exp 1 j¼1
x i w ij ð10Þ
Where y symbolizes the output layer node (In this investigation, we have occupied only
one output node, many nodes can be used). Error E amongst experimental and cal-
culated data can be deliberate as (11).
Error ¼ 0:5ðd yÞ2 ð11Þ
Where d represents the observed data propagation from the output layer and a hidden
layer that is represented in Eqs. (12) and (13) respectively.
dy ¼ ðd yÞð1 yÞ ð12Þ

dy ¼ #j d #j ð1 yÞdy wj1 ; j ¼ 1; . . .. . .::J ð13Þ
The weight adjustment between hidden and output layers and input and the hidden
layer can be carried out using below amounts (14–15), correspondingly:
Dwij ¼ ady #j ; i ¼ 1; . . .. . .::I; j ¼ 1; . . .. . .; J ð14Þ
Dwnij ¼ ady #j ; j ¼ 1; . . .. . .; J ð15Þ
Where a signifies learning rate, moreover momentum can be dignified using Eqs. (16)
and (17):
Dwnij ¼ ady #j þ bDwn1

j1 ; j ¼ 1; . . .. . .; J ð16Þ
Dwnij ¼ ady #j þ bDwn1

j1 ; i ¼ 1; . . .. . .:; I; j ¼ 1; . . .. . .; J ð17Þ
Where n specifies iterations of error back-propagation; and b characterizes momentum

constant. The training process in the flat region of the error surface and avoids fluc-
tuations in the weights are accelerated by using this momentum method.
There are various kinds of activation functions, such as linear, tan-sigmoid, and
logarithmic sigmoid, sigmoid, and so forth, which can be used in different layers of
ANN. In the suggested work, we have used the tan-sigmoid function in the hidden
layer and linear function in the output layer. The choice of the tan-sigmoid function in
the hidden layer is supported out since it is the most suitable activation function and its
performance is measured superior as related to former activation functions. Likewise,
we have used the linear function in the output layer since it is a regression problem and
henceforth we required to put on linear function in the output layer. The linear and
sigmoid functions are characterized mathematically in Equalities (18) and (19)
correspondingly.
X ð xÞ ¼ linearðxÞ ð18Þ
2
;ð x Þ ¼ 1 ð19Þ
ð1 þ 2e2x Þ
The paper proposed novel framework for Passive Fault Tolerant Control for SISO
uncertain level control system consist of machine learning algorithm and conventional
PID controller. The control action estimate and modeling have always remained a
challenging task under system parameter uncertainty and unpredicted faulty situation.
To grab the challenge in this research paper, a robust and additional flexible control
scheme has been proposed for the prediction of control action and take appropriate
control action in SISO uncertain level control system. The estimation has been con-
ceded out with FFBPNN using simple data, normalized data, and data with statistical
moments of control output/manipulated variable in SISO uncertain level control system
for actuator, sensor, system component (leak) faults and system parameter uncertainty.
However, simulation results are absent from this paper for the validation of the prosed
FTC framework using neural network.
References
1. Patel, H.R., Shah, V.A.: Fault detection and diagnosis methods in power generation plants -
the indian power generation sector perspective: an introductory review. PDPU J. Energy
Manage. 2(2), 31–49 (2018)
2. Patel, H.R., Shah, V.A.: A framework for fault-tolerant control for an interacting and non-
interacting level control system using AI. In: Proceedings of the 15th International
Conference on Informatics in Control, Automation and Robotics, pp. 180–190 (2018).
https://doi.org/10.5220/0006862001800190
3. Patel, H.R., Shah, V.A.: A passive fault-tolerant control strategy for a non-linear system: an
application to the two tank conical non-interacting level control system. Maskay 9(1), 1–8
(2019)
4. Patel, H.R., Shah, V.A.: Fuzzy logic based passive fault tolerant control strategy for a single-
tank system with system fault and process disturbances. In: Proceedings of the 5th
International Conference on Electrical and Electronics Engineering (ICEEE), Istanbul,
Turkey, pp. 257–262 (2018)
5. Patel, H.R., Shah, V.A.: Performance comparison of passive fault tolerant control strategy
with PI and fuzzy control of single-tank level process with sensor and system fault. Am.
J. Eng. Appl. Sci. 12(2), 236–246 (2019)
6. Bhandare, D., Kulkarni, N. R.: Performances evaluation and comparison of PID controller
and fuzzy logic controller for process liquid level control. In: 15th International Conference
on Control, Automation and Systems, Busan, pp. 1347–1352 (2015). https://doi.org/10.
1109/iccas.2015.7364848
7. Bonivento, C., Isidori, A., Marconi, L., Paoli, A.: Implicit fault-tolerant control: application
to induction motors. Automatica 40(3), 355–371 (2004)
8. Patel, H.R., Shah, V.A.: Fault tolerant control systems: a passive approaches for single tank
level control system. J. Instr. Control Eng. 6(1), 11–18 (2018)
9. Patel, H.R., Shah, V.A.: Passive fault tolerant control system using feed-forward neural
network for two-tank interacting conical level control system against partial actuator failures
and disturbances. IFAC-Papers OnLine 52(14), 141–146 (2019)
10. Patel, H.R., Shah, V.A.: Stable fault tolerant controller design for takagi-sugeno fuzzy
model-based control systems via linear matrix inequalities: three conical tank case study.
Energies 12(11), 2221 (2019)
11. Patel, H.R., Shah, V.A.: Integrated design of model-based passive fault-tolerant control for
nonlinear systems based on PID and fuzzy control. In: Pant, M., Sharma, T.K., Verma, O.P.,
Singla, R., Sikander, A. (eds.) Soft Computing: Theories and Applications. Advances in
Intelligent Systems and Computing, pp. 155–169. Springer, Singapore (2020)
12. Patel, H.R., Shah, V.A.: Comparative study between fractional order PIkDl and integer order
PID controller: a case study of coupled conical tank system with actuator faults. In: 4th
Conference on Control and Fault Tolerant Systems, pp. 390–396. IEEE (2019)
13. Luzar, M., Witczak, M., Korbicz, J., Witczak, P.: Neural-network based robust FTC:
application to wind turbines. In: Rutkowski, L., Korytkowski, M., Scherer, R.,
Tadeusiewicz, R., Zadeh, L.A., Zurada, J.M. (eds.) Artificial Intelligence and Soft
Computing. ICAISC 2014. LNCS, pp. 97–108. Springer, Cham (2014)
14. Chen, M., Mei, R.: Actuator fault tolerant control for a class of nonlinear systems using
neural networks. In: 11th IEEE International Conference on Control and Automation. IEEE,
Taichung (2014)
15. Qikun, S., Bin, J., Peng, S.: Neural network-based fault tolerant control scheme against un-
modeled fault. Stud. Syst. Decis. Control 91, 163–190 (2017)
16. Patel, H.R., Shah, V.A.: Actuator and system component fault tolerant control using interval
type-2 Takagi-Sugeno fuzzy controller for hybrid nonlinear process. Int. J. Hybrid Intell.
Syst. 15(3), 143–153 (2019)
17. Fayaz, M., Kim, D.: A prediction methodology of energy consumption based on deep
extreme learning machine and comparative analysis in residential buildings. Electronics 7,
222 (2018)
Forecasting the Day-Ahead Prices in Electricity
Spot Market of Turkey by Using Artificial
Neural Networks
Berna Tektaş(&), Aygülen Kayahan Karakul, and Rozi Mizrahi
Faculty of Economics and Administrative Sciences,

Izmir Kâtip Celebi University, Cigli, Izmir, Turkey
{berna.tektas,aygulen.kayahan,
rozi.mizrahi}@ikcu.edu.tr
Abstract. Artificial Neural Networks (ANN) which is a part of artificial

intelligence is an effective tool of modelling the relations between variables
without any restrictive assumption of dataset. However, it is not widely used in
social sciences as in the other engineering or natural sciences. The purpose of
this study is to perform an application of ANN in modelling the relationships of
variables in the electricity stock market EXIST which matches the supply and
demand of Generation Companies (GenCo) and electricity distribution compa-
nies in Turkey. The output variable is the price of electricity in the stock market,
input variables are formed as the historical data of prices and days of the weeks.
Performance of network has been determined by multiple performance
indicators.
Keywords: Artificial Neural Networks Electricity price forecast Electricity

wholesale market
1 Introduction
Since the deregulation of the electricity markets and the establishment of energy
exchanges for the wholesale of electricity, the electricity price forecasts have become a
fundamental input to decision-making mechanism of all players of the energy exchanges
[1–3]. Forecasting the volatile spot prices of electricity with a reasonable accuracy can
provide generation companies (GenCos), distribution companies (DisCo), or large
industrial consumers a rational bidding strategy in the day-ahead market (DAM). Also
the share of regulated markets such as the day ahead market and the balancing power
market in electricity wholesale increases day by day, and, the prices of bilateral contracts
and energy derivatives have become based on these spot market prices. Therefore, the
players of the electricity market can use forecasts of spot electricity prices as an input for
their mid or short termed production or consumption planning optimization models.
Zareipour et al. [4] showed that a 1% improvement in the mean absolute percentage
error (MAPE) of short term electricity price forecasting would result in between 0.1 and
0.35% decrease in operation costs of electricity companies.
Forecasting the Day-Ahead Prices in Electricity Spot Market 1049
Because of its distinctive features such as non-storability, the requirement of supply

and demand equilibrium and demand inelasticity, the electricity prices generally more
volatile than the prices of any other commodity or financial assets [5]. Besides, factors
such as transmission bottlenecks, weather conditions, fuel prices or equipment outages
can also cause high volatility, extreme spikes and multiple seasonality, in the electricity
prices [6]. All these specific factors and features of the electricity makes its price
forecasting a challenging task.
In last 20 years, many different of methods and ideas are applied for electricity price
forecasting. Weron [7] reviewed electricity price forecasting literature and classified
these models into five groups. These groups consist of multi-agent models (e.g. Nash-
Cournot framework, supply function equilibrium, strategic production cost models and
agent-based simulation models), fundamental models (e.g. parameter-rich fundamental
models and parsimonious structural models), reduced-form models (e.g. jump-diffusion
models, Markov regime-switching models), statistical models (e.g. similar- day and
exponential smoothing methods, regression models, AR-type time series models,
ARX-type time series models, threshold autoregressive models, heteroscedasticity and
GARCH type models), and computational intelligence models (e.g. Feed-forward
neural networks, recurrent neural networks, fuzzy networks, support vector machines).
Each method has its own pros and cons. For instance, multi-agent models can take the
behavior of the rivals into account however, their accuracy are not high.
In this paper, we describe application of artificial neural networks for short-term
(day-ahead) electricity price forecasting of Energy Exchange Istanbul (EXIST) in
Turkey.
The remainder of this paper is structured in the following manner. Section 2 pro-
vides an overview of Turkish Electricity Market as well as insight to the problem
context. Section 3 describes the used artificial neural network (ANN) methodology and
the proposed framework for electricity price forecast model. Section 4 provides a
comprehensive analysis of the develop ANNs for forecasting day-ahead prices in
EXIST. Section 5 concludes the paper, offering perspectives relevant to further
research.
2 Electricity Stock Market in Turkey
Energy Exchange Istanbul (EXIST) electricity stock market of Turkey officially

established on 12.03.2015. The electricity requirement is balanced one day before the
physical delivery of the electricity in the DAM [8–10]. Both the suppliers (e.g. Gen-
Cos) and buyers (e.g. DisCos and eligible big customers) give their bids consisting
quantity and price pairs until 11:00 a.m. Deregulated electricity markets aim to meet
the required electricity demand with the lowest price. Therefore, the market operator
(EXIST) clears the market and determines the clearance price according to the inter-
section of supply and demand curves until 2:00 p.m.
Electricity markets differ from country to country for several reasons. Firstly, they
differ in the share of deregulated markets that they have. In the Turkish electricity
market, DAM (*60%) and bilateral contracts (BCs) (*35%) are the main part of the
electricity wholesale market. The other 5% share of the Turkish electricity wholesale
1050 B. Tektaş et al.
market is balance market (BM) (*2%) and imbalance (*3%). Thus, EXIST controls
*62–63% of the Turkish electricity market.
Kölmek and Navruz [11] forecast spot prices in Turkish electricity market with
both artificial neural network and ARIMA model. As [11], Ozguner et al. [12] propose
an ANN model to forecast the hourly prices in Turkish electricity market and compare
forecast performances of ANN with multiple linear regression. Ugurlu et al. [13] used
Gated Recurrent Units for electricity price forecasting and compared the results with
the Recurrent Neural Networks. Findings of these studies [11–13] show that ANN has
a little better performance than time series. Ozyildirim and Beyazit [13] compare
machine-learning method having radial basis function with the multiple linear
regression for electricity price prediction in Turkish electricity market. The results of
their study shows that almost there is no difference between the prediction perfor-
mances of the models.
Although Turkish day-ahead electricity market has an improving literature, still
there are limited numbers of research. This study is one of the pioneering studies for the
price forecasts in Turkish electricity market. As a nonlinear forecasting method, an
ideally constructed neural network is proposed in this paper and the MAPE and R2
values shows that the proposed neural network model gave good results.
3 Methodology
3.1 Artificial Neural Networks (ANN)

ANN is a modelling method and data processing technology that simulates the bio-
logical neural neurons. It searches the relations between variables with the help of the
learning algorithms and activation functions as in the cognitive functions of nervous
system. The components of ANNs as nodes and layers are shown in the Fig. 1.
ANN models the relationships between the variables without any a-priori assump-
tions as in the classical traditional methods.
Fig. 1. Feed forward neural network with three layers [14].

ANN can be grouped into sub groups according to their various characteristics.
According to structure of connections, there are feedback and feed-forward networks.
When the information is transformed from inputs to outputs direction, it is called the
feed forward neural networks while the feedback networks use the directed circles and
joints between neurons. Feed forward neural networks trained by back propagation is
called Multilayer Perceptron (MLP) [15]. MLP has input, hidden and output layers and
degree of accuracy of modelling depends on the hidden layers. It uses linear, sigmoid
or softmax functions for activation of hidden and output layers. Input data set is divided
randomly into three sets: Training, testing and validation. Learning and activation
functions are used to set up a model ensuring that the Sum of Errors are minimized.
According to learning model ANN can be classified into three as supervised, unsu-
pervised and reinforcement.
The degree of accuracy of model can be calculated with the coefficient of deter-
mination R2 and the Forecasting performance of network is determined by error cal-
culation methods. There are many kinds of error calculation methods in forecasting
models. In this research the Mean Absolute Percentage Error (MAPE) is used since it
gives the error with percentage deviation from real values, its formula is given in the
Eq. (1) where xt are real values, ^xt are estimated values and n is the size.
Pn xt ^xt
t¼1 ð xt Þ
MAPE ¼ 100 ð1Þ
n
4 Application
The data set is obtained from the official website of EXIST and it consist of the hourly
data from 01.01.2017 to 01.01.2020. Literature review show that the historical prices
are frequently used to estimate the day-ahead prices [16–18], and the expected load
(System potential demand – SPD) is another important indicator in the neural network
[16–18]. Temporal effect of days, hours, months and seasons are also taken into
account. The variables that used in the research are given in the Table 1.
Table 1. Inputs and output of the research

Category Variable
System Marginal Price: P (d, h) Output
Price for same hour of 1 day before: P (d-1, h) Input
Price for same hour of 3 days before: P (d-3, h) Input
Price for same hour of 7 days before: P (d-7, h) Input
System Potential Demand (SPD) Input
Days of week (1,…,7): d Input
Hours of day (1,…,24): h Input
Months of year (1,…,12): m Input
Seasons of year (1,…,4): s Input
Outliers are extracted and the dataset is divided into two as Training set and Testing
set with the proportions 90% and 10% respectively. MATLAB 2020a is used to verify
the ANN analysis. Firstly it is aimed to determine the most appropriate model con-
sisting the most relevant variables. Results are given in Table 2.
Table 2. Results of models

R
Training Test Validation All
Model 1 P (d-1, h), P (d-3, h) 0.83 0.84 0.83 0.84
Model 2 P (d-1, h), P (d-3, h), P (d-7, h) 0.86 0.85 0.86 0.86
Model 3 P (d-1,h), P (d-3,h), P (d-7,h), SPD 0.88 0.88 0.88 0.88
Model 4 P (d-1, h), P (d-3, h), P (d-7, h), SPD, h 0.89 0.89 0.88 0.89
Model 5 P (d-1, h), P (d-3, h), P (d-7, h), SPD, h, d 0.88 0.89 0.89 0.88
Model 6 P (d-1, h), P (d-3, h), P (d-7, h), SPD, h, d, m 0.88 0.89 0.89 0.88
Model 7 P (d-1, h), P (d-3, h), P (d-7, h), SPD, h, d, m, s 0.89 0.88 0.88 0.89
According to results the first three models are eliminated and the analysis is con-
tinued with the models of 4, 5, 6 and 7 which has relatively better results in R values.
Feed forward neural networks trained by back propagation on the testing sets of each
model. Trainlm used for training function, LearnGDM is used for adaption learning
function Tansig function is used for transfer function and the network is trained until
the Mean Square Error is minimized. Then to determine the best number of hidden
layers and best the number of neurons in each hidden layers many trials on ANN
applied and the best results on R are given in the Table 3.
Table 3. Results on different architectures of models

R
x-10-1 x-20-1 x-10-10-1 x-10-20-1 x-20-20-1
Model 4 0.89 0.88 0.87 0.89 0.90
Model 5 0.89 0.89 0.90 0.90 0.91
Model 6 0.89 0.90 0.90 0.90 0.91
Model 7 0.89 0.90 0.90 0.88 0.91
In Table 3, x is the number of input neurons. For example for model 4 x = 5 which
are P (d-1, h), P (d-3, h), P (d-7, h), SPD, h. According to results given in the Table 3,
the models that has an architecture of the x-20-20-1 are more successful in all models.
So the architecture is selected as it consist of 2 hidden layers with 20 neurons in each
hidden layers. After optimizing the architecture of network each model is trained and
the trained network is simulated on test data. The predicting ability i.e. the performance
of models are compared with the MAPE and R2 The Results are given in the Table 4.
Table 4. Performance of models

Model 4 Model 5 Model 6 Model 7
MAPE (%) 10.01 9.95 10.72 10.10
R2 0.81 0.82 0.82 0.82
According to the results given in Table 4, the Model 5 performs best with MAPE %
9,95 and R2 is 0,82. The architecture of Model 5 and the regression results of Model 5
while training is given in the Fig. 1 and Fig. 2 respectively (Fig. 3).
Fig. 2. The architecture of 5th model
Fig. 3. The results of model 5 while training

5 Conclusion
In this research the day ahead prices of electricity stock market of Turkey has been
estimated by using ANN. 7 different models has been set up and the four of them has
been selected to continue to research by deciding on the R values. For four models the
networks are trained and the models are optimized by using many different combina-
tions of training, adapting learning and activation functions and different architectures.
The performance of models are compared with R2 and MAPE values that the network is
simulated on the test data. The predicting performance of four models are nearly 10%
given in the Table 4. For model 4, having the architecture of 5-2-2-1, MAPE is 10,01%
with R2 is 0,81. For the best model, Model 5, having an architecture of 6-2-2-1, MAPE
is 9,95% and R2 as 82%. For Model 6, having architecture of 7-2-2-1 the performance
indicators are 10,72% of MAPE and 82% of R2. For the last model Model7, having
architecture of 8-2-2-1 MAPE is 10,10% and the R2 is 82%.
As a result the R2 value as 82% means the 82% of change in electricity prices can
be explained by the input variables of the models, i.e. the historical prices, SPD, and
temporal effects of hours, days, months and seasons. MAPE of 10,10% indicates that
the percentage deviation of forecasted values from real values is 10,10. For further
research, adding other input variables to our Models of this research can be helpful to
investigate the cause of the rest of change in electricity prices.
References
1. Bunn, D.W. (ed.): Modelling Prices in Competitive Electricity Markets. Wiley, Chichester
(2004)
2. Eydeland, A., Wolyniec, K.: Energy and Power Risk Management. Wiley, Hoboken (2003)
3. Weron, R.: Modeling and Forecasting Electricity Loads and Prices: A Statistical Approach.
Wiley, Chichester (2006)
4. Zareipour, H., Canizares, C.A., Bhattacharya, K.: Economic impact of electricity market
price forecasting errors: a demand-side analysis. IEEE Trans. Power Syst. 25, 254–262
(2010)
5. Ugurlu, U., Tas, O., Gunduz, U.: Performance of electricity price forecasting models:
evidence from Turkey. Emerg. Mark. Financ. Trade 54(8), 1720–1739 (2018). https://doi.
org/10.1080/1540496X.2017.1419955
6. Shrivastava, N.A., Panigrahi, B.K.: A hybrid wavelet-elm based short term price forecasting
for electricity markets. Int. J. Electr. Power Energy Syst. 55, 41–50 (2014)
7. Weron, R.: Electricity price forecasting: a review of the state-of-the-art with a look into the
future. Int. J. Forecast. 30(4), 1030–1081 (2014)
8. Sivrikaya Tektas, B.: A decision support model for long-term investment planning of
electricity generation from the perspective of generation companies. Ph.D. Dissertation,
Istanbul Technical University, Turkey (2016)
9. Sivrikaya Tektas, B.: Long-termed investment planning model for a generation company
operating in both bilateral contract and day-ahead markets. Int. J. Inf. Decis. Sci. 8(1), 24–52
(2016)
10. Tektas-Sivrikaya, B., Cebi, F., Turan, H.H., Kasap, N., ve Delen, D.: A fuzzy long-term
investment planning model for the GenCo in a hybrid electricity market considering climate
change impacts. Inf. Syst. Front. 19(5), 975–991 (2017)
11. Kölmek, M.A., Navruz, İ.: Forecasting the day-ahead price in electricity balancing and
settlement market of Turkey by using artificial neural networks. Turk. J. Electr. Eng.
Comput. Sci. 23, 841–852 (2015)
12. Ozguner, E., Tor, O.B., Güven, A.N.: Probabilistic day-ahead system marginal price
forecasting with ANN for the Turkish electricity market. Turk. J. Electr. Eng. Comput. Sci.
25, 4923–4935 (2017)
13. Ugurlu, U., Oksuz, I., Tas, O.: Electricity price forecasting using recurrent neural networks.
Energies 11, 1255 (2018). https://doi.org/10.3390/en11051255
14. Pao, H.T.: Forecasting electricity market pricing using artificial neural networks. Energy
Convers. Manag. 48, 907–912 (2007)
15. Arjona, D.G., Perez, G.L., Gonzales, A.G.: Non-linear QSAR modeling by using multilayer
perceptron feedforward neural networks trained by back-propagation. Talanta 56, 79–90
(2002)
16. Szkuta, B.R., Sanabria, L.A., Dillon, T.S.: Electricity price short-term forecasting using
artificial neural networks. IEEE Trans. Power Syst. 14(3), 851–857 (1999)
17. Yamin, H.Y., Shahidehpourb, S.M., Li, Z.: Adaptive short-term electricity price forecasting
using artificial neural networks in the restructured power markets. Electr. Power Energy
Syst. 26, 571–581 (2004)
18. Aggarval, S.K., Sahini, L.M., Kumar, A.: Electricity price forecasting in deregulated
markets: a review and evaluation. Electr. Power Energy Syst. 31, 13–22 (2009)
Intelligent Learning
Using Formal Concept Analysis Tools
in Road Environment-Type Detection
A. Boulmakoul1(B) , Z. Fazekas2 , L. Karim1,3 , G. Cherradi1 , and P. Gáspár2

1
Casablanca, Morocco
2
Institute for Computer Science and Control (SZTAKI), Budapest, Hungary
[email protected]
3
LISA Laboratory, ENSA Berrechid, Hassan 1st University, Settat, Morocco
[email protected]
Abstract. This paper presents an urban road environment-type (RET)

detection algorithm for an ego-vehicle that is equipped with an auto-
matic traffic sign recognition (TSR) system. The RETs considered in
the paper are the downtown, the residential, and the business/industrial
areas. The Galois-lattice (GL) often used in formal concept analysis
(FCA) is employed to specify the RET detection problem in a formal
manner. With the help of the GL, a correspondence between the traffic
sign (TS) and crossroad (CR) categories, on the one hand, and the RETs,
on the other, is established and represented. This correspondence is also
characterized by means of a fuzzy matrix structure, and its alpha-cuts.
After an FCA-based learning process, in the detection phase, i.e., while
the ego-car is driven along the route, the maximum couplings between
RETs and TS/CR sets are searched for in an on-the-fly manner. The
proposed RET detection method is suitable for real-time implementa-
tion. The paper underlines the importance and applicability of GLs in
classification problems.
Keywords: Road environment-type detection · Fuzzy sets ·

Galois-lattice · Formal concept analysis · Knowledge discovery and
representation
1 Introduction
A smart vehicle is capable of detecting its environment, and navigate along

a route either with the help of the driver, i.e., being driven by the driver, or
indeed all by its own, i.e., in total autonomy. The available technologies rely
on various input data streams. These streams can be generated by one or more
camera, one or more infrared sensor, one or more LiDAR, one or more radar,
etc. Current research developments produce innovative sensor fusion solutions
that build upon computer vision, sensor networks, machine learning and artificial
intelligence techniques, including deep learning approaches, see e.g., [12].
A traffic sign (TS) based urban road environment-type (RET) detection sys-
tem was proposed in [5]. The RET detection system gathers TS data along the
route with the help of the TSR system, and employs a stochastic change detec-
tion method in respect of the collected TS data to identify RET change-points,
i.e., certain road locations, along a route. Various variants of TS-based RET
detection system was proposed later, e.g., [6–8] and [9]. These detection meth-
ods rely on different computing techniques ranging from the shallow artificial
neural networks (ANNs) to the heuristic rule-based decision making techniques.
Some of these approaches use additional input data types, such as types and
location of the encountered crossroads (CRs) along the route. Clearly, numerous
other input data-types could be employed for the purpose, e.g., number of lanes
along the route, width of the actual road/lane.
Galois lattices (GL) are algebraic structures related to the order theory, see
e.g., [3,4]. Galois structures and concept lattices are key structures of the formal
concept analysis (FCA), [10,13], which is a widely used data mining method-
ology for formal contexts composed of objects, which are in turn described by
attributes. A good introduction to GLs is given by [1] via a fairly common edu-
cational example: evaluation of class-test results by a teacher. The educational
and test-design conclusions are also interesting for an educational practitioner,
including those working in the higher education.
Closer to our present topic of detecting the road environment, in partic-
ular, the road environment meant in a socio-economic sense and viewed at a
corresponding scale, a knowledge-based agricultural landscape analysis frame-
work was proposed in [11]. The GL-based framework presented there includes a
concept lattice structure of topological relations for qualitative spatial reason-
ing. The main objective was to recognize landscape models on land-use maps
extracted from satellite images. These models are abstract models describing
agricultural spatial structures as sets of spatial entities and qualitative spatial
relations between these entities. The map-regions, or zones, can represent certain
agricultural environments, such as crops, meadows, forest, buildings, etc.
In the present paper, a GL is employed to capture the correspondence
between traffic signs (TSs) and crossroads categories (CRs), on the one hand, and
urban RETs, on the other. The three urban RETs considered are the downtown,
residential, and business/industrial areas. These characterize the road environ-
ments around the ego-car, and provide a hint for the driver about the traffic
safety risks associated with such areas. The detection approach presented here
takes into account only certain – i.e., relevant for the purpose – TS types. Fur-
thermore, for a TS to be considered in the computations, it needs to be visi-
ble/detectable from the ego-car moving along its route.
As a first processing step, fuzzy sets are created based on the occurrence
probabilities of the TS types within the urban RETs. Then, the correspondence
between TSs and urban RETs – by means of the alpha-cuts applied to the fuzzy
matrix built in the previous step – are established. Then, a learning process –
based on a FCA method – explores and highlights the maximum couplings that
Using Formal Concept Analysis Tools 1061
exsist between a TSs and the RETs. In the next processing step, these maximum
couplings are used in the inference of the actual RET. A new, efficient closed
sets exploration procedure for characterizing the individual RETs is proposed in
the paper to support the inference.
The remainder of the paper is structured as follows. Section 2 provides a brief
recall of the fuzzy concepts used. In Sect. 3, an FCA-based RET detection app-
roach is described, while in Sect. 4, the software architecture of the implemented
prototype, as well as the methodology of its deployment is covered. A simple,
but illustrative example is given to show the main processing steps. Finally, in
Sect. 5, the conclusions are drawn.
2 Preliminaries
This work focuses on the algebraic calculations – carried out within and with the
help of GL structures – to reason about the relationships between the elements of
a domain, e.g., various TSs, CRs and the considered RETs. It also makes use of
the fuzzy sets and related concepts. The notations and the necessary definitions
for understanding the rest of the paper are given in this section. We recall the
definitions of fuzzy set and GL concepts based on [2,14,15].
2.1 Fuzzy Concepts
Definition 1: If X is a collection of objects denoted generically by x, then a

in X is a set of ordered pairs.
fuzzy set A
= {x, μ (x)|x ∈ X}
A (1)
A
μA(x) is the membership function, or grade – in some publications also

referred to as the degree of compatibility, or the degree of truth – of x in A

that maps X to the membership space M (When M ={0, 1}, A is crisp set and
μA(x) is identical to the characteristic function of non-fuzzy set).
The (crisp) set of elements that belong to the fuzzy set A at least to the
degree α is called the α − level set:
Aα = {x ∈ X | μA(x) ≥ α} (2)
Aα
= {x ∈ X | μA(x) > α} is called “strong α − level set” or “strong
α − cut”. Throughout this paper, the following notations have been used. R+ =
[0, ∞); X = {x1 , x2 , ..., xn } is the universal set; F(X) is the class of all fuzzy sets
of X; μA (xi ) : X −→ [0.1] is the membership function of A ∈ F(X); Ac ∈ F is
the complement of A ∈ F.
2.2 Formal Concept Analysis Concepts

Definition 2: A formal context k := (G, M, I) consists of two sets G and M
and a relation I between G and M . The elements of G are called the objects
and the elements of M are called the attributes of the context. In order to
express that an object g is in a relation I with an attribute m,we write gIm or
(g, m) ∈ I and read it as “the object g has the attribute m” [4,10,13].
Definition 3: For a set A ⊆ G of objects we define
A = {m ∈ M | gIm f or all g ∈ A} (3)
(the set of attributes common to the objects in A). Correspondingly, for a set B
of attributes we define
B = {g ∈ G | gIm f or all m ∈ B} (4)
(the set of objects which have all attributes in B ).
Definition 4: a formal concept of the context (G, M, I) is a pair (A, B) with

A ⊆ G, B ⊆ M, A = B and B = A. We call A the extent and B the intent
of the concept (A, B). B(G, M, I) denotes the set of all concepts of the context
(G, M, I).
Proposition: if (G, M, I) is a context, A, A1 , A2 ⊆ G are sets of objects and

B, B1 , B2 are sets of attributes. Then :
1) A1 ⊆ A2 ⇒ A2 ⊆ A1 1, ) B1 ⊆ B2 =⇒ B2 ⊆ B1

2) A ⊆ A 2, ) B ⊆ B

3) A = A 3, ) B = B

4) A ⊆ B ⇐⇒ B ⊆ A ⇐⇒ A × B ⊆ I.
Definition 5: If (A1 , B1 ) and (A2 , B2 ) are concepts of a context, (A1 , B1 ) is

called a subconcept of (A2 , B2 ), provided that A1 ⊆ A2 (which is equivalent
to B2 ⊆ B1 ). In this case, (A2 , B2 ) is a superconcept of (A1 , B1 ), and we
write (A1 , B1 ) (A2 , B2 ). The relation is called the hierarchical order(or
simply order) of the concepts. The set of all concepts of (G, M, I) ordered in this
way is denoted by B(G, M, I) and is called the concept lattice of the context
(G, M, I).
2.3 Fuzzy Galois Lattice

Inaccurate and fuzzy values either measured in physical systems, or detected
in, calculated for biological, economic, or social systems inspired great deal of
research among others in the field of fuzzy FCA [2], particularly with regard
to the generalization of closure operators concerning fuzzy relations and the
underlying problem of fuzzy formal concepts generation.
The FCA consists in inducing formal concepts <Objects, Attributes> from

a formal binary context. These formal concepts are based on the following def-
inition: “An object belongs to the Objects subset, if it has all the attributes”
[10]. Since the relation considered in the formal context is Boolean, an object
has an attribute altogether, or does not have it at all. Indeed, in real problems
the FCA is brought to use formal contexts of various nature, obtained by mea-
surements, observations, judgments, etc., where the relation between an object
and an attribute can be more uncertain, and less specific. Or even gradual, or
imprecise. Various approaches have been proposed for the extension of the FCA
to the fuzzy ‘worls’.
3 RET Detection by Means of FCA

3.1 Notation
Hereinafter, we describe some useful notations
for the rest
of the paper.
m,i
Ei = {spatial object of layer (i)} = ej |j = 1..n which defines a layer.
The set of all layers: E = ∪i Ei .
ϑi (m, ε) = {o ∈ Ei |δ (m, o) ≤ ε L (m, o) true } all the objects of the
layer Ei which are at a distance from the mobile (i). L (m, o) defines a spatial
predicate between the mobile m and the object o.
C = Road Environment − T ype (RET ) = {σi |i = 1..m } .
Example:
C = {downtown,
residential
area, business |industrial area , . . .}
m,j m,j
θi,k = L σi , ek true
The context is built by a learning process, e.g., a set of vehicles col-
lects information concerning the RETs and the CR categories along certain
routes. Induction type urban area explained by the other variables (according
to layer) :C × Ei , ou C ×j=k
j=1 ⊕Ej . This correspondence will be designated by
{:layers}
: m being the set of topics considered for the context. Induction type
urban area explained by the other variables (according to selected layers). We
{:layers} {:layers}
define =j=k
j=1 ⊕m , where ⊕ denotes an algebraic operation on
1; if [∧(γi ,σj )α]
matrices (t-conorm of fuzzy relations sets) Λα (γi , σj ) = {0; otherwise
Where:
ET S = {γ1 , γ2 , γ3 , γ4 }; ECR = {γ5 , γ6 , γ7 , γ8 , γ9 }; C = {σ1 , σ2 , σ3 }.
Hereafter we illustrate this process by infered fuzzy sets:

(D) σ 1 = γ1/0.22, γ2/0.18, γ3/0.07, γ4/0.02, γ5/0.28, γ6/0.13, γ7/0.01, γ8/0.001, γ9/0.05

(R) σ 2 = γ1/0.03, γ2/0.025, γ3/0.08, γ4/0.04, γ5/0.32, γ6/0.08, γ7/0.005, γ8/0.002, γ9/0.02

(IC) σ 3 = γ1/0.04, γ2/0.017, γ3/0.055, γ4/0.02, γ5/0.20, γ6/0.015, γ7/0.001, γ8/0.002, γ9/0.04
Table 1. The probability of occurrence of TS’s and CR’s – along a 50 m path-length

– in the three urban RE’s considered from [6].
σ1 σ2 σ3
RET Downtown Residential Industrial/Commercial
TS
γ1 22% 3% 4%
γ2 18% 2,5% 1,7%
γ3 7% 8% 5,5%
γ4 2% 4% 2%
γ5 28% 32% 20%
γ6 13% 8% 1,5%
γ7 1% 0,5% 0,1%
γ8 0,1% 0,2% 0,2%
γ9 5% 2% 4%
4 The FCA-based RET Detection Process

The analysis of formal concepts is an analytical practice based on graph theory
and strongly linked to Galois lattices [3]. It is a theory of mathematical algebra
that makes it possible to formalize concepts as basic units of human thought
and to analyze data in attribute-object form [10]. Created in the early 80’s and
over the last three decades, it has become a popular engineering tool for human-
centered knowledge representation and data analysis with many applications in
many fields. In this section, we rely on the GL foundations, [3,4], and illustrate
with a simple example the use and the interpretation of its elements. The main
definitions regarding Galois correspondences and GLs are presented in [3]. Other
results on GL redefined in the field of FCA appear in [10]. Herein, we use the
broad presentation given in [13].
4.1 RET Detection Process
1. Pre-processing: Fuzzy and alpa-cuts context construction,

2. Generation of closed sets and calculation of associated fuzzy sets,
3. For each detection vector do,
4. associate closed sets with extent (TSs) containing it,
5. For each closed set compute the fuzzy cardinality,
6. Select the closed set with maximum fuzzy cardinality and associate the intent
(RET) with the detection vector.
The data table below represents an example of the RET as a fuzzy formal context
(from Table 1)
We have:
(D) σ 1 = {γ1/0.22, γ2/0.18, γ3/0.07, γ4/0.02, γ5/0.28, γ6/0.13, γ7/0.01, γ8/0.001, γ9/0.05}
(R) σ 2 = {γ1/0.03, γ2/0.025, γ3/0.08, γ4/0.04, γ5/0.32, γ6/0.08, γ7/0.005, γ8/0.002, γ9/0.02}
(IC) σ 3 = {γ1/0.04, γ2/0.017, γ3/0.055, γ4/0.02, γ5/0.20, γ6/0.015, γ7/0.001, γ8/0.002, γ9/0.04}
Fig. 1. GL of the context given by Table 1, thresholds level: 0.04
Table 2. RET-fuzzy formal context illustration
Fuzzy closed set Fuzzy set– extent/intent

[γ3 , γ4 , γ5 , γ6 /R] {γ1/0, γ2/0, γ3/0.08, γ4/0.04, γ5/0.32, γ6/0.08, γ7/0, γ8/0, γ9/0}
[γ3 , γ5 , γ6 /D, R]
{γ1/0, γ2/0, γ3/0.07, γ4/0, γ5/0.28, γ6/0.13, γ7/0, γ8/0, γ9/0} ∩
{γ1/0, γ2/0, γ3/0.08, γ4/0, γ5/0.32, γ6/0.08, γ7/0, γ8/0, γ9/0}
= {γ1/0, γ2/0, γ3/0.07, γ4/0, γ5/0.28, γ6/0.08, γ7/0, γ8/0, γ9/0}
[γ3 , γ5 /R, D, IC]
{γ1/0, γ2/0, γ3/0.07, γ4/0, γ5/0.28, γ6/0, γ7/0, γ8/0, γ9/0} ∩

{γ1/0, γ2/0, γ3/0.08, γ4/0, γ5/0.32, γ6/0, γ7/0, γ8/0, γ9/0} ∩
{γ1/0, γ2/0, γ3/0.055, γ4/0, γ5/0.20, γ6/0, γ7/0, γ8/0, γ9/0}
= {γ1/0, γ2/0, γ3/0.055, γ4/0, γ5/0.20, γ6/0, γ7/0, γ8/0, γ9/0}
[γ1 , γ2 , γ3 , γ5 , γ6 , γ9 /D] {γ1/0.22, γ2/0.18, γ3/0.07, γ4/0, γ5/0.28, γ6/0.13, γ7/0, γ8/0, γ9/0.05}
[γ1 , γ3 , γ5 , γ9 /D, IC]
{γ1/0.22, γ2/0, γ3/0.07, γ4/0, γ5/0.28, γ6/0, γ7/0, γ8/0, γ9/0.05} ∩
{γ1/0.04, γ2/0, γ3/0.055, γ4/0, γ5/0.20, γ6/0, γ7/0, γ8/0, γ9/0.04}
= {γ1/0.04, γ2/0, γ3/0.055, γ4/0, γ5/0.20, γ6/0, γ7/0, γ8/0, γ9/0.04}
[γ1 , γ2 , γ3 , γ4 , γ5 , γ6 , γ7 , γ8 , γ9 /∅] {γ1/0, γ2/0, γ3/0, γ4/0, γ5/0, γ6/0, γ7/0, γ8/0, γ9/0}
Table 2 gives the results of browsing within the GL. For each detection vector,
we associate the closed (intent/extent) set of the GL, whose intent contains it,
or with which the intersection with the intent is not empty. If several closed sets
are found, the fuzzy intersection of the intents is performed to determine the
extent (Table 3).
Table 3. RET detection process results
Detected sequence (50m) Closed set Fuzzy Sets RET

γ1 γ2 γ3 [γ1 , γ2 , γ3 , γ5 , γ9 /D] {γ1/0.22, γ2/0.18, γ3/0.07} D
γ2 γ3 γ4 [γ3 , γ4 , γ5 , γ6 /R] {γ3/0.08, γ4/0.04} /R[0.12] D
[γ1 , γ2 , γ3 , γ5 , γ9 /D] {γ2/0.18, γ3/0.07} /D[0.25]
γ4 γ5 [γ3 , γ4 , γ5 , γ6 /R] {γ4/0.04, γ5/0.32} /R R
γ3 γ9 [γ1 , γ2 , γ3 , γ5 , γ9 /D] {γ3/0.07, γ9/0.05} /D[0.12] D
[γ1 , γ3 , γ5 , γ9 /D, IC] {γ3/0.055, γ9/0.04} /D, IC[0.095]
γ1 γ5 [γ1 , γ2 , γ3 , γ5 , γ9 /D] {γ3/0.07, γ9/0.05} /D[0.12] D
[γ1 , γ3 , γ5 , γ9 /D, IC] {γ3/0.055, γ9/0.04} /D, IC[0.095]
4.2 FCA Process and Architecture
In this section, we describe the use of the closed GL associated with the rela-
tionship expressing the dependencies between the TSs and the CRs, on the one
hand, and RET, on the other hand. The exploration process starts with the col-
lection of various data from the road environments observed within the urban
network. This relatively large volume of data is then exploited to form probabil-
ities and matrix structures defining the dependencies, or couplings between the
elements being in relation. This step requires the choice of objects and attributes
in the (TS, CR) x RET repository. These structures materialize as fuzzy con-
texts, which will be subsequently thresholded (i.e., binarized according to some
thresholds) to create binary Galois correspondences. Then these contexts will be
used to generate GLs and thus lead to the exploration according to the holistic
exploration practice as presented in Fig. 1. It shows an illustrative example of
the process and the corresponding software architecture.
5 Conclusion
In this paper, an FCA approach was applied to the problem of urban RET
detection. Our purpose was to characterize the road environment – swishing
past around an ego-car – based on TS and CR occurrences detected by a TSR
system. The results obtained by browsing GL closed sets provide insight to the
detection intricacies. The results of the pilot-study was rather promising, but
large-scale tests are required to thoroughly validate the approach.
References
1. Andor, C., Joo, A., Mero, L.: Galois-lattices: a possible representation of knowledge
structures. Eval. Educ. 9(2), 207–215 (1985)
2. Antoni, L., Krajči, S., Krı́dlo, O.: Representation of fuzzy subsets by Galois con-
nections. Fuzzy Sets Syst. 326, 52–68 (2017)
3. Barbut, M., Monjardet, B.: Ordre et Classification. Algèbre et Combinatoire,
Tomes I et II, Paris, Hachette (1970)
4. Birkhoff, G.: Lattice theory, American Mathematical Society Colloquium Publica-
tions, vol. XXV, 1st ed., 1940, 3rd ed. (1967)
5. Fazekas, Z., Balázs, G., Gerencsér, L., Gáspár, P.: Inferring the actual urban road
environment from traffic sign data using a minimum description length approach.
Transp. Res. Proc. 27, 516–523 (2017)
6. Fazekas, Z., Balázs,G., Gerencsér, L., Gáspár, P.: Detecting change in the urban
road environment along a route based on traffic sign and crossroad data. In: Intelli-
gent Transport Systems – From Research and Development to the Market Uptake,
LNICST, Social Informatics amp; Telecommunications Engineering, vol. 222, pp.
252–262. Springer, (2018)
7. Fazekas, Z., Balázs, G., Gáspár, P.: ANN-based classification of urban road envi-
ronments from traffic sign and crossroad data. Acta Polytech. Hung. 15, 83–100
(2018)
8. Fazekas, Z., Balázs, G., Gáspár, P.: Building upon modularity in artificial neural
networks. ERCIM News 116, 30–31 (2019)
9. Fazekas, Z., Balázs, G., Gáspár, P.: A heuristic approach to road environment-
type detection from traffic sign data. In: Proceedings of the 4th IEEE Conference
on Control and Fault Tolerant Systems, Casablanca, Morocco, pp. 288–293. IEEE
(2019)
10. Ganter, B., Wille, R.: Formal Concept Analysis: Mathematical Foundations.
Springer, New York (1997). ISBN 3540627715
11. Le Ber, F., Napoli, A.: A Galois lattice for qualitative spatial reasoning and repre-
sentation. In: Proceedings of the ECAI Workshop on Advances in Formal Concept
Analysis for Knowledge Discovery in Databases, Lyon, France, 12 p. (2002)
12. Vishnukumar, H.J., Butting, B., Müller, C., Sax, E.: Machine learning and deep
neural network—artificial intelligence core for lab and real-world test and vali-
dation for ADAS and autonomous vehicles: AI for efficient and quality test and
validation. In: Intelligent Systems Conference, London, UK, pp. 714-721. IEEE
(2017)
13. Wille, R.: Concept lattices and conceptual knowledge systems. Comput. Math.
Appl. 23(6–9), 493–515 (1992)
14. Wolff, K.E.: Concepts in fuzzy scaling theory: order and granularity. Fuzzy Sets
Syst. 132(1), 63–75 (2002)
15. Zimmermann, H.J.: Fuzzy Set Theory and Its Applications, 4th edn. Springer,
Heidelberg (2001)
An Intelligent Overtaking Assistant System
for Autonomous Vehicles
Ersin Armağan(&) and Tufan Kumbasar
Control and Automation Engineering Department, Istanbul Technical University,

Istanbul, Turkey
{armagane,kumbasart}@itu.edu.tr
Abstract. Autonomously overtaking cars is one of the most challenging

functionality when developing self-driving vehicles due to its safe crucial nat-
ure. This study presents a novel intelligent overtaking assistant based on fuzzy
logic and its deployment into TORCS environment. For achieving this goal, the
developed intelligent overtaking assistant decides when to overtake, and cal-
culates the desired position and speed values to accomplish the overtaking
action by using the relative speed and position information. Thus, to estimate the
positions and speed of the other vehicles, we designed and deployed a multi
object-tracking algorithm. Intelligent overtaking assistant consists of four fuzzy
logic systems that have modular cascade structure. The fuzzy logic based
overtaking assistant decides “when to start or end the overtaking action” and
“whether speed adjustment is needed or not”, and calculates “the desired lateral
position on the road” and “how much speed increase is needed during over-
taking action”. The efficiency of the intelligent overtaking assistant is examined
in TORCS environment and the presented results clearly demonstrate that its
performance is highly efficient in various road structures.
Keywords: Overtaking Autonomous vehicle Fuzzy logic Decision making
1 Introduction
In literature, the tactical decision making for overtaking autonomously has been widely
handled as it is one of the most challenging problems of autonomous cars [1–10]. The
tactical decision making of overtaking and generation of the throttle, brake and steering
angle control inputs are solved with if-then rules [1], covariance matrix adaptation
evolutionary strategy [2], and mixed observable Markov decision process [3]. Rein-
forcement learning is also employed to learn overtaking behaviors in a similar way with
humans learn to drive [4–6]. On the other hand, a fuzzy logic based system is designed
to execute an autonomous overtaking for different cases in two-way road with an
oncoming vehicle from the other direction [7]. Another fuzzy logic based overtaking
system, which is capable of path tracking and lane changing, is presented and examined
when a lane is free and the road is long enough to overtake [8]. There are also fuzzy
logic based overtaking systems that have opponent blocking strategies with various
difficulty levels [9, 10].
An Intelligent Overtaking Assistant System for Autonomous Vehicles 1069
In this study, an Intelligent Overtaking Assistant (IOA) system for autonomous

vehicles is proposed and deployed in the realistic The Open Racing Car Simulator
(TORCS). The IOA autonomously decides when to overtake and calculates the ref-
erence speed and lateral position to overtake the opponent vehicle. In the design of the
IOA, we utilize four fuzzy systems for separate tasks such as generating overtaking
decision signal and speed adjustment decision signal, calculating suitable lateral
position to overtake and how much speed increase is needed during overtaking. We
start with presenting the perception structure of the IOA as the extracted information is
directly used via the fuzzy systems. Then, we present the structures of the fuzzy
systems that define the IOA. The performance of the IOA is evaluated on two different
tracks that have different characteristics such as straight and curvy road. The results
clearly show not only how the IOA performs overtaking but also the efficiency of IOA.
This paper is organized as follows. Section 2 presents details on the structure and
design of the IOA. In Sect. 3, performance of experimental results to validate the IOA.
The driven conclusions and future work are presented in Sect. 4.
2 Intelligent Overtaking Assistant
In this section, we present the structure of the proposed IOA that is shown in Fig. 1.
The IOA is capable to perceive the environment and to generate the desired lateral
position and speed reference values ðpd ; vd Þ for a successful overtaking, accordingly.
Fig. 1. Intelligent overtaking assistant.
The tactical decision making mechanism of the IOA is composed of a Fuzzy Logic
based Overtaking Position Assistant (FOPA) and Fuzzy Logic based Overtaking
Position Regulator (FOPR) that generate the pd value while the Fuzzy Logic based
Overtaking Velocity Assistant (FOVA) and Fuzzy Logic based Overtaking Velocity
Regulator (FOVR) calculate the vd value. The design of the fuzzy systems is accom-
plished by converting expert knowledge into fuzzy rules, and thus one might claim that
the IOA is mimicking an expert driver. The outputs of IOA are tracked with low-level
controllers that generate the throttle, brake, gear and steering angle values of the vehicle
as described in [12, 13].
1070 E. Armağan and T. Kumbasar
2.1 Perception
In this section, we define the inputs of the IOA and present how they are calculated and
estimated. The proposed IOA system uses and processes the longitudinal distance
between ego and opponent vehicles (d), lateral distance between ego and opponent
vehicles (d), lateral position of the ego vehicle on the track (pego Þ and percentage
relative speed (vrel %) that is defined as:

vrel % ¼ vego vopp =vego 100 ð1Þ
where vego and vopp are the speeds of ego and opponent car, respectively. The inputs
signals are illustrated in Fig. 2.
Fig. 2. The inputs of IOA.
In TORCS, the ego vehicle is equipped with a 360° sensor with a resolution of 1°
which is referred as the opponent sensor [1, 11]. The opponent sensor provides
information on the closest distance from the ego vehicle to opponent vehicle (dc ) and
the relative segment information of opponent vehicle which can be easily transformed
to angle value (s). From these perceptions, the lateral and longitudinal distances (d; d)
are calculated via trigonometric operations in a straightforward manner. The speed and
position of the ego vehicle ðvego ; pego Þ are directly extracted from TORCS.
In order to track the lateral and longitudinal distances (d; d), a Kalman filter that is
defined with a constant acceleration model is designed. In the case of multiple vehicles
on the track, it is necessary to associate the vehicles with their previous states. Thus, to
tackle this issue, we used the Mahalanobis distance [14] for association.
The Maha-
d
lanobis distance (cdist ða; bÞ) between current position vector (~ r¼ ) and previous
d

d
position vector (~ rprev ¼ prev ) can be described as:
dprev
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
T
cdist ~ rprev ¼
r;~ ~r ~ rprev S1 ð~ r ~ rprev Þ ð2Þ

where S is the covariance matrix of ~ r and ~
rprev . A small cost value cdist ~
r;~
rprev is
interpreted as a higher probability that these two are the same, and thus they can be
associated (i.e. it means it is the same opponent vehicle). We also estimate velocity of
opponent (vopp ) from the Kalman filter via the longitudinal distance d.
2.2 Desired Lateral Position Generation for Overtaking

The desired lateral position of the proposed IOA is generated as follows:
pd ¼ pf a þ pego ð1 aÞ ð3Þ
where pego is the current lateral position of the ego vehicle and a 2 ½0; 1 is the tactical
decision signal generated by FOPA. If a ¼ 0, the ego vehicle stays on its current
position pd ¼ pego . If the overtaking conditions are met (a [ 0), reference lateral
position (pd ) is tuned by taking into account the generated output pf of FOPR.
The tactical decision-making is accomplished with FOPA that processes the inputs
vrel % and d to generate overtaking decision signal (a). The inputs of FOPA are defined
with the linguistic variables Negative Big (NB), Negative Small (NS), Zero (Z),
Positive Small (PS) and Positive Big (PB) as shown in Fig. 3. The rules are presented
in Table 1 where ‘Zero (Z) = 0’ means that ego vehicle should continue driving
without overtaking, ‘Small (S) = 0.5’ means that vehicle starts or ends the overtaking
action and ‘Big (B) = 1’ means that vehicle continues to overtake. The rules are
generated by considering various overtaking situations. For example, if the ego vehicle
is approaching to opponent (d is PS) and relative speed is bigger than zero (vrel % is
PB), then the overtaking action should start (a is S).
Fig. 3. Membership functions of FOPA.
Table 1. Rules of FOPA.

vrel d
NB NS Z PS PB
NB Z Z B Z Z
NS Z Z B Z Z
Z Z Z B Z Z
PS Z S B B S
PB Z S B B S
After overtaking is initiated, i.e. a [ 0, pego is adjusted to overtake the opponent by

the FOPR as given in (3). FOPR calculates the suitable lateral position (pf ) for over-
taking by using the inputs pego and d. The input MFs are presented in Fig. 4 while the
rule base is as given in Table 2. Note that, the input pego is normalized into universe of
discourse of the antecedent MFs via the width of the road because tracks have different
road widths. The output MFs are defined with the following linguistic terms: Left2
(L2), Left1 (L1), Center (C), Right1 (R1) and Right2 (R2). As shown in Fig. 2, L2
represents left edge of road and R2 represents right edge of road. In the design of rules,
we took account various overtaking cases. As an example for defining the rules of
FOPR, if the opponent vehicle is moving in front of the ego vehicle (d is C) and ego
vehicle is moving at right lane (pego is R1), then the ego vehicle should switch to left
lane (pf is L2). On the other hand, if the ego vehicle is on left lane (pego is L1), then the
ego vehicle should switch to right lane (pf is R2). If the opponent vehicle is moving in
front of the ego vehicle (d ¼ C) and ego vehicle is moving at center of the road
(pego ¼ C), ego vehicle can overtake the opponent by going left lane (pf ¼ L2).
Fig. 4. Membership functions of FOPR.
Table 2. Rules of FOPR.

d pego
L2 L1 C R1 R2
L2 L1 C R1 R2 R2
L1 C R1 R2 R2 R2
C C R1 L2 L1 C
R1 L2 L2 L2 L1 C
R2 L2 L2 L1 C R1
2.3 Desired Longitudinal Velocity Generation for Overtaking

The desired speed of the proposed intelligent overtaking assistant is defined as:
vd ¼ ðvf þ vego Þb þ vego ð1 bÞ ð4Þ
where vego is the current speed of ego vehicle, b 2 ½0; 1 is the overtaking speed
decision signal generated by FOVA and vf is the necessary speed change that is
generated by the FOVR. If b ¼ 0, vehicle continuous with vd ¼ vego while for b [ 0
the reference speed is adjusted by the FOVR as given in (4).
The tactical decision-making mechanism FOVA uses vrel % and d to generate
decision signal b. The inputs of FOVA are defined with the linguistic variables NB,
NS, Z, PS and PB as shown in Fig. 5. The rules of FOVA are tabulated in Table 3
where ‘Z = 0’ means that ego vehicle should continue with vego , ‘S = 0.5’ means that
vehicle starts to increase its speed to overtake the opponent and ‘B = 1’ represent that
vehicle should increase its speed as much as possible. The rules are created by con-
sidering like following overtaking conditions. If the ego vehicle is parallel to opponent
(d is Z) and relative speed is bigger than zero (vrel % is PS), then the speed of the
vehicle should be increased to pass the opponent fast (b is B).
Fig. 5. Membership functions of FOVA and FOVR.
Table 3. Rules of FOVA and FOVR.

vrel d
NB NS Z PS PB
NB Z Z B Z Z
NS Z Z B Z Z
Z Z Z B Z Z
PS B B B B S
PB B B B B S
After speed adjustment is decided, i.e. b [ 0, FOVR calculates how much speed
increment (vf ) is required for overtaking in addition to vego . The input MFs are defined
with the NB, NS, Z, PS and PB as shown in Fig. 5. The rule base is given in Table 3
where ‘Z = 0’ means that ego vehicle should not increase its speed, ‘S = 5’ means that
speed increase should be small and ‘B = 10’ represent that vego should be increased a
lot in the overtaking action. The rules of FOVR can be easily interpreted, for instance if
the ego vehicle is approaching to opponent (d is PS) and relative speed is bigger than
zero (vrel % is PS) then vego is increased to complete the overtaking (vf ¼ B).
We provide results conducted on two tracks of TORCS to examine how the IOA is
performing overtaking in a straight track (Motorway) in a curvy track (Ruudskogen).
The overtaking performance of IOA in a straight track and curvy track is illustrated
in Fig. 6 and Fig. 7, respectively. It can be clearly concluded from Fig. 6a and Fig. 7a,
the overtaking performance of proposed IOA is highly efficient in both track types. We
can also observe from the shaded areas in Fig. 6b and Fig. 7b that as the ego vehicle is
approaching to the opponent vehicle, both the tactical decision signals converge to 0.5
to initiate the overtaking action. Then, since a; b [ 0, the ego vehicle starts to increase
its speed as shown in Fig. 6c and Fig. 7c, and aligns itself on the left lane of track to
accomplish the overtaking action. We can also observe from the velocity plots obtained
in curvy track that the speed of the vehicle is lower than the one obtained from straight
track in order to overtake while passing a curvature.
Fig. 6. Performance of IOA on a straight track: (a) x-y positions (b) tactical decisions
(c) velocities.
Fig. 7. Performance of IOA on a curvy track: (a) x-y positions (b) tactical decisions
(c) velocities.
In this study, we proposed a novel solution approach to the overtaking problem of

autonomous cars by presenting IOA system composed of fuzzy systems. The fuzzy
logic based IOA system uses the perceptual data to predict the information of the
opponent car in order to decide when to overtake and to calculate the reference speed
and lateral position for a satisfactory overtaking performance. We evaluated the per-
formance in the realistic TORCS environment. The presented results clearly show the
efficiency of the proposed overtaking approach.
As for our future work, we plan to extend the case studies with multiple vehicles
with the inclusion of traffic rules to evaluate the IOA in a complex environment.
References
1. Onieva, E., Pelta, D.A., Alonso, J., Milanés, V., Pérez, J.: A modular parametric architecture
for the TORCS racing engine. In: 2009 IEEE Symposium on Computational Intelligence and
Games, pp. 256–262. IEEE, Milano (2009)
2. Bonyadi, M.R., Michalewicz, Z., Nallaperuma, S., Neumann, F.: Ahura: a heuristic-based
racer for the open racing car simulator. IEEE Trans. Comput. Intell. AI Games 9(3), 290–304
(2016)
3. Sezer, V.: Intelligent decision making for overtaking maneuver using mixed observable
markov decision process. J. Intell. Transp. Syst. 22(3), 201–217 (2018)
4. Kaushik, M., Prasad, V., Krishna, K.M., Ravindran, B.: Overtaking maneuvers in simulated
highway driving using deep reinforcement learning. In: 2018 IEEE Intelligent Vehicles
Symposium (iv), pp. 1885–1890. IEEE, Changshu (2018)
5. Huang, H.H., Wang, T.: Learning overtaking and blocking skills in simulated car racing. In:
2015 IEEE Conference on Computational Intelligence and Games (CIG), pp. 439–445.
IEEE, Tainan (2015)
6. Loiacono, D., Prete, A., Lanzi, P.L., Cardamone, L.: Learning to overtake in TORCS using
simple reinforcement learning. In: IEEE Congress on Evolutionary Computation, pp. 1–8.
IEEE, Barcelona (2010)
7. Perez, J., Milanes, V., Onieva, E., Godoy, J., Alonso, J.: Longitudinal fuzzy control for
autonomous overtaking. In: 2011 IEEE International Conference on Mechatronics, pp. 188–
193. IEEE, Istanbul (2011)
8. Naranjo, J.E., Gonzalez, C., Garcia, R., De Pedro, T.: Lane-change fuzzy control in
autonomous vehicles for the overtaking maneuver. IEEE Trans. Intell. Transp. Syst. 9(3),
438–450 (2008)
9. Cardamone, L., Lanzi, P.L., Loiacono, D., Onieva, E.: Advanced overtaking behaviors for
blocking opponents in racing games using a fuzzy architecture. Expert Syst. Appl. 40(16),
6447–6458 (2013)
10. Onieva, E., Cardamone, L., Loiacono, D., Lanzi, P.L.: Overtaking opponents with blocking
strategies using fuzzy logic. In: Proceedings of the 2010 IEEE Conference on Computational
Intelligence and Games, pp. 123–130. IEEE, Dublin (2010)
11. TORCS: The Open Racing Car Simulator. http://torcs.sourceforge.net/. Accessed 20 Feb
2020
12. Armagan, E., Kumbasar, T.: A fuzzy logic based autonomous vehicle control system design
in the TORCS environment. In: 2017 10th International Conference on Electrical and
Electronics Engineering (ELECO), pp. 737–741. IEEE, Bursa (2017)
13. Armagan, E., Kumbasar, T.: TORCS oyun ortamında bulanık mantık tabanlı akıllı bir
otonom araç kontrol sistemi tasarımı. Pamukkale Üniv. Mühendislik Bilimleri Dergisi 24(8),
1435–1442 (2018)
14. Mahalanobis, P.C.: On the generalised distance in statistics. Proc. Natl. Inst. Sci. India 2(1),
49–55 (1936)
Predicting Movie Ratings with Machine
Learning Algorithms
Sandy Çağlıyor1(&) and Başar Öztayşi2

1
Haliç University, Beyoglu, 34445 Istanbul, Turkey
[email protected]
2
Istanbul Technical University, Macka, 34367 Istanbul, Turkey
Abstract. The fact that a film is a hedonic product makes it difficult to assess its
quality before consumption, therefore consumers who want to reduce uncer-
tainty need various quality signals in their decision-making processes. In recent
years, adding to movie-related information, user reviews or ratings on online
movie databases have become important quality signals, where many movie
viewers use these sites to decide which movie to watch or whether or not to
watch a certain movie. In this study, it is attempted to estimate the rating and
popularity of a movie by using the main product features as the origin, pro-
duction year, actor and plot. A database containing 8943 movies shot between
2000 and 2019 from the website sinemalar.com is formed with the help of a web
crawler Latent Dirichlet allocation topic extraction is applied to plots and
assigned topics obtained from LDA analyzes, along with other movie-related
attributes are used to predict the rating class and popularity class of a movie by
employing machine learning algorithms such as random forest, gradient
boosting tree and decision tree. Using the random forest algorithm attribute
statistics, based on their contribution to the predictive power of the model the
relative variable importance is also examined.
Keywords: Movie rating Machine learning Latent Dirichlet Allocation
1 Introduction
The cinema industry has been quite remarkable for academicians in terms of its
complex and dynamic structure, consequently having many unsolved problems and the
excess of available data [1]. While the motion picture literature up to the 80’s generally
dealt with issues such as theoretical, historical and critical approaches, or sociological
psychological cultural evaluation of films, it gained a new dimension in the 80 s with
the introduction of mathematical modeling into the field. Since then, numerous studies
involving various mathematical modeling approaches have been carried out.
In the 2000s, the widespread use of the Internet and the increase in user-generated
content brought new challenges and opportunities to this domain [2]. Researchers who
had to work with limited data in the past now have the opportunity to work with large
databases thanks to IMDB (Internet Movie Database), Box Office Mojo, Wikipedia,
Rotten Tomatoes, Criticker, and many other sites. Besides, while researchers were able
to obtain information coming from the audiences via more conventional techniques
1078 S. Çağlıyor and B. Öztayşi
such as questionnaires or focus groups, these online sources have accelerated infor-
mation flow from the consumer and increased the amount of information. However,
these opportunities have brought along challenges such as the correct processing and
correct analyzing and correct interpretation of these increased data as well as emerging
new variables.
One of the new variables that emerge with online movie database sites is the user-
generated rating. Ratings are important in two ways. On the one hand, it is a valuable
indicator in terms of understanding the consumer as it shows the audience’s reaction to
the movie, on the other hand, it is a signal that reduces the uncertainty for other
consumers who have not seen that movie yet. Today many cinema viewers check the
score of a movie on various sites when deciding whether to watch the movie.
Due to this importance, in this study, it is aimed to develop a model predicting such
ratings. In addition, in line with the idea that how many people voted for a movie is an
important indicator, an attempt was made to estimate not only the ratings but also the
number of votes. The film’s production year, country of origin, genre, actor/actress
information and assigned topics obtained from plots by applying LDA analysis were
used and three different machine learning algorithms, Random Forest (RF), Gradient
Boosting Tree (GBT) and Decision Tree (DT) were employed.
In the literature, user comments and social media contents were frequently analyzed
using various text analysis techniques and their impact on film’s success were exam-
ined. Although the plot is not influenced by personal ideas and experiences, it has an
important place in film marketing because of its direct effect on consumers’ the
decision-making process, yet it has rarely been used in prediction models [3, 4].
Another gap in the literature is the scarcity of data source variety. Since it is the
largest online database and provides prepared datasets, IMDb has been chosen as the
data source in the vast majority of studies in this field, but movies are cultural products
and tastes may vary from country to country [5]. Based on the idea that country-specific
studies will enrich the literature, we preferred to use a local source sinemalar.com.
The remaining part of the paper proceeds as follows: Sect. 2 briefly reviews the
literature forecasting of movie rating prediction. Section 3 deals with the variables that
are used in this study and our methodology. Section 4 presents the findings of the
research and finally, Sect. 5 includes a discussion of the findings, limitations, and
implications of the findings to future research into this area.
2 Literature Review
In the literature on the motion picture industry, how to formulate the motion-picture
success has been subject to considerable debate. While some studies focused on artistic
merits, some preferred to study commercial appeal, and some studies examined the
relationships between the two [6]. The indicators often used to express the success of a
movie were box office performance or expert evaluative judgments. With the wide-
spread use of the Internet viewer ratings and comments on the movie, review platforms
have become an important indicator of movie success as well.
Predicting Movie Ratings with Machine Learning Algorithms 1079
The vast majority of the rating studies in the literature focus on rating prediction for
single users due to the increase in movie recommendation systems [7–9]. Cohort
models have not been studied as much as single-user models. However, models built
using pre-production data are important for determining the marketing strategies of the
movies [10].
There are different data sources that have been used in studies that predict the
ratings of films. For instance, while Hsu, Shen, and Xie [11] used metadata features
such as genre, country, runtime, director in their model, Demir, Kapralova, and Lai
used Google Search frequencies [12]. Social media and microblogs are also the sources
used in rating prediction [13–15]. Since the viewers generally rate and comment after
watching a movie, models developed from social media or reviews can provide good
results in the post-production stage, but they are not suitable for the pre-production
phase.
In the motion picture industry, the timing of a model is important, as well as the
accuracy and comprehensiveness of the model. Considering the size of the investments
made and the cost of filming, pre-production decision support systems are very
important for determining marketing strategies such as correct distribution or adver-
tisement. Ning et al. [10] also highlighted this issue and used variables that are
obtainable at the preproduction stage such as MPAA (The Motion Picture Association
of America) rating, genre, player, budget, and plot information.
3 Methodology
This study consists of four stages: the formation of the dataset, applying text analysis to
the plots, testing the performance of prediction models and finally evaluating the
contribution of independent variables to the models’ predictive performance.
For the formation of the dataset, first information such as the rating, the number of
voters, country of origin, production year, the ratings and the number of voters, the first
three actors or actresses, and the plot of the movie are extracted with the help of a
parser. Then the raw data is cleaned and transformed. Rows with too many missing
values were excluded and finally, a dataset consisting of 8943 movies is obtained.
Studies showed topic modeling techniques such as Latent Dirichlet Allocation
(LDA) can better reveal what a plot is about [4, 10]. Therefore in the second stage of
the study, to generate topics, a LDA based model is used, where the textual corpus of
the plots is the input. First, to eliminate the terms that do not contain content, common
text preprocessing steps likes: stop word filtering, punctuation erasing, case conversion
performed. As in the other steps, ZemberekNLP (Natural Language Processing)
Library was used for stemming and POS tagging. Since they were very diverse at first,
to clean the tags to a certain degree some manual interventions were also needed. Based
on the plots 20 topics were generated.
After assigning the topics to the related documents along with other movie-related
variables, the rating and number of ratings are tried to be estimated by employing
Random Forest (RF), Gradient Boosting Tree (GBT) and Decision Tree (DT) algo-
rithms. The partition ratio is taken as 80:20. Average Percentage Hit Rate (APHR) is
used to test the performance of the different models.
Nmber of Samples Correctly Classified

APHR ¼ ð1Þ
Total Number of Samples
Once the model is developed, using the random forest’s attribute statistics, in other
words how often a single attribute is used at different levels of the tree, the contribution
of each variable to the predictive performance of the model is examined. A variable
being selected repeatedly as best split means that this variable is most likely an
informative attribute. As a measure for variable importance, the number of splits at
each level is divided with its candidates at the same level and results are summed. This
score basically tells relative to the other attributes – which are the most predictive ones.
This analysis is conducted for both ratings and the number of ratings, to observe if
there is a difference between the factors affecting the ratings or the number of the
ratings submitted.
3.1 Variables
Independent Variable. As is mentioned before, in this study we choose two different

independent variables: the ratings and the number of ratings that are submitted. Based
on the premise that a movie rated by more people can be interpreted as being watched
by more people or having higher product awareness, therefore like rating, the number
of ratings is also an important indicator, which might indicate whether a movie liked or
not. We approached the subject from a classification perspective. Based on the fre-
quency distribution and the values ratings were arbitrarily classified into three and the
number of rating into four.
Genre. The content or genre of a movie is often examined as an independent variable
in many studies in the movie industry domain [16, 17]. While fewer definitions were
used to describe the genre of a film in the past, more specific classifications are now
made. In this study, instead of using the first genre, we took the first three genres named
at sinemalar.com and devised genre variables in a 1-0 binary fashion. Instead of
showing each genre as a dummy variable, each genre was designed as a binary variable
that is assigned the value 1 if said genre is within the list. We obtained 31 genres from
sinemalar.com website. When the genre distribution is examined, it is seen that the vast
majority of the films belong to drama, comedy or action genre (see Fig. 1).
Actor. The actress effect is a controversial issue both in how to formulate it correctly and
how much impact it has on a film’s success [18]. However, for many movie watchers, the
well-known actress is seen as a quality signal. Many films usually employ three lead
actors or actresses. In most of the posters or movie review sites, the names of the top three
actors stand out. Focusing on only the first lead actor may not fully represent the star
power. Except that the second or third player can be a quality signal of a movie, in order to
take into account, the synergy between the actors, they should be included. From the actor
and actresses’ ratings and the number of ratings, nine different variables are generated.
Firstly, for each actor or actress, the ratings and number of ratings are indicated as
continuous variables. In order to capture the synergy or “ensemble effect” the ratings of
the three actors were averaged and denoted as a different variable. Additionally, if the two
18%
16%
14%
12%
10%
8%
6%
4%
2%
0%
Biograhy
Crime
Adventure
Other
Horror
Historyy
Drama
Comedy
Animasyon
Youth
Dance
Sci-Fi
AcƟon
Sports
Thriller
Family
EroƟc
RomanƟc
Fig. 1. Distribution of movie genres in the sample
actors previously acted together in another movie, this was also shown with a different
binary variable.
Year. The year variable indicates the production of a movie. The year variable indi-
cates the production of a movie. Some movies are distributed directly to online
streaming or VOD channels without being released to movie theaters. Also, the pos-
sibility that a movie can be watched not only in the movie theater but also on the VOD
or online streaming channels, it is thought that instead of the release date production
year will be a more accurate movie-related attribute.
Country of Origin. Literature suggests that a movie’s country of origin has an effect
on its theatrical demand [19]. Considering that each country has its own cinema style,
the country of origin is an important attribute about the film. Literature relating to the
country of origin suggests that knowledge about the country of origin influences
consumers’ interests [20, 21]. Most of the movies in our dataset are the US and Turkey
originated (see Fig. 2).
6000
5000
4000
3000
2000
1000
0
Fig. 2. Distribution of movies based on country of origin

4 Results
The results are summarized in Table 1. The results reveal that RF and GBT provide
similar performances. Both models are able to predict the approximately 55% of the
classes correctly. Decision tree on the other hand provides a lower accuracy.
Table 1. Prediction results

Algorithm Rating number Rating class
RF 56.96% 55.56%
GBT 57.74% 56.72%
DT 48.80% 47.23%
When then attribute statistics of random forest are examined it is seen the top 3
factors, which relatively contribute more to the rating predicting model are year, origin
and the average rating of three actors or actresses. In the rating number predicting
model, on the other hand, the number of ratings for each actor or actresses are the
variables with the highest sum of the split to candidate ratio. When in terms are genres
evaluated for rating prediction horror, drama and thriller relatively contribute more to
the model, where to number of rating model adventure, action, and sci-fi. In terms of
topic contribution to rating prediction model Topic 0. In both models Topic 1 and
Topic 2's relative contributions are small.
5 Conclusion
In this study, we try to develop a model forecasting the rating and number of ratings of
movies by using historical information of movies shot between 2000 and 2019.
A database is formed by web-scraping sinemalar.com consisting of 8943 movies. In
addition to variables from the website, LDA analysis is performed on the plots of the
movies. It has been observed that most of the topics contribute to both prediction
models, though not as much as the production year, country of origin or the actors.
There are some limitations in this article that offer opportunities for future research.
Firstly, in this study, the target value classification was made by trial and error, a more
systematic classification can positively affect the performance of models. Apart from
that, creating a more elaborate model by adding other movie-related features (director,
specifications, studio, budget, result) can both help to increase accuracy and provide a
deeper understanding of factors that can affect ratings. Along with other movie-related
variables, user-generated comments can be examined too.
References
1. Eliashberg, J., Elberse, A.A., Leenders, M.A.A.: The motion picture industry: critical issues
in practice, current research, and new research directions. J. Mark. Sci. 25(6), 638–661
(2006)
2. Waldfogel, J.: Cinematic explosion: new products, unpredictabilty and realized quality in the
digital era. J. Ind. Econ. 64(4), 755–772 (2016)
3. Hung, Y., Guan, C.: Winning box office with the right movie synopsis. Eur. J. Mark. 54(3),
594–614 (2020)
4. Lash, M.T., Zhao, K.: Early predictions of movie success: the who, what, and when of
profitability. J. Manag. Inf. 33(3), 874–903 (2016)
5. Lorenzen, M.: Internationalization vs. globalization of the film industry. J. Ind. Innov. 14(4),
349–357 (2007)
6. Holbrook, M.B., Addis, M.: Art versus commerce in the movie industry: a two-path model
of motion-picture success. J. Cult. Econ. 32(2), 87–107 (2008)
7. Lund, J., Ng, Y.K.: Movie recommendations using the deep learning approach. In:
Proceedings of the IEEE19th Conference on Information, pp. 47–54 (2018)
8. Mukta, M.H.S., Khan, E.M., Ali, M.E.: Predicting movie genre preferences from personality
and values of social media users. In: Proceedings of 11th Conference on Social Media,
pp. 624–627 (2017)
9. Khopkar, S., Nikolaev, A.G.: Predicting long-term product ratings based on few early ratings
and user base analysis. Electron. Commer. Res. 21, 38–49 (2017)
10. Ning, X., Wang, L., Yac, X.B.: Rating prediction via generative convolutional neural
networks based regression. PatternRecognit. Lett. 132, 1–9 (2018)
11. Hsu, P.-Y., Shen, Y.-H., Xie, X.-A.: Predicting movies user ratings with Imdb attributes. In:
International European Conference on Information Retrieval, pp. 444–453. Springer,
Heidelberg, (2014)
12. Demir, D., Kapralova, O., Lai, H.: Predicting IMDB movie ratings using Google Trends.
Dept. Elect. Eng. Stanford Calif. 2, 1–5 (2012)
13. Oghina, A., Breuss, M., Tsagkias, M.: Predicting IMDB movie ratings using social media,
pp. 503–507 (2012)
14. Kesharwani, M., Rakesh, B.: Movie rating prediction Twitter sentiment analysis. J. Adv.
Comput. Technol. 5, 2347–2804 (2017)
15. Özkaya, A.E., Sert M. : Movie rating prediction using ensemble learning and mixed type
attributes. In: 25th Signal Processing and Communications Applications Conference (SIU),
pp. 1–4. IEEE (2017)
16. Eliashberg, J., Sawhney, M.S.: A parsimonious model for forecasting gross box-office
revenues of motion pictures. J. Mark. Sci. 15, 113–131 (1996)
17. Lovallo, D., Clarke, C., Camerer, C.: Robust analogizing and the outside view two empirical
test of case-based decision making. Strategic Manage. J. 33, 496–512 (2012)
18. Kim, T., Hong, J., Kang, P.: Box office forecasting using machine learning algorithms based
on SNS data. J. Forecast. 31(2), 364–390 (2015)
19. Wallace, W., Seigerman, A., Holbrook, M.B.: The role of actors and actresses in the success
of films: how much is amovie star worth. J. Cult. Econ. 17(1), 1–27 (1993)
20. Verlegh, P., Steenkamp, J.: Review and meta-analysis of country-of-origin. J. Econ. Psychol.
20, 521–546 (1999)
21. Gazley, A., Clark, A., Sinha, A.A.: Understanding preferences for motion pictures. J. Bus.
Res. 64(8), 854–861 (2011)
Trajectory Tracking of a Quadcopter
Using Adaptive Neuro-Fuzzy Controller
with Sliding Mode Learning Algorithm
Hasan Kemik, Mehmet Berkin Dal, and Yesim Oniz(B)
Department of Mechatronics Engineering, Istanbul Bilgi University, Istanbul, Turkey

[email protected], [email protected], [email protected]
Abstract. In this work, the trajectory tracking control of an unmanned

aerial vehicle (UAV) has been accomplished using adaptive neuro-fuzzy
controllers. The update rules of the proposed controller have been derived
based on the sliding mode control theory, where a sliding surface has been
generated utilizing the parameters of the neuro-fuzzy controller to direct
the error towards zero in a stable manner. To assess the effectiveness
of the proposed control scheme, Parrot AR.Drone 2.0 has been utilized
as the test platform, on which conventional PID and fuzzy logic con-
trollers have been also implemented to provide means for comparing the
performance of the proposed controller. Different reference trajectories
have been generated for the real-time experimental studies, in which the
discrepancies from these trajectories are used to determine the input sig-
nals to be applied to the proposed controllers. The analytical claims have
been justified by the obtained results from the real-time experiments in
the presence of large nonzero initial errors.
1 Introduction
There is a growing interest in drones because of their simple mechanical struc-
ture and high maneuverability. Especially within the last two decades, drones have
been started to be used in many diverge areas ranging from delivery to damage
assessment after disasters. Although drones are offering many promising features
that will promote their usage in numerous different sectors in an ever-increasing
manner, they pose several challenges in terms of the control, as they are inher-
ently unstable and have nonlinear coupled dynamics. There are numerous on-going
research projects to address these shortcomings [4,8,10], and in the present paper
an adaptive neuro-fuzzy control method that is capable of dealing with both cou-
pled and non-linear characteristics of the drones is suggested.
In the earlier studies on drone control methods, model based controllers have
been utilized extensively [1,2,6]. However, the performance of these controllers
may significantly degrade in the presence of uncertainties in system parameters
and disturbance. The main reason for this can be stated as that the perfor-
mance of such controllers is highly dependent on the accuracy and precision
Trajectory Tracking of a Quadcopter 1085
of the system’s mathematical model, which is generally derived under various

assumptions for nearly ideal conditions. However, in the presence of uncertain
parameters and unmodeled dynamics the performance of such controllers may
significantly degrade. Model free methods including fuzzy logic controllers (FLC)
and the neural networks (NN) can be used to overcome this shortcoming.
In fuzzy logic, the membership functions are constructed based on the expert
knowledge, and the outputs are computed using these membership functions.
One of the main shortcomings of fuzzy logic is that if the required expert knowl-
edge is missing or incomplete, the resulting control action could fail in achiev-
ing the desired performance specs. To address this shortcoming, hybrid control
schemes combining fuzzy logic with neural networks can be used, in which the
learning capability of the neural networks has been utilized in determining the
parameters of the membership functions of the fuzzy logic.
The parameter adaptation in neuro-fuzzy controllers has been generally
accomplished using gradient-based methods. However, the stability of the learn-
ing is not guaranteed in these algorithms because of the entrapment of the local
minima. In this study, variable structured systems-based learning algorithms
have been employed to assure stable online tuning of the parameters.
This paper is organised as follows: Sect. 2 starts with a brief description of the
general structure of the AR.Drone. The details of the proposed controller along
with the derivation of the adaptation laws simulated are presented in Sect. 3.
Section 4 is devoted the obtained results from the real-time experimental studies.
Concluding remarks are given in Sect. 5.
2 The AR.Drone Platform

As shown in Fig. 1, drones possess six degrees of freedom, three of which are
translational and three of which are rotational. Similar to the aeroplanes, the
term pitch is used to describe rotational action of movement along the y-axis,
whereas roll is corresponding to a rotational action of movement along the x-
axis. Furthermore, in the drones translational movement along the x-axis can be
generated by adjusting the pitch value, and translational movement along the
y-axis is resulting from the rolling movement.
For the experimental studies, AR.Drone 2.0 has been selected as the test
platform because of its low cost and easily availability. The proposed control
algorithms have been implemented in Python and the resulting control signals
have been sent to the drone via WiFi connection. The following set of variables
can be fetched from the sensor board of the Ar.Drone
z : the altitude of the drone relative to the ground

vx : the linear velocity of the drone in x-direction
vy : the linear velocity of the drone in y-direction
θ : pitch of the drone
φ : roll of the drone
ψ : yaw of the drone
1086 H. Kemik et al.
Fig. 1. Overall structure of AR.Drone
Connection to the AR.Drone’s internal flight controller is restricted, where

only the following set of motion control commands scaled to the interval [−1, 1]
can be set:
uż : linear velocity command in the z-direction

uψ̇ : yaw rate
uθ : pitch command
uφ : roll command
where uθ and uφ are used to control the translational velocities of the drone
along the x- and y-directions, respectively.
3 Adaptive Neuro-Fuzzy Controller

Control of the drone can be considered a challenging task, since the performance
of a controller is highly dependent on the accuracy of the system’s mathematical
model, and this a rather difficult task for the drones. In addition, the data com-
ing from the sensors are inevitably noisy and comprise significant uncertainty.
Under these conditions, model-free solutions may be considered as a feasible
alternative, where an approximate model is usually sufficient to meet the per-
formance requirements. Neural networks (ANN) [5] and fuzzy logic controllers
(FLC) [3,7] can be called among the most widely used model-free approaches.
The use of hybrid controllers combining fuzzy systems with neural networks
has been a popular approach in engineering field for many years to solve control
and identification problems [11]. This kind of structure combines the learning
ability of the neural networks with the linguistic rule interpretation capability
of fuzzy systems. In this paper, the development of an adaptive neuro-fuzzy
architecture with two inputs and one output is considered for the trajectory
tracking problem of a drone. The structure of the controller is presented in
Fig. 2.
Fig. 2. Structure of neuro-fuzzy inference system
In the presented structure, each incoming signal is fuzzified by using Gaussian

membership functions using the following equations:
2
(x1 − cAi )
2
x2 − cBj
μAi (x1 ) = exp − 2 μBj (x2 ) = exp − 2 (1)
σA i
σB j
where i = 1, .., n and j = 1, .., n. In (1) c and σ terms are standing for the mean
and standard deviation of the corresponding Gaussian membership function,
respectively.
The obtained membership function values are used in the premise part of
the IF-THEN rule base of Takagi-Sugeno type inference system. The consequent
part fij of the rules consists of a linear function of the input parameters. The
general form of the rules Rij can be written as follows:
Rij : If x1 is Ai and x2 is Bj , then fij = ai x1 + bj x2 + dij
where ai , bj and dij are some constants that will be updated using the learning
algorithm.
To compute the strength of the rule Rij , T -norm is used for the membership
functions of the premise part as follows:
Wij = μAi (x1 )μBj (x2 ) (2)
The output signal of the fuzzy-neural network is denoted by τn (t) and it can
be considered as the weighted sum of the strehths of the rules:

I
J
τn (t) = fij Wij (3)
i=1 j=1
3.1 The Sliding Mode Learning Algorithm

The membership functions’ parameters of the premise part, namely cAi , cBj ,
σAi and σBj , along with the parameters of the consequent part (ai , bj and
dij ) will be updated such that the output of the network τn (t) will be of the
desired value. The adaptation of the aforementioned parameters is generally
accomplished by using the gradients of an error function, which is commonly
described as the difference between the desired and actual output values of the
network. On the other hand, learning algorithms relying on the gradients suffer
from the possibility of getting stuck into one of the local minima. To alleviate
this problem, in this paper the use of a learning algorithm relying on sliding
mode control theory is preferred, in which the main objective is to direct the
states of the system onto a predetermined plane and keep them moving along
this plane to the global minimum of the system.
In the sliding mode control (SMC) theory, generally the discrepancies
between the desired and actual values of the system’s output and its deriva-
tives are employed to construct the sliding surface.
s (e, ė) = ė + λe (4)
with λ being a positive constant determining the slope of the sliding surface,
and e = τn − τd , where τd being desired output of the system.
Once the sliding surface is reached, to keep the system’s states on it the
following conditions should be satisfied:
s (e, ė) = 0 and ṡ (e, ė) = 0 (5)

For the case, when the system’s states start outside of the sliding surface, they
should be forced to approach this surface at least asymptotically. This requires
lim ṡ < 0 and lim ṡ > 0 (6)

s→0+ s→0−
which is commonly referred to as the reachability condition. Lyapunov stabil-

ity method can be used to enforce it [9], where a positive Lyapunov function
candidate with a negative definite derivative is required.
1 2 ∂s
V (s) = s ≥ 0 and V̇ (t, s) = s ≤0 (7)
2 ∂t
If a Lyapunov function candidate satisfying both conditions can be deter-
mined, then it is guaranteed that the system states will approach the sliding
surface asymptotically and remain on it for all succeeding times.
In this study, the following Lyapunov function candidate has been selected:
1 2 1 1 2
V (e(t)) = s = e2 (t) = (τn (t) − τd (t)) = 0 (8)
2 2 2
Differentiating the selected V (e(t)) results in the following condition:
V̇ (e(t)) = eė = e(τ̇n − τ̇d ) ≤ 0 (9)

where chain rule has been employed to find the time derivative of τn .
dτn ∂τn dfij ∂τn dcAi ∂τn dcBj ∂τn dσAi ∂τn dσBj
= + + + + (10)
dt ∂fij dt ∂cAi dt ∂cBj dt ∂σAi dt ∂σBj dt
dτn dfij 2fij Wij (x1 − cAi ) dcAi 2fij Wij (x2 − cBj ) dcBj
= Wij + 2 + 2
dt dt σA i
dt σB j
dt
2fij Wij (x1 − cAi )2 dσAi 2fij Wij (x2 − cBj )2 dσBj
+ 3 + 3
σAi dt σBj dt
Assuming τd is constant, i.e. τ̇d = 0, if the adaptation rules for the parameters
are selected as:
dfij
= −αsgn(e) (11)
dt
dcAi
= −αfij (x1 − cAi )sgn(e) (12)
dt
dcBj
= −αfij (x1 − cBj )sgn(e) (13)
dt
dσAi
= −αfij σAi sgn(e) (14)
dt
dcBj
= −αfij σBj sgn(e) (15)
dt
the derivative of the Lyapunov candidate function with the selected adaptation
laws yields:

(x1 − cAi )2 (x2 − cBj )2
V̇ = −αesgn(e)Wij 1 + 4fij 2
2 − 2 (16)
σA i
σB j
which implies that V̇ < 0 ∀e = 0.

To verify the effectiveness of the proposed learning algorithm, real-time exper-
iments have been carried out on the drone for different trajectories. For each
translational axis of the drone, a separate neuro-fuzzy controller has been
employed. The inputs of these controllers have been selected as the discrep-
ancies from the target trajectories and their derivatives in x-, y- or z-directions.
Five membership functions have been assigned to each input. And initially these
membership functions have been distributed uniformly to cover the entire input
range.
In the first set of the experiments, a squared shape trajectory in the xy-plane
has been generated. In another set, a S-shaped trajectory has been constructed
in the xz-plane. At the beginning of each experiment, the parameters of the
consequent part of the fuzzy rules have been initialized randomly. The learning
rate α for each NFC has been selected using trial-and-error method (α(x) = 2.73,
α(y) = 0.73, α(z) = 2.53), and the sampling time has been set to 0.025 s to enable
a stable communication with the drone over Wi-Fi. Along with the neuro-fuzzy
controller, PID and conventional FLCs have been also used to provide means
for comparison. All experiments have been carried out indoors in a controlled
environment. The results of these experiments have been presented in Figs. 3
and 4 and the RMSE values are given in Table 1.
5 120
Reference Reference
PID PID
115
FLC FLC
0 NFC NFC
110
-5 105
100
-10
Y(cm)
Z(cm)
95
-15
90
-20 85
80
-25
75
-30 70
-5 0 5 10 15 20 25 30 35 -2 0 2 4 6 8 10 12 14 16
X(cm) X(cm)
Fig. 3. Square-shaped trajectory Fig. 4. S-shaped trajectory
Table 1. RMSE values for x-, y-, and z-directions
Square-shaped trajectory in xy-plane S-shaped trajectory in xz-plane

RMSE value in x-direction y-direction z-direction x-direction y-direction z-direction
PID 0,66 1,77 3,49 1,77 0,80 7,91
FLC 1,72 2,1 1,34 1,52 1,32 3,27
NFC 2,27 1,24 1,05 1,14 0,71 3,03
Regarding the presented results, it can be observed that the proposed NFC
can track the given trajectory closely, which can be also inferred from the RMSE
values provided in the Table 1. The NFC controller generally provides the small-
est amount of the error. The only exception is the error in the x-direction for
the square-shaped trajectory. For the NFC, the largest deviations from the ref-
erence trajectory occur at the corners, at which there is a sudden change in one
direction that requires the readjustment of the network parameters. The time
required for the adaptation and delays associated with the communication might
be the reason that RMSE value of NFC is larger, compared to PID and FLC.
5 Conclusion
In this paper, a learning algorithm relying on sliding mode control theory has
been adopted to tune the parameters of a neurofuzzy controller. In the derivation
of the adaptation laws, Lyapunov stability method has been used. The developed
algorithm has been tested for various trajectories in real time, whose results
justify the stability and robustness of the learning process. Encouraged by the
experimental results, a new study is to be launched in which fuzzy type-1 sets
will be replaced by fuzzy type-2 sets.
References
1. Alaimo, A., Artale, V., Milazzo, C.L.R., Ricciardello, A.: Pid controller applied to
hexacopter flight. J. Intell. Robot. Syst. 73(1–4), 261–270 (2014)
2. Das, A., Subbarao, K., Lewis, F.: Dynamic inversion with zero-dynamics stabilisa-
tion for quadrotor control. IET Control Theory Appl. 3(3), 303–314 (2009)
3. Das, H., Mazumdar, A.S., Dey, R., Roy, L.: Experimental implementation of fuzzy
vision-based tracking control of quad-rotor. In: International Workshop Soft Com-
puting Applications, pp. 297–313. Springer (2016)
4. Hua, M.D., Hamel, T., Morin, P., Samson, C.: A control approach for thrust-
propelled underactuated vehicles and its application to vtol drones. IEEE Trans.
Autom. Control 54(8), 1837–1853 (2009)
5. Rohan, A., Rabah, M., Kim, S.H.: Convolutional neural network-based real-time
object detection and tracking for parrot AR drone 2. IEEE Access 7, 69575–69584
(2019)
6. Salih, A.L., Moghavvemi, M., Mohamed, H.A., Gaeid, K.S.: Modelling and PID
controller design for a quadrotor unmanned air vehicle. In: 2010 IEEE International
Conference on Automation Quality and Testing Robotics (AQTR), vol. 1, pp. 1–5.
IEEE (2010)
7. Santoso, F., Garratt, M.A., Anavatti, S.G.: A self-learning TS-fuzzy system based
on the c-means clustering technique for controlling the altitude of a hexacopter
unmanned aerial vehicle. In: 2017 International Conference on Advanced Mecha-
tronics, Intelligent Manufacture, and Industrial Automation (ICAMIMIA), pp. 46–
51. IEEE (2017)
8. Tweedale, J.W.: Fuzzy control loop in an autonomous landing system for
unmanned air vehicles. In: 2012 IEEE International Conference on Fuzzy Systems
(FUZZ-IEEE), pp. 1–8. IEEE (2012)
9. Utkin, V.I., Guldner, J., Shi, J.: Sliding Mode Control in Electromechanical Sys-
tems, vol. 9. CRC Press, Boca Raton (1999)
10. Varga, M., Zufferey, J.C., Heitz, G.H.M., Floreano, D.: Evaluation of control strate-
gies for fixed-wing drones following slow-moving ground agents. Robot. Auton.
Syst. 72, 285–294 (2015)
11. Yager, R., Zadeh, L.: Fuzzy Sets, Neural Networks and Softcomputing. Thomson
Learning, New York (1994)
Influencer Identification System Design Using
Machine Learning Techniques
Elvira Israfilova(&), Armagan Arslan, Nihan Yildirim,

and Tolga Kaya

{israfilova15,arslanar,yildirimni,
kayatolga}@itu.edu.tr
Abstract. Being one of the most effective spreading methods, word-of-mouth

networking moved to its digital version - social media platforms. Social media
represents a digital platform which allows users to interact with each other in
different ways. Audio, video, image or text content are examples of such inter-
actions. Widespread use of social media platforms brings new ways to marketing
solutions in order to reach potential audience, especially thorough those users who
actively interact with and have an influence on their followers. Brands which
utilizes abilities of such users of social media are involved in field of influencer
marketing. One of the challenges in this rapidly growing field is finding person
(influencer) which will match requirements of the brand. Rising number of
influencers and social media platforms leads to increasing data and difficulties in
finding appropriate influencer. The purpose of this study is to provide design of an
influencer identification system utilizing machine learning algorithms. Although
the number of Influencers rising, they are still in minority comparing to all users of
social media. Thus, focusing on solving problem of imbalanced classification,
performances of Logistic Regression, Linear Discriminant Analysis, K-Nearest
Neighbors, Support Vector Machines and different tree based methods were
compared and Random Forest method is selected to be used in the system.
Keywords: Influencer Influencer identification Social media Twitter

Random forest Classification algorithms Engagement rate Decision support
system
1 Introduction
Social media has become a massive part of today’s daily life of individuals and
companies. Moreover, it become a tool where individuals share their thoughts, feelings,
and news on a certain topic [1]. Therefore, people learn news, others opinions and other
issues from each other using social media. The intended use of social media has effect
on rapidly expanding in marketing activities [2]. However, it is crucial to distinguish
social media marketing and influencer marketing. Social media marketing signifies that
brands/businesses uses their own social media accounts to encourage themselves and/or
their products. Whereas, influencer marketing, being a subdomain of marketing
(affiliate marketing) builds a communication channel via influencers [2].
Influencer Identification System Design 1093
Those users of social media who have ability to influence others and produce
messages of high social quality, importance and recognition are said to be influencers,
opinion leaders, domain experts [3]. These individuals are responsible for triggering
long cascades of influence that convince their peers to perform a similar action (for
instance, buying a product). Targeting these influencers usually leads to a vast spread
of information across the network [4]. It is argued that influencers have multiple
interests and tend to be early adopters in markets. Moreover, they are trusted by others
and have a large social network [5].
There are different categorizations of influencers depending on the number of fol-
lowers. Narassiguin and Sargent classified influencers as micro, regular, rising, mid,
macro and mega depending on the number of their followers, mentioning that regular and
rising influencers are more desirable for the business needs [6]. Due to deep knowledge in
the specific field micro-influencers are known as most engaged ones. Micro-influencers
built trusted relationships with followers through the high level of interaction with them,
hence they have potential to be as the influencers of the future. Depending on aim of the
brand campaign (new product release, increasing trust to product etc.) businesses should
properly select right type of influencer for the campaign [7].
The purpose of this study is to provide a design of an influencer identification
system utilizing machine learning algorithms. Performances of Logistic Regression,
Linear Discriminant Analysis, K-Nearest Neighbor, Support Vector Machines and
various tree based methods were compared and tree based Random Forest method was
selected to be used in the system.
Paper is organized as follows: Next section provides a methodology of the system
design. At the third section different machine learning algorithms compared and
method with the best performance proposed for using in the system. Conclusions about
Influencer Identification System described in the final section of the paper.
2 Data and Methodology
We performed primary and secondary market research in order to identify most

important issues related to the topic of the study. We interviewed eight experts in the
field of digital, social media and influencer marketing both in Turkey and abroad. There
are three CEO of different social media agencies, two Digital Marketing Specialist, one
Head of Data Analytics in media agency, and one Digital Business Developer and
Management Consultant with four to ten years of the experience in the related field
were interviewed. As secondary market research numerous influencer marketing
platforms, sector related reports and academic articles were studied. Following issues
are to be considered in the system. Firstly, social media account age of potential
effective influencer in Turkey must be not older than three and half years (except
celebrities). Additionally, an option for selecting influencers according to the aim of
campaign should be added to the user interface, since influential power on followers
differs with the type of influencer [8]. Micro-influencers are effective in increasing trust
to the brand/product or service, while macro-influencers have power of spreading
product/services to wide auditory [9]. Furthermore, followers are the key target of the
marketers, thus the consistency between influencer’s interests and his/her audience
1094 E. Israfilova et al.
interests have to be evaluated. It should be mentioned that social media platforms are
highly dynamic and periodical database update is required in this context.
As Oztaysi et al. stated, “system design is the process of designing the architecture,
modules, interfaces and data for a system to satisfy specified requirements” [10].
Influencer Identification system is a decision support system (DSS) which supposes
existence of following functions: information inflow, data storage, data analysis and
meaningful output for decision makers. A database, software system and user interface
are the main three components of DSS [11].
Four most effective social media platforms [12] proposed to be integrated to the
system: Instagram, YouTube, Facebook and Twitter. Each platform has differently
named but logically the same user and followers related features.
Information inflow to the system supposed to be directly from social media plat-
forms. Each of them has its own Application Programming Interface (API) which
allows creation of applications that access data and features of other applications,
services or operating system [13]. In most cases, data is kept in JSON format and then
if needed, it can be converted to the CSV format, which both are suitable for data
analysis and applying machine learning algorithms in the R-Studio environment.
Database management systems (DBMS) can be divided in two categories – SQL
based and NoSQL based DBMS. Social media data is highly dynamic and can be
supposed as Big Data since the number of “sharing activity” almost uncountable.
NoSQL database has some advantages over SQL databases. Being non-relational
database, entities in NoSQL database can be duplicated, which in consequence
increases speed of processing. In addition, there is no need for predefined schemas
which makes it useful in the conditions of constantly changing API. Moreover, NoSQL
based DBMS’s are less costly and deliver faster performance than SQL based ones.
Finally, SQL based databases do not support JSON data format. After detailed
benchmarking of existing NoSQL based DBMS among couchbase, CouchDB and
MongoDB, the best option occurred as MongoDB.
For data analysis and classification model construction and performing we have
selected R language. R-Studio is open-source software with large number of packages
for performing data analysis and modeling can be integrated to the MongoDB. R-
Studio also provides cloud and server services which can be utilized by the system.
Result of the system components analysis shown in the data flow diagram in the Fig. 1.
3 Influencer Identification Algorithm Modelling
We choose Twitter for the classification algorithm modeling since it has an open-source
API for developers. Considering Influencer as domain expert we determined hashtag
related to the chosen topic. Hashtag is the type of label for the certain content [14]. It
used for finding sharing related to the content. On the time of the experiment Turk-
ish TV Series was released at the Netflix platform, thus we selected Netflix related
hashtags to retrieve data. We made a research on most used hashtags related to the
Netflix in Turkey and find out following hashtags: #Netflix, #netflixtürkiye, #netflix,
#netflixturkey, #netflixTürkye, #netflixturkiye.
Fig. 1. Data flow diagram for influencer identification system
We retrieved data from Twitter for three days at the different time and time intervals
in December 2019. Dimension of retrieved raw dataset is 12757 observations over 90
variables. Some of variables was excluded from dataset due to absence of any data in
column, while some of them was useless at this stage of modeling (ex. text data, web
page link etc.) Excluding users without any followers, final dataset has dimension of
9712 observations over 18 variables.
Identification of Influencer is the two-class classification problem. Users divided
into two classes: Influencer (I) and Non-Influencer (NI). Relying on academic and
social media experts’ knowledge, we added following features to the dataset in order to
assign class to each user and train the model: account age, follower to following ratio,
average daily tweet count and average number of followers gained in daily basis. For
each observation we calculated these derived features. Thresholds for Influencers are as
follows:
– Account age is less or equal to three and half years;
– Follower to following ratio more than one,
– Average daily followers gain number between one and eight hundred and
– Average daily tweet number less than one hundred.
Users who simultaneously meet these requirements supposed to be an Influencers.
After assigning classes to users these derived four features was excluded from dataset
and final dataset consists of 9679 observations (I = 33, NI = 9679) over 18 variables
(features). Meaning of each feature given in the table below (Table 1).
Classification methods in machine learning are used in order to predict classes of
each observation in a dataset using trained model. In this paper we aim to classify each
observation (user) as Influencer (I) or Non-Influencer (NI). We applied Logistic
Regression, Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), Sup-
port Vector Machines (SVM) and different tree based methods. These methods are
briefly described below [15].
Table 1. Features used in influencer identification classification modeling.

Variable Meaning Type of
variable
source Utility used to post the Tweet Categorical
is_quote Identification whether tweet was quoted or not Boolean
is_retweet Identification whether tweet was retweeted or not Boolean
quoted_favorite_count Number of tweets liked by Twitter users who has Numeric
quoted the original tweet.
quoted_retweet_count Number of tweets are currently retweeted by Numeric
Twitter users who has quoted the original tweet
quoted_followers_count number of followers the account of user quoted Numeric
tweet currently has
quoted_friends_count Number of friends the account of user quoted Numeric
tweet currently has
quoted_statuses_count Number of tweets posted by user quoted current Numeric
tweet
retweet_favorite_count Number of tweets liked by Twitter users who has Numeric
retweeted the original tweet
retweet_retweet_count Number tweets are currently retweeted by Twitter Numeric
users who has retweeted the original tweet
retweet_followers_count Number of followers the account of user retweeted Numeric
tweet currently has
retweet_friends_count Number of friends the account of user retweet Numeric
tweet currently has
retweet_statuses_count Number of tweets posted by user quoted current Numeric
tweet
followers_count Number of followers of tweet owner Numeric
friends_count Number of friends of tweet owner Numeric
listed_count Number of public lists that this user is a member Numeric
of
statuses_count Number of Tweets issued by the user Numeric
favourites_count Number of Tweets this user has liked Numeric
Logistic Regression is one of the classification methods which assigns probability

that response belongs to one of the predetermined categories. Distribution of proba-
bilities is between 0 and 1 and most important tuning parameter is threshold for
probability [15]. LDA also estimates probabilities of certain response but not in direct
manner as it Logistic Regression method does. It models distribution of predictors
separately in each of the response classes and then utilizes Bayes’ Theorem to flip these
around into estimates for overall probabilities. Being a non-parametric method of
machine learning, KNN method used both for classification and regression problems.
This method firstly identifies K number of neighbors close to observation we want to
classify. Probability to assign the one of a class related to the fracture of points around
this observation which belongs to the same class. Most important tuning parameter for
KNN is K value, thus it is useful to find such value of K utilizing which model
performs better. SVM is an extension of Support Vector Classifier which suppose that
classes separated by linear boundary. However, SVM can produce non-linear (poly-
nomial, radial) boundary. SVM approach is to find such hyperplane and determine such
boundary margin which will clearly separate classes in classification problem [15].
Finally, tree based algorithms are visual represented algorithms. Each visualization
consists of nodes and branches. Each node of tree represents the point of decision based
on each input variable and through these nodes classifier assigns classes to each
observation.
Classification models focuses on predicting major class in the dataset [16]. In our
study there are two classes – Influencer, which is minor and Non-Influencer, which is
major class in the dataset. Minor class is class of prime interest of this study. In order to
balance class distribution resampling methods should be applied. We applied combined
resampling methods in order to construct model for imbalanced dataset. Cateni et al.’s
study proposes a combined resampling method which includes principles of both
oversampling and undersampling [16]. Proposed model follows structure, where N is
number of observations to be removed or added to the dataset:
N ¼ round½ðk1 n0 Þðk0 n1 Þ ð1Þ
Table 2. Comparison of algorithms for influencer identification

Classifier Predicted Actual Balanced model accuracy
class class (%)
I NI
Logistic regression I 29 1168 87.9
NI 4 8511
Linear discriminant analysis I 30 3434 77.7
NI 3 6245
SVM/Linear I 32 1578 90
cost = 0.50 NI 1 8101
SVM/Polynomial I 27 1080 85.3
degree = 3, scale = 0.1, c = 1 NI 6 8599
K-nearest neighbors I 26 236 88.2
K=5 NI 7 9443
Random forest I 29 24 93.8
mrty = 13 NI 4 9655
AdaBoost classification tree I 25 6 87.8
niter = 50, NI 8 9673
method = AdaBoost.M1
Conditional tree I 27 225 89.7
mincriterion = 0.5 NI 6 9454
Neural networks I 15 2270 61
size = 5, decay = 0 NI 18 7409
Target percentages of samples for major and minor classes given by parameters k0
and k1 respectively, sum of which equals 1 [16]. However, proposed approach takes in
to consideration that overall class distribution should be sustained and minor class
should not prevail over major one. When two subsets (one consists of undersampled
major class, the other one consists of oversampled minor class) are built, they merged
to create new resampled training dataset. This dataset we used in training model, while
model evaluation is performed using initial imbalanced dataset.
Resulting contingency matrices for each model shown in Table 2. Besides the
overall model accuracy, rate of correctly assigned Influencers is one of the most
important metrics for this problem. Thus we evaluated model relying on balanced
model accuracy instead on overall model accuracy, which calculates each class pre-
diction separately and then takes mean value of them:
Balanced model accuracy ¼ ðTP=P þ TN=NÞ=2 ð2Þ
We find out that model which utilizes Random Forest method with number of
predictors sampled for splitting at each node mrty = 13, outperforms other methods
thus we propose this method to be used in the system.
4 Conclusion
In this study, different machine learning techniques was applied to the problem of
influencer identification and model with the highest performance level was proposed to
be used in the designed system. Imbalanced dataset requires applying resampling
methods in order to obtain balance between major and minor classes and construct
model. We used combined resampling method which includes both oversampling and
undersampling techniques. Then, we trained model utilizing Logistic Regression,
Linear Discriminant Analysis, K-Nearest Neighbor, Support Vector Machines and
different tree based methods. After the comparison of models, we proposed the model
based on Random Forest algorithm with 93.8% balanced class prediction accuracy. We
designed system in such a way that there is overall consistency between data structure,
database management system and analytics software. Periodical database updating
procedures, social media data volume and the data structure (format) require NoSQL
based database management system which can be integrated with data analytics soft-
ware - R-Studio.
In previous research and practices, the mostly adapted models which provided high
accuracy can be listed as Random Forrest and SVM [1, 17]. Our study, support the
model in general, however differ from these by applying different resampling method
which resulted in more reliable results for imbalanced dataset. It can be concluded that
the tree based Random forest model with combined resampling method for imbalanced
datasets served as the one with the highest accuracy.
In future studies, proposed model can be improved by adding Sentiment Analysis
and developing a new Engagement Rate model, which will enable the usage of weights
for different interaction types in social media.
References
1. Şimşek, M., Kabakuş, A.T.: Twitter Üzerindeki Etkili Bireylerin Makine Öğrenmesi
Sınıflandırma Algoritmaları İle Tespiti. Gazi Mühendislik Bilimleri Dergisi 4(3), 183–195
(2018)
2. Kemp, S.: Digital 2019: Q3 global digital statshot. Essential insights into how people around
the world use the internet, mobile devices, social media and e-commerce. https://
datareportal.com/reports/digital-2019-q4-globaldigital-statshot. Accessed 09 Mar 2020
3. Razis, G., Anagnostopoulos, I., Zeadally, S.: Modelling influence with semantics in social
networks: a survey. ACM Comput. Surv. 53(1), 1–38 (2020)
4. Dave, K., Bhatt, R., Varma, V.: Identifying influencers in social networks. In: 5th
International Conference on Weblogs and Social Media 2011, pp. 1–9. AAAI, Palo Alto
(2011)
5. Kiss, C., Bichler, M.: Identification of influencers—measuring influence in customer
networks. Decis. Support Syst. 46(1), 233–253 (2008)
6. Narassiguin, A., Sargent, S.: Data science for influencer marketing: feature processing and
quantitative analysis. arXiv:1906.05911v1 (2019)
7. The State of Influencer Marketing 2020: Benchmark Report. https://influencermarketinghub.
com/influencer-marketing-benchmark-report-2020/. Accessed 09 Mar 2020
8. Schouten, A., Janssen, L., Verspaget, M.: Celebrity vs. influencer endorsements in
advertising: the role of identification, credibility, and product-endorser fit. Int. J. Advert. 38,
258–281 (2019)
9. Coulter, K.S., Roggeveen, A.: “Like it or not” consumer responses to word-of-mouth
communication in on-line social networks. Manage. Res. Rev. 35(9), 878–899 (2012)
10. Onar, S.C., Oztaysi, B., Kahraman, C., Öztürk, E.: A decision support system to optimize
debt collection assignments. In: Kahraman, C., Cebi, S., Cevik Onar, S., Oztaysi, B., Tolga,
A., Sari, I. (eds.) Intelligent and Fuzzy Techniques in Big Data Analytics and Decision
Making. INFUS 2019. Advances in Intelligent Systems and Computing, vol. 1029. Springer,
Cham (2020)
11. Launden, K.C., Launden, J.P.: Management Information Systems: Managing the Digital
Firm, 15th edn. Pearson, London (2018)
12. Ahmad, I., The most popular social media platforms of 2019. https://www.
digitalinformationworld.com/2019/01/most-popular-global-social-networks-apps-
infographic.html. Accessed 09 Mar 2020
13. Bucher, T.: Objects of intense feeling: the case of the Twitter API. Comput. Cult. 3 (2013)
14. Small, T.: What the hashtag. Inf. Commun. Soc. 14(6), 872–895 (2011)
15. James, G., Witten, D., Hastie, T., Tibshirani, R.: An Introduction to Statistical Learning.
Springer Texts in Statistics. Springer, New York (2013)
16. Cateni, S., Colla, V., Vannucci, M.: A method for resampling imbalanced datasets in binary
classification tasks for real-world problems. Neurocomputing 135, 32–41 (2014)
17. Lee, D.J., Han, J., Chambourova, D., Kumar, R.: Identifying Fashion accounts in social
networks. In: Proceedings of ACM KDD 2017 Machine learning meets fashion Workshop,
Nova Scotia, Canada (2017)
Efficient and Real-Time Face Recognition
Based on Single Shot Multibox Detector
Youngshin Ahn1, Sumi Kim2, Fei Chen1, and Jaeho Choi1(&)

1
Department of ECE, Jeonbuk National University, CAIIT,
Jeonju, Republic of Korea
[email protected]
2
Seoyeong University, Gwangju, Republic of Korea
Abstract. In this paper we present an efficient and real-time human face

detection and recognition method based on human body region of interests
(ROIs) provided by the single shot multibox detector (SSD). The SSD is a state-
of-art general purpose object detector that can detect all kinds of items in the
image data and provides detection probabilities. On the other hand, the his-
togram of oriented gradients (HOG) is another superb detector that is specially
designed for human face detection. However, it takes much time to scan the
whole image data in order to get the face features. Hence, the issue to us is to
reduce the computation time spent for searching human faces and to cope with
scalability of the object sizes. Here, in our method, we place the SSD in front of
the HOG. The SSD is used to make the ROIs of the human bodies, not the faces
importantly, so that the image data containing the human body ROIs only are
processed by the HOG. In this way, the HOG can save much time to produce the
ROIs of human faces. Then, the feature vectors for the human face ROIs are
computed in order to train and also to recognize the people’s identities by using
a deep learner. The computer simulations are performed to verify the proposed
system using several well-known data bases. The performance evaluation is
done in terms of speedup and accuracy as the multiplicity and scalability of
people changes. The results show us that the proposed system performs effi-
ciently and robustly than that of the conventional system without SSD, and
advantageously it comes with better real-time feasibility.
Keywords: Real-time face recognition Single shot multibox detector (SSD)

Histogram of oriented gradient (HOG)
1 Introduction
Recently, researches concerning face detection are rising with much attention and
applications using face detection algorithms are making huge changes in modern life
styles [1]. Among many important applications of face detection, clearly, facial
recognition is the outstanding one. When human faces are present, the face recognizer
actually attempts to establish whose face it is. The basic process begins with a com-
puter application that captures human faces from digital images or from video frames.
Correlation or similarity metrics can be used to compare the target human face to
images in a database of stored records. By any means, it cannot be hundred percent
Efficient and Real-Time Face Recognition 1101
accurate, however, it can provide a strong chance that a person’s face matches someone
in the database [2].
On the other hand, the human face recognition may not be so much about face
recognition at all. It can be more about face detection. In other words, one needs to find
the faces in the image; face detection is the first step in automatic facial recognition.
Accurately detecting human faces in arbitrary natural scenes is a challenging problem.
Some researches claim that the human face detection is the most important process;
when faces can be located exactly in any scene, the recognition step afterwards is not so
complicated anymore [3].
Moreover, concerning the methods for face recognition, Davis [4] has distributed
the DLIB library [5] for implementing a face recognition algorithm. Also, to make face
recognition easier, Adam [6] has distributed the FACE RECOGNITION library for
Python.
Our research focus is to reduce the human face detection time in such a way that
one can efficiently make recognition of individual faces at real-time possible. In our
efforts, the well-known algorithms related to single shot multibox detector (SSD) and
histogram of oriented gradient (HOG) are adopted for efficient face detection.
The SSD is a state-of-art general purpose object detector that can detect all kinds of
items in the image data and provides detection probabilities. On the other hand, the
HOG is another superb detector that is specially designed for human face detection.
However, the HOG takes much time to scan the whole image data in order to get the
face features. Hence, the issue to us is to reduce the computation time spent for
searching human faces and to cope with scalability of the image data sizes. Here, in our
method, we place the SSD in front of the HOG. The SSD is used to make the ROIs of
the human bodies, not the faces, so that the image data containing the human body
ROIs only are processed by the HOG. In this way, the HOG can save much time to
produce the ROIs of human faces. Then, the feature vectors for the human face ROIs
are computed in order to train and also to recognize the people’s identities by using a
deep learner.
In the following, the fundamentals and their literatures are mentioned in Sect. 2 and
the proposed method is described in Sect. 3. The simulations and results are provided
in Sect. 4 and the conclusion is made in Sect. 5.
2 Related Work
Over the years, the face detection methods have been developed and improved. First
there is a model-based face detection method. For example, Froeba [7] uses the edge-
based method; Viola [8] uses weak classifier cascades. Viola uses simple Haar features
which check the difference of brightness of area. Adam uses the HOG method [9] and
the deep learning method [10, 11] is also used.
Moreover, following face detection, it is necessary to find out the pose of the face
and then rotate and adjust the image so that the eyes and mouth are centered as much as
possible. Kazemi [12] finds the poses of a face and makes the human face detection
more accurate.
1102 Y. Ahn et al.
In addition, concerning real-time processing of object detection and recognition,

recently, there are several deep learning methods such as YOLO [13] and SSD [14].
3 Proposed Method
In this paper, our attention is on the reduction of face detection time as a means to
achieve a real-time face recognition system. In this effort, we attempt to accelerate the
face detection performance of the HOG by accommodating the SSD method in front of
the HOG. Especially, in the case of images taken by a high resolution camera, the face
detection using the HOG itself takes too much time and the system latency becomes
apparent. Therefore, for fast and real-time face recognition, one can exploit the SSD,
which has a very efficient object detection capability, in order to assist the HOG for
human face detection. In the proposed method, the SSD is actually used to extract the
human body, not the human face importantly, and the human body window is passed to
the HOG as a region of interest (ROI) for easier and faster localization of the human
face. Later in the simulation experiments, the proposed method shows much better
performance when the one deals with high resolution images.
Now, let us look at the face recognition implemented in the library [6]. First, it
encodes a picture using the HOG algorithm to create a simplified version of the image.
Using this simplified image, it finds the part of the image that most looks like a face.
Here, the HOG is a vector that divides an area into cells of a certain size. The HOG is a
histogram of the direction of edge pixels for each cell, and it connects the histogram
values to find a face-like region. Second, it figures out the pose of the face by finding
the main landmarks in the face [12]. Once those landmarks are found, one uses them to
warp the image so that the eyes and mouth are centered. Third, it passes the adjusted
and centered face image through a neural network, that knows how to measure features
of the face [15] and saves those 128 measurements. OPENFACE [16] are used to make
those measurements. Finally, it uses all the faces measured in advance and finds the
person who has the measurements closest to the measurements of the target face with
some uncertainty.
In the above procedures of face recognition, the first step computing the HOG is
computation intensive and it becomes a computational bottleneck as the size, resolu-
tion, and frame speed of the digital image data increases. In order to circumvent this
bottleneck, Adam has tried to reduce the data size in the second step and implemented
modified HOG using Python [6]. The first modification is to reduce the resolution of
the image data by half along horizontal and vertical directions making the image data
size down to the quarter of the original one. The second place of modification is on the
reduction of video frames utilized. Instead of using all of the video frames, here, by
down-sampling by two, it only uses odd numbered frames. In these two modifications,
the size of the image data to compute the HOG can be reduced by eight times in
comparison to the conventional one [9]. Hence, one can expect much faster detection of
human face ROIs in the HOG at the cost of some recognition degradation due to loss of
image data resolution.
As mentioned above, the reduction of image data resolution can speed up the face
recognition, but at the same time it comes with a loss of recognition accuracy. In this
paper, we propose an enhanced version of face recognition method that can achieve the
recognition accuracy as good as the one in the conventional HOG [9] while main-
taining the detection speed-up relatively comparable to the one by Adam [6]. The
procedures of the proposed method are illustrated in Fig. 1. The key point of the
proposed method lies in the newly adopted SSD scheme that is used to locate the
human object, first. The SSD is very efficient object detection method and it is pro-
gramed to produce a region of interest (ROI) of an object, and we are only concerned
with human body ROIs. The ROIs of human bodies preserve the original image data
resolution but their data size is generally smaller in comparison to the image data
frame. The ROIs of human bodies are passed to the HOG and the human faces are
detected based the ROIs of human bodies only. In this way, HOG’s computational load
can be reduced by a large amount and can circumvent the bottleneck apparent in [6]. In
other words, those two initial steps in the proposed method involving human body
detection using SSD and human face detection using HOD allow for a faster face
recognition without sacrificing image data resolution as well as recognition accuracy.
Video Frame (Input Image)

Object Detection Using SSD
Face Detection Using HOG
Transform and Crop
Deep Neural Network
Representation
Face Recognition
Fig. 1. Overview of the proposed face detection and recognition procedures.
4 Simulations Results
In this section, the simulation experiments are described. The Python with the Jupyter
Notebook is used and the codes are run on the Window-10 PC with Intel Core-8265
1.6 GHz CPU and 8 GB RAM. The input image data is obtained by using Webcam
with a frame rate of 30 fps. The image frame size is 640 480 for the low resolution
and 1280 720 for the high resolution.
In the first set of simulations, the performance of the proposed method is evaluated
in terms of the face recognition time in second with respect to the size of the human
body ROI defined by n. The parameter n determines the ratio between the image data
size and the human body ROI. For example, when n = 4, the size of the human body
ROI box is four time small than that of the original image and it is well illustrated in
Fig. 2. The simulation results are summarized in Table 1 and it shows that the time
spent to recognize a human face decreases as the size of the ROI decreases as expected.
At the low input image resolution of 480 640, the size of the ROI does not
1104 Y. Ahn et al.
Table 1. Face recognition time with respect to the size of the human body ROI.
n Low resolution High resolution
Size of ROI Time (sec) Size of ROI Time (sec)
1 480 640 0.6133 720 1280 1.0604
2 339 453 0.4923 509 905 0.7192
3 227 370 0.4438 415 739 0.6042
4 240 320 0.4373 360 640 0.5443
5 215 286 0.4276 321 572 0.5092
Fig. 2. The relationship between n and the size of ROI in the image size of w h.
significantly affect speed; however, at the high resolution of 720 1280, the size of the
ROI has a significant impact on the face recognition speed.
For the second set of experiments, the proposed face recognition method is com-
pared to the conventional one described in [6] and the results are summarized in
Table 2. As illustrated in the table the proposed method output-performs the conven-
tional method in both of input image data resolution cases. While it looks like the time
gain is relatively small in the low resolution, however, the time gain becomes appar-
ently bigger in the high resolution cases. It is important to note that the table is the
worst case performance of the proposed method that as the size of the human body ROI
decreases the time gain gets much bigger.
By evaluating the two sets of preliminary experiments, one can recognize that the
proposed method is a very efficient and it becomes a time saving face recognition
scheme.
Table 2. Comparison of face recognition time between the proposed method and the
conventional one in [15].
Resolution Size of ROI Existing method Proposed method Improvement
High 720 1290 1.0604 0.6020 1.76 times
Low 480 640 0.6133 0.5307 1.15 times
Fig. 3. Face recognition result of a person with detection accuracy in green and the person’s
identity name tag in red.
Finally, Fig. 3 visually shows a detection and face recognition result. There are
colored boxes, object types, and percentage numbers. The green-boxed human body
ROI with person: 1.00 indicates the human body ROI produced by SSD with 100%
accuracy. The potted plant is also detected with 85% accuracy. The red-boxed human
face ROI is produced by HOG and the object type name obama is the name of the
person identified by the deep learning recognition system.
5 Conclusion
In this paper, the efficient and real-time human face recognition method has been
presented. It is shown that by using SSD in conjunction to HOG the face recognition
can be efficiently performed at the higher speed while the recognition accuracy is well
maintained. The performance of the proposed system becomes faster and more efficient
as the input image data resolution gets higher and the size of the human body ROI gets
smaller. The simulation results have shown us that the speed up can be as small as 44%
in comparison to the conventional face recognition system using a HOG. The higher
speed-up can be achieved at the higher image data resolution. In sum, the proposed
method can be efficiently implanted for real-time application of human face recognition
in modern life where image data size gets bigger and bigger. In the proposed method,
while the face detection can be done efficiently and saves the time but the face
recognition needs an improvement. It can be by exploiting the tracking algorithm that
once the face is recognized, the tracking algorithm can be applied and save the pro-
cessing time. We will look into its possibility in the future study.
Acknowledgement. This work has been supported partly by BK21 and Jeonbuk National
University 2020.
1106 Y. Ahn et al.
References
1. WIKIPEDIA Homepage. https://en.wikipedia.org/wiki/Face_detection. Accessed 2 Mar
2020
2. FACEFIRST Homepage. https://www.facefirst.com/blog/face-detection-vs-face-recognition.
Accessed 2 Mar 2020
3. FACE DETCTION Homepage. https://facedetection.com. Accessed 2 Mar 2020
4. Davis, E.K.: Dlib-ml: a machine learning toolkit. J. Mach. Learn. Res. 10, 1755–1758 (2009)
5. DLIB Homepage. http://dlib.net/. Accessed 2 Mar 2020
6. ADAM’S LIBRARY Homepage. https://github.com/ageitgey/face_recognition. Accessed 2
Mar 2020
7. Froeba, B., Kueblbeck, C.: Real-time face detection using edge-orientation matching. In:
Proceedings of the Audio- and Video-based Biometric Person Authentication, pp. 78–83
(2001)
8. Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple features. In:
Proceedings of the CVPR, pp. 511–518 (2001)
9. Dalal, N., Triggs, B.: Histograms of oriented gradients for human detection. In: CVPR 2005,
pp. 886–893 (2005)
10. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image
recognition. In: International Conference on Learning Representations (2015)
11. ADAM GEITGEY Homepage. https://medium.com/@ageitgey/machine-learning-is-fun-
part-4-modern-face-recognition-with-deep-learning-c3cffc121d78. Accessed 2 Mar 2020
12. Kazemi, V., Sullivan, J.: One millisecond face alignment with an ensemble of regression
trees. In: CVPR 2014, pp. 1867–1874 (2014)
13. Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: unified, real-time
object detection. arXiv preprint arXiv:1506.02640 (2015)
14. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C., Berg, A.: SSD: single shot
multibox detector. LNCS, vol. 9905, pp. 21–37 (2016)
15. Schroff, F., Kalenichenko, D., Philbin, J.: FaceNet: a unified embedding for face recognition
and clustering. arXiv preprint arXiv:1503.03832 (2015)
16. OPENFACE Homepage. https://cmusatyalab.github.io/openface. Accessed 2 Mar 2020
Machine Learning-Based Robust Feedback
Observer for Fault Diagnosis in Bearing
Farzin Piltan and Jong-Myon Kim(&)
School of Electrical Engineering, University of Ulsan,

Ulsan 680-749, South Korea
Abstract. Rolling element bearing (REB) represent a class of nonlinear and

multiple-degrees-of-freedom rotating machines that have pronounced coupling
effects and can be used in various industries. The challenge of understanding
complexity in a bearing’s dynamic behavior, coupling effects, and sources of
uncertainty presents substantial challenges regarding fault diagnosis (FD) in a
REB. Thus, a proposed FD algorithm, based on an TSK fuzzy multi structure
feedback observer, is represented. Due to the effect of the system’s complexities
and uncertainties for FD, a feedback observer (FO) is proposed. To address the
FO drawbacks for FD in the REB such as robustness, the multi structure
technique is represented. In addition, the TSK fuzzy algorithm is applied to the
multi structure FO (MSFO) to increase the performance of signal estimation and
reliability. In addition, the energy residual signals are generated and the machine
learning technique known as a support vector machine (SVM) adaptively
derives the threshold values that are used for classification the faults. The
effectiveness of the proposed technique is validated using a Case Western
Reverse University (CWRU) vibration dataset.
Keywords: Roller bearing element Fault diagnosis Feedback observer

Variable structure technique TSK fuzzy algorithm Support vector machine
1 Introduction
The applications of the Rolling Element Bearings (REBs) have been increasing in
recent years, including steel mills, paper mills, and various power generators. The REB
is a complex mechanical component that has various nonlinear parameters. Therefore,
the complexities of the tasks and the nonlinear parameters in REBs have led to
numerous complications associated with fault detection and classification. Numerous
types of faults have been represented in REBs, which can be divided into four main
groups: inner fault (IF), outer fault (OF), ball fault (BF), and cage fault [1].
Different procedures have been recommended for FD in REBs [1–3]. Hybrid-based
FD is considered as a reliable, stable, and robust technique. Several techniques can be
represented as hybrid techniques for FD that are reported in [4–6]. In this research, a
hybrid-based observer is represented for fault diagnosis in the REB. The main concept
of observation-based FD is REB modeling. System modeling is categorized as the
following groups: a) mathematical and b) system identification. Apart from the accu-
racy and reliability of the mathematical-based system modeling, complexity is the main
drawback of this technique [5]. To address these issues, system identification algo-
rithms (e.g., Auto-Regressive with eXogenous inputs (ARX), ARX-Laguerre, and
intelligent-based ARX-Laguerre) have been reported in [6, 7].
The observer is used to estimate the bearing’s vibration signals in normal and
abnormal conditions. Various observers have been proposed for signal estimation such
as proportional-integral observer (PIO), proportional multiple integral (PMI) observer,
sliding mode observer (SMO), and feedback observer (FO) [8]. However, the PIO and
PMI are easy to implement but, robustness and accuracy are the main drawbacks [6, 8].
The SMO is robust, stable, and reliable but this technique is suffering from high-
frequency oscillation and accuracy [3]. To address the issues of high-frequency
oscillation, FO has been described in [8].
In this research, the hybrid technique is used to FD in the REBs. Firstly, the fuzzy
ARX-Laguerre is proposed to modeling the REB based on the vibration signals. Next,
the FO is proposed for vibration signal estimation. Moreover, the multi-structure
technique is used to improve the robustness of FO. In addition, to increase the accuracy
in the multi-structure FO (SFO), TSK fuzzy algorithm is represented. Beyond the
above, the fuzzy ARX-Laguerre fuzzy-structure feedback observer (PM) is used to
accurate signal estimation. After generating the energy residual signal, the SVM is
represented for FD. The block diagram of the proposed technique is represented in
Fig. 1. This paper is organized as follows. The second section outlines the fuzzy ARX-
Laguerre REB modeling based on the vibration signals. The proposed technique for FD
is represented in the third section. In the fourth section, the proposed method for FD in
the REB is analyzed. Conclusions are provided in the last section.
2 Fuzzy ARX-Laguerre Technique for REB Modeling
In complicated systems such as REBs, system modeling is a significant challenge,

especially in uncertain conditions. Hence, to modeling the REB, the ARX-Laguerre
technique is represented. Regarding Fig. 1, the fuzzy ARX-Laguerre technique has
three main parts: a) ARX system modeling, b) increase the robustness by ARX-
Laguerre technique, and c) improve the accuracy of nonlinear system modeling by
fuzzy ARX-Laguerre method. The ARX technique for REB modeling based on
vibration signals can be represented by Eq. 1 [6, 7].
d dUi
X
YM
i X
YMA ðkÞ ¼ CYM ðiÞYMA ðk iÞ þ CUi ðiÞUi ðk iÞ; ð1Þ
i¼1 i¼1

where YML ðk Þ; Ui ðkÞ; ðCx ðiÞ; CYM ðiÞÞ, and dUi ; dYMi are the REB’s model based on
ARX technique, reference signal for REB’s modeling, REB’s model parameters, and
the order of the REB modeling, respectively. The REB is a nonlinear and complex
component and the ARX technique is a linear algorithm. Therefore, to improve the
power of noise rejection (robustness), the ARX-Laguerre procedure is represented for
REB modeling.
Machine Learning-Based Robust Feedback Observer 1109
Fig. 1. Proposed method for REB fault detection and classification.
The definition of ARX-Laguerre technique is [7]:
iX iUi 1
Y 1 X
YMAL ðk Þ ¼ n;YM ðk; cY ÞYMAL ðk iÞ þ
CY ðiÞO n;Ui k; cU Ui ðk iÞ;
CUi ðiÞO
AL i
n¼0 n¼0
ð2Þ

n;YM ðk; cY Þ and O
Here, YMAL ðk Þ; O n;Ui k; cU are the REB’s model based on ARX-
AL i
Laguerre technique, the REB’s function of orthonormal, and the REB’s reference
signal orthonormal function, respectively. To increase the REB modeling accuracy, the
fuzzy technique can be introduced. The two inputs fuzzy technique is used to modeling
the nonlinearity of REB. The fuzzy rule-base is represented as the following equation:
if \F:I[ is \F:LV[ and \S:I[ is \S:LV[ then \out[ is \out:LV[; ð3Þ
Here, \F:I[; \F:LV[; \S:I[; \S:LV[; \out[ and \out:LV[ are first input of
fuzzy, the first linguistic variable of the fuzzy, the second input of fuzzy, the second
linguistic variable of fuzzy, the fuzzy output, and the linguistic variable for fuzzy
output, respectively. Therefore, the state-space formulation for fuzzy ARX-Laguerre
REB’s modeling is represented as the following equation.

XFAL ðk þ 1Þ ¼ rX XFAL ðkÞ þ rY YFAL ðkÞ þ rf Yf ðkÞ þ rU Ui ðkÞ
: ð4Þ
YFAL ðk þ 1Þ ¼ ðbÞT XFAL ðk Þ

Here, XFAL ðkÞ; YFAL ; Yf and rX ; rY ; rU ; rf ; b are the state of the model for REB based
on fuzzy ARX-Laguerre technique, the output of the model for REB based on fuzzy
ARX-Laguerre technique, the nonlinear signal model based on fuzzy algorithm, and
fine-tuning coefficients, respectively.
3 Fault Detection and Classification of REB Based

on Machine Learning-Based Robust Feedback Observer
The fuzzy ARX-Laguerre technique for vibration signal modeling the normal condition
of REB is represented in the previous part. Now, to improve the estimation accuracy
the fuzzy ARX-Laguerre fuzzy-structure feedback observer (PM) is proposed. This
technique is used to improve the accuracy of REB’s vibration signal estimation for fault
detection and classification. Therefore, firstly the fuzzy ARX-Laguerre feedback
observer is represented as the following definition.
8
>
> XFO ðk þ 1Þ ¼ rX XFO ðkÞ þ rY YFO ðkÞ þ rf Y^f ðk Þ þ rU Ui ðkÞ þ u
^ FO ðkÞ
>
>
>
> þ r e
< 1 FO
YFO ðk þ 1Þ ¼ ðbÞT XFO ðk Þ ; ð5Þ
>
> eFO ¼ XFAL ðk Þ XFO ðk Þ
>
>
>
> erFO ¼ Yo ðkÞ YFO ðkÞ
:
ef FO ¼ Yo ðkÞ ðYFO ðkÞ þ Y^f ðkÞÞ
To improve the accuracy uncertainties estimation of the ARX-Laguerre feedback

observer, the uncertainties and unknown conditions (e.g., motor speed variation) can be
estimated based on the following equation.
^ FO ðk þ 1Þ ¼ u
u ^ FO ðkÞ þ r2 erFO þ r3 ef FO : ð6Þ
Here, XFO ðk Þ; YFO ðk Þ; Y^f ; u

^ FO ðkÞ; Yo ðkÞ and ðr1 ; r2 ; r3 Þ are the estimation state of
the model for REB based on fuzzy ARX-Laguerre feedback observer (FALFO), the
estimation of the REB signal based on FALFO, the nonlinear part of signal estimation
based on fuzzy technique, uncertainties (motor speed variation) estimation based on the
FALFO to minimize the error in the normal condition, original signal of the REB, and
the coefficient of the FALFO, respectively. Regarding Fig. 1, the FALFO is suffering
from robustness especially in unknown conditions. To solve this obstacle, the fuzzy
ARX-Laguerre structure feedback observer (FALSFO) is represented. The structure
(variable structure) algorithm can be used to increase the robustness.
8
>
> XSFO ðk þ 1Þ ¼ rX XSFO ðkÞ þ rY YSFO ðkÞ þ rf Y^f ðkÞ þ rU Ui ðk Þ þ u
^ SFO ðk Þ
>
>
>
> þ r e
< 1 SFO
YSFO ðk þ 1Þ ¼ ðbÞT XSFO ðkÞ ; ð7Þ
>
> eSFO ¼ XFAL ðk Þ XSFO ðkÞ
>
>
>
> erSFO ¼ Yo ðkÞ YSFO ðk Þ
:
ef SFO ¼ Yo ðkÞ ðYSFO ðk Þ þ Y^f ðk ÞÞ
The uncertainties can be estimated by FALSFO based on the following equation.
^ SFO ðk þ 1Þ ¼ u
u ^ SFO ðkÞ þ r2 erSFO þ r3 :sgn ef SFO : ð8Þ
^ SFO ðkÞ and r3 are the estimation state of the model for
Here, XSFO ðk Þ; YSFO ðkÞ; u
REB based on FALSFO, the estimation of the REB signal based on FALSFO,
uncertainties estimation based on FALSFO to minimize the error and increase the
robustness in the normal condition, and the coefficient, respectively. Moreover, to
improve the signal estimation accuracy, the T-S fuzzy algorithm may be defined
regarding the following equation.
^ f ð k þ 1Þ ¼ u
if eFSFO ðkÞ is Th then u ^ f ðkÞ þ rF eFSFO : ð9Þ
Hence, the fuzzy ARX-Laguerre fuzzy-structure feedback observer (PM) can be

represented by the following equation.
8
>
> XFSFO ðk þ 1Þ ¼ rX XFSFO ðkÞ þ rY YFSFO ðkÞ þ rf Y^f ðk Þ þ rU Ui ðkÞ þ u
^ FSFO ðkÞ
>
>
>
> þ r e
< 1 FSFO
YFSFO ðk þ 1Þ ¼ ðbÞT XFSFO ðk Þ
;
>
> eFSFO ¼ XFAL ðkÞ XFSFO ðk Þ
>
>
>
> erFSFO ¼ Yo ðk Þ YFSFO ðk Þ
:
ef FSFO ¼ Yo ðkÞ ðYFSFO ðkÞ þ Y^f ðkÞÞ
ð10Þ
In addition to increase the uncertainties estimation based on PM, the Eq. (11) can
be represented.
^ FSFO ðk þ 1Þ ¼ u
u ^ FSFO ðkÞ þ r2 erFSFO þ r3 :sgn ef FSFO þ r4 u
^ f ðkÞ: ð11Þ

^ f ðkÞ; rf ; r4 ; XFSFO ðkÞ; YFSFO ðkÞ and u
Here, eFSFO ðkÞ; Th; u ^ FSFO ðkÞ are the error
estimation for REB based on PM, the threshold value to tuning the estimation accuracy,
the T-S fuzzy estimation behavior, coefficients, estimation state of the model for REB
based on PM, the estimation of the REB signal based on PM, uncertainties estimation
based on PM to minimize the error and increase the robustness in the normal condition,
respectively. Regarding Fig. 1, the proposed PM technique can be minimized the error
of the signal estimation for REB. So, the residual signal ðr ðk ÞÞ that can be defined
based on the following equation is minimized in the normal condition.
r ðkÞ ¼ Yo ðkÞ YFSFO ðkÞ: ð12Þ
The feature of energy ðE Þ is extracted from the residual signal to improve the
performance of fault detection and classification, and can be defined by the following
equation.
X
n
¼
E r ðk Þ2 : ð13Þ
k¼1
Based on the difference between the amplitude of the energy residual signals, the
decision-making unit can detect and classify the faults easily. To detect and classifi-
cation the faults, a machine learning algorithm based on the SVM technique is pre-
sented in this research. This procedure has a powerful mathematical background due to
the availability of the diverse kinds of kernels that can be used for training. Therefore,
the soft margin SVM is defined by the following equation [9–11]:

wi ;T qðqi Þ þ c wi vi : ð14Þ
where ðqi ; wi Þ; ð;; cÞ; qðqi Þ; and vi are the inputs, outputs, features, and maximum
distance, respectively.
4 Dataset, Results and Discussion
In this research, the Case Western Reserve University (CWRU) bearing dataset [11] is
used to analyze the effectiveness of the proposed algorithm. The vibration sensor
collects data in various conditions (e.g., normal, inner fault, outer fault, and ball fault).
The data classification of the CWRU bearing dataset is represented in [6]. The energy
of residual signals for normal and abnormal conditions is illustrated in Fig. 2.
Fig. 2. Energy residual signal for fault classification based on the proposed algorithm.
Regarding the Fig. 2, the energy of residual signals used for differentiating normal,
ball fault, inner fault, and outer fault signals are well differentiable for fault detection
and classification based on the proposed algorithm. Figure 3 demonstrates that all of
the methods (i.e., FALFO, FALSFO, and proposed method) show high fault detection
accuracies (i.e., differentiating between normal and abnormal states). Regarding fault
identification (i.e., differentiating between types of faults), it can be seen that the
proposed algorithm using the PM + SVM resulted in the smallest numbers of mis-
classified samples for all of the signal classes in comparison with its counterparts. In
addition, we compare our proposed method with state-of-the-art FALFO and FALSFO
for performance analysis. As shown in Tables 1, the proposed method for REB fault
diagnosis outperforms the state-of-the-art FALSFO method and FALFO technique,
yielding average performance improvements of 3%, and 6.2% for three faults,
respectively.
Fig. 3. Confusion matrices for proposed method, FALSFO, and FALFO based on the SVM
method.
Table 1. The average of fault detection and classification accuracy when the torque load is 0–3
hp.
Algorithms (Proposed method) FALSFO FALFO
Crack diameters (in) 0.007 0.014 0.021 0.007 0.014 0.021 0.007 0.014 0.021
Normal 100% 100% 100% 100% 100% 100% 96% 96% 96%
BF 95% 96.5% 96.2% 90% 92% 93.6% 86% 88% 90%
OF 94% 95% 97% 92% 93% 93.3% 85% 89% 91%
IF 93% 92% 92.8% 88% 88% 86.8% 83% 82% 84%
ACA 95.5% 95.9% 96.5% 92.5% 93.3% 93.5% 87.5% 88.8% 90.3%
5 Conclusions
In this paper, a nonlinear hybrid approach based on the machine learning-based

advanced observer is presented for detection and classification REB faults. The fuzzy
ARX-Laguerre technique is proposed for REB modeling in normal condition based on
the vibration signal. Next, the nonlinear feedback observer is used to estimate the REB
vibration signal. The multi-structure technique is used to reduce the issue of the motor
speed variant. In addition, the TSK fuzzy technique is proposed to reduce the error of
estimation. Moreover, the SVM is represented to detect and classification the REB
faults. The effectiveness of the proposed observer is validated using a publicly available
vibration dataset of CWRU. Beyond the above, the proposed algorithm improves the
fault classification accuracy from 87.5% in the FALFO to 96.5%. In our future work,
we would like to verify the applicability of this technique to other prognostics and
health management problems, such as gearbox fault identification, and pipeline leak
detection and localization.
Acknowledgements. This research was financially supported by the Ministry of Trade, Industry
& Energy (MOTIE) of the Republic of Korea and Korea Institute for Advancement of Tech-
nology (KIAT) through the Encouragement Program for The Industries of Economic Coopera-
tion Region (P0006123).
References
1. Gao, Z., Cecati, C., Ding, S.X.: A survey of fault diagnosis and fault-tolerant techniques—
part I: fault diagnosis with model-based and signal-based approaches. IEEE Trans. Ind.
Electron. 62, 3757–3767 (2015)
2. Cecati, C.: A survey of fault diagnosis and fault-tolerant techniques—part II: fault diagnosis
with knowledge-based and hybrid/active approaches. IEEE Trans. Ind. Electron. 62, 3768–
3774 (2015)
3. Piltan, F., Kim, J.M.: Bearing fault diagnosis by a robust higher-order super-twisting sliding
mode observer. Sensors 18, 1128 (2018). https://doi.org/10.3390/s18041128
4. Guo, X., Chen, L., Shen, C.: Hierarchical adaptive deep convolution neural network and its
application to bearing fault diagnosis. Measurement 93, 490–502 (2016)
5. Lou, X., Loparo, K.A.: Bearing fault diagnosis based on wavelet transform and fuzzy
inference. Mech. Syst. Signal Process. 18, 1077–1095 (2004)
6. Piltan, F., Prosvirin, A.E., Jeong, I., Im, K., Kim, J.M.: Rolling-element bearing fault
diagnosis using advanced machine learning-based observer. Appl. Sci. 9(24), 5404 (2019)
7. Bouzrara, K., Garna, T., Ragot, J., Messaoud, H.: Decomposition of an ARX model on
Laguerre orthonormal bases. ISA Trans. 51, 848–860 (2012)
8. Piltan, F., Kim, J.M.: Bearing fault diagnosis using an extended variable structure feedback
linearization observer. Sensors 18(12), 4359 (2018)
9. Prosvirin, A., Kim, J., Kim, J.M.: Efficient rub-impact fault diagnosis scheme based on
hybrid feature extraction and SVM. In: Bhatia, S.K., Tiwari, S., Mishra, K.K., Trivedi, M.C.
(eds.) Advances in Computer Communication and Computational Sciences, pp. 405–415.
10. Elangovan, K., Krishnasamy, T.Y., Mohan, R., Iwase, M., Takuma, N., Wood, K.: Fault
diagnosis of a reconfigurable crawling-rolling robot based on support vector machines. Appl.
Sci. 7, 1025 (2017)
11. Adams, M.L.: Analysis of rolling element bearing faults in rotating machinery: experiments,
modeling, fault detection and diagnosis. Ph.D. Thesis, Case Western Reserve University,
Cleveland, OH, USA (2001)
Malfunction Detection on Production Line
Using Machine Learning: Case Study
in Wood Industry
Kağan Özgün(&), Sami Can Aklan, Ahmet Tezcan Tekin,

and Ferhan Çebi
Istanbul Technical University Management Engineering Department,

Istanbul, Turkey
[email protected]
Abstract. The concept of the Internet of Things, especially in the last decade,
has created the opportunity to place sensors in every event and location that can
be tracked to collect data via these sensors. Collecting data from sensors is not a
stand-alone solution. After the problem of data collection and storage of large
amounts of data collected has been overcome, it has been made easier by per-
forming analytical operations with this data. The machine learning algorithms
and methods used in the robotics sector are used in different fields to make
productions, process and machine groupings by making various estimations for
the industry with more complex algorithms or clustering operations with the
collected data. Within the scope of this project, it is aimed to monitor the con-
dition of the machines on the production line with the data collected from the
machines used in the production process and to make fault detection on the
machines by using the machine learning methods for the maintenance and repairs
of the machines before they break down, produce faulty products and stop the
production line. In this study, anomaly detection methods which are proposed in
the literature were performed to data which was collected by sensors. Also, the
artificial neural network was applied to the dataset. The results show us these
technics can be used in the manufacturing sector for fault detection.
Keywords: Malfunction detection Big data Internet of Things Machine

Learning Production lines
1 Introduction
With the rapidly developing technology in recent years, the size and production cost of
the computer components has decreased considerably and these components have
become more accessible than ever before. Thanks to these developments, we have
started to use more powerful products that have diminished in size in many points in
our lives with lower costs. The concept of the Internet of Things, especially in the last
decade, enables the connection of electronic devices with each other, facilitate data
transfer between them [1]. Therefore, it is possible to place sensors in every event and
location that can be tracked and collect data via these sensors. However, collecting data
is not the solution itself. The first issue is that fast and continuous data flow leads data
Malfunction Detection on Production Line Using Machine Learning 1117
to accumulate in large volumes in a shorter time, therefore, it gives rise to the problem
of data storage. The second issue is what to do with the collected data in such sizes.
Because the data is raw and it is meaningless without processing. At this point, Big
Data systems have entered our lives. Large data systems have facilitated the storage of
data from a wide range of devices [2]. On the other hand, developments in computer
science make it easier to process data and make analytical operations with it. Machine
Learning algorithms and methods used in the robotics are implemented in different
fields to make various estimations with more complex algorithms, clustering operations
with the data collected about the products, processes and machines [3].
Within the scope of this project, it is aimed to monitor the condition of the
machines on the production line with the data collected from the machines used in the
production process and to make fault detection on the machines by using the machine
learning methods for the maintenance and repairs of the machines before they break
down, produce faulty products and stop the production line.
In the modelling section, properties of the project data, train and test splits tech-
nique for algorithms described and related works examined. Then proposed method-
ology section include details of the algorithms that used for malfunction detection.
Finally results of the algorithms, evaluation of the results and future works declared in
the conclusion section.
2 Modelling
Anomaly detection is referred to as detecting items, events or observations that do not

conform to an expected pattern and exhibit significantly different behavior than the
majority of the rest [4]. Most of the time, anomalies or outliers are associated with
something that goes wrong and they are pointers of some future problematic situation.
This would mean a cyber-attack for cybersecurity companies, fraudulent transaction for
banks or malfunctioning equipment for manufacturing companies. Either way, ana-
lyzing the data and defining what the normal and anomaly is play a critical role in
taking preventive actions and avoiding undesired outcomes in the future. As IoT
applications are getting more spread and connectivity of devices increases, collecting
data for related fields is not as challenging as before. However, it is a quite difficult task
to analyze them with traditional algebraic and statistical methods due to the ever-
increasing volume of the data [5]. An approach of using Machine Learning (ML) al-
gorithms is quite a handful for analyzing data and promising faster and more reliable
results than traditional methods to detect anomalies.
Several studies in the literature are focusing on anomaly detection on the pro-
duction line and using different methods. Some of these studies use different types of
data and others use different statistical methods and machine learning algorithms. In the
study conducted by Gian Antonio Susto and colleagues, it was aimed to detect anomaly
in semiconductor production and compare the results of the studies. Three different
methods were used in this study. These are Principal Component Analysis (PCA),
Angle Based Outlier Detection (ABOD) and Local Outlier Factor (LOF) methods.
They used simple statistical methods and clustering methods instead of sophisticated
machine learning algorithms. The PCA method is based on marking the collected data
1118 K. Özgün et al.
as usual in regular operation and comparing the incoming new data with the typically
marked data. The ABOD method puts the collected data in a graph drawn as value and
time and marks the usual data with rows drawn at certain angles over the data con-
sidered normal. Data outside the plotted area are classified as abnormal. The LOT
method is based on the density of the points drawn on the graph. It compares each point
with neighboring points, marking points that are long-distance than a specified average
value as abnormal. As a result of the study, LOF method was found to be the most
successful on the data set used [6].
In another case study by Alexander Grab et al., It was aimed to determine the
anomaly on the machine by using the temperature and power consumption values
obtained from the industrial reflow oven and Generic Anomaly Detection for Pro-
duction Lines (GDAPL) algorithm. To use the GDAPL algorithm, which is named by
them, it is necessary to cluster on the data first. After clustering, segmentation was
performed because they needed values that they could generally base on. Then, they
developed the time series based GADPL algorithm. This algorithm is based on the
nearest neighbor method. Since the study is performed with only two different input
data, it is simple and more suitable for machine-oriented anomaly detection [7].
Another study by Mitsubishi Electric Research Laboratories is based on unsuper-
vised machine learning methods. LOF, One-Class SVM and Auto-Encoder Neural
Network (Auto-Encoder NN) methods were used in this study. Because all the algo-
rithms selected for the study are based on unsupervised machine learning; the input
data were the same. Auto-Encoder NN algorithm showed the most successful results
when the results collected over the established models were compared according to
error rates [8].
MDF manufacturer company who wants to take advantage of the recent develop-
ments in the information technologies is selected for the project. MDF manufacturing
process has six main steps. These are chipping, refining, resonating, drying, pressing
and cooling. Sensors are placed throughout these six steps, on 21 different machines.
Measurements are taken about the operations of the machines and conditions of the
products with 5 min of intervals, 24 h a day, during a month. A raw data, consisting of
9959 observations and 356 variables, are collected. The observations are not labelled as
normal or anomaly. Some of these variables are rpm level of a machine, the humidity
level of a product or temperature of water discharge, etc. After the data collected, it is
cleaned for the sake of the analysis. Some data entry mistakes are adjusted, observa-
tions with missing values are removed. After that, the data is divided into three parts:
Validation set, the training set and test set. The validation set is created under the
assumption that the machines operate correctly under optimum conditions and the first
24 h of observations are accepted as normal or non-anomaly. Training and test set are
created as 75% and 25% of data, respectively. For the analysis, four different methods
are used to detect if any observation is an anomaly or not. Mahalanobis Distance,
Isolation Forrest, One-Class SVM and ANN are applied on the production line as a
whole and then, on the 21 machines separately.
Before the preparation of the models, data cleaning methods are applied to the
dataset. Type of variables was checked whether they are suitable for modelling.
Observations that states the condition of variable as On/Off or Open/Closed are con-
verted into 0/1 to run mathematical operations or excluded completely. The number of
missing values in the variables are examined. If the amount of missing values in 1
variable is over the threshold value, that variable is excluded from the dataset. Other
missing values are filled according to the mean of a variable.
On the other hand, all 356 variables have quite different units and ranges. There-
fore, min-max normalization is applied to dataset and observations were scaled
between 0 and 1. Finally, the variables with 0 variances are excluded from the dataset
and 356 variables are reduced to 329 after the data cleaning and preparation process.
Data examined in this project is not labelled as normal and anomaly. Therefore, we
select a method that commonly used in unsupervised algorithms, depending on that
method the 290 rows of first-day data considered as normal. In the Mahalanobis
Distance Algorithm, we used a different method to evaluate the results. Since this
method calculates the variance of rows, it compares the result value with a threshold
and makes anomaly detection. Therefore, in this method, we used the threshold value
of the first day with the correct admission data when comparing the accuracy of the
results. For this reason, validation was performed with all test results instead of 290
rows. In the first stage of the evaluation process, we compared the results of the study
with the first-day data and the test data. In making this comparison, we found it
appropriate to create a confusion matrix. The confusion matrix is a matrix created by
comparing the data that the test results classify as normal and abnormal with the first-
day data which is the validation value.
Confusion matrix divides the data to four different clusters. True positives (TP) are
data that predicted as positive and positive, false positives (FP) are predicted positive
and negative, true negatives (TN) are predicted as negative and true negative and false
negatives are predicted negative and actual positive results [9]. The confusion matrix
alone is not very meaningful, but a few different metrics have been formed with the
results of this matrix (Table 1).
Table 1. Confusion matrix.

Predicted
Actual True positives False negatives
(TP) (FN)
False positives True negatives
(FP) (TN)
F1 Score ¼ ð2 Precision RecallÞ=ðPrecision þ RecallÞ ð1Þ
F1 Score ¼ f2 ½TP=ðTP þ FPÞ ½TP=ðTP þ FNÞg=

ð2Þ
f½TP=ðTP þ FPÞ þ ½TP=ðTP þ FNÞg
F1 Score is a metric calculated by using precision and recall values. F1 score shows
that positive predicted results are actual positives and model does not miss out the
positives and predicts them negatively. A drawback of that metric is it gives the same
weight to precision values and recall values. In summary, the algorithm provides a
successful result while the F1 Score converges to 1.0 [10].
3.1 Mahalanobis Distance Approach

Mahalanobis Distance is the distance that measured between a point and a distribution.
It was first introduced by P. C. Mahalanobis in 1936 and used for multivariate anomaly
detections and classification on unbalanced datasets [11]. It is a method that takes the
means of the vectors and their inverse covariance matrixes into account. It measures
distance relative to the centroid, a central point which is considered as an overall mean
for multivariate data. The centroid is a point in multivariate space where all means from
all variables intersect. The larger the MD, the further away from the centroid the data
point is. And further away the point, the more possible the point is an anomaly.
Formulation of MD as follows:

D2 ¼ ðX lÞ C1 ðX lÞ ð3Þ
where; D2 is the square of the Mahalanobis Distance, X is the observed vector, a row in
the dataset, m is the vector mean, column mean of each column, C−1 is the inverse
covariance matrix of the independent variables.
The method firstly used for all production line. Mahalanobis distances of obser-
vations are calculated for each observation. Distribution of MD values derived from the
training set is analyzed to set a threshold value. The first point the notice that all the
MD values are larger than 0, as its nature, and the distribution is positively skewed.
Hence, it is decided to set the threshold value at the .90 quantile of the distribution.
After the threshold value is set from the training set, it is used to identify whether an
observation in the test set is an anomaly according to its MD. An observation with MD
higher than the threshold value is labelled as an anomaly. This approach is used
because the observations in the dataset were not labelled as normal or anomaly in the
first place. Then, the distribution of the MD values is plotted with MD being on the y-
axis and time-variable being on the x-axis. Here it is analyzed that where and when the
significant deviations are started, how long it takes to turn back to its average value.
After that, the same approach is followed for the validation set. Under the assumption
of the first 24 h of observations are accepted as normal, MD is calculated from the first
290 observations. Different from the training set, the threshold is set as the maximum
value of the MD. Observations from the test set are labelled again according to
threshold value found from the validation set. Finally, both results are compared and a
confusion matrix is created to evaluate the model accuracy.
3.2 Auto-Encoder Neural Network Approach

Auto-Encoders are basic learning circuits which expect to change contributions to
yields with the least conceivable measure of mutilation [12]. While reasonably
necessary, they play a significant job in AI. Auto-Encoders were first presented during
the 1980s by Hinton and the PDP gathering to address the issue of back-propagation
without an instructor, by utilizing the information as the educator. Together with
Hebbian learning rules, ANN give one of the principal ideal models for solo learning
and for starting to address the riddle of how synaptic changes prompted by nearby
biochemical occasions can be composed in a self-sorted out way to deliver worldwide
learning [13].
Using this algorithm in clustering operations, we have an artificial neural network
such that the number of input nodes are same as number of variables, 2 nodes in the
middle layer and the number of nodes in the output is still the number of input
variables. Thanks to this network, we have grouped our dataset. In determining the
parameter of the algorithm, we had to intervene at several points for each study. These
varied in the structure of the artificial neural network, the number of nodes, the number
of training cycles required for the learning of the algorithm and the amount of data to
be used in each training.
3.3 Isolation Forrest Approach

Isolation Forest is a model-free calculation that is computationally effective, can
undoubtedly be adjusted for use with parallel processing ideal models, and has been
demonstrated to be exceptionally successful in recognizing inconsistencies. The main
advantage of the algorithm is that it does not depend on building a profile for infor-
mation with an end goal to discover tests that don’t fit in with this profile [14].
Isolation Forest algorithm is a commonly used method for the detection of
anomalies and outliers and is based on decision trees. We chose to use this algorithm in
our study because we think it is an appropriate method for our case study. As this
method is tree-based, we could not distinguish it as an anomaly and potential anomaly,
but only on the detection of anomalies on the data. As with ANN, the application was
performed on data sets separated by test and train. The model created with train data
was run with test data and the results were collected.
3.4 One-Class Support Vector Machine Approach

Support vector machines are methods that depend on the Structural Risk Minimization
principle from statistics. SVM’s were introduced to solve the two-class pattern recog-
nition problem. A different version of SVM has been proposed for handle classification
problems using one-class data. This new method, named as one-class SVM [15].
The One-Class SVM method is a successful classification technique. Primarily, there
are many examples of anomaly and outlier detection. For this reason, we used this
method, considering that it may be suitable for our application. As in Isolation Forest,
this algorithm is only intended to detect anomalies. We could not detect any potential
anomaly because they did not work with any threshold values. The algorithm was run
with a single configuration for the overall production line and machine-based operations.
As with ANN, the application was performed on data sets separated by test and train. The
model created with train data was run with test data and the results were collected.
4 Conclusion
In this study, we aimed to make malfunction detection on machines by using sensor

data on the production line of MDF manufacturing. During our research, we had the
opportunity to apply four different methods that listed as Mahalanobis Distance, Auto-
Encoder Neural Network, Isolation Forest and One-Class SVM (Table 2).
Table 2. Production line results.

Algorithm Accuracy Precision Recall Specificity F1 score
Mahalanobis Distance 0.993 0.991 1 0.930 0.995
Auto-Encoder NN 0.996 0.996 1 0 0.998
Isolation Forrest 0.920 1 0.920 – 0.959
One-Class SVM 0.989 1 0.989 – 0.994
When we look at the results of our model studies for the detection of malfunction in
the production line, we concluded that Auto-Encoder NN is the most successful
algorithm that obtains the highest scores according to both accuracy and F1 score
metric. Table 3 summarizes the results of the algorithms with 21 machine data. The
Count Score on F1 value indicates on how many machines the algorithm has the
highest F1 score. The Number of Fail metric shows how many models this algorithm
does not produce results or produces an accuracy score of 0.5 or less. According to
these results, we found that the most suitable and successful algorithm for detecting
malfunction on a machine basis is One-Class SVM.
When we evaluated the methods we used as a result of our studies and trials, we
found that the most successful algorithm for detecting a malfunction in the production
line is the Auto-Encoder Neural Network. In our machine-based studies, One-Class
SVM algorithm was the method that produces the most accurate results in different
machines.
Table 3. Final results of the machine based models.

Algorithm Average Average F1 Count score on Number of
accuracy score F1 fail
Mahalanobis 0.797 0.834 0 3
Distance
Auto-Encoder NN 0.836 0.903 4 0
Isolation Forrest 0.839 0.883 4 2
One-Class SVM 0.871 0.901 13 1
As a result of this study, we have developed an application to help identify mal-

function in production systems and to perform maintenance and repair at the right time
without affecting the speed and flow of the production line. This study can be used as
an adjunct decision application to make the production line work more efficiently. In
case of malfunction detection for the production line in general, the site manager can
check whether there is a widespread problem on the belt. Furthermore, it will be able to
provide control based on the machine, with models that work based on machinery.
As a result, this study, which will work with continuously collected sensor data,
will be useful for increasing the production line efficiency by performing maintenance
and repair works on production lines at the right time. Besides, problems that cannot be
seen with the human eye can be identified in advance with the studies performed. The
results of this study will increase the accuracy rate of malfunction detection process by
working with more data and comparing the results with long-term production data.
As the future work, models that used in this study, can be re-run and compared
again with long-term data and this application can be made ready for use and integrated
into various frontend tools.
References
1. Perwej, Y., Omer, M.K., Sheta, O.E., Harb, H.A.M., Adrees, M.S.: The future of Internet of
Things (IoT) and its empowering technology. Int. J. Eng. Sci. Comput. 9(3), 20192–20203
(2019)
2. Dai, H.N., Wang, H., Xu, G., Wan, J., Imran, M.: Big data analytics for manufacturing
internet of things: opportunities, challenges and enabling technologies. Enterp. Inf. Syst. 1–
14 (2019)
3. Deshpande, V., George, L., Badis, H.: PulSec: secure element based framework for sensors
anomaly detection in industry 4.0. IFAC Proc. 52(13), 1204–1209 (2019)
4. Li, S.: Anomaly detection for dummies. https://towardsdatascience.com/anomaly-detection-
for-dummies-15f148e559c1. Accessed 02 July 2019
5. Chamikara, M.A.P., Bertok, P., Khalil, I., Liu, D., Camtepe, S.A., Atiquzzaman, M.: A
Trustworthy privacy preserving framework for machine learning in industrial IoT systems.
IEEE Trans. Ind. Inf. 16, 6092–6102 (2020)
6. Susto, G.A., Terzi, M., Beghi, A.: Anomaly detection approaches for semiconductor
manufacturing. Proc. Manuf. 11, 2018–2024 (2017)
7. Graß, A., Beecks, C., Soto, J.A.C.: Unsupervised anomaly detection in production lines. In:
Machine Learning for Cyber Physical Systems, pp. 18–25 (2019)
8. Liu, J., Guo, J., Orlik, P.V., Shibata, M., Nakahara, D., Mii, S., Takáč, M.: Anomaly
detection in manufacturing systems using structured neural networks. In: 13th World
Congress on Intelligent Control and Automation (WCICA), China, pp 175–180 (2018)
9. Luque, A., Carasso, A., Martin, A., Heras, A.: The impact of class imbalance in classification
performance metrics based on the binary confusion matrix. Pattern Recogn. 91, 216–231
(2019)
10. Basile, V., Bosco, C., Fersini, E., Nozza, D., Patti, V., Pardo, F.M.R., Rosso, P., Sanguinetti,
M.: Multilingual detection of hate speech against immigrants and women in Twitter. In:
Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019).
Association for Computational Linguistics (2019)
11. Selva, P.: Mahalanobis distance – understanding the math with examples (2019)
12. Chen, R., Wu, W., Qi, H., Wang, J., Wang, H.: A stacked autoencoder neural network
algorithm for breast vancer diagnosis with magnetic detection electrical impedance
tomography. IEEE Access 8, 5428–5437 (2020)
13. Baldi, P.: Autoencoders, Unsupervised learning and deep architectures. In: Proceedings of
the 2011 International Conference on Unsupervised and Transfer Learning workshop, vol.
27, pp. 37–50 (2012)
14. Hariri, S., Kind, M.C., Brunner R.J.: Extended isolation forrest. ArXiv, abs/1811.02141
(2019)
15. Fernández Hilario, A., García López, S., Galar, M., Prati, R.C., Krawczyk, B., Herrera, F.:
Learning from Imbalanced Data Sets. Springer, Heidelberg (2019)
Classification of Breast DCE-MRI Images
via Boosting and Deep Learning Based
Stacking Ensemble Approach
Ahmet Haşim Yurttakal1(B) , Hasan Erbay2 , Türkan İkizceli3 ,

Seyhan Karaçavuş4 , and Cenker Biçer5
1
Computer Technologies Department, Technical Sciences Vocational School,
Bozok University, Yozgat, Turkey
[email protected]
2
Computer Engineering Department, Engineering Faculty,
University of Turkish Aeronautical Association, 06790 Etimesgut, Ankara, Turkey
3
Haseki Training and Research Hospital, Department of Radiology,
University of Health Sciences, İstanbul, Turkey
4
Kayseri Training and Research Hospital, Department of Nuclear Medicine,
University of Health Sciences, Kayseri, Turkey
5
Statistics Department, Arts and Science Faculty, Kırıkkale University,
Kırıkkale, Turkey
Abstract. The radiomics features are capable of uncovering disease

characteristics to provide the right treatment at the right time where
the disease is imaged. This is a crucial point for diagnosing breast can-
cer. Even though deep learning methods, especially, convolutional neural
networks (CNNs) have demonstrated better performance in image clas-
sification compared to feature-based methods and show promising per-
formance in medical imaging, but hybrid approaches such as ensemble
models might increase the rate of correct diagnosis. Herein, an ensemble
model, based on both deep learning and gradient boosting, was employed
to diagnose breast cancer tumors using MRI images. The model uses
handcrafted radiomic features obtained from pixel information breast
MRI images. Before training the model these radiomics features applied
to factor analysis to optimize the feature set. The accuracy of the model
is 94.87% and the AUC value 0.9728. The recall of the model is 1.0
whereas precision is 0.9130. F1-score is 0.9545.
Keywords: Breast cancer · Radiomic · Gradient boosting · Deep

learning · Stacked ensemble
1 Introduction
Breast cancer, a general health problem, worsens day by day due to sedentary
daily life habits. It is the most common cancer among women, comprising 23%
of all female cancers all over the world [19]. In Western countries, one in every
1126 A. H. Yurttakal et al.
eight to nine women has breast cancer at some point during their lifetime [4], also
among the deaths from cancer in the western world, after women’s lung cancer
breast cancer comes second [4]. Thus, early detection of a breast cancer tumor,
before physical symptoms develop, is crucial not only in the treatment process
but also in reducing death rates. Mammography is a valuable tool in identify-
ing breast cancer in an early stage, and it is the only test tool to date proven
to reduce deaths due to breast cancer [5,7,10,25]. To decrease false-negative
diagnosis in mammography, biopsy recommended for lesions with greater than
a 2% chance of having suspected malignant tumors [24] and, among them, less
than 30 percent found to have malignancy [1,23]. To reduce unnecessary biop-
sies, recently, Magnetic Resonance Imaging (MRI) also used for the diagnosis
of breast cancer [20,28] since it has an excellent capability for soft tissue imag-
ing and the most sensitive technique in detecting breast diseases, besides, it
does not contain potentially dangerous radiation [15]. But interpreting MRI
images is both time-consuming and requires reader experience. In recent years,
computer-assisted systems are emerging with different goals in many fields, espe-
cially diagnosing in medicine such as to detect abnormal lesions and determine
tissue characterization in medical images [12,14].
On the other hand, radiomic features, i.e. quantitative imaging features [26],
have the superior capability of providing crucial information regarding tumor
identification. Radiomics has emerged from oncology but can be applied to
other imaged diseases. Recently, many successful studies conducted in the field
of radiomic feature-based classification. Gu et al. [13] analyzed performance
radiomic features based on the prediction of Ki-67 in non-small cell lung cancer.
Brunese [6] proposed Voting ensemble classifier for brain cancer detection using
radiomic features. Alzubi et al. [3] investigated the classification of lung cancer
using a boosted neural network. Glioblastoma [21], cervical cancer [29], pancre-
atic cystic lesions [9] are some of the other important studies in the literature.
It is worthy to note that the selection of the optimal set of radiomics features is
still a challenging problem to guarantee a reliable and robust prediction. Ensem-
ble learning is a paradigm that enables creating the model with multi-machine
learning algorithms to improve accuracy. It is a committee-based approach and
aims to get a vote from each member. Even though training time is higher than
traditional learners, ensemble learning more accurate than a single learner.
In this study, 105 different radiomic features were extracted from the thresh-
old based segmented tumorous region. Factor analysis was applied to the dataset
to select the optimal features to guarantee a reliable and robust prediction.
Then, Stacking Ensemble Model built on Gradient Boosting Machine (GBM)
and Deep Learning (DL) was employed to diagnose breast cancer tumors using
Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) images.
The model is promising in terms of performance criteria such as accuracy, error
rate, precision, specificity, and precision, and it is competitive compared to the
literature.
Classification of Breast DCE-MRI Images 1127
The rest of the paper is organized as follows. In Sect. 2, the dataset is detailed,
and the methods used in the study presented. Section 3 describes simulation
results as well as the effectiveness of the proposed model via some statistical met-
rics. Finally, Sect. 4 concludes the paper with a brief summary and the paper’s
contribution.
2 Materials and Methods
2.1 Dataset
The dataset used in the current study is the same dataset used in [30]. It is a
private dataset and contains MRI images obtained from Haseki Training and
Research Hospital in İstanbul, Turkey. The institutional research ethics commit-
tee at Haseki Training and Research Hospital approved this retrospective study.
The dataset consists of breast MRI images of 200 tumorous regions among them,
98 are benign (17.63 ± 5.79 mm in size) and 102 malignant (29.80 ± 9.88 mm
in size). Patients’breast MRI images and data system lesion characteristics were
assessed by two radiologists independently, and lesioned regions were classified
into benign and malign. Later, the diagnosis of all lesions was verified with a
core-needle biopsy.
The proposed approach is composed of two main stages. During the first
stage, data processed and features selected. To be more precise and describe
in detail, the region of interests (ROIs) determined, each ROI pre-processed,
segmented, features extracted selected. And the second stage is the mixed stacked
ensemble classification model, which includes some classification algorithms.
2.2 Data Pre-processing and Splitting
The tumorous regions in the MRI images were cropped manually as square
shapes to obtain ROIs. Then, each ROI resized to 50 × 50. After that, ROIs
normalized so that pixel values range between 0 and 1. Later, each pre-processed
ROIs denoised via deep neural network DnCNN [31]. The final set of ROIs
randomly split into training, validation, and test sets. Table 1 shows the splitting.
Table 1. Training and test sets statistics
Cases Training Validation Test Total

Benign 58 17 23 98
Malign 70 16 16 102
Total 128 33 39 200
2.3 Segmentation
Segmentation is a crucial step in advanced image analysis applications to match

the grayscale ROI image to a fixed value to obtain a binary image. Here, fixed
thresholding performed for segmentation. Any pixels larger than the pre-specified
threshold value between 0 and 1 in grayscale set to 1 in a binary image. Pixels
smaller than or equal to the pre-specified value set to 0 in the binary image.
2.4 Radiomic Features Extraction
Feature extraction is the most important step in data-driven methods, like in

the medical data process. It is the process of obtaining qualitative and quan-
titative information from clinical images that can characterize the diagnosis of
the disease [22]. Open-source pyradiomic python package was performed during
feature extraction [26].
The package provides the necessary infrastructure for the extraction of
Radiomics features from medical imaging. It allows feature extractions both in
2D and 3D. The set of features divided into several families, but the intensity-
based, histogram-based, morphological, and textural based features processed.
Here, a total of 105 different radiomic features were extracted from seven matri-
ces using breast MRI images and masks for benign and malignant classification.
2.5 Factor Analysis
The main aim of applying factor analysis was to reduce the complexity of the
dataset, in other words, to eliminate those features that don’t have an adequate
effect in discriminating benign and malignant lesions. Statistically speaking, it
reduces correlated variables into a lesser number of uncorrelated factors. By
doing the factor analysis, all the observations in the dataset effectively utilized by
eliminating the effects of correlated features. Here, SPSS software was employed
to process the factor analysis to the dataset. In the end, it discovered that only 10
features had 99% representation for the dataset. Namely, the number of features
reduced from 105 to 10 with a good representation of the dataset.
2.6 Proposed Ensemble Model
Recall that ensemble learning is a paradigm that enables creating the model with
multi-machine learning algorithms to improve accuracy. It is a committee-based
approach and aims to get a vote from each member. Multiple learners, called
based/weak learners, are trained to solve the same problem. Stacking is also one
of the ensemble learning approaches. The stacked model is an algorithm class
that includes a second level “Meta Learner” training to find the best combination
of base classification models [16]. Here, the mixed stacked ensemble model was
used and it contains some classification algorithms. An open-source H2O python
package performed during classification [2].
Gradient Boosting (GB) and Deep Learning (DL) used as the base models,
but Generalized Linear (GL) used as a meta learner. GB creates parallel regres-
sion trees, fully distributed across all features of the data set [11]. Deep learning
was the multi-layered feed-forward neural network type [8]. Stochastic gradient
descent (SGD) preferred as optimizer and learning rate was 0.005. Moreover,
elastic regularization was used to prevent overfitting. Two hidden layers and one
output layer were used in the deep learning architecture. Hidden layer size was
200 × 200. The activation function of the hidden layers was Rectifier and the
activation function of the output layer was Softmax.
The model architecture implemented in Python environment. SPSS program
also used for statistical operations. A total of 200 tumor breast MR images
randomly allocated, 60% for training, 20% for validation, 20% for testing. The
overall performance of the proposed ensemble model measured for regression and
classification separately. The regression model’s effectiveness obtained by evalu-
ating Mean Squared Error (MSE), Root Mean Squared Error (RMSE), but, the
classification model’s effectiveness was by evaluating Gini Coefficient, Absolute
Matthews Correlation Coefficient (MCC), F1 Score, Accuracy, Logloss, Area
Under ROC Curve (AUC), and Area Under Precision-Recall Curve (AUCPR).
Figures 1 and 2 show the ability to distinguish benign and malign tumorous
lesions. According to Fig. 1 presents the validation set predictions. Figure 2 the
proposed ensemble model mispredicted two false-positive cases during testing,
i.e. testing with unseen data.
Fig. 1. Stacked ensemble confusion Fig. 2. Stacked ensemble confusion

matrices-validation matrices-test
The proposed model obtained 94.87% accuracy by correctly predicting 37

of 39 cases encountered from the test set. Only 2 cases misclassified. The AUC
value of the model is 0.9728 whereas MCC is 0.9008. The precision, recall, F1
score values are 0.9130, 1.00, 0.9545 respectively. Detailed performance statistics
are given in Table 2.
Table 2. Stacked ensemble model performance
Metrics Values
MSE 0.109168411760
RMSE 0.330406434199
LogLoss 0.376175538966
AUC 0.972826086957
AUCPR 0.893640683255
Accuracy 0.948717948718
Precision 0.913043478261
Recall 1.0
F1 score 0.954545454545
Computer-assisted systems are emerging with different goals in many fields, espe-
cially diagnosing in medicine. These systems are important in early diagnosis,
rehabilitation, and treatment planning, so reliability and robustness are among
the must requirements. Herein, an ensemble model for diagnosing breast MRI
was proposed, the model capable of distinguishing benign and malignant tumors,
even when the tumor biologic features reflect differences. There exist similar sys-
tems in the literature.
Besides, Whitney et al. [27] developed a system to differentiate benign and
malign in luminal A-type lesions. In the study, a total of 654 DCE-MRI lesions
was segmented with fuzzy C-means and, 38 radiomic features were extracted.
As result, the LDA classifier achieved AUC score of 72.9%. Liang et al. [17]
estimated Ki-67 status using T2-weighted and, contrast-enhanced T1-weighted
images of breast cancer tumors. The dataset composed of 318 patients, among
them 200 included in the training set and 118 in the validation set. The extracted
radiomic features were classified with LR and 72% accuracy obtained. Ma et
al. [18] analyzed the relationship between radiomic features and Ki-67. A total
of 56 radiomic features were extracted from images segmented with an active
contours algorithm. NB, KNN and SVM classifiers were employed. Among these
NB showed the best performance in AUC, accuracy, sensitivity, and specificity
values by achieving 77.3%, 75.7%, 77.7%, 76.9%, respectively.
In this study, a computerized breast cancer diagnosis system is proposed.

The system uses only breast DCE-MRI images, namely quantitative biomarkers
are the input and, biochemical markers were ignored. In the development of the
system, radiomic features were used. The feature set was optimized by factor
analysis before training. An ensemble-based approach was preferred. To the best
of our knowledge, the model is the first in the literature in classifying tumorous
breast lesions using the breast DCE-MRI image set. The model was trained and
tested with the comparatively large lesions (benign 17.63 ± 5.79 mm and malign
29.80 ± 9.88 mm). The performance values are encouraging. The AUC value of
the model was 97.28% and the accuracy 94.87%. The precision, recall, F1 score
values were 0.9130, 1.00, 0.9545 respectively. It is aimed to develop the model
by applying different classification algorithms.
References
1. Adler, D., Helvie, M.: Mammographic biopsy recommendations. Curr. Opin.
Radiol. 4(5), 123–129 (1992)
2. Aiello, S., Click, C., Roark, H., Rehak, L., Lanford, J.: Machine learning with
python and H2O. In: Lanford, J. (ed.) vol. 20, p. 2016 (2016)
3. ALzubi, J.A., Bharathikannan, B., Tanwar, S., Manikandan, R., Khanna, A.,
Thaventhiran, C.: Boosted neural network ensemble classification for lung cancer
disease diagnosis. Appl. Soft Comput. 80, 579–591 (2019)
4. Aydıntuğ, S.: Meme kanserinde erken tanı. Sted 13(6), 226–229 (2004)
5. Berg, W.A.: Benefits of screening mammography. JAMA 303(2), 168–169 (2010)
6. Brunese, L., Mercaldo, F., Reginelli, A., Santone, A.: An ensemble learning app-
roach for brain cancer detection exploiting radiomic features. Comput. Methods
Programs Biomed. 185, 105134 (2020)
7. Cady, B., Michaelson, J.S.: The life-sparing potential of mammographic screening.
Cancer 91(9), 1699–1703 (2001)
8. Candel, A., Parmar, V., LeDell, E., Arora, A.: Deep learning with H2O. H2O. AI
Inc. (2016)
9. Dalal, V., Carmicheal, J., Dhaliwal, A., Jain, M., Kaur, S., Batra, S.K.: Radiomics
in stratification of pancreatic cystic lesions: machine learning in action. Cancer
Lett. 469, 228–237 (2019)
10. Feig, S.A., D’Orsi, C.J., Hendrick, R.E., Jackson, V.P., Kopans, D.B., Monsees, B.,
Sickles, E.A., Stelling, C.B., Zinninger, M., Wilcox-Buchalla, P.: American college
of radiology guidelines for breast cancer screening. AJR Am. J. Roentgenol. 171(1),
29–33 (1998)
11. Friedman, J., Hastie, T., Tibshirani, R.: The Elements of Statistical Learning.
Springer Series in Statistics, vol. 1. Springer, New York (2001)
12. Gity, M., Arabkheradmand, A., Taheri, E., Shakiba, M., Khademi, Y., Bijan, B.,
Sadaghiani, M.S., Jalali, A.H.: Magnetic resonance imaging features of adenosis in
the breast. J. Breast Cancer 20(1), 116 (2017)
13. Gu, Q., Feng, Z., Liang, Q., Li, M., Deng, J., Ma, M., Wang, W., Liu, J., Liu, P.,
Rong, P.: Machine learning-based radiomics strategy for prediction of cell prolifer-
ation in non-small cell lung cancer. Eur. J. Radiol. 118, 32–37 (2019)
14. Hamidinekoo, A., Denton, E., Rampun, A., Honnor, K., Zwiggelaar, R.: Deep learn-
ing in mammography and breast histology, an overview and future trends. Med.
Image Anal. 47, 45–67 (2018)
15. Kuhl, C.: The current status of breast MR imaging part I. choice of technique,
image interpretation, diagnostic accuracy, and transfer to clinical practice. Radi-
ology 244(2), 356–378 (2007)
16. Van der Laan, M.J., Polley, E.C., Hubbard, A.E.: Super learner. Stat. Appl. Genet.
Mol. Biol. 6(1) (2007)
17. Liang, C., Cheng, Z., Huang, Y., He, L., Chen, X., Ma, Z., Huang, X., Liang,
C., Liu, Z.: An MRI-based radiomics classifier for preoperative prediction of Ki-67
status in breast cancer. Acad. Radiol. 25(9), 1111–1117 (2018)
18. Ma, W., Ji, Y., Qi, L., Guo, X., Jian, X., Liu, P.: Breast cancer Ki67 expression
prediction by DCE-MRI radiomics features. Clin. Radiol. 73(10), 909-e1 (2018)
19. Özmen, V.: Breast cancer in the world and turkey. J. Breast Health 4(2), 6–12
(2008)
20. Prince, J.L., Links, J.M.: Medical Imaging Signals and Systems. Pearson Prentice
Hall, Upper Saddle River (2006)
21. Qian, Z., Li, Y., Wang, Y., Li, L., Li, R., Wang, K., Li, S., Tang, K., Zhang, C.,
Fan, X., et al.: Differentiation of glioblastoma from solitary brain metastases using
radiomic machine-learning classifiers. Cancer Lett. 451, 128–135 (2019)
22. Rizzo, S., Botta, F., Raimondi, S., Origgi, D., Fanciullo, C., Morganti, A.G., Bel-
lomi, M.: Radiomics: the facts and the challenges of image analysis. Eur. Radiol.
Exp. 2(1), 1–8 (2018)
23. Sickles, E.A.: Periodic mammographic follow-up of probably benign lesions: results
in 3,184 consecutive cases. Radiology 179(2), 463–468 (1991)
24. Sickles, E.A.: Screening for breast cancer with mammography. Clin. Imaging 15(4),
253–260 (1991)
25. Smith, R.A., Cokkinides, V., Eyre, H.J.: American cancer society guidelines for
the early detection of cancer, 2003. CA: Cancer J. Clin. 53(1), 27–43 (2003)
26. Van Griethuysen, J.J., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan,
V., Beets-Tan, R.G., Fillion-Robin, J.C., Pieper, S., Aerts, H.J.: Computational
radiomics system to decode the radiographic phenotype. Cancer Res. 77(21), e104–
e107 (2017)
27. Whitney, H.M., Taylor, N.S., Drukker, K., Edwards, A.V., Papaioannou, J.,
Schacht, D., Giger, M.L.: Additive benefit of radiomics over size alone in the dis-
tinction between benign lesions and luminal a cancers on a large clinical breast mri
dataset. Acad. Radiol. 26(2), 202–209 (2019)
28. Wollins, D.S., Somerfield, M.R.: Q and a: magnetic resonance imaging in the detec-
tion and evaluation of breast cancer. J. Oncol. Pract. 4(1), 18–23 (2008)
29. Wormald, B.W., Doran, S.J., Ind, T.E., D’Arcy, J., Petts, J., deSouza, N.M.:
Radiomic features of cervical cancer on T2-and diffusion-weighted MRI: prognos-
tic value in low-volume tumors suitable for trachelectomy. Gynecol. Oncol. 156,
107–114 (2019)
30. Yurttakal, A.H., Erbay, H., İkizceli, T., Karaçavuş, S.: Detection of breast cancer
via deep convolution neural networks using MRI images. Multimed. Tools Appl.
79, 1–19 (2019)
31. Zhang, K., Zuo, W., Chen, Y., Meng, D., Zhang, L.: Beyond a gaussian denoiser:
residual learning of deep CNN for image denoising. IEEE Trans. Image Process.
26(7), 3142–3155 (2017)
Machine Learning Approach for Automatic
Categorization of Service Support Requests
on University Information Management
System
Aytuğ Onan1(&), Erdem Atik2, and Adnan Yalçın2

1
Faculty of Engineering and Architecture, Department of Computer
Engineering, İzmir Katip Çelebi University, 35620 İzmir, Turkey
[email protected]
2
Graduate School of Natural Sciences, Division of Software Engineering,
İzmir Katip Çelebi University, 35620 İzmir, Turkey
{erdem.atik,adnan.yalcin}@ikcu.edu.tr
Abstract. The bug reports and service support requests provided by users to
software developers serve as an important source of information for software
maintenance. For large software projects, the number of users involved may be
high. As a result, large number of service support requests may be generated. In
this study, we present a machine learning approach for automatic categorization
of service support requests on university information management system of
Izmir Katip Celebi University. We have collected a text corpus, which contains
17,831 bug reports and service support requests. On the corpus, preprocessing
stages, such as, tokenization, stop word filtering and stemming have been
employed. To represent text documents, TF-IDF term weighting scheme has
been utilized in conjunction with 1-gram model. In the empirical analysis, five
conventional classification algorithms (i.e., Naïve Bayes, k-nearest neighbor
algorithm, C4.5 decision tree algorithm, random forest algorithm and support
vector machines) have been taken into consideration. The experimental results
indicate that the presented machine learning scheme can yield promising results
on assigning service support requests to one of the related modules, as student
information system, personnel information system, electronic document man-
agement system and scientific research projects system. We achieved a classi-
fication accuracy of 92.26% with support vector machines.
Keywords: Machine learning Text mining Bug reports Software

requirements
1 Introduction
The bug reports and service support requests provided by users to software developers
serve as an essential source of information for software maintenance. Users notify soft-
ware developers in software maintenance by bug reports, regarding every part of a
software product involving corrective maintenance [1]. For large software projects, the
number of users involved may be high. As a result, large number of service support
requests may be generated.
1134 A. Onan et al.
With the recent progresses in information and communication technologies,

immense quantity of unstructured text documents become available on information
systems. Text classification is the subfield of text mining in which text documents have
been assigned into one or more predefined class labels. Text classification is an
essential task in library science, information science and computer science [2]. Many
tasks from various fields can be modelled as a text classification task, such as, news
filtering, document organization, retrieval, sentiment analysis and spam filtering [3, 4].
In this study, we present a text classification based approach for automatic cate-
gorization of service support requests on university information management system of
Izmir Katip Celebi University. We have collected a text corpus, which contains 17,831
bug reports and service support requests. On the corpus, pre-processing stages, such as,
tokenization, stop word filtering and stemming have been employed. To represent text
documents, TF-IDF term weighting scheme has been utilized in conjunction with 1-g
model. In the empirical analysis, five conventional classification algorithms (i.e., Naïve
Bayes, k-nearest neighbour algorithm, C4.5 decision tree algorithm, random forest
algorithm and support vector machines) have been taken into consideration. The
experimental results indicate that the presented machine learning scheme can yield
promising results on assigning service support requests to one of the related modules,
as student information system, personnel information system, electronic document
management system and scientific research projects system. We achieved a classifi-
cation accuracy of 92.26% with support vector machines.
The rest of this paper is structured as follows: In Sect. 2, related work has been
presented. In Sect. 3, the methodology of the study has been briefly given. In Sect. 4,
the experimental results have been presented. Finally, Sect. 5 presents the concluding
remarks of the study.
2 Related Work
As far as is known, there is no earlier study to determine the relevant modules on

service support software requests with the use of text classification techniques. How-
ever, there are several earlier research contributions on text classification for similar
purposes on different fields.
Alshalabi et al. [5] employed machine learning techniques to automatically classify
Malay texts. In this study, two feature selection algorithms (i.e., information gain and
chi-square) and two machine learning algorithms (i.e., k-nearest neighbor algorithm
and Naïve Bayes algorithm) have been employed. In another study, Hmeidi et al. [6]
classified text documents into three predefined topics, as economical, political and
sports.
In another study, Bardız [7] developed a machine learning based system to rec-
ommend appropriate hospital branch for patients based on their complaints, by taking
the earlier records of patients provided by the doctors in the hospital management
ystem. Similarly, Delen and Crossland [8] presented a machine learning based
approach on text collections of business organizations. The information obtained from
text documents containing customer complaints and satisfaction has been utilized to
improve the product development, error tracking, and guarantee period.
Machine Learning Approach for Automatic Categorization 1135
In another study, Sancar [9] utilized text classification methods for demand recog-
nition and steering. In this scheme, text documents have been processed by optical
character recognition method and their sentences were analysed with the use of machine
learning classifiers, such as, Naïve Bayes, multinomial Naïve Bayes, support vector
machines, k-nearest neighbour algorithm, artificial neural networks and decision tree
algorithm. For the compared configurations, the highest predictive performance has
been achieve by multinomial Naïve Bayes algorithm, with a classification accuracy of
87.6%. Parlak and Uysal [10], in their study, addressed the issue of classifying medical
documents according to diseases. A single-label multi-class dataset was created in the
database of medical terms referred as MEDLINE. In the empirical analysis, three dif-
ferent text classification algorithms (namely, C4.5, Bayesian network and random forest
algorithm) have been utilized. Among these algorithms, the Bayesian network has
achieved the highest classification accuracies.
3 Methodology
In this section, the main methodology of the study (dataset collection, preprocessing
and text classification algorithms) has been presented.
3.1 Dataset Collection and Preprocessing

The general stages employed in this study outlined below:
Determination of text documents: We have collected a text corpus, which contains
17,831 bug reports and service support requests. The requests were obtained from
university information management system of Izmir Katip Celebi University (ÜBYS).
Four predefined categories have been utilized to collect text documents. The categories
include student information system, personnel information system, electronic document
management system and scientific research projects system.
Organization and processing of text documents: Since punctuation marks are not
informative for construction of classification model, they have been eliminated. Sim-
ilarly, the numerical characters and personnel information have been cleared. Then,
normalization has been obtained on words, so that typographical errors have been
eliminated. In this way, the number of features (words) have been minimized. The
separation process is applied to the root of the words in normalized texts. The purpose
of this process is to provide the simplification of words with the same root but different
suffixes. Conjunctions, cursors, stereotypes, abbreviations, etc. are taken into consid-
eration and the words are made to be the simplest. In order to determine the root form
of the words, the Zemberek library, which considers the rule-based structure of
Turkish, was preferred [11].
Then, stop word filtering process was applied. Thus, the words that should not be
included in the evaluation were excluded. Some data sets were determined for stop-
words filtering, and a common pool was created from the combination of these sets. By
removing the data contained in this pool from the texts, unnecessary words, con-
junctions, prepositions, etc. words have been deleted. After this process, the same
method was applied for the Turkish name list (Ali, Ahmet, Mehmet, etc.) and the
names in the texts were eliminated.
1136 A. Onan et al.
Creating term matrices on the cleaned raw data: The modules to which the requests
belong are used as class labels. Cleaned texts are considered as related documents. To
represent text documents, TF-IDF term weighting scheme has been utilized in con-
junction with 1-g model.
3.2 Text Classification Algorithms

In the empirical analysis, five supervised learning algorithms (namely, k-nearest
neighbour algorithm, Naïve Bayes algorithm, random forest algorithm, C4.5 decision
tree algorithm, and support vector machines) have been considered. In addition,
majority voting ensemble of five supervised learning algorithms have been taken into
consideration. The rest of this section briefly discuss the text classification algorithms
utilized in the empirical analysis:
K-nearest neighbour (KNN) is a learning algorithm based on an instance for
supervised learning tasks, including classification and regression tasks. In this scheme,
in the training set, the class label for an instance was calculated based on the similarity of
the instance with its closest neighbours. All instances were stored in this system, and the
class mark was defined at the time of classification based on the k- neighbours [12, 13].
Naïve Bayes (NB) is a text classification algorithm based on Bayes’ theorem. The
algorithm is based on conditional independence assumption. The assumption regards as
if the attributes are conditionally independent with regard to the class label. This
simplifies the necessary computations involved. The algorithm is a very efficient
algorithm for machine learning tasks, such as text classification [14].
Random forest algorithm [15] is a supervised learning algorithm based on bagging
and random subspace algorithm. In the algorithm, decision trees that have been con-
structed on bootstrap samples of the training data. The ensemble diversity has been
achieved by random projection of features. In this way, the model can deal with data
sets with noisy or irrelevant feature sets.
C4.5 decision tree algorithm is an important decision tree induction algorithm [16]. In
this algorithm, information gain has been utilized as test attribute selection measure. In
each iteration, the algorithm selects an attribute with the highest information gain [16].
Support vector machines are linear algorithms for classification and regression tasks
[17]. For the two-class classification tasks, the algorithm finds a hyperplane in higher
dimensional space so that instances of two class have been well separated as much as
possible. The algorithm can yield promising results for text mining tasks [13].
In addition, five aforementioned supervised learning algorithms have been com-
bined into an ensemble by the simple majority voting scheme.
4 Experimental Procedure and Results
In this section, evaluation measures, experimental procedure and experimental results

have been presented.
4.1 Evaluation Measures

Classification accuracy (ACC) is the proportion of true positives and true negatives
obtained by the classification algorithm over the total number of instances as given by
Eq. 1:
TN þ TP
ACC ¼ ð1Þ
TP þ FP þ FN þ TN
where TN denotes number of true negatives, TP denotes number of true positives, FP

denotes number of false positives and FN denotes number of false negatives.
4.2 Experimental Procedure

In the empirical analysis, 10-fold cross validation has been utilized. In this scheme, the
original dataset is randomly divided into ten mutually exclusive folds. Training and
testing process are repeated ten times and each part is tested and trained ten times. The
results reported in this section are the average results for 10-folds. In the empirical
analysis, TF-IDF term weighting scheme has been utilized in conjunction with 1-g
model. The classification algorithms have been implemented via WEKA and the
default parameters of it has been utilized for the classification algorithms.
Table 1. Classification accuracy values obtained by text classifiers.

Classification algorithm Classification accuracy
K-nearest neighbor algorithm (KNN) 80.78
Naive Bayes algorithm (NB) 78.20
Random forest algorithm (RF) 87.38
C4.5 decision tree algorithm 88.54
Support vector machines (SVM) 92.26
Majority voting ensemble 90.68
4.3 Experimental Results

In the empirical analysis, five conventional classification algorithms (i.e., Naïve Bayes,
k-nearest neighbor algorithm, C4.5 decision tree algorithm, random forest algorithm
and support vector machines) have been taken into consideration. In Table 1, the
empirical results obtained by conventional classification algorithms for automatic
identification of service support requests on university information management system
have been presented in terms of classification accuracy values.
The experimental results indicate that the presented machine learning scheme can
yield promising results on assigning service support requests to one of the related
modules, as student information system, personnel information system, electronic
document management system and scientific research projects system. The highest
predictive performance among the compared conventional classification algorithms
1138 A. Onan et al.
have been achieved by support vector machines with a classification accuracy of

92.26%. Majority voting ensemble scheme has achieved the second highest predictive
performance. The third highest predictive performance has been achieved by C4.5
decision tree algorithm.
5 Conclusion
The bug reports and service support requests provided by users to software developers
serve as an important source of information for software maintenance. For large
software projects, the number of users involved may be high. As a result, large number
of service support requests may be generated. In this study, we present a machine
learning approach for automatic categorization of service support requests on university
information management system of Izmir Katip Celebi University. We have collected a
text corpus, which contains 17,831 bug reports and service support requests. On the
corpus, preprocessing stages, such as, tokenization, stop word filtering and stemming
have been employed. To represent text documents, TF-IDF term weighting scheme has
been utilized in conjunction with 1-g model. In the empirical analysis, five conven-
tional classification algorithms (i.e., Naïve Bayes, k-nearest neighbor algorithm, C4.5
decision tree algorithm, random forest algorithm and support vector machines) have
been taken into consideration. The experimental results indicate that the presented
machine learning scheme can yield promising results on assigning service support
requests to one of the related modules, as student information system, personnel
information system, electronic document management system and scientific research
projects system. We achieved a classification accuracy of 92.26% with support vector
machines.
References
1. Bettenburg, N., Premraj, R., Zimmermann, T., Kim, S.: Duplicate bug reports considered
harmful really? In: Proceedings of 2008 IEEE International Conference on Software
Maintenance, pp. 337–345. IEEE, New York (2008)
2. Han, J., Kamber, M.: Data Mining: Concepts and Techniques. Morgan Kaufmann
Publishers, Burlington (2006)
3. Aggarwal, C.C., Zhai, C.X.: Mining Text Data. Springer, Heidelberg (2012)
4. Onan, A., Korukoğlu, S., Bulut, H.: Ensemble of keyword extraction methods and classifiers
in text classification. Expert Syst. Appl. 57, 232–242 (2016)
5. Alshalabi, H., Tiun, S., Omar, N., Albared, M.: Experiments on the use of feature selection
and machine learning methods in automatic malay text categorization. Proc. Technol. 11(1),
748–754 (2013)
6. Hmeidi, I., Al-Ayyoub, M., Abdulla, N.A., Almodawar, A.A., Abooraig, R., Mahyoub, N.
A.: Automatic Arabic text categorization: a comprehensive comparative study. J. Inf. Sci. 41
(1), 114–124 (2015)
7. Bardız, A.: Medikal metin ayrıştırma ve sınıflandırma. In: Ulusal Biyomedikal Cihaz
Tasarımı ve Üretimi Sempozyunu, pp. 76–80 (2017)
8. Delen, D., Crossland, M.D.: Seeding the survey and analysis of research literature with text
mining. Expert Syst. Appl. 34(3), 1707–1720 (2008)
9. Sancar, Y.: Metin madenciliği kullanılarak talep tanıma ve yönlendirme sistemi. Atatürk
Üniversitesi Fen Bilimleri Enstitüsü Bilgisayar Mühendisliği Anabilim Dalı Yüksek Lisans
Tezi (2016)
10. Parlak, B., Uysal, A.K.: Tıbbi dokümanların hastalıklara göre sınıflandırılması. In:
Proceedings of 23rd Signal Processing and Communications Applications Conference,
pp. 1–4. IEEE, New York (2015)
11. Akın, A.A., Akın, M.D.: Zemberek: an open source NLP framework for Turkic languages.
Structure 10, 1–5 (2007)
12. Aha, D.W., Kibler, D., Albert, M.K.: Instance-based learning algorithms. Mach. Learn. 6,
37–66 (1991)
13. Onan, A.: Mining opinions from instructor evaluation reviews: a deep learning approach.
Comput. Appl. Eng. Educ. 28(1), 117–138 (2020)
14. Lewis, D.D.: Naïve (bayes) at forty: the independence assumption in information retrieval.
In: European Conference on Machine Learning, pp. 4–15. Springer, Berlin (1998)
15. Breiman, L.: Random forests. Mach. Learn. 45(1), 5–32 (2001)
16. Gehrke, J.: The Handbook of Data Mining. Lawrence Erlbaum Associates, New Jersey
(2003)
17. Vapnik, V.: The Nature of Statistical Learning Theory. Springer, Heidelberg (2013)
Comparative Study of Different Machine
Learning Models for Remote Sensing
Bathymetry Inversion
Shen Wei1,2(&), Ji Qian1,2, Rao Yali1,2, and Meng Ran1,2

1
Shanghai Ocean University, Shanghai, China
[email protected]
2
Shanghai Engineering Research Center of Estuarine and Oceanographic
Mapping, Shanghai, China
Abstract. Water depth is an essential element of oceanographic research and

marine surveying. Bathymetry inversion based on remote sensing is a time-
effective, low-cost, and wide-coverage solution for shallow sea. Using World-
ViewII multi-spectral remote sensing imagery and laser sounding data, Back
Propagation neural network model (BP), random forest model (RF) and extreme
learning machine model (ELM) were used to inverse water depth Surrounding the
Chinese Ganquan island, and the inversion accuracy was compared and evaluated.
The results show that among the BP, RF and ELM, the RF has the highest water
depth inversion accuracy. The root mean square error (RMSE) of the check point is
0.85, the mean absolute error (MAE) is 0.60, and the mean relative error (MRE) is
3.54%. The coefficient determination R2 reaches 0.97; within the range of 0–10 m
and 15–20 m water depth, the inversion of the ELM is the best; in the range of
10–15 m water depth, the RF has the better inversion effect.
Keywords: Machine learning Remote sensing bathymetry inversion Back

propagation neural network model Random forest model Extreme learning
machine model
1 Introduction
Water depth is an important topographical element. The measurement of depth is helpful

to better understand the topography and geomorphology of shallow sea, and also plays
an important role in carrying out economic and military activities in offshore area [1].
Remote sensing technology is considered a time-effective, low-cost, and wide-coverage
solution. It is also a supplement and improvement to traditional bathymetric measure-
ment methods and techniques. The definition of depth inversion is to use the remote
sensing imagery and some necessary information to calculate the water depth [2].
At present, the models for depth inversion mainly include theoretical models, semi-
theoretical and semi-empirical models, and statistical models. Among them, semi-
theoretical and semi-empirical models that commonly include linear regression model,
dual-band ratio model, and neural network mode are widely used [3]. Lyzenga [4] et al.
proposed a linear relationship between water depth and radiance based on the under-
lying reflection model to obtain water depth information. Stumpf [5] and others
Comparative Study of Different Machine Learning Models 1141
proposed to obtain the water depth value based on the linear regression relationship
between the ratio between the wave bands and the water depth, which to some extent
eliminated the effects of different substrates in shallow seas. Wang [6] discussed the
regression relationship between the first-order differential of reflectivity and water
depth through different models and methods, and proved that water depth and reflec-
tance on water surface have the best correlation in the near infrared band. In order to
improve the accuracy of depth inversion, many scholars began to use neural networks
to perform water depth inversion, and to establish the non-linear relationship between
water depth and reflectivity to retrieve water depth. Sandidge [7] and others used BP
(Back Propagation) neural network to establish a correlation model between remote
sensing reflectance and measured water depth. Anctil [8] et al. used neural networks to
capture the complex changes in water depth fluctuations, and analyzed the effects of
water depth fluctuation models established by input delay neural networks, recurrent
neural networks, and RBF(Radial Basis Function), and concluded that recurrent neural
networks are more suitable for inversion of water depth. Deng [9] and others estab-
lished water depth inversion models of BP and RBF, and compared with single-band
and multi-band models. Zheng [10] and others also established BP and RBF artificial
neural network water depth inversion models, and concluded that the RBF model has a
simpler structure, lower sample requirements, and high inversion accuracy.
Taking Ganquan Island sea area as an example, this paper uses WorldViewII
remote sensing imagery and laser sounding to establish water depth inversion model of
BP neural network, RF, and ELM, and evaluates the accuracy, aim to find out the
suitable depth invertion model and parameters. This paper including introduction,
machine learning model, remote sensing depth inversion experiment, conclusion and
references. In this paper, I refer to a large number of literature materials, combined with
the ideas I want to express.
2 Machine Learning Model
2.1 BP Neural Network Model

BP (Back Propagation) neural network [11] is a multi-layer feedforward neural network
trained according to error back-propagation. The main characteristics of this network
are forward signal feedback and backward propagation of errors. In forward transfer,
the input signal is processed layer by layer from the input layer through the hidden
layer to the output layer. The neuron status of each layer only affects the next layer. If
the actual output does not match the expected output, transfer to back propagation,
invert the output error layer by layer through the hidden layer, and divide the error to all
neurons in each layer. To adjust the network weights and thresholds, after continuous
training, the predicted output value of the BP neural network approaches the expected
output value.
1142 S. Wei et al.
2.2 Random Forest

Random Forest (RF) is a supervised learning algorithm. It is an integrated learning
algorithm that uses decision trees as learners. In the algorithm, multiple prediction
models can be generated at the same time, and the prediction accuracy can be improved
by comprehensively analyzing the prediction results of each model [12]. Perform self-
sampling for each decision tree, and use the sample data outside the bag to analyze
errors.
Assume that there are M sample units and N variables in the training data. The
construction process of the random forest model is as follows:
1) Randomly extract M samples from the original training data and sample them to
generate a training set. The uncollected sample data constitutes the out-of-bag data
set and generates a decision tree.
2) At each node of each tree, randomly select n < N variables and use them as
candidate variables for segmenting the node. The variable tree at each node is
consistent.
3) Each tree is recursively segmented until all training samples of the node belong to
the same class, and the segmentation is no longer performed.
2.3 Extreme Learning Machine

Extreme Learning Machine (ELM) is a type of machine learning algorithm based on
feedforward neuron network. It was proposed by associate professor Huang Guangbin
of Nanyang Technological University in 2004 [13]. Its main feature is that the hidden
layer nodes parameters can be given randomly or artificially given without adjustment,
and the learning process only needs to calculate the output weight. It can be seen in
paper [14, 15] that the ELM application works well. ELM has the advantages of high
learning efficiency and strong generalization ability. It is widely used in classification,
regression, clustering, feature learning and other problems. It is a simple, easy and
effective new algorithm for feedforward neural network of single hidden layer. There is
no need to adjust the bias of input weights and hidden elements in the calculation
process [16]. The basic idea is as follows:
Suppose there are N ðxi ; yi Þ, xi ¼ ½xi1 ; xi2 ; . . .; xin T 2 Rn ; yi ¼ ½yi1 ; yi2 ; . . .; yin T 2
R ; gð xÞ is activation function, the number of hidden layers is L, The ELM model can
m
be expressed as:
XL XL
i¼1
b i g i xj ¼ b g w i xj þ bi ¼ yj
i¼1 i
ð1Þ
j = 1, 2, …, N; wi ¼ ½wi1 ; wi2 ; . . .; win T represents the weight between the input node
and the i-hidden layer node; bi ¼ ½bi1 ; bi2 ; . . .; bim T represents the weight vector
between the i-hidden layer node and the output node; bi represents the threshold of the
i-hidden layer node. The matrix of the ELM model is:
Hb ¼ Y ð2Þ
2 3 2 3
yT1 b1 T
6 yT 7 6 T7
6 27 6 b2 7
Y¼6 7; b ¼ 6
6
7
7;
6 .. 7 6 ... 7
4 . 5 4 5
yTN bTL
2 3
gðw1 x1 þ b1 Þ gðw2 x1 þ b2 Þ ... gðwL x1 þ bL Þ
6 gðw1 x2 þ b1 Þ gðw2 x2 þ b2 Þ ... gðwL x2 þ bL Þ 7
6 7
H¼6
6 .. .. .. ..
7
7
4 . . . . 5
gðw1 xN þ b1 Þ gðw2 xN þ b2 Þ . . . gðwL xN þ bL Þ
When the activation function g(x) is infinitely differentiable, the weights w and
b can be randomly selected and remain unchanged during training. Therefore, training
ELM can be transformed into the least square solution problem of calculating Hb = Y.
The ELM algorithm has three steps:
1) Determine the number of hidden layer neurons L;
2) Select an infinitely differentiable function as the activation function g(x) of the
hidden layer neurons, and then calculate the
output matrix H of the hidden layer;
3) Calculate output layer weight b^ b^ ¼H Y .
þ
3 Remote Sensing Depth Inversion Experiment
3.1 Research Area and Data

This paper selects Chinese Ganquan Island sea area as the research area. The range of
the research area is 118° 34′41″ E to 111° 35′38″ E, 16° 29′56″ N to 16° 31′17″ N. The
remote sensing imagery used is a high-resolution WorldViewII multispectral remote
sensing imagery acquired at 03:33 (GMT) on April 2, 2014. There are four bands are
blue, green, red and near infrared, with 2 m spatial resolution. Radiation calibration,
atmospheric correction, flare correction and other pre-processing are performed on the
study area. 494 water depth points were extracted from the laser sounding data mea-
sured by the Leica hawk eye system. 353 water depth points extracted from the laser
data are used to establish the inversion model and the remaining 141 are used for
accuracy inspection, as shown in Fig. 1. The contemporaneous tide data were down-
loaded from the China Maritime Service Network (http://ocean.cnss.com.cn/).
1144 S. Wei et al.
Fig. 1. Laser points data (a DTM from laser data, b red depth points are used to establish the
inversion model and the green depth points for accuracy inspection)
3.2 Experiment Procedure

In order to ensure the applicability of the model, a part of all measured points is
randomly selected as the control points for modeling, and the remaining points are used
as check points for model verification. The WorldViewII multispectral imagery bands
B1, B2, B3, B4 and the ratio between the bands B1/B2, B1/B3, B1/B4, B2/B3, B2/B4,
and B3/B4 as the independent variable, taking the spectral reflectance and reflectance
ratio of each band of the imagery as the input value of the model, using the corre-
sponding measured water depth value as the output value of the model, using the control
points to establish the model, and using the check points to verify the model. According
to the evaluation index of the model, the most applicable model is judged.
In the process of building the model, BP neural network needs to set a large number
of parameters, in order to obtain the best inversion model, it is necessary to continu-
ously change the number of neurons, the hidden layer function and the output function.
RF and ELM have the ability to learn independently, and the best inversion model can
be obtained by automatically adjusting the parameters of samples. The BP neural
network model is implemented using Matlab programming [17], through continuous
training and adjustment, the final determined parameters are: training function is
trainlm, input layer nodes are 10, hidden layer nodes are 5, and the maximum number
of training times is 1000, the training requires an accuracy of 0.00004 and a learning
rate of 0.1. RF is realized by programming in R language. It is determined that the
number of regression trees in the algorithm is 1000, and the number of variables
required to segment nodes is 6. The ELM model is implemented on the Matlab plat-
form. The input weight matrix is the input matrix of the training set, the offset matrix is
the output matrix of the training set, 20 hidden layer neurons, the transfer function is
the Sigmoidal function, and the type selection regression.
3.3 Water Depth Inversion and Results Analysis

The BP neural network model, the RF model and the ELM model were used to fit the
check points, and the inversion value of water depth was calculated. The errors between
the inversion value and the measured value of each model were compared, and the
accuracy of the model was verified and analyzed. For quantitative comparison of the
model inversion result, the determination coefficient (R2) evaluate the advantages and
disadvantages of the model, the root mean square error (RMSE), mean relative error
(MRE) and the mean absolute error (MAE) three statistical parameters as the depth of
the inversion accuracy of evaluation index, the depth of the model inversion value
regression analysis with water depth measured values, comparing the inversion pre-
cision of the model.
Fig. 2. 1:1 relationship between water depth inversion and real value
From Fig. 2, we can see that the overall matching are relatively compact, and the
inversed water depths by the three models have a significant correlation with the
measured depths. The predicted value of the BP neural network model in the 0–5 m
water depth range is consistent with the trend of the measured value, and the error is
small; within the depth range of 5–15 m, the error increases; the maximum error is in
the range of 15–20 m water depth. The overall fitting of both the RF model and the
ELM model is good. In the water depth range of 0–10 m and 15–20 m, the inversion
error of the RF model is greater than that of the ELM model; within the water depth
range of 10–15 m, the inversion error of the ELM model is greater than that of RF
model.
1146 S. Wei et al.
Pn 2 P
Y Þ n1 ðYi Pi Þ2
1 ðYi
R2 ¼ Pn ð3Þ
1 ðYi Y Þ
2
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn 2
i¼1 ðPi Yi Þ
RMSE ¼ ð4Þ
n
Pn
jðPi Yi Þj
MAE ¼ i¼1 ð5Þ
n
Pn Pi Yi

1 Yi
MRE ¼ 100% ð6Þ
n
where, Pi is the predicted depth of the i-th checkpoint; Yi is the measured water depth at
the i-th checkpoint; Y is the average of the measured water depth; n is the number of
checkpoints. The larger the R2, the greater the correlation between the actual water
depth and the inversion water depth, and the better the inversion effect. The smaller the
RMSE is, the smaller the fluctuation of depth inversion error is and the better the
inversion effect is. The smaller MAE is, the smaller the absolute value of the error
between the measured water depth and the inversion water depth is, and the better the
inversion effect is. The smaller the MRE, the more credible the inversion is (Table 1).
Table 1. Inversion results of water depth inversion model

Model R2 RMSE MAE MRE
BP 0.95 1.01 0.73 8.47%
RF 0.97 0.851 0.60 3.54%
ELM 0.97 0.866 0.65 7.30%
From the table analysis, it can be seen that the decision coefficient R2 of BP neural
network, RF and ELM model is large on the whole, reaching above 0.95. However, the
R2 of BP neural network model is the smallest, and RMSE, MAE and MRE are the
largest among the three models. The R2 of the RF model is 0.97, the correlation
between measured water depth and inversion water depth is better, and the values of
RMSE, MAE and MRE are all smaller than those of the other two models, so the RF
model has a better effect on water depth inversion. Although R2 of ELM model is
slightly smaller than RF and RMSE and MAE are slightly larger than RF, there is a
large difference between MRE of the ELM and RF. The overall accuracy evaluation of
the RF model is higher than that of the other two bathymetry inversion models. By
comparing the inversion bathymetric and the measured bathymetric, the R2 of the
checkpoint is 0.97, the RMSE is 0.85, the MAE is 0.60, the MRE is 3.54%, the
regression fitting effect is better, and the model’s fitting degree is higher.
4 Conclusion
This paper uses WorldViewII remote sensing data and laser sounding data to establish
the shallow water depth inversion models of BP neural network, random forest, and
extreme learning machine, and horizontal comparison was made with the nonlinear
regression model used by predecessors. The results show that, on the whole, the
nonlinear regression model based on machine learning is ideal for water depth inver-
sion. For the machine learning model, the input layer, weight and function all have an
impact on the water depth inversion effect of the model. Therefore, it is necessary to
make repeated experiments to get the most suitable inversion model. For different water
depth ranges, the model with the best depth inversion effect is also different, which may
be because the factors affecting each water depth range are different. In the depth range
of less than 5 m, there are more suspended sediment and artificial wastes near the
shore, which affect the inversion effect. In the range of water depth greater than 15 m,
as the water depth gradually increases, the light penetration in the water becomes
weaker, thus affecting the inversion effect.
The conclusion of this manuscript uses a single water body type, the results have
certain limitations. there are still some shortcomings. When establishing the model
using machine learning, the number of hidden nodes is determined based on the
training error, so it is difficult to find the real optimal solution, which will affect the
inversion effect to some extent. In the future research, we should optimize the selection
and number of parameters, optimize the model, and improve the applicability of the
model.
References
1. Shao, S.: Research on shallow water depth extraction of the GF-satellite remote sensing.
Harbin Engineering University (2019)
2. Zhang, L., Mou, X.Y., Ji, H.L.: Research on water depth inversion in reservoir area based on
multi band remote sensing data. J. Hydraulic Eng. 49(5), 639–647 (2018)
3. Cao, B., Qiu, Z.G., Cao, B.C.: Comparison among four inverse algorithms of water depth.
J. Geomat. Sci. Technol. 4, 388–393 (2016)
4. Lyzenga, D.R.: Passive remote sensing techniques for mapping water depth and bottom
features. Appl. Opt. 17(3), 379 (1978)
5. Stumpf, R.P., Sinclair, M.: Determination of water depth with high-resolution satellite
imagery over variable bottom types. Limnol. Oceanogr. 48(1), 547–556 (2003)
6. Wang, J.J., Tian, Q.J.: Study on shallow coastal water depth derived method based on
hyperspectral remote sensing. Sci. Geogr. Sinica 6, 843–848 (2007)
7. Sandidge, J.C., Holyer, R.J.: Coastal Bathymetry from Hyper-spectral Observation of Water
Radiance. Remote Sens. Environ. 65(3), 341–352 (1998)
8. Anctil, F., Coulibaly, P.: Wavelet analysis of the interannual variability in southern Quabec
streamflow. J. Clim. 17(1), 163–173 (2004)
9. Deng, Z.D., Ye, X., Guan, H.J.: Remote sensing of water depth based on RBF neural
network. J. PLA Univ. Sci. Technol. 14(1), 101–106 (2017)
10. Zheng, G.Z., Le, X., Wang, H.P.: Inversion of water depth from WorldView-02 satellite
imagery based on BP and RBF neural network. Earth Sci. 42(12), 2345–2353 (2017)
1148 S. Wei et al.
11. Wang, Y.J., Zhang, Y.: Study on remote sensing of water depth based on bp artificial neural
networks. Ocean Eng. 4, 37–42 (2005)
12. Liu, Y.L.: A review of random forest. Nankai University, Tian Jin (2008)
13. Huang, G.B, Zhu, Q.Y., Siew, C.K.: Extreme learning machine: a new learning scheme of
feedforward neural networks (2005)
14. Eshtay, M., Faris, H., Obeid, N.: Improving extreme learning machine by competitive swarm
optimization and its application for medical diagnosis problems. Expert Syst. Appl. 104,
134–152 (2018)
15. Cao, W., Ming, Z., Wang, X.: Improved bidirectional extreme learning machine based on
enhanced random search. Memet. Comput. 11(5), 1–8 (2017)
16. Wu, Z.Q., Mao, Z.H., Wang, Z.: Research on remote sensing inversion of shallow water
depth based on extreme learning machine. Hydrogr. Surv. Chart. 39(3), 11–15 (2019)
17. Wang, X.C., Shi, F., Yu, L.: 43 Cases of MATLAB Neural Network Analysis. Beijing
University of Aeronautics and Astronautics Press, BeiJing (2013)
Container Terminal Workload Modeling Using
Machine Learning Techniques
Üstün Atak1(&) , Tolga Kaya2 , and Yasin Arslanoğlu1

1
Maritime Faculty, Istanbul Technical University, 34940 Istanbul, Turkey
{ustun.atak,arslanoglu}@itu.edu.tr
2
[email protected]
Abstract. Container terminals are complex facilities which serve to maritime

transportation. Effectivity and productivity of the terminals are very crucial for
international transportation. Often, terminal managers deal with logistic prob-
lems with high costs. Reducing operation cost and time are the key elements for
a smooth and secure process. To minimize the port stay time of the vessels, each
part of the operation should be optimized. In order to serve container vessels
efficiently, the port management should analyze the cargo handling process
considering the ship and container characteristics. In this scope, the purpose of
this study to model quay crane handling time in traditional container terminals
based on port operations data using machine learning techniques. To conduct
operational efficiency analysis, we have analyzed container terminal operations
using alternative regression models based on operations data of more than
400.000 handling movements in a traditional terminal in Turkey. Results sug-
gest that the efficiency of the terminal can be increased in planned rush hours
such as the periods before the mealtime and the vessels’ operation completion.
Keywords: Container terminal Operational efficiency analysis Regression

Machine learning Quay crane handling
1 Introduction
Maritime transportation is one of the important factors to the development of world trade.
Different plan of logistics can assure more economical and safer transportation. The
working behavior of the crew and the port personnel are key elements to accomplish safe
port to port transportation. Despite all the necessary actions are taken and the trainings are
carried out, accidents occur due to possible human errors. Failure to follow pre-defined
rules results in several ship-related accidents. These accidents can range from the slightest
minor work accident to large-scale marine contamination [1]. These accidents can also
cause fatal injuries and damage to the ship’s structures [2]. On the other hand, many
accidents not only affect the marine and land ecosystem, but also people’s habitats [3].
Continuous, fast and reliable port service is crucial for the sustainability of global
maritime transport and world trade. In this aspect, technological developments contribute
to the smooth operation of port services. These developments provide the increase of
transported cargo amount and the shortening the port stay time. One of the major
1150 Ü. Atak et al.
problems of human supervised systems is human error. The human error factor can affect
the container terminal operations such as time losses, failure of cargo operation, or worse
serious injury and accidents. To handle these challenges, one of the operational solutions
is to analyze container workload with machine learning techniques. Machine learning
methods are used to investigate complex and nonlinear systems with the data obtained
from the systems. These techniques allow making predictions and inferences by its ability
to analyze and copy the human intelligence. In this study, in order to optimize workload of
container terminal, Multiple Linear Regression (MLR), Support Vector Machine
regression (SVM), K-Nearest Neighbors (KNN), Regression Trees (RT), Random Forest
(RF) and boosting methods are used for analyzing operational data acquired from actual
movement of equipment and container in focused terminal. The aim of the study is to
analyze operational performance within the scope of workload efficiency for container
terminals using machine learning techniques.
The rest of the paper is organized as follows: in the second section a literature review
is given. The third section is devoted to the methodology of the study. In Sect. 4, the
case study for the container terminal is given. Finally, conclusions are discussed.
2 Literature
In the field of operational improvement such as optimization and accident/risk analysis,

machine learning applications are among the most useful successful methods applied in
the maritime domain [4]. The studies about operational improvement by using machine
learning methods have been studied more frequently for the last decade. For instance,
liner passenger vessel; the coordinate of the ship, speed, wind characteristics have ana-
lyzed such as how each variable affects the energy efficiency of ship propulsion system
[5]. Moreover, regression models in maritime are studied frequently in different areas
such as efficiency ship operation [6], estimation of freight demand [7], hinterland con-
nection [8], port cargo performance prediction [9] and security disaster planning [10].
On the other hand, machine learning techniques are essential for port management
and optimization of container terminals. A machine learning model of terminal tractors
is studied with the method called the Q-Learning technique and it has been improved
and compared to the previous terminal tractor assignments [11]. Tijjani and Bucak [12]
explored the best solution of container loading process and use the port area in the best
way by using Q-Learning method, as well. As a result of the study, port stay times of
container ships are reduced by optimization of the results. The study conducted by
Kozan and Preston [13] aims to organize the container transfers using Genetic Algo-
rithms. It is confirmed that the port served the ships in a shorter time with the applied
method. The study conducted in Singapore Port about customer feedbacks shows that a
prototype system with a machine learning application could fulfill to identify the
customer loyalty factors [14]. Similarly, workload of terminal which could be the root
cause of any risk related accidents is another topic of maritime domain in the scope of
container terminal management. To give an example study is about the capacity of
container ships which can sail with an enormous amount of cargoes and is concluded
with the impact of motherships to berth and quay crane workload [15]. A hierarchical
scheme for yard crane workload management study proposes that a combination of
Container Terminal Workload Modeling 1151
simulation and optimization to minimize waiting time for assigned vehicle jobs [16].
The study about stowage planning of large containership proves that trade-off between
ship stability and crane intensity could be useful to improve work order quality and
computational efficiency in the terms of workload of container terminal [17]. A solution
for space organizing problems at container terminals is proposed with simulation
studies in the scope of workloads and scheduling methods of quay cranes [18].
3 Methodology
The first method of study, MLR creates a linear equation to model data. In other words,
MLR is a statistical method that analyses explanatory variables to predict the response
variable’s outcome. The strength sides of the method are (1) the most common
approach for numeric data modeling, (2) easy implementation of any data (3) gives a
solution about estimation which is the strength and size of the relationships. The
weaknesses of the method are following: (1) too strong data assumptions, (2) pre-form
the model by a user, (3) missing data gap, (4) extra processing of categorical data,
(5) knowledge of statistics to solve the model [19]. The notation of method is:
Y ¼ b0 þ b1 X1 þ b2 X2 þ . . . þ bp Xp þ 2 ð1Þ
The second method, SVM regression uses almost the same principle as SVM
classification as a nonparametric method. The main idea is to fit error within a pre-
defined threshold limit. The strength sides of the method are (1) accuracy, (2) better
work on small datasets, (3) more efficient with the help of a subset training points. The
weak sides of the methods are (1) computational time for large datasets and (2) less
effective on noisier datasets. The mathematical formula of SVM as follows [20]:
1 T Xn
K Ai ; AT u þ b yi
f ðxÞ ¼ min u u þ b2 þ C ð2Þ
ðu;bÞ2Rn þ 1 2 i¼1
The other method is KNN. The KNN could be used for both classification and
regression problems likewise SVM algorithm. KNN is used in estimation and recog-
nition as a non-parametric method. The strength sides of the method are (1) to interpret
the prediction easily; (2) non-parametric, (3) faster in compare with other machine
learning algorithms. The weaknesses of the method are (1) high computational needs,
(2) high memory requirements, (3) costly prediction stage, (4) sensitivity for outliers.
The mathematical formula;
X
^f ðx0 Þ ¼ 1 y_ ð3Þ
K XI_ 2N0 I
In addition, tree-based methods for regression such as RT, RF and boosting could
be used for both regression and classification problems. These methods are used in
order to compare results of MLR, SVM and KNN. Further details regarding tree-based
methods could be found at different books in the literature [21–23].
4 Case Study – Container Terminal
In this section, the non-automated container terminal operation data in Turkey is

analyzed by six different machine learning methods as described before. The container
handling time is dependent variable of the model with respect to independent variables
as vessel name, order, length, crane, container weight and container’s full or empty
status. The container terminal served 294 vessels from feeder to mothership with four
gantry cranes in five years’ time period. Total container box movement number which
is the base data for our study is 424.501. There are 7 independent variables about
container movement and operational data. The dataset has total 424.501 rows and 22
rows in the beginning. The regression data remain 162.668 rows after evaluating of
missing data to create suitable set. At the same time, container weight and crane
assignments are multiplied with each other to create a new subset. The mentioned data
set further evaluated for over and misleading waiting period times. The observation is
found 94.990 rows for regression. Randomly selected 80% of data is used as a training
set and the remaining 20% is used for the test set.
In this scope, data analyses are focused on operational benchmarking. The process
of benchmarking is to calculate fastest handling time for container vessels. First, all
mentioned machine learning methods are used to model handling time and calculate the
lowest Root Mean Square Error (RMSE). The maritime domain assumption for con-
tainer handling crane movement capacity is calculated in order to reach the best
operation data for sample container terminal [24]. The optimal movement capacity for
container handling which is 15–25 lifts/h is used as a base for container handling time
calculation. In the beginning, all data derived from terminal is directly used for data
analyzing after removing missing points or outliers. Later on, all models are evaluated
in the scope of RMSE metrics. The best method which has the lowest RMSE is used for
further evaluation of container terminal workload analyzes.
5 Findings
“R programming language/R Studio” with several packages such as “tidyverse, caret,

pls, MASS, randomForest, rpart, e1071 and Metrics” are used with Intel i7 6700
processor on “MS Win 10 Pro” environment. The data set is analyzed with MLR,
SVM, KNN, RT, RF and boosting methods. The models are compared with respect to
RMSE. Results can be seen in Table 1.
MLR model has RMSE as 112, SVM, RT, boosting regression have almost the
same RMSE with MLR. Similarly, RF regression generates a very close result as 120.
On the other hand, KNN regression has the worst performance. When we apply the
maritime domain assumption that optimal movement capacity for container handling
15–25 lifts/h, the RMSE for MLR is 26. So, we take the base RMSE as maritime
domain assumption to analyze container terminal workload with respect to regression
methods.
Table 1. Comparison of RMSE for mL. methods.

Method MLR SVM KNN RF RT Boosting
RMSE 112 113 4998 120 113 114
Furthermore, MLR model is evaluated in more detail to reach the relevant hours of
terminal operation workload. The findings show that just before the meal time which is
12:00 has the fastest on container handling times for five years period compared with
other times in a day as in Table 2.
Table 2. Coefficients of container handling time.

Coefficient Estimate Std. Error t value Pr (>|t|)
Intercept 206.1 113.6 1.837 0.069
Container weight −0.001 0.001 −11.15 0.001***
Crane #1 −8.505 2.659 −3.199 0.001 **
Tallyman #4 −7.420 2.814 −2.637 0.008 **
Tallyman #6 −12.8 3.125 −4.096 0.001***
Tallyman #7 −15.43 3.147 −4.902 0.001***
Tallyman #15 −22.1 2.995 −7.379 0.001***
Tallyman #18 −16.14 3.393 −4.757 0.001***
Tallyman #19 −10.82 2.906 −3.722 0.001***
Tallyman #20 −8.229 3.071 −2.679 0.007**
Tallyman #21 −55.11 15.5 −3.555 0.001***
Tallyman #25 −8.528 3.041 −2.805 0.005**
Empty container −23.22 2.391 −9.714 0.001***
Hour – 00:00 8.118 2.453 3.310 0.001***
Hour – 08:00 11.62 2.466 4.713 0.001***
Hour – 12:00 −8.274 2.381 −3.475 0.001***
As seen on Table 2 which indicates the statistically significant coefficients (indi-

cated with three asteriks), working hour 12:00 has a favorable effect on container
handling time. These results show that when container terminal working at the men-
tioned time period, container handling time could be observed as minimum or optimum
compared to other times. To evaluate these results in detail, port watch times and meal
times are obtained from the terminal management office. The mealtime for terminal
workers stated as 12:00 and workers give more effort to finish all operations before
mealtime in order not to postpone vessel departure time one more hour. On the other
hand, shift times which are 8–16, 16–00, 00–08, the results show that there are lots of
time losses at 00:00 and 08:00 h in the view of the regression coefficient of terminal
operation data. This proves that when a worker reaches at the end of his/her shift,
he/she tends to work less efficiently.
6 Conclusion
The container vessels can carry almost 20.000 boxes with different types of cargoes.
From motherships to feeder vessels they need special and efficient cargo handling
procedures. Stakeholders are forced to carry containers quickest and cheapest way. To
complete voyages, vessel operators are instructed to sail at optimum speed depends on
weather and sea conditions. Since there are several factors such as route planning,
weather routing, stowage planning, etc. to determine voyages at sea in the view of
voyage optimization, there should be another option to optimize container terminal
operations with the help of operational data analyzes.
In this study, the non-automated container terminal is analyzed with six different
machine learning techniques (1) MLR, (2) SVM, (3) KNN, (4) RT, (5) RF and
(6) boosting methods. Dependent variable that container handling time is found by
independent suitable terminal operational data. Within the mentioned methods, MLR is
found the best fit model in the view of RMSE. On the other hand, the model of study is
related to the maritime domain assumption of crane handling abilities. The proposed
model shows that to ensure safe and efficient container terminal operation, workers can
work more efficiently in point of movement figures and handling times.
In the future studies, container terminal models could be enriched so as to use more
operational data. Moreover, different machine learning methods and ensemble tech-
niques could be used to improve the performances of container terminal operational
workload models.
References
1. Lecue, M., Darbra, R.M.: Accidents in European ports involving chemical substances:
characteristics and trends. Saf. Sci. 115, 278–284 (2019)
2. Cao, X., Lam, J.S.L.: A fast reaction-based port vulnerability assessment: Case of Tianjin
Port explosion. Transp. Res. Part A: Policy and Pract. 128, 11–33 (2019)
3. Helle, I., Jolma, A., Venesjarvi, R.: Species and habitats in danger: estimating the relative
risk posed by oil spills in the northern Baltic Sea. Ecosphere 7(5), e01344 (2016)
4. Tirunagari, S.: Data mining of causal relations from text: analysing maritime accident
investigation reports. arXiv preprint arXiv:1507.02447 (2015)
5. Petersen, J.P., Jacobsen, D.J., Winther, O.: Statistical modelling for ship propulsion
efficiency. J. Mar. Sci. Technol. 17, 30–39 (2012)
6. Uyanik, T., Arslanoglu, Y., Kalenderli, O.: Ship fuel consumption prediction with machine
learning. In: 4th International Mediterranean Science and Engineering Congress, pp. 1–7
(2019)
7. Patil, G.R., Sahu, P.K.: Estimation of freight demand at Mumbai port using regression and
time series models. KSCE J. Civ. Eng. 20(5), 2022–2032 (2016)
8. Feng, L.: Regression analysis on the hinterland connection of peripheral ports in the China
Bohai Sea regions. Int. J. Adv. Manage. Sci. 2, 159–162 (2013)
9. Jayaprakash, P.O., Gunasekaran, K., Muralidharan, S.: Comparison of artificial neural
network models and multiple linear regression models in cargo port performance prediction.
Adv. Mater. Res. 403–408, 3570–3577 (2012)
10. Paul, J.A., Maloni, M.J.: Modeling the effects of port disasters. Maritime Econ. Logist. 12,
127–146 (2010)
11. Fotuhi, F., Huynh, N., Vidal, J.M., Xie, Y.: Modeling yard crane operators as reinforcement
learning agents. Res. Transp. Econ. 42, 3–12 (2013)
12. Tijjani, S., Bucak, I.O.: An approach for maximizing container loading and minimizing the
waste of space using Q-learning. In: 2013 The International Conference on Technological
Advances in Electrical, Electronics and Computer Engineering, TAEECE 2013, pp. 235–238
(2013)
13. Kozan, E., Preston, P.: Genetic algorithms to schedule container transfers at multimodal
terminals. Int. Trans. Oper. Res. 6, 311–329 (1999)
14. Ng, K.S., Liu, H., Kwah, H.B.: A data mining application: customer retention at the Port of
Singapore Authority (PSA). SIGMOD Rec. (ACM Spec. Interest Group Manage. Data) 27,
522–525 (1998)
15. Martin, J., Martin, S., Pettit, S.: Container ship size and the implications on port call
workload. Int. J. Shipp. Transp. Logist. 7, 553–569 (2015)
16. Guo, X., Huang, S.Y.: Dynamic space and time partitioning for yard crane workload
management in container terminals. Transp. Sci. 46, 134–148 (2012)
17. Fan, L., Low, M.Y.H., Ying, H.S., Jing, H.W., Min, Z., Aye, W.C.: Stowage planning of
large containership with tradeoff between crane workload balance and ship stability. World
Congress on Engineering 4–6 July 2012, London, UK, vol. 2182, pp. 1537–1543 (2010)
18. Jeong, Y.-H., Kim, K.-H., Woo, Y.-J., Seo, B.-H.: A simulation study on a workload-based
operation planning method in container terminals. Ind. Eng. Manage. Syst. 11, 103–113
(2012)
19. Lantz, B.: Machine learning with R (2013)
20. Yang, Y., Fuli, R., Huiyou, C., Zhijiao, X.: SVR mathematical model and methods for sale
prediction. J. Syst. Eng. Electron. 18, 769–773 (2007)
21. James, G., Witten, D., Hastie, T., Tibshirani, R.: An Introduction to Statistical Learning.
Springer, New York (2013)
22. Raschka, S., Mirjalili, V.: Python Machine Learning: Machine Learning and Deep Learning
with Python, Scikit-Learn, and TensorFlow 2. Packt Publishing Ltd., Birmingham (2019)
23. Lantz, B.: Machine learning with R: expert techniques for predictive modeling (2019)
24. Thoresen, C.A.: Port Designer’s Handbook, 3rd edn. Thomas Telford Ltd., London (2014)
Usage of Machine Learning Algorithms
for Flow Based Anomaly Detection System
in Software Defined Networks
Muhammet Fatih Akbaş1(&), Cengiz Güngör2, and Enis Karaarslan3

1
Information Technologies Department, İzmir Kâtip Çelebi University,
İzmir, Turkey
[email protected]
2
Computer Engineering Department, Tokat Gaziosmanpaşa University,
Tokat, Turkey
[email protected]
3
Computer Engineering Department, Muğla Sıtkı Koçman University,
Muğla, Turkey
[email protected]
Abstract. Computer networks are becoming more complex in the number of

connected nodes and the amount of traffic. The growing number and increasing
complexity of cyber-attacks makes network management and security a chal-
lenge. Software defined networks (SDN) technology is a solution that aims for
efficient and flexible network management. The SDN controller(s) plays an
important role in detecting and preventing cyber-attacks. In this study, a flow-
based anomaly detection system running on the POX controller is designed.
A comparative analysis of the supervised machine algorithms is given to choose
the optimum anomaly detection method in SDN based networks. NSL-KDD
dataset is used for training and testing of the classifiers. The results show that
machine learning algorithms have great potential in the success of flow-based
anomaly detection systems in the SDN infrastructure.
Keywords: Software Defined Networks Flow-based anomaly detection

system Machine learning
1 Introduction
Software Defined Networks (SDN) allows dynamic and flexible management of

complex and large scale computer networks by separating switching and filtering
processes. The ability to program the network traffic flows; is based on dividing the
control and data planes. The control component is moved to a centralized controller.
High-level network policies can be created on the controller and these policies are
converted into flow rules and inserted into the flow table of the switches. Switches are
only responsible for the transmission of packets according to the flow rules created [1].
The controller has control over the topology of the whole network and can respond
quickly to the anomalies in the network traffic. This characteristic of SDN enables the
development of more effective network security solutions. SDN infrastructure is also
Usage of Machine Learning Algorithms 1157
open to security threats which are described in [2, 3]. Artificial intelligence (AI) based
network management approach and SDN security is also possible [4].
Anomaly/intrusion detection systems used to detect threats in traditional network
infrastructures are generally located on the network perimeter; internet service provi-
ders, backbone devices, or firewall devices. This can be extended to all endpoints in
SDN [5]. There are several anomaly detection system approaches such as packet-based
or flow-based. Packet-based anomaly detection systems have some performance dis-
advantages as they look at the payload information of the packet. Faster anomaly
detection systems can be developed with flow-based systems.
An anomaly detection system can be added to the SDN controller to make the SDN
infrastructure more secure against the network attacks. In this study, a flow-based
anomaly detection system running on the POX controller is designed. This study
focuses on the supervised machine learning algorithms such as Support Vector
Machine (SVM), K-Nearest Neighborhood (K-NN), and Decision Trees. A compara-
tive analysis of these algorithms is given. The attacks; Denial of Service (DoS) and
Distributed Denial of Service (DDoS) which have a disruptive effect on SDN infras-
tructure is focused on. Sub-datasets are extracted from the NSL-KDD [6] dataset for
each attack category.
In the second section of this study, related works are given. In the third section, the
design of the flow-based anomaly detection system is described, the NSL-KDD dataset
and performance evaluation metrics are explained. In the fourth section, the perfor-
mance of the selected machine learning algorithms is analyzed.
2 Related Works
Studies on SDN security are mainly focused on intrusion/anomaly detection. Detecting

and preventing DDoS attacks are one of the challenges of network security. DDoS
attacks against the controller can cause the entire network to crash. Machine learning
techniques are analyzed to protect the controller and switches in the SDN infrastructure
against intrusion and DDoS attacks in [7]. Self-organizing maps (SOM) were used for
the detection of DDoS attacks in [8]. In this study, switches in the network are mon-
itored periodically over the NOX controller. Flows are constantly collected from all
switches and flow-based features related to DDoS attacks are extracted. SOM is trained
with using these features and the traffic is classified as normal or anomaly. In [9], the
SVM classifier was used to detect DDoS attacks in the SDN environment. In [10], a
metaheuristic method called the binary bat algorithm is used for feature extraction and
entropy-based classifier for traffic classification. In [11], a network-based intrusion
detection and prevention system that depends on advantages of the SDN approach was
proposed. C4.5 algorithm is used for traffic classification. This approach has the
potential to reduce the cost and decrease the latency of the system. In [12], a deep
learning based intrusion detection system is proposed that uses the NSL-KDD dataset.
Different learning rates have been tested to optimize the model. In [13], curve fitting,
pattern recognition and time series methods are used as an artificial neural network
model for SDN. The trained model gave the best accuracy rate with the pattern
recognition method. In [14], an intrusion detection system was developed using a fuzzy
1158 M. F. Akbaş et al.
approach. This approach has a high detection rate and low false alarm rate. In [15], an
intelligent intrusion detection system that uses machine learning algorithms is proposed
for 5G networks. Random forests method is used for selecting a characteristic subset of
traffic features. Classification of the flows is made by combining K-Means++ and
AdaBoost algorithms.
3 Flow Based Anomaly Detection System for SDN
Intrusion/Anomaly detection systems generally use a deep packet inspection

(DPI) approach to detect anomalies in network traffic. Implementing the packet-based
approach in encrypted network traffic brings some difficulties and restrictions. Espe-
cially, inspecting the packet payload information is computationally costly and this
causes a certain delay in high-speed networks. Flow-based detection systems are a new
approach which is preferred to detect anomalies in high-speed networks. Flow-based
detection systems only look at the packet header information. In this way, more effi-
cient anomaly detection systems can be developed in terms of performance. Figure 1
shows the design and anomaly detection process of the flow-based anomaly detection
system running on the POX controller for SDN based networks. The flow collector
module periodically collects flow records from the flow tables of switches. The features
are extracted from the collected flows in the feature extractor module which is used for
anomaly detection and then these features are sent to the classifier. The classifier in the
anomaly detection module decides whether the traffic is an anomaly by using the six
features indicated with bold font in Table 1. These features are described in Table 3. If
an anomaly is detected, an alarm is sent to the cyber security administrator and a flow
rule is inserted into the flow table of the switches to block attack traffic. If no anomaly
is detected, the flow collection process continues. OpenFlow protocol is used while
collecting flow statistics and response against malicious flows.
In this study, the NSL-KDD dataset was used which is an improved version of the
KDD CUP 99 dataset. The NSL-KDD dataset has 42 features, one of which is a class
label and many redundant records are removed from the KDD CUP 99 dataset. These
features are grouped under four categories as shown in Table 1. Attacks in training and
test datasets are classified under four main categories as DoS, network scanning (Probe),
gaining unauthorized access (Remote to Local, R2L) and obtaining root privileges (User
to Root, U2R). Attacks indicated with bold font in Table 2 are not in the training dataset
but only in the test dataset. We can consider these attacks as zero-day attacks.
There are 125973 records in the NSL-KDD training dataset and 22544 records in
the test dataset. When the distribution of the records is analyzed, approximately 99% of
the records in the training dataset are normal (53%), DoS (36%) and Probe (9%) traffic
and approximately 86% of the records in the test dataset are normal (43%), DoS (33%)
and Probe (10%) traffic. The R2L and U2R attack categories have a very low per-
centage. The confusion matrix is used to measure the performance of machine learning
classification models. The table shows the number of the actual values in the dataset
and the number of true and false predictions of the classification model.
Performance evaluation metrics such as accuracy, precision, recall, F1-score are
measured using the values in this table. The abbreviations used in the equations are;
True Positive (TP), True Negative (TN), False Positive (FP) and False Negative (FN).
Fig. 1. Design of flow-based anomaly detection system and anomaly detection process
Accuracy: It is the ratio of the number of correct predictions to the number of all
predictions. It shows the overall performance of the model.
Accuracy ¼ ðTP þ TNÞ=ðTP þ TN þ FP þ FNÞ ð1Þ
Precision: It is the ratio of the number of correctly classified attack records to the total
number of records predicted as attack.
Precision ¼ TP=ðTP þ FPÞ ð2Þ
Recall: It is the ratio of the number of correctly classified attack records to the total
number of attack records.
Recall ¼ TP=ðTP þ FNÞ ð3Þ
F1-Score: It is the harmonic average of precision and recall performance evaluation

metrics.
F1Score ¼ ð2 Precision RecallÞ=ðPrecision þ RecallÞ ð4Þ

Table 1. Categories of features in the NSL-KDD

Category Features
Basic features F1, F2, F3, F4, F5, F6, F7, F8, F9, F10
Content based features F11, F12, F13, F14, F15, F16, F17, F18, F19, F20, F21, F22
Time based features F23, F24, F25, F26, F27, F28, F29, F30, F31
Host based features F32, F33, F34, F35, F36, F37, F38, F39, F40, F41
Table 2. Categories of attacks

Category Training Test
DoS back, land, neptune, pod, smurf, apache2, back, land, mailbomb, neptune,
attacks teardrop pod, processtable, smurf, teardrop,
udpstorm, worm
Probe ipsweep, nmap, portsweep, satan ipsweep, mscan, nmap, portsweep, saint,
attacks satan
R2L ftp_write, guess_passwd, imap, ftp_write, guess_passwd, httptunnel, imap,
attacks multihop, phf, spy, warezclient, multihop, named, phf, sendmail,
warezmaster snmpgetattack, snmpguess, spy,
warezclient, warezmaster, xlock, xsnoop
U2R buffer_overflow, loadmodule, perl, buffer_overflow, loadmodule, perl, ps,
attacks rootkit rootkit, sqlattack, xterm
Table 3. Selected features

Features Description
duration Length of time duration of the connection
protocol_type Protocol used in the connection
src_bytes Number of data bytes transferred from source to destination in single
connection
dst_bytes Number of data bytes transferred from destination to source in single
connection
count Number of connections to the same destination host as the current
connection in the past two seconds
srv_count Number of connections to the same service (port number) as the current
connection in the past two seconds
4 Comparative Analysis of Machine Learning Algorithms
In this study, sub-datasets were extracted from the NSL-KDD dataset for each attack
category. The binary classification (normal or anomaly) was made for each sub-dataset.
SVM, K-NN and Decision Trees are used as supervised machine learning algorithms for
classification. Table 4, Table 5 and Table 6 shows the obtained results for each machine
learning algorithm. Decision Tree algorithm shows the best performance with 91.206%
in the detection of the DoS attacks and the K-NN algorithm shows the best performance
with 90.191% in the detection of the probe attacks. Successfully results were obtained
because the characteristics of DoS and probe attacks are related to the selected flow-
based features. The accurate results could not be obtained for the R2L and U2R sub-
datasets. Because there were almost no records predicted as attack. The main reason for
this is that due to the very low number of records belong to R2L and U2R attack classes
and the emergence of sub-datasets with the unbalanced class distribution. Also, the
characteristics of R2L and U2R attacks are mostly related to content based features and
their extraction is very difficult and complex computationally. Within the scope of the
study, relevant features for the nature of SDN were chosen to develop real-time
applications in the SDN environment. These features can be extracted through the POX
controller using the flow data in the flow table of the switches.
Table 4. Obtained results with SVM algorithm

Metrics Sub-datasets extracted from
NSL-KDD dataset
DoS vs. Normal Probe vs. Normal
Test Training Test Training
Accuracy 0.85725 0.96998 0.87883 0.94229
Precision 0.87547 0.97044 0.87500 0.94028
Recall 0.85725 0.96998 0.87883 0.94229
F1-score 0.85273 0.96986 0.86575 0.93944
Table 5. Obtained results with K-NN algorithm

NSL-KDD dataset
Accuracy 0.88661 0.99666 0.90191 0.97515
Precision 0.89848 0.99666 0.89810 0.97854
Recall 0.88661 0.99666 0.90191 0.97515
F1-score 0.88411 0.99666 0.89666 0.97590
Table 6. Obtained results with decision tree algorithm

NSL-KDD dataset
Accuracy 0.91206 0.99493 0.89581 0.96234
Precision 0.91890 0.99493 0.89592 0.96373
Recall 0.91206 0.99493 0.89581 0.96234
F1-score 0.91078 0.99493 0.88532 0.96005
5 Conclusion
The results show that machine learning algorithms have a great potential in the success
of the flow-based anomaly detection systems in SDN infrastructure. A comparative
analysis of the supervised machine learning algorithms for flow-based anomaly
detection is given in this study. Six features are selected from the flow table of the
switches for anomaly detection. The Decision Tree algorithm gives the highest accu-
racy rate with 91.206% in the detection of DoS attacks and the K-NN algorithm shows
the highest accuracy rate with 90.191% in the detection of the probe attacks. Future
work will include more detailed studies on the flow-based features and ensemble
learning methods. A solution which combines clustering and classification algorithms
will be studied for achieving better results.
References
1. Kim, H., Feamster, N.: Improving network management with software defined networking.
IEEE Commun. Mag. 51(2), 114–119 (2013)
2. Kreutz, D., Ramos, F.M.V., Verissimo, P.: Towards secure and dependable software-defined
networks. In: Second ACM SIGCOMM Workshop on Hot Topics in Software Defined
Networking, pp. 55–60. Association for Computing Machinery, Hong Kong (2013)
3. Akbaş, M.F., Karaarslan, E., Güngör, C.: A preliminary survey on the security of software-
defined networks. Int. J. Appl. Math. Electron. Comput. 4(Special Issue-1), 184–189 (2016)
4. Akbaş, M.F., Karaarslan, E., Güngör, C.: Yazılım Tanımlı Ağların Güvenliğinde Yapay
Zekâ Tabanlı Çözümler: Ön İnceleme [Artificial intelligence based solutions in the security
of software defined networks: preliminary review]. In: International Artificial Intelligence
and Data Processing Symposium, Malatya, Turkey, pp. 496–501 (2016)
5. Mehdi, S.A., Khalid, J., Khayam, S.A.: Revisiting traffic anomaly detection using software
defined networking. In: Sommer, R., Balzarotti, D., Maier, G. (eds.) Recent Advances in
Intrusion Detection 2011. LNCS, vol. 6961, pp. 161–180. Springer, Heidelberg (2011)
6. Tavallaee, M., Bagheri, E., Lu, W., Ghorbani, A.A.: A detailed analysis of the KDD CUP 99
data set. In: 2009 IEEE Symposium on Computational Intelligence for Security and Defense
Applications, pp. 1–6, IEEE, Ottawa (2009)
7. Ashraf, J., Latif, S.: Handling intrusion and DDoS attacks in software defined networks
using machine learning techniques. In: 2014 National Software Engineering Conference,
pp. 55–60. IEEE, Rawalpindi (2014)
8. Braga, R., Mota, E., Passito, A.: Lightweight DDoS flooding attack detection using
NOX/OpenFlow. In: IEEE Local Computer Network Conference, pp. 408–415. IEEE,
Denver (2010)
9. Kokila, R.T., Selvi, S.T., Govindarajan, K.: DDoS detection and analysis in SDN-based
environment using support vector machine classifier. In: 2014 Sixth International
Conference on Advanced Computing, pp. 205–210. IEEE, Chennai (2014)
10. Sathya, R., Thangarajan, R.: Efficient anomaly detection and mitigation in software defined
networking environment. In: 2015 2nd International Conference on Electronics and
Communication Systems, pp. 479–484. IEEE, Coimbatore (2015)
11. Le, A., Dinh, P., Le, H., Tran, N.C.: Flexible network-based intrusion detection and
prevention system on software-defined networks. In: 2015 International Conference on
Advanced Computing and Applications, pp. 106–111. IEEE, Ho Chi Minh City (2015)
12. Tang, T.A., Mhamdi, L., McLernon, D., Zaidi, S.A.R., Ghogho, M.: Deep learning approach
for network intrusion detection in software defined networking. In: 2016 International
Conference on Wireless Networks and Mobile Communications, pp. 258–263. IEEE, Fez
(2016)
13. Abubakar, A., Pranggono, B.: Machine learning based intrusion detection system for
software defined networks. In: 2017 Seventh International Conference on Emerging Security
Technologies, pp. 138–143. IEEE, Canterbury (2017)
14. Shalini, S., Shafreen, N.A., Sathiya, P.L., Vetriselvi, V.: Intrusion detection system for
software-defined networks using fuzzy system. In: Mandal, J., Saha, G., Kandar, D., Maji, A.
(eds.) Computing and Communication Systems. LNNS, vol. 24, pp. 603–620. Springer,
Singapore (2018)
15. Li, J., Zhao, Z., Li, R.: Machine learning-based IDS for software-defined 5G network. IET
Netw. 7(2), 53–60 (2018)
Speech Analysis with Deep Learning
to Determine Speech Therapy for Learning
Difficulties
Nogol Memari1(&), Saranaz Abdollahi2, Sonia Khodabakhsh3,

Saeideh Rezaei4, and Mehrdad Moghbel5
1
Binary University of Management and Entrepreneurship, Puchong, Selangor,
Malaysia
[email protected]
2
Vali-e-Asr University of Rafsanjan, Rafsanjan, Iran
3
Universiti Tunku Abdul Rahman (UTAR), Kampar, Perak, Malaysia
4
UCSI University, Kuala Lumpur, Malaysia
5
Universiti Teknologi Malaysia, Kuala Lumpur, Malaysia
Abstract. The vocal tract movements involved in human speech makes the
vocalisation of a complex array of coordinated and meaningful acoustic utter-
ances possible. At the same time, it is hypothesized that related cognitive dis-
orders can potentially interfere with the neurological, pre-articulatory and fine
motor controls required for these fine movements. By leveraging the cognitive
complexity of speech production, it is possible to detect a range of different
disorders. Computer screening systems can be considered as an efficient
approach for the early diagnosis and screening of voice disorders. For achieving
the highest detection rate possible, a hybrid machine learning-based approach is
proposed by combining Deep Learning with AdaBoost classifier. First, a set of
acoustic features will be extracted using traditional features associated with the
presence of autism, such as fundamental frequency descriptors. Then, a deep
learning framework will be utilized for extracting additional acoustic contextual
descriptors not definable using traditional feature extraction methods. Finally,
the most informative features will be selected using a minimal-redundancy
maximal-relevance feature selection approach with an AdaBoost classifier
analysing all the selected features and informing the operator regarding the
patient’s condition.
Keywords: Speech analysis Autism detection Deep learning Learning

difficulties Social communication disorder Cognitive speech disorder
1 Introduction
The era of “big data” has been associated with a substantial rise in the amount of data
produced and stored which in turn has made it possible to leverage these large datasets
in developing machine learning-based assistive tools [1]. One such field is the research
regarding the underlying cognitive functions involved in human speech where the
complexity of the speech production makes the detection of a range of different
Speech Analysis with Deep Learning to Determine Speech Therapy 1165
pathologies such as Autism spectrum disorder (ASD) and Dysphonia. Considered as a

neuro-developmental disorder with a prevalence rate of 1 in 45, ASD is characterised
by deficits in social communication and social interaction, in addition to restricted and
repetitive patterns of behaviour [2]. Thought to be caused by unhealthy abuse of vocal
cords, Dysphonia detrimentally affects the voice quality, pitch and loudness with
approximately 10% of the population suffering from its complications [3]. The cor-
relation between voice pathologies and acoustic features extracted from the speech
patterns and voice has been clinically established and used for detecting various dis-
orders [4–6].
The rest of the paper is organised as follows: Sect. 2 presents a review of speech
pathology detection methods. Section 3 describes the methodology adopted in this
manuscript for comparing different features along with an introduction to the features
used in this study and Sect. 4 draws the conclusion.
2 Related Works
The acoustic signals produced by humans carry a large amount of information such as
their approximate age, gender, emotional and cognitive states. In general, it is
hypothesised that many disorders that can affect the cognitive process of humans often
interrupt the planning, neurological and pre-articulatory functions of the brain which
are required for carrying the fine vocal tract movements during speech (see Fig. 1).
However, due to the lack of trained professionals and the lack of public knowledge
regarding the matter, many individuals affected with speech disorders do not seek
diagnosis and treatment. Utilising the speech signals in combination with computa-
tional paralinguistic and machine learning, different cognitive disorders such as ASD
can be detected with relatively high accuracy.
Fig. 1. An illustration of the anatomical structures and muscles utilized during speech [7].
1166 N. Memari et al.
The early diagnosis of ASD is essential for maximising the effectiveness of therapy
along with mitigating the patient’s stress while interacting with other people. Features
computed from the patient’s acoustic signal such as articulation, loudness, pitch, and
rhythm have consistently shown encouraging results in computerised detection of ASD
[8–10]. To promote early diagnosis of ASD, a challenge was proposed in 2013 known
as the Autism Sub-Challenge to classify the occurrence of autism in children [11]. The
Child Pathological Speech Database (CPSD) was developed to provide speech
recordings of children who were either: (i) Typically Developing; (ii) diagnosed as
having a Pervasive Developmental Disorders (PDD) such as Autism; (iii) diagnosed as
having a Pervasive Developmental Disorders Non-Otherwise Specified (NOS); or,
(iv) diagnosed as having a specific language impairment such as Dysphasia (DYS). For
the first time in an autism detection challenge, a participating team proposed using a
Deep Neural Network (DNN) [12]. Their proposed system consisted of two hidden
layers of stacked Restricted Boltzmann Machines (RBMs). As the system was used as
part of an ensemble of classifiers, their proposed system was able to achieve an
Unweighted Average Recall (UAR) for the DNN with 92.2% which was better than
human classification. The challenge was won by a knowledge-driven system leveraged
traditional machine learning method based on traditional features, including Funda-
mental Frequency (F0), Harmonic to-Noise Ratio (HNR), shimmer, and jitter [13].
Their proposed SVM classifier achieved UAR of 93.6%. Huang and Hori [14] also
proposed using DNN coupled with feature normalisation and dimensionality reduction
approaches, including Principal Component Analysis (PCA) and Linear Discriminant
Analysis (LDA). Their proposed approach was able to achieve a UAR of 92.9%.
Nakai et al. [15] proposed using single-word utterances for identifying children
with ASD and compared their results with the clinical judgment of speech therapists.
Their dataset was comprised of 30 children with ASD and 51 children with typical
development with all the participants being between the ages of three and ten years old
with no comorbid disorders. A set of 24 features were computed for each single-word
utterance and used by an SVM classifier for identifying the presence of ASD. Their
proposed SVM based method showed better accuracy compared to the accuracy
achieved by ten speech therapists that used the same single-word utterances for
identifying patients with ASD. Their proposed method achieved a sensitivity and
specificity of 81% and 73%, respectively, whereas the average sensitivity and speci-
ficity of the therapists was 54% and 80%, respectively.
3 Methodology
Identifying the behavioural and biological features which can be used for early
detection of ASD has been gaining attention as computerised detection systems have
shown to provide comparable accuracy to human classification. In general, ASD often
manifests as linguistic peculiarities such as echolalia, out of context phrasing,
Abnormal prosody as well as pronoun and role reversal which are detectable in speech
patterns [16, 17]. Besides, acoustic features derived from the speech pattern such as
articulation, loudness, pitch, and rhythm have also proven beneficial in early detection
of ASD [18] along with speech-based interaction and therapy systems for improving
the social skills of children suffering from ASD [19].
For the construction of the dataset, voice samples from the ‘‘Saarbruecken Voice
Database’’ (SVD) [20] that contains 2041 recordings (627 male and 727 female) from
healthy and un-healthy individuals provided by the Institute of Phonetics of the
University of Saarland are used with the recordings being sampled at 50 kHz and 16-bit
resolution classified into 71 classes including functional and organic disorders. The
dataset consists of recordings of vowels /a/, /i/, /u/ coupled with an appropriate sentence.
Preferred by the majority of voice assessment systems, the use of vowels for evaluating
the patient’s voice avoids the common linguistic artefacts [3] with the vocalisation of the
vowel /a/ being the most commonly used in practice.
This study will focus on constructing a novel computer-assisted diagnosis system
(CAD) for assisting with the screening of the population for signs of ASD combining
traditional machine learning with Deep Learning. In addition to classification, DNN
can be used as a means of learning intrinsic representations from complex speech data
where the resulting DNN representations are input to a traditional classifier in an
attempt to facilitate the learning process. The features used in this study will be
extracted using the traditional feature extraction methods and the newly introduced
Deep Learning-based feature extraction techniques. The Features will then be com-
bined and the most informative features will be selected using a minimal-redundancy
maximal-relevance feature selection approach [21]. Finally, an AdaBoost classifier will
be used for evaluation of the patient’s condition, as illustrated in Fig. 2.
Fig. 2. The proposed feature extraction and classification steps.

3.1 Deep Learning

In this section, deep learning concept is briefly introduce as and is by no means a
comprehensive discussion of deep learning. For a more in-depth overview of deep
learning, please refer to [22]. The core component of a Deep Neural Network (DNN) is
the artificial neuron unit, in general, DNN can be considered as an extension of ANN
which (in general) have two or more layers. In summary, these neurons are used to
either amplify or attenuate the amplitude of the signal received from other adjacent
neurons. This is done by computing a weighted sum of inputs received from other
neurons using a (typically) non-linear activation function and creating the transformed
output signal as illustrated in Fig. 3. The output of a neuron can be computed as:

y ¼ u xT x þ b ð1Þ
The stacking of these neurons in a side-by-side structure forms a single layer network.
Compared to traditional ANN, the advantage of deep learning is the stacking of single-
layer networks to create a multilayered pipeline of non-linear transformations that is
capable of learning representations (features) for a given problem using various
abstraction levels. Although the concept of the neural network was conceived in the
1960s (with major advances in the later years), the break-through that arguably sparked
the current interest in neural networks (and deep learning) was the introduction of the
structured, layer-by-layer network architecture proposed in Deep Belief Networks
(DBN) and Stacked Autoencoders (SAE). Constructed by stacking many layers of
Restricted Boltzmann Machines (RBMs) or feedforward autoencoders and often trained
using backpropagation over the entire network for determining the most appropriate
weights. The introduction of Rectified Linear Units (ReLUs) as activation function was
of particular importance for tasks such as image processing which often have a limited
amount of training data. The output of the ReLU can be computed as:

y ¼ max 0; xT x þ b ð2Þ
Fig. 3. A representation of an artificial neuron with various connections and weights. Based on
the weights, the neuron eighter attenuates or amplifies the received signals and passes the sum
through an activation function.
The ReLU enables the neurons to have a constant gradient when the addition of the
inputs is positive with the non-linearity being observed by setting the output to zero
when the sum of the inputs is a negative value. The use of ReLU’s makes for a more
efficient training due to ReLU not suffering from the difficulties associated with
‘vanishing gradient’ observed in other activation functions. GPU based training and
implementation of deep neural networks (DNN) have also contributed greatly to the
wide adaption of DNN technologies as the networks can train considerably faster. More
recent advances regarding the use of deep learning for tasks related to speech pro-
cessing and computational paralinguistics have been focused on using Convolutional
Neural Networks (CNN) for learning non-descriptive feature representations directly
from the raw acoustic signals.
3.2 Traditional Statistics Based Features

Traditional statistics based features extracted from the speech signal have shown to be
highly valuable as they can be used effectively for the detection of disorders. In this
study, he most commonly used statistical features will be combined with features
extracted by the DNN in hopes of achieving high accuracy in identifying patients
suffering from disorders.
Utilising expert knowledge, acoustic features are used for providing compact and
discriminant representations of the acoustic signal with openSMILE feature extraction
toolkit [23] utilised for extraction of the two most widely used computational par-
alinguistic feature sets for detection of ASD knows as the extended Geneva Mini-
malistic Acoustic Parameter Set (eGeMAPS) and the large-scale Interspeech 2013
Computational Paralinguistics Challenge feature set (ComParE) [23]. Successfully
utilised for classifying signs of ASD in children voice samples [10, 24], eGeMAPS
[25] quantifies the distribution of low-level descriptors (LLDs) into an 88-dimensional
acoustical representation of an utterance representing spectral, cepstral, prosodic and
voice quality utilising the statistical mean and the coefficient of variation. ComParE
[23] feature set also utilises LDDs for computation of a feature set representing pro-
sodic, spectral, cepstral and voice quality similar to eGeMAPS but computes 6373
features which makes feature selection an important aspect when using these features.
Fundamental Frequency (F0) features are used to quantify the vibration rate of the
vocal folds, making them an important representer for laryngeal function. Jitter is used
for quantifying cycle-to-cycle changes in fundamental frequency by measuring the
instabilities in the oscillation pattern of the vocal folds. The shimmer is used for
quantifying cycle-to-cycle changes in speech amplitude by measuring the instabilities
in the oscillation pattern of the vocal folds. Harmonic to Noise Ratio (HNR) represents
the ratio of signal information over noise which is caused by turbulent airflow resulting
from an incomplete vocal fold closure in patients suffering from various disorders. Mel-
Frequency Cepstral Coefficients (MFCC) are used to quantify any possible damages to
the vocal tract (independent of the vocal folds) which might have been damaged as a
result of voice pathologies. In this study, a total of 13 MFCC coefficients will be used.
First and second derivatives of the cepstral coefficient are used to quantify the dynamic
behaviour of the acoustic signal.
4 Conclusions
With the availability of large datasets and increased computational capacity, the
healthcare sector has started to experience a slew of machine learning-based methods
for detecting various pathologies in patients using image and signal based processing of
patient data. On such field is the non-invasive detection of cognitive disorders such as
ASD using voice data gathered from the patients, which have shown to be as accurate
as the manual examination of a patient by an expert. In this study, a hybrid method
combining the effectiveness of deep learning-based feature extraction with traditional
acoustic features is proposed. The main strength of the proposed method is the simple
feature extraction and classification steps which makes it a speedy method and also
makes it possible to implement the whole system on portable devices such as smart-
phones for real-time analysis of patient’s condition. In future, based on ever-increasing
computational capability of mobile silicon chips, the traditional features can be directly
used in a deep learning network alongside the raw acoustic signal which not only will
eliminate the separate feature selection and classification steps, it might also increase
the classification accuracy.
References
1. Memari, N., Ramli, A.R., Saripan, M.I.B., Mashohor, S., Moghbel, M.: Retinal blood vessel
segmentation by using matched filtering and fuzzy C-means clustering with integrated level
set method for diabetic retinopathy assessment. J. Med. Biol. Eng. 39(5), 713–731 (2019)
2. Zablotsky, B., Black, I., Maenner, J., Schieve, A., Blumberg, J.: Estimated prevalence of
autism and other developmental disabilities following questionnaire changes in the 2014
National Health Interview Survey. Natl. Health Stat. Report 13, 1–20 (2015)
3. Verde, L., De Pietro, G., Sannino, G.: Voice disorder identification by using machine
learning techniques. IEEE Access 6, 16246–16255 (2018)
4. Cunningham, S.P., Green, P.D., Christensen, H., Atria, J.J., Coy, A., Malavasi, M., Rudzicz,
F.: Cloud-based speech technology for assistive technology applications (CloudCAST). In:
AAATE Conference, pp. 322–329 (2017)
5. Cummins, N., Scherer, S., Krajewski, J., Schnieder, S., Epps, J., Quatieri, F.: A review of
depression and suicide risk assessment using speech analysis. Speech Commun. 71, 10–49
(2015)
6. Schuller, B.: Can affective computing save lives? Meet mobile health. Computer 5, 13
(2017)
7. Cummins, N., Baird, A., Schuller, W.: Speech analysis for health: current state-of-the-art and
the increasing impact of deep learning. Methods 151, 41–54 (2018)
8. Bone, D., Li, M., Black, P., Narayanan, S.: Intoxicated speech detection: a fusion framework
with speaker-normalized hierarchical functionals and GMM supervectors. Comput. Speech
Lang. 28(2), 375–391 (2014)
9. Oller, D.K., Niyogi, P., Gray, S., Richards, J.A., Gilkerson, J., Xu, D., Warren, F.:
Automated vocal analysis of naturalistic recordings from children with autism, language
delay, and typical development. Proc. Natl. Acad. Sci. 107(30), 13354–13359 (2010)
10. Ringeval, F., Marchi, E., Grossard, C., Xavier, J., Chetouani, M., Cohen, D., Schuller, B.:
Automatic Analysis of Typical and Atypical Encoding of Spontaneous Emotion in the Voice
of Children (2016)
11. Schuller, B., Steidl, S., Batliner, A., Vinciarelli, A., Scherer, K., Ringeval, F., Mortillaro, M..
The INTERSPEECH 2013 computational paralinguistics challenge: social signals, conflict,
emotion, autism. In: Proceedings INTERSPEECH 2013, 14th Annual Conference of the
International Speech Communication Association, Lyon, France (2013)
12. Lee, Y., Hu, Y., Jing, H., Chang, F., Tsao, Y., Kao, C., Pao, L.: Ensemble of machine
learning and acoustic segment model techniques for speech emotion and autism spectrum
disorders recognition. In: INTERSPEECH, pp. 215–219, August 2013
13. Asgari, M., Bayestehtashk, A., Shafran, I.: Robust and accurate features for detecting and
diagnosing autism spectrum disorders. In: Interspeech, pp. 191–194 (2013)
14. Huang, C.L., Hori, C.: Classification of children with voice impairments using deep neural
networks. In: 2013 Asia-Pacific Signal and Information Processing Association Annual
Summit and Conference, pp. 1–5 (2013)
15. Nakai, Y., Takiguchi, T., Matsui, G., Yamaoka, N., Takada, S.: Detecting abnormal word
utterances in children with autism spectrum disorders: machine-learning-based voice
analysis versus speech therapists. Percept. Motor Skills 124(5), 961–973 (2017)
16. Carpenter, M., Tomasello, M., Striano, T.: Role reversal imitation and language in typically
developing infants and children with autism. Infancy 8(3), 253–278 (2005)
17. Deng, J., Cummins, N., Schmitt, M., Qian, K., Ringeval, F., Schuller, B.: Speech-based
diagnosis of autism spectrum condition by generative adversarial network representations.
In: Proceedings of the 2017 International Conference on Digital Health, pp. 53–57 (2017)
18. Le Couteur, A., Haden, G., Hammal, D., McConachie, H.: Diagnosing autism spectrum
disorders in pre-school children using two standardised assessment instruments: the ADI-R
and the ADOS. J. Autism Dev. Disord. 38(2), 362–372 (2008)
19. Mower, E., Black, M.P., Flores, E., Williams, M., Narayanan, S: Rachel: design of an
emotionally targeted interactive agent for children with autism. In: 2011 IEEE International
Conference on Multimedia and Expo (2011)
20. Martínez, D., Lleida, E., Ortega, A., Miguel, A., Villalba, J.: Voice pathology detection on
the Saarbrücken voice database with calibration and fusion of scores using multifocal toolkit.
In: Advances in Speech and Language Technologies for Iberian Languages, pp. 99–109.
21. Peng, H., Long, F., Ding, C.: Feature selection based on mutual information criteria of max-
dependency, max-relevance, and min-redundancy. IEEE Trans. Pattern Anal. Mach. Intell.
27(8), 1226–1238 (2005)
22. Nassif, A.B., Shahin, I., Attili, I., Azzeh, M., Shaalan, K.: Speech recognition using deep
neural networks: a systematic review. IEEE Access 7, 19143–19165 (2019)
23. Eyben, F., Weninger, F., Gross, F., Schuller, B.: Recent developments in open smile, the
Munich open-source multimedia feature extractor. In: Proceedings of the 21st ACM
International Conference on Multimedia, pp. 835–838 (2013)
24. Schmitt, M., Marchi, E., Ringeval, F., Schuller, B.: Towards cross-lingual automatic
diagnosis of autism spectrum condition in children’s voices. In: Speech Communication; 12.
ITG Symposium (2016)
25. Eyben, F., Scherer, K.R., Schuller, B.W., Sundberg, J., André, E., Busso, C., Truong, K.:
The Geneva minimalistic acoustic parameter set (GeMAPS) for voice research and affective
computing. IEEE Trans. Affect. Comput. 7(2), 190–202 (2015)
A Compromise-Based New Approach
to Learning Fuzzy Cognitive Maps
Miraç Murat(&) and Umut Asan

{murat18,asanu}@itu.edu.tr
Abstract. Fuzzy Cognitive Maps (FCMs), first introduced by Kosko, are

graph-based knowledge representation tools. In order to improve the efficiency,
robustness and accuracy of FCMs, different learning approaches and algorithms
have been introduced in the literature. The algorithms aim to revise the initial
knowledge of experts and/or extract useful knowledge from historical records in
order to yield learned weights. One considerable drawback of FCM is that, in its
original form, it often yields the same output under different initial conditions.
Since the results of the learning algorithms are highly dependent on the rea-
soning mechanism (i.e. updating function) of FCMs, this drawback also affects
the performance and accuracy of these algorithms. Therefore, problems
including (conflicting) multiple initial vectors, multiple weight matrices and
multiple desired final state vectors have received only limited attention. In order
to address this issue and provide a better modeling framework for this type of
problems, a compromise-based new fuzzy cognitive mapping approach based on
particle swarm optimization is suggested. To justify the effectiveness and
applicability of the proposed approach, an illustrative example is provided.
Keywords: Learning Fuzzy Cognitive Maps Multiple input data Particle

swarm optimization Compromise Mean squared deviation
1 Introduction
Fuzzy cognitive mapping, first proposed by Kosko [1], is a soft-computing technique

that is used to model and analyze the dynamic behavior of complex systems. It allows
performing qualitative simulations in order to analyze possible changes and to observe
whether the system converges toward a steady state. The accuracy and robustness of
these analyses highly depend on the causal connections (i.e. weights) determined in the
fuzzy cognitive map [2]. Although using expert knowledge to generate FCMs is a
powerful tool, it is not always possible to find the relevant experts or experts may
provide conflicting opinions because of their different personal views, expertise and
backgrounds [3]. Moreover, the construction of a real FCM often involves dealing with
a large amount of concepts and connections, which is a time-consuming and tedious
task [2]. To improve the efficiency, robustness and accuracy of FCMs in modeling and
prediction, various learning approaches and algorithms have been introduced in the
A Compromise-Based New Approach to Learning Fuzzy Cognitive Maps 1173
literature. The algorithms aim to revise the initial knowledge of experts and/or extract
useful knowledge from historical data in order to yield learned weights.
Most of the learning approaches proposed in the literature are based on the learning
principles developed in the field of artificial neural networks [2]. The learning algo-
rithms for FCMs can be broadly categorized into three types: Hebbian-based, error-
driven and hybrid learning algorithms [4].
One considerable drawback of the reasoning mechanism of FCMs is that, in its
original form, it often yields the same output (steady state values) under different initial
conditions. This issue was recently demonstrated in [5, 6] for a variety of FCM con-
figurations used for prediction. Also, Napoles et al. [7] showed for binary classification
problems that algorithms ensuring convergence yield the same unique fixed-point
attractor for all input sequences. Using algorithms that produce an equilibrium state
independent of the initial conditions will be unable to discriminate among patterns that
have different semantics [8]. Therefore, this type of reasoning mechanisms will perform
well in control scenarios, but are not appropriate for prediction scenarios.
Since the results of the learning algorithms are highly dependent on the reasoning
mechanism of FCMs, this drawback also affects the performance and accuracy of these
algorithms. Therefore, problems including (conflicting) multiple initial vectors, mul-
tiple weight matrices and multiple desired final state vectors have received only limited
attention. In order to address this issue and provide a better modeling framework for
this type of problems, a compromise-based new fuzzy cognitive mapping approach
based on particle swarm optimization is proposed. The suggested approach integrates a
recently developed reasoning mechanism [5] into the learning algorithm to produce
steady state values sensitive to initial conditions. Using multiple inputs (initial state
values, initial weight matrices and desired state values) based on experts’ knowledge
and/or historical records, the suggested approach computes an optimal weight matrix
that best fits the problem. To justify the effectiveness and applicability of the proposed
approach, an illustrative example on brand positioning is provided. The results
demonstrate that the suggested approach provides a more accurate calculation and,
thereby, more distinctive and reliable results.
2 Fuzzy Cognitive Maps
2.1 Classic Approach

As mentioned above, FCMs allow performing qualitative simulations to analyze the
dynamic behavior of a system. In this iterative process, the activation degrees of the
next step are calculated using the following reasoning mechanism ([9]; for a simpler
form of the updating function see Kosko [10]):
0 1
ðt Þ
Xn
f @A i
ðt1Þ
j ¼ 1; Aj wji A
ðt1Þ
Ai ¼ þ ð1Þ
j 6¼ i
1174 M. Murat and U. Asan
ðt Þ
where Ai denotes the activation degree of concept i at iteration step (t), wji denotes the
sign and intensity of the direct causal link from concept j to concept i and f is a
monotonically non-decreasing threshold function (e.g. sigmoid and hyperbolic tangent
function).
The simulation
requires defining an initial state vector
ð0Þ ð0Þ
Að0Þ ¼ A1 ; A2 ; . . .; Aðn0Þ , which can be provided by experts or obtained from his-
torical data. The updating mechanism is repeated until either the system converges to a
fixed-point attractor or a predefined maximum number of iterations is reached (i.e. a
limit cycle or chaotic behavior is present) [10].
2.2 Alternative Approach

A critical issue on the convergence of FCMs is the insensitivity of the steady state
values to initial conditions. As Asan and Kadaifci [5] showed in a recent study, for a
variety of FCM configurations the classical reasoning mechanism of FCMs yielded the
same output for different initial state vectors. To address this issue [5, 6] proposed an
improved FCM approach, which reasoning mechanism is summarized below:
Calculation of the Reachability Matrix: To predict the long-term behavior of the
system, the indirect causal connections between the nodes are needed to be considered.
To do this the reachability matrix is calculated by raising the direct relationship matrix
to its successive powers and summing them up:
rij ¼ wij þ w2ij þ w3ij þ þ wTij ð2Þ
where rji denotes the ith row and jth column element of the reachability matrix and
T denotes the iteration step at which the system converges to a steady state.
Calculation of the Steady State Values: The steady state values are obtained using
Eq. 3:
Xn ð0Þ
aif ¼ j ¼ 1 ðaj rji Þ ð3Þ
j 6¼ i
ð0Þ
where aif denotes the steady state value of the ith concept, aj represents the initial state
value of the jth concept. Finally, to enable the comparison of the final concept values
alternative normalization methods can be used.
3 Particle Swarm Optimization
The new approach, proposed to find the weight matrices that lead to desired steady
states, is integrating a new reasoning mechanism (described above) with the Particle
Swarm Optimization (PSO) algorithm. PSO is a heuristic population-based technique,
which belongs to the class of swarm intelligence algorithms. Since PSO requires only a
few parameters to be tuned and is relatively simple and easy to apply [11], it is
frequently used in earlier FCM learning studies [12, 13].
Each swarm intelligence algorithm has a population, named a swarm, consisting of
individuals named a particle. Each particle, which essentially represents a potential
solution, systematically probes the search space and learns the experiences of the
swarm to reach an optimal solution [11]. The position, that is a nominated solution by
the particle, is updated by the memories of particles and swarm. The particle retains its
best position pbest ever attained in its memory and the swarm retains the best position
of the particles gbest in its memory. By communicating gbest to all the particles, the
position of each particle is updated according to an adaptive velocity and consequently,
a new generation is produced [11].
Assume a D-dimensional search space, S RD , and a swarm consisting of M
particles. The current position and velocity of the ith particle at kth iteration are as
follows:
The current position, Xik ¼ xki;1 ; xki;2 ; . . .xki;D , is a D-dimensional vector, where
xki;d 2 ½ld ; ud , 1 d D and ld and ud are the lower and upper limits for the dimension d.

The current velocity, Vik ¼ vki;1 ; vki;2 ; . . .vki;D , is also a D-dimensional vector. Each
velocity at the current iteration is restricted to its maximum value of vkmax;d .
At each iteration, the positions and velocities and thence the swarm are updated by
the Eqs. 4 and 5:
h i
Vik þ 1 ¼ v xVik þ c1 r1 Pki Xik þ c2 r2 Pkgi Xik ð4Þ
Xik þ 1 ¼ Xik þ Vik þ 1 ð5Þ
where Pi is the best previous position of the ith particle, namely pbest, and Pgi is the
global best position ever encountered, namely gbest. v and x are parameters to control
the magnitude of the velocity and are called constriction factor and inertia weight
respectively. c1 and c2 are cognitive and social acceleration constants respectively, and
r1 , r2 are uniform random vectors on the interval [0, 1]. These constants and random
vectors control the convergence speed.
The quality of a nominated solution by a particle is evaluated through a specified
fitness function. And hence the fitness function leads to determine the values Pi and Pgi
by the Eqs. 6 and 7.

Pi : f ðXi Þ Pi
Pi ¼ ð6Þ
Xi : f ðXi Þ\Pi

Pg 2 fP0 ; P1 ; . . .; PM gf Pg ¼ minðf ðP0 Þ; f ðP1 Þ; . . .; f ðPM ÞÞ ð7Þ
where f is the fitness function.

In this study, the fitness function f is considered as an FCM simulation that allows
it to reach a steady state. This FCM simulation explores the adequacy of the system
abstraction provided by a particle. A particle provides the system abstraction through
its D elements that each one represents the relationship between any two out of
n concepts offered by the experts.
4 The Proposed Learning Fuzzy Cognitive Maps Approach
The proposed approach aims at building an FCM that transforms diverse expert
opinions into a compromised one taking various initial-desired state pairs into con-
sideration. To find the FCM architecture yielding the compromised solution, the PSO
algorithm is used. PSO searches the solution space for a single optimum FCM archi-
tecture that can enable the system to converge for each initial-desired state pair.
Besides, the total deviation from expert opinions is minimal for the optimum FCM
architecture.
The proposed objective function considers the sum of two deviations: i) deviation
from the experts’ opinions and ii) deviation from the desired state(s). Deviation from
experts’ opinions is calculated using the differences between the weight values in the
candidate FCM architecture and the corresponding values obtained from experts. More
formally, the mean squared deviation for a candidate architecture is calculated as
shown
1 XE XC XC l 2
MSEW ¼ 2
n ij wij ð8Þ
LC k¼1 i¼1 j¼1
where nlij is the relationship between the concepts Ci and Cj expressed by the lth expert
ðl ¼ 1; 2; . . .; LÞ. The deviation from the desired state is calculated for each of the
initial-desired state pairs of the candidate FCM architecture. The mean squared devi-
ation between the desired state and the final state of FCM, where the candidate
architecture converges from the initial state, is calculated by Eq. 9.
1 XS XC d f
2
MSEC ¼ a ij a ij ð9Þ
SC i j
where adi: and ai:f indicate the desired and final state for the ith input ði ¼ 1; 2; . . .; SÞ,
respectively.
By incorporating the PSO steps into the new reasoning mechanism of FCM, each
candidate preliminary solution evolves into new candidate solutions. The steps of the
proposed approach are presented below:
Step 1. Initialize positions and velocities of the particles
Once the initial-desired state pairs are determined by experts or obtained from
historical records, positions and velocities of the particles are initialized. More for-
mally, to start evaluating candidate solutions, preliminary FCM architectures are ran-
domly created in line with the relationships provided by the experts. In addition, to
evolve the candidate solutions into better ones the velocities are randomly initialized.
Step 2. Evaluate the fitness value of each particle

Step 2.1. By using the position of each particle, an FCM is simulated for each state
pair until the system reaches its peculiar convergence point.
Step 2.2. MSEW and MSEC are calculated through Eqs. 8 and 9. The sum of the
MSEs are set as the fitness value of the particle.
Step 3. Determine Local Best values. For each particle, the local best fitness is
assigned as the current fitness and the local best position is assigned as the current
position.
Step 4. Determine Global Best particle. The parameters of the present particle
having minimum fitness are assigned to the Global Best particle.
Step 5. Update each particle. The position and velocity values of each particle are
updated according to Global Best particle. After getting updates as shown in Eqs. 4 and
5, M new candidate solutions are acquired.
Step 6. Reevaluate the fitness value of each particle. By using the position of each
new particle, FCM simulations are performed once more as described in Step 2.
Step 7. Determine new Local Best and fitness values. Current fitness value is
assigned as the Local Best fitness value if present one is better than the Local Best.
After updates for all particles, the Global Best particle is determined.
Step 8. Check the stopping criteria. If the maximum iteration number or possible
minimum fitness value (0 for the present study) are reached, the algorithm ends.
5 Illustrative Example
The proposed method is illustrated using data from a product positioning study con-
ducted by Asan and Kadaifci [5] in the entertainment industry. The data consists of the
opinions of the experts and the perceptions (current state) and preferences (desired
state) of the consumers. Experts are asked to express their opinions on the causal
relations among 13 different positioning attributes believed to express the perceptions
and preferences of game consoles. How game consoles are perceived on the basis of
these positioning attributes is provided by hardcore gamers. Table 1 provides the
perceptions and preferences of five hardcore gamers on one game console brand with
respect to the 13 attributes (for the complete data the reader should refer to [6]. In each
vector, the value one indicates that an attribute is believed to be hold by the product and
will be considered as activate (in the initial state vector).
Table 1. The status perceptions of five consumers

Current status perception Desired status perception
C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13
Co.1 1 1 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 1 .5 0 0 1 1
Co.2 0 1 1 0 1 0 0 0 1 0 0 0 1 0 0 0 .5 0 0 0 1 0 1 0 1 1
Co.3 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 1 1
Co.4 0 0 1 1 0 0 0 0 1 0 1 0 1 0 0 0 .5 0 0 0 1 .5 0 1 1 1
Co.5 0 1 1 0 0 0 1 0 0 1 1 0 1 0 0 0 0 0 0 0 1 .5 1 0 1 1
The proposed method tries to find the optimal weight matrix that has minimal
distance to all expert opinion matrices. By using these matrices and the perceptions of
five consumers, the optimal FCM weight matrix that is expected to converge to the
desired state from the associated initial states (perceptions of consumers) is calculated.
So, the FCM can be used for occasions that differ according to each consumer.
The proposed reasoning mechanism first computes the indirect relationships to be
used in the FCM simulations. The initial state values are converted into its associated
desired values by separately considering indirect causal relations of all possible lengths.
The indirect relationship matrices (of length t) are obtained by raising the candidate
weight matrix to its powers (tth) and the system state is calculated at each FCM iteration
as given in Eq. 10.
ðtÞ
Ai ¼ C W t ð10Þ
ðtÞ
where Ai denotes the activation degree of the concepts at iteration t. The proposed
approach produces the optimal weight matrix by running 1000 iterations with 100
particles. These optimal weights allow to convert the initial states into the desired ones
for each consumer as shown in Fig. 1. It can be seen from the figure that the proposed
method enables both to find compromised weights and to converge each initial state to
its desired one. The results show that by considering the indirect relations in the
proposed way, it is more likely to reach the desired system state. Furthermore, the
results are sensitive to the initial conditions. In Fig. 1-c, the activation values that reach
almost one or zero appear to be close to the desired state values expressed a priori by
consumer three. The remaining plots also demonstrates accurate predictions.
1 1 1
0.9 0.9 0.9
0.8 0.8 0.8
0.7 0.7 0.7
0.6 0.6 0.6

Concepts Values
Concepts Values
Concepts Values
0.5 0.5 0.5
0.4 0.4 0.4
0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0 0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9
Iteration Iteration Iteration
a) b) 1
c)
1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
Concepts Values
Concepts Values
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
Iteration Iteration
d) e)
Fig. 1. Simulation plots obtained through compromised matrix for five consumers
The proposed particle swarm intelligence based approach helps to determine

compromised weight values that allows to simulate diverse prediction scenarios. The
determine weight values in this way can eliminate information loss that might occur
when the matrices are aggregated through traditional methods. The results of the
illustrative example supports this argument by demonstrating that a compromise-based
solution (FCM weight matrix) can satisfy consumers with minimum deviation.
6 Conclusion and Further Research
One considerable drawback of FCM is that, in its original form, it often yields the same
output under different initial conditions. Since the results of the learning algorithms are
highly dependent on the reasoning mechanism of FCMs, this drawback also affects the
performance and accuracy of these algorithms. In order to address this issue and
provide a better modeling framework, a compromise-based new fuzzy cognitive
mapping approach based on particle swarm optimization is proposed. The effectiveness
and applicability of the new approach is justified by an example from the positioning
literature. Comparisons with current learning approaches remain as a further research.
References
1. Kosko, B.: Fuzzy cognitive maps. Int. J. Man Mach. Stud. 24, 65–75 (1986)
2. Papageorgiou, E.I.: Learning algorithms for fuzzy cognitive maps—a review study. IEEE
Trans. Syst. Man Cybern. Part C Appl. Rev. 42, 150–163 (2011)
3. Salmeron, J.L., Mansouri, T., Moghadam, M.R.S., Mardani, A.: Learning fuzzy cognitive
maps with modified asexual reproduction optimisation algorithm. Knowl. Based Syst. 163,
723–735 (2019)
4. Felix, G., Nápoles, G., Falcon, R., Froelich, W., Vanhoof, K., Bello, R.: A review on
methods and software for fuzzy cognitive maps. Artif. Intell. Rev. 52, 1707–1737 (2019)
5. Asan, U., Kadaifci, C.: An improved fuzzy cognitive mapping method. In: International
Conference on Intelligent and Fuzzy Systems, pp. 115–122. Springer, Cham (2019)
6. Asan, U., Kadaifçi, Ç.: A new product positioning approach based on fuzzy cognitive
mapping. J. Fac. Eng. Archit. Gazi Univ. 35, 1047–1061 (2020)
7. Nápoles, G., Bello, R., Vanhoof, K.: How to improve the convergence on sigmoid fuzzy
cognitive maps? Intell. Data Anal. 18, S77–S88 (2014)
8. Nápoles, G., Salmeron, J.L., Froelich, W., Falcon, R., Espinosa, M.L., Vanhoenshoven, F.,
Bello, R., Vanhoof, K.: Fuzzy cognitive modeling: theoretical and practical considerations.
In: Intelligent Decision Technologies 2019, pp. 77–87. Springer, Singapore (2020)
9. Stylios, C.D., Groumpos, P.P.: Modeling complex systems using fuzzy cognitive maps.
IEEE Trans. Syst. Man Cybern. A Syst. Humans 34, 155–162 (2004)
10. Kosko, B.: Hidden patterns in combined and adaptive knowledge networks. Int. J. Approx.
Reason. 2, 377–393 (1988)
11. Li, L., Liu, F.: Group Search Optimization for Applications in Structural Design. Springer,
Heidelberg (2011)
12. Parsopoulos, K.E., Papageorgiou, E.I., Groumpos, P.P., Vrahatis, M.N.: A first study of
fuzzy cognitive maps learning using particle swarm optimization. In: The 2003 Congress on
Evolutionary Computation, 2003. CEC 2003, pp. 1440–1447. IEEE (2003)
13. Parsopoulos, K.E., Papageorgiou, E.I., Groumpos, P.P., Vrahatis, M.N.: Evolutionary
computation techniques for optimizing fuzzy cognitive maps in radiation therapy systems.
In: Genetic and Evolutionary Computation Conference, pp. 402–413. Springer, Heidelberg
(2004)
Fuzzy Analytics
Fuzzy Analytics of Spaces’ Dynamic Use
in Smart Buildings
Azedine Boulmakoul1(&), Abdellah Daissaoui2,3, Ahmed Lbath3,

and Lamia Karim1,4
1
Casablanca, Morocco
2
EMSI, LPRI, Casablanca, Morocco
[email protected]
3
[email protected]
4
ENSA Berrechid, LISA Laboratory, Hassan 1st University, Settat, Morocco
Abstract. The applications of indoor spatial information target two types of

services: Management of building components with its indoor facilities and use
of indoor space. In this article, we consider the second category, which includes
the representation requirements of spatial components such as rooms and hall-
ways, and constraints such as doors. Indoor location-based services, indoor
route analysis or indoor geo-tagging services are also addressed in this work.
Space management is of increasing interest to buildings operators. Indoor
monitoring can increase the efficiency of the use of available space while
reducing costs. By means of WiFi devices, the localization of persons is pos-
sible. The movement person’s profiles with smartphones are recorded and
evaluated anonymously. As soon as data is available over a longer period,
forecasts can also be made. In this work, we propose an anonymous person
tracking system to ensure the temporal analysis of spaces occupation. We use
fuzzy set theory to achieve this goal. Data from people’s movements will be
analyzed to generate time-dependent partitions. These partitions will be moni-
tored over time to detect invariants and occupation space patterns. For the
purpose of intelligent buildings modelling, we use the OGC IndoorGML
repository to model the indoor spaces.
Keywords: Smart budding Indoor spatial analysis Fuzzy clustering

Occupancy dynamic analytics
1 Introduction
Analysis of the behavior or navigation of people in smart buildings has progressed

rapidly in recent years. Applications conventionally focus more on the visualization
and monitoring of building occupancy and the monitoring of passengers traffic. Data
collection on user behavior in a smart building can be done through several
technologies. The most requested techniques use sensors, beacons and WiFi devices.
The data collected from the sensors is transferred to database servers in the cloud.
Several services access this data via a gateway [2, 3]. Various services are provided to
building managers and users. These services go a long way toward lighter asset
management, achieving cost reductions, improving the user experience, and supporting
sustainability efforts. Detecting the occupation of space distinguishes two categories.
The first uses practices that require smartphone-like geolocation equipment. The other
category, detection is based on a passive approach based on surveillance of areas or
spaces instead of identifying devices by the use of cameras [5–8]. Data from the
detection of people in spaces is used to generate dynamic classifications. Precisely
these data are spatio-temporal and machine-learning algorithms allow producing sev-
eral clustering according to homogeneous periods. In this article, we introduce an
algorithmic process in which several clustering could easily be combined. In this
process, a fuzzy context matrix is established. From this fuzzy context, we construct a
fuzzy similarity matrix, used to make hierarchical groupings. The max-min transitive
closure of this similarity produces a new partitions hierarchy.
The rest of this paper is organized as follows. Section 2 reviews related works of
building meta-model. Section 3 presents preliminaries on fuzzy sets. It then proposes a
new method for combining groups gathered from several clustering processes. The
method is based on the generation of a correspondence, used to produce a new fuzzy
similarity matrix. This fuzzy matrix will be used to generate a new clustering. Finally,
Sect. 4 concludes the paper.
2 Indoor Objects in IndoorGML
The OGC® IndoorGML [4, 12, 13] standard proposes to establish a common scheme
for indoor navigation applications. It models topology and semantics of indoor spaces
(see Fig. 1, 2 and 3). It defines the following information on indoor space: (1) Context
and navigation constraints, (2) Spatial subdivisions and types of connectivity between
spaces, (3) Geometric and semantic properties of spaces and connectivity, (4) Navi-
gation networks (logical and metric) and their relationships. The difference between
indoor and outdoor space is that an interior space is composed of complex constraints
such as corridors, doors, stairs, elevators, etc. The semantic representation of the
interior space is based on the semantics of the cells. The interior space can be broken
down into different cells. The subdivision of the cells can represent the topography of a
building, the RFID coverage available; indicate the security zones or the public/office
zones, etc. Each cell then receives semantics in relation to its use. For example, in a
topographic space, it is possible to have classroom, ‘door’, and ‘window’, in the RFID
space - ‘Spot point A’, ‘Spot point B’ etc. and in a security area – controlled area’
“operations area”, “Restricted area”.
Fuzzy Analytics of Spaces’ Dynamic Use in Smart Buildings 1185
Fig. 1. (a) Topographic space, (b) Adjacency graph.
Fig. 2. (a) Adjacency Graph with navigable and non-navigable edges, (b) Connectivity Graph,
(c) Accessibility Graph.
Fig. 3. UML diagram of IndoorGML Navigation module

Fig. 4. Simplified model
Figure 4, models a time series of sensor data associated with an indoor space of a
smart building. In this work, time series concerns the detection of people and consti-
tutes the entry to the process of dynamic structural analytics [15].
3 Fuzzy Analytics
In this section, we present a new fuzzy analytics process to study the combination of
multiple partitions from clustering methods. First, we recall main of the definitions
relating to fuzzy sets and we give a fuzzy similarity calculation formula used for this
work.
3.1 Fuzzy Sets [16]
If X is a collection of objects denoted generically by x, then a fuzzy set A ~ in X is a set

~
of ordered pairs A ¼ x; lA~ ð xÞ jx 2 X :
lA~ ð xÞ is called the membership function or grade (also degree of compatibility or
degree of truth) of x in A ~ that maps X to the membership space M (When M = {0,1}, A ~
is crisp set and lA~ ð xÞ is identical to the characteristic function of nonfuzzy set). The
~
(crisp) set of elements that belong to thefuzzy set A at least to the degree a is called the
a − level set: Aa ¼ x 2 X jlA~ ð xÞ a .

Aa ¼ x 2 X jlA~ ð xÞ a is called “strong a-level” or “strong a-cut”. Throughout
this paper, the following notations have been used. R þ ¼ ½0; 1Þ; X ¼ fx1 ; x2 ; . . .; xn g
is the universal set; F ð X Þ is the class of all fuzzy sets of X; lA~ ðxi Þ: X ! [0.1] is the
membership function of A 2 F ð X Þ; Ac 2 F is the complement of A 2 F .
3.2 Similarity Between Fuzzy Sets
We define two similarity measures S : F 2 ! R þ between fuzzy sets A and B as

follows [14, 16]:
Pi¼n
i¼1 minðlA ðxi Þ; lB ðxi ÞÞ
SðA; BÞ ¼ Pi¼n ð1Þ
i¼1 maxðlA ðxi Þ; lB ðxi ÞÞ
3.3 Clustering Dynamics

In this section, we propose a process of partitions combination [1, 9, 11] according to
an algorithmic framework in which multiple partitions could easily be consolidated. In
this framework, the similarity-based description matrices of input partitions clustering
are aggregated into a fuzzy similarity matrix in which the final hierarchy could be
established. In order to easily illustrate this process.
Process Steps
The stages of this process are developed as follows: (1) Create the fuzzy context,
(2) Create the fuzzy similarity matrix by means of Eq. (1), (3) generate the recovery
hierarchy [10], (4) create the max-min transitive closure of the fuzzy similarity matrix,
(5) generate the partitions hierarchy [10].
We offer an example to trace the above process deployment. Let A, B and C be
three partitions given below:
A = {A1:{234}, A2:{1}}, B = {B1:{123}, B2:{4}}, C = {C1:{12}, C2:{3}, C3:
{4}}. Clusters proposed here are supposed obtained from the data of the indoor spaces
of a smart building.
The matrix given in Table 1 (step 1), illustrates the creation of context used by the
consolidation process.
Table 1. Partitions fuzzy context C

Clusters ID Set 1 2 3 4
A1 {2,3,4} 0 1/3 1/3 1/3
A2 {1} 1 0 0 0
B1 {1,2,3} 1/3 1/3 1/3 0
B2 {4} 0 0 0 1
C1 {1,2} 1/2 1/2 0 0
C2 {3} 0 0 1 0
C3 {4} 0 0 0 1
The generic element given in the context of Table 1, is calculated as follows:
1
Cði; jÞ ¼ Ii ð jÞ ð2Þ
ki k
Where Ii ð jÞ is equal to 1 if i belongs to set j, 0 otherwise. kik denotes the cardinal of

the set i.
Table 2 (step 2) proposes the fuzzy similarity matrix produced for the context given
in Table 1.
Table 2. Fuzzy similarity matrix

Clusters ID 1 2 3 4
1 1 0,4 0,1 0
2 1 0,32 0,73
3 1 0,75
4 1
Figure 5 (step 3) provides covers tree given by the similitude analysis of the fuzzy
similarity matrix.
Fig. 5. Fuzzy similarity relation’s hierarchical clustering dendrogram
The max-min transitive closure of the fuzzy similarity relation is given below (see
Table 3, step 4):
Table 3. Max-min transitive closure

Clusters ID 1 2 3 4
1 1 0,4 0,4 0,4
2 1 0,73 0,73
3 1 0,75
4 1
Figure 6 (step 5) provides partitions tree given by the similitude analysis of the
fuzzy max-min transitive similarity matrix.
Fig. 6. Hierarchical clustering dendrogram of fuzzy similarity relation’s Max-Min closure
Obtained clustering will be the vectors of the exploration of indoor spaces, to drive
monitoring and optimization strategies.
4 Conclusion
Persons movement data mining in a smart building has received several attentions in
recent years and many clustering methods have been introduced. These methods
generate several time-dependent partitions. These partitions will be monitored over
time to detect invariants and patterns of occupancy. In this work, we have proposed a
new approach to combine the results of groupings given by clustering algorithms. The
method is based on the generation of a fuzzy context. The idea of producing a fuzzy
similarity is used to combine temporal clusters. This work developed the approach to
combine groups gathered from clustering processes. An important step is planned to
carry out the experiments necessary to test our approach.
References
1. Ayad, H., Kamel, M.: Finding natural clusters using multi-clusterer combiner based on
shared nearest neighbors. In: Windeatt, T., Roli, F. (eds.) Proceedings of 4th International
Workshop on Multiple Classifier Systems, Lecture Notes in Computer Science, vol. 2709,
pp. 166–175. Springer, Guildford (2003)
2. Abade, B., Abreu, D.P., Curado, M.: A non-intrusive approach for indoor occupancy
detection in smart environments. Sensors 18, 3953 (2018). https://doi.org/10.3390/
s18113953
3. Candanedo, L., Feldheim, V.: Accurate occupancy detection of an office room from light,
temperature, humidity and CO2 measurements using statistical learning models. Energy
Build. 2016(112), 28–39 (2016)
4. Kang, H.-K., Li, K.-J.: A standard indoor spatial data model- OGC IndoorGML and
implementation approaches. ISPRS Int. J. Geo Inf. 6(4), 1–25 (2017)
5. Lu, H., Guo, C., Yang, B., Jensen, C.S.: Finding frequently visited indoor POIs using
symbolic indoor tracking data. In: Proceedings of the 19th International Conference on
Extending Database Technology (EDBT 2016), pp. 449–460 (2016)
6. Huh, J.H., Seo, K.: An indoor location-based control system using bluetooth beacons for IoT
systems. Sensors 17, 2917 (2017)
7. Jin, M., Jia, R., Spanos, C.J.: Virtual occupancy sensing: using smart meters to indicate your
presence. IEEE Trans. Mob. Comput. 2017(16), 3264–3277 (2017)
8. Cecílio, J., Duarte, K., Martins, P., Furtado, P.: RobustPathFinder: handling uncertainty in
indoor positioning techniques. In: Procedia Computer Science, vol. 130, pp. 408–415
(2018). ISSN 1877-0509
9. Kuncheva, L.I.: Combining Pattern Classifiers: Methods and Algorithms. Wiley, New York
(2004). ISBN: 0-471-66025-6
10. Klir, G.J., Clair, U.S., Yuan, B.: Fuzzy Set Theory: Foundations and Applications. Prentice
11. Mirzaei, A., Rahmati, M.: A novel hierarchical-clustering-combination scheme based on
fuzzy-similarity relations. IEEE Trans. Fuzzy Syst. 18(1), 27–39 (2010)
12. OGC, OGC CityGML Encoding Standard, Document No. 12-019 (2012). http://www.
opengeospatial.org/standards/citygml. Accessed 02 Jan 2020
13. OGC, OGC IndoorGML, Document No. 14-005r4 (2014). http://www.opengeospatial.org/
standards/indoorgml. Accessed 02 Jan 2020
14. Wang, W.-J.: New similarity measures on fuzzy sets and on elements. Fuzzy Sets Syst. 85
(3), 305–309 (1997)
15. Li, X., Yan, Y., Tu, J.: The simplification of computer simulated persons (CSPs) in CFD
models of occupied indoor spaces. Build. Environ. 93(2), 155–164 (2015). ISSN 0360-1323
16. Zimmermann, H.-J.: Fuzzy Set Theory—and Its Applications. Springer, Dordrecht (2001).
ISBN 978-94-010-3870-6
Object Monitoring Under Z-Information Based
on Rating Points
Olga M. Poleshchuk(&)
Moscow Bauman State Technical University, Moscow, Russia

[email protected]
Abstract. The paper developed models for rating evaluation of objects and
monitoring the functioning of these objects under Z-information. The rating of
objects is determined in the framework of qualitative characteristics. When con-
structing rating models, statistical information or a survey of experts is used. The
first and second components of the Z-numbers are values of semantic spaces with
the properties of completeness and orthogonality. Operations with Z-numbers are
based on weighted points that aggregate the information of numbers over all a-cuts
of their components. Recognition of the states of objects according to the results of
their rating evaluation is based on the distance between the Z-numbers, determined
on the basis of weighted points. The distance is determined taking into account the
first and second components of the Z-numbers as well as the first component,
which is multiplied by the weighted point of the second component. Monitoring
the functioning of objects is carried out on the basis of a comparison of the first
components of the Z-numbers, and in case of their equality on the basis of the
second components of the Z-numbers.
Keywords: Z-number Rating point Monitoring Semantic space

Weighted point
1 Introduction
When we receive information, we understand that, as a rule, it is not completely

reliable. When constructing various models, we are dealing with various measure-
ments, estimates and measures, which usually cannot be represented as a clear number
with absolute reliability. This objective reality significantly complicates decision-
making procedures due to the risk of big mistakes, because in order to make the right
and accurate decision, we need to evaluate the reliability of information. Since for
many years the reliability of the incoming information was not properly evaluated and
accordingly not taken into account [1], this gap was the main problem of decision -
making, which was eliminated only in 2011 after the introduction of Z-number by
Professor Lotfi Zadeh [2].
Professor Zadeh combined in a Z-number the fuzzy value of a real-valued uncertain
variable and the fuzzy estimate of the reliability of this value. Information with
Z-numbers has been called Z-information. The concept of Z-number, which made it
possible to increase the information content of incoming data and evaluate them from
1192 O. M. Poleshchuk
different perspectives, opened up huge potential for solving problems in problem areas
with the active participation of experts and opened the way for describing real infor-
mation in a form understandable for a computer. The definition of a Z-number sum-
marized and generalized many years of researching into the fuzzy sets theory. It was an
invaluable contribution for modelling processes and objects of the real world in the
conditions of taking into account the uncertainty of the real situation, human thinking
and the complexity of the real environment.
A significant number of works in recent years are devoted to approaches to the
formalization of expert information based on Z-numbers, analysis, prediction and
decision-making based on Z-information.
In [3], an approach was developed to operate with Z-numbers, based on the con-
version of Z-numbers into fuzzy numbers using their expectations. In [4, 5], the authors
developed a general approach to operating with Z-numbers (discrete and continuous
numbers). In [6], a model was developed for aggregating the opinions of an expert
group using the t-norm and t-conorm. In [7], Z-numbers are used for reasoning.
Today there is no unique approach to ranking Z-numbers [8]. In [9], the authors
propose ranking the Z-numbers by the first component, and comparing the second
components when the first components are equal. The authors [10] propose a new
approach to ranking Z-numbers using an extension of the expected utility function.
This approach is based on direct operation with Z-numbers, which is rather compli-
cated. The disadvantage of this approach is the inability to determine the membership
function of the final result. The authors of [11] correct this drawback by operating with
Z-numbers using a-cuts of its components.
The authors [9] use the first and second components of the Z-number in conjunction
with the weighted first component to determine the distance between the Z-numbers.
The weights of all three components are taken equal. In [12] it is proposed to weigh the
first component by defuzzifying the second component according to the well-known
the center of gravity method [13]. In [14], the distance between Z-numbers is deter-
mined using the Jaccard similarity measure. In [15], the distance between Z-numbers is
determined on the basis of weighted segments, which are aggregating segments for
Z-numbers, determined using all their a-cuts.
At present, there are no models for determination of rating points of objects and
models for monitoring the functioning of objects under Z-information. Eliminate this
gap is the purpose of this paper.
The rest of this paper is organized as follows: Sect. 2 gives the basic concepts and
definitions. Section 3 defines expert evaluation models under Z-information. Section 4
proposes monitoring of objects based on ratings under Z-information. Section 5 gives
conclusions.
2 Basic Concepts and Definitions

According to the [16] a fuzzy set A~ (type-I fuzzy set) is a pair x; l ~ ð xÞ ; x 2 X, where
A
~ X - universal set of A.
lA~ ð xÞ : X ! ½0; 1- membership function of A, ~
Object Monitoring Under Z-Information Based on Rating Points 1193
According to the [16] linguistic variable is fX; T ð X Þ; U; V; Sg, where X- is a name

of a variable; T ð X Þ ¼ Xl ; l ¼ 1; m - a term-set of variable X (a set of terms or names
of linguistic values X); V - is a syntactical rule that gives names of the values of a
linguistic variable X; S - is a semantic rule that gives to every fuzzy variable with a
name from T ð X Þ a corresponding fuzzy subset of a universal set U. Each of values X is
a fuzzy variable with a value from a universal set U. A linguistic variable with a fixed
set of terms is named a semantic space [16].
As a result of the theoretical studies and practical applications of semantic spaces,
the requirements to the membership functions ll ð xÞ; l ¼ 1; m of terms Xl ; l ¼ 1; m have
_
been formulated [17–19]: 1. The sets U l ¼ fx 2 U : ll ðxÞ ¼ 1g corresponding to terms
Xl ; l ¼ 1; m are points or intervals. 2. The corresponding membership functions
_ _
ll ð xÞ; l ¼ 1; m increase to the right of U l and decrease to the left of U l . 3. Membership
functions have maximum two first type discontinuity points. 4. For every
Pm
x 2 U ll ðxÞ ¼ 1. In [18] semantic scopes with properties 1–4 have been named Full
l¼1
Orthogonal Semantic Spaces (FOSS).
A Z-number is an ordered pair of fuzzy numbers, denoted Z ¼ ðA; RÞ, where A is a
fuzzy number with membership function lA ð xÞ : X ! ½0; 1 (fuzzy constraint on values
of a real-valued uncertain variable X) and R is a fuzzy number with membership
function lR ð xÞ : ½0; 1 ! ½0; 1 , which is fuzzy constraint on the measure of reliability
of the first component such as confidence, sureness, strength of belief, and probability
or possibility [2].
3 Creation of Expert Evaluation Models Under Z-

Information
The author has developed expert evaluation models for qualitative and quantitative
characteristics in linguistic terms, which will be used to formalize the first and second
components of Z-numbers. To create a model based on statistical information, expert
evaluations of qualitative characteristic X in a verbal scale with levels Xl ; l ¼ 1; m,
m 2 are used. Statistical information is represented by a sample of expert evaluations
of volume N. Suppose the objects nl ; l ¼ 1; m are assigned to a level Xl ; l ¼ 1; m.
Denote by ql ; l ¼ 1; m the minimum of nNl ¼ dl ; l ¼ 1; m and nlNþ 1 ¼ dl þ 1 ; l ¼ 1; m 1.
To formalize fuzzy numbers corresponding to terms Xl ; l ¼ 1; m, functions ll ð xÞ; l ¼
1; m are used whose graph is a trapezoid or a triangle. The parameters of these func-
tions can be specified in different ways, for example, the abscissas of their vertices or
the abscissas of the upper vertices and the lengths of the wings (left and right). In this
paper, we use the second option for setting parameters. Then [19, 20]
l1
P Pl
l1 ð xÞ ¼ 0; c1 n21 ; 0; n1 , ll ð xÞ ¼ di þ nl1
2 ; d i nl
2 ; n l1 ; n l , lm ð x Þ ¼
i¼1 i¼1
1 dm nm1 nm1
2 ; 1 dm þ 2 ; nm1 ; 0Þ l ¼ 2; m 1.
The second model is based on a direct survey of experts. The expert determines
with absolute certainty (membership function value equal to unity) typical values
1 2
sl ; sl for each term Xl ; l ¼ 1; m on the universal set U ¼ ½0; 1. Then [21]

s1 s2 s1 s2 s1 s2
l1 ð xÞ ¼ 0; s21 ; 0; 2 2 1 , ll ð xÞ ¼ s1l ; s2l ; l 2 l1 ; l þ 12 l ; l ¼ 2; m 1; lm ð xÞ ¼
1 s1 s2
sm ; 1; m 2 m1 ; 0Þ.
If in the second model the characteristic is quantitative, then the universal set
changes depending on the range of values of this characteristic. The approach to the
construction of membership functions does not change. If the range of values of a
quantitative characteristic is a limited set, then the universal set can be transformed into
a segment ½0; 1 by simple arithmetic operations.
4 Rating Points for Objects Under Z-Information

and Monitoring Based on Ratings
Consider a set of N objects for which qualitative characteristics Xv ; v ¼ 1; k are

evaluated based on linguistic scales with levels Xlm , l ¼ 1; mm ; m ¼ 1; k. Together, the
evaluated characteristics have a significant impact on the characteristic Y - «The suc-
cess of the functioning of objects», which is evaluated based on the linguistic scale
«Extremely unsuccessfully», «Unsuccessfully», «Moderately successfully», «Rela-
tively successfully», «Extremely successfully». Based on the Sect. 3 methods, we
construct the FOSSs with the names Y, Xv ; v ¼ 1; k, term-sets Yi ; i ¼ 1; 5, Xlm , l ¼
1; mm ; m ¼ 1; k respectively and membership functions li ð xÞ; i ¼ 1; 5,
llm ð xÞ; l ¼ 1; mm ; m ¼ 1; k. The fuzzy numbers corresponding to the terms Xlm , l ¼
1; mm ; m ¼ 1; k and Yi ; i ¼ 1; 5 are denoted respectively by X ~lm ; l ¼ 1; lm ; m ¼ 1; k and
~
Yi ; i ¼ 1; 5. U ¼ ½0; 1 as universal sets of all FOSSs is selected.
Let consider FOSS R with name «Reliability of information» and linguistic terms:
U - «Unlikely», NVL - «Not very likely», L - «Likely», VL - «Very likely», EL -
«Extremely likely» with membership functions lU ; lNVL ; lL ; lVL ; lEL correspondingly.
The fuzzy numbers corresponding to the lU ; lNVL ; lL ; lVL ; lEL are denoted by
~ i ; i ¼ 1; 5.
R

We will call Z-numbers X ~ i ; l ¼ 1; mm ; m ¼ 1; k; i ¼ 1; 5 the estimates of objects
~lm ; R
n n
within the framework of the characteristics Xv ; v ¼ 1; k. We denote by X ~m ; R
~m ; m ¼
1; k; n ¼ 1; N the estimate of the n-th object within the framework of the characteristic Xm .

Fuzzy number X ~mn with membership function lnm ð xÞ ¼ anm1 ; anm2 ; anmL ; anmR , n ¼ 1; N; m ¼
1; k is equal to one of fuzzy numbers X ~lm ; l ¼ 1; mm ; m ¼ 1; k, anm1 ; anm2 - the abscissas of the
vertices of the upper base of the trapezoid, which is a graph of the membership function
lnm ð xÞ, anmL ; anmR - the lengths of the left and right wings of the trapezoid respectively. Fuzzy
n
number R ~ nm ; n ¼ 1; N; m ¼ 1; k with membership function knm ð xÞ ¼ rm1 ; rm2 ; rmL
n
;;
n
rmR Þ, n ¼ 1; N is equals to the one of fuzzy numbers R ~ i ; i ¼ 1; 5. Fuzzy rating of the n-th
object within the framework of characteristics Xv ; v ¼ 1; k is defined as a Z-number:

Zn ¼ A ~n; R~ n ; n ¼ 1; N, where A ~ n ¼ x1 X ~1n . . . xk X ~kn with membership

P
k Pk P
k P
k
function ln ðxÞ ¼ xm anm1 ;
xm anm2 ; xm anmL ; ~ n ¼ r1n ; r2n ; ;
xm anmR ; n ¼ 1; N, R
m¼1 n m¼1 m¼1 m¼1 n n
rLn ; rRn Þ; r1n ¼ max r1m
n
; r2 ¼ max r2m
n
; rLn ¼ max rLm n
; rR ¼ max rRm [9]. The weights
m m m m
of the characteristics, ranked as their importance decreases, are found by the formula
xm ¼ 2ðkkm þ 1Þ
ðk þ 1Þ ; m ¼ 1; k.
In [22] the definition of weighted point H of fuzzy number A ~ with membership
function lA ð xÞ ¼ ða; aL ; aR Þ is given:
R1 A1a þ A2a
2ada Z1
0
2
ð aR aL Þ
H ¼ ¼ ða ð1 aÞaL þ a þ ð1 aÞaR Þada ¼ a þ
R1 6
2ada 0
0
Based on the definition of weighted point [22], in [23] the definition of weighted
segment ½H1 ; H2 for fuzzy number A ~ with membership function lA ð xÞ ¼
ð a1 ; a2 ; aL ; aR Þ is given:
R1 2a1 ð1aÞaL R1
H1 ¼ 2 2ada ¼ a1 16 aL ; H2 ¼ 2a2 þ ð21aÞaR 2ada ¼ a2 þ 16 aR :
0 0
We will determine the weighted point of fuzzy number A ~ with membership function
lA ð xÞ ¼ ða1 ; a2 ; aL ; aR Þ as the midpoint of the weighted segment:
a1 þ a2 ð aR aL Þ
H ¼ þ
2 12
Determine the weighted point an ; n ¼ 1; N for the fuzzy number A~ ¼ x1 X ~1n

k n
P P
k
~kn
. . . xk X with membership function ln ðxÞ ¼ xm anm1 ; xm anm2 ;
m¼1 m¼1
P
k P
k P
k P
k
xm anm1 þ xm anm2 xm anmR xm anmL
P
k P
k
xm anmL ; xm anmR Þ; n ¼ 1; N: an ¼ m¼1
2
m¼1
þ m¼1
12
m¼1
; n ¼ 1; N.
m¼1 m¼1
We will call an ; n ¼ 1; N as a crisp rating point of the n-th object. Determine the weighted

~ n ¼ r1n ; r2n ; rLn ; rRn : #n ¼ r1 þ r2 þ rR rL . Let us
n n n n
point #n ; n ¼ 1; N for the fuzzy number R
2 12
determine the distance between two Z-numbers Z1 ¼ A ~1; B ~ 1 ; Z2 ¼ A ~2; B~ 2 with
weighted points ða1 ; b1 Þ; ða2 ; b2 Þ:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
d ðZ1 ; Z2 Þ ¼ 13 ða1 a2 Þ2 þ 13 ða1 b1 a2 b2 Þ2 þ 13 ðb1 b2 Þ2 . The distance defined in
this way is a partial case of a more general distance, which is given in [12].

Consider Z-numbers Zi5 ¼ Y~i ; R ~ 5 ; i ¼ 1; 5 and rating of the n-th object

Zn ¼ A ~n; R~ n ; n ¼ 1; N. Calculate d ðZn ; Zi5 Þ; i ¼ 1; 5; n ¼ 1; N. If, d ðZn ; Zl5 Þ ¼
max d ðZn ; Zi5 Þ; i ¼ 1; 5; n ¼ 1; N, then the state of the n-th object is determined by the l-
i¼1;5
th level of the scale: «Extremely unsuccessfully», «Unsuccessfully», «Moderately
successfully», «Relatively successfully», «Extremely successfully».
When monitoring the functioning of objects based on their rating, we will use the
1 1 2 2 2
ranking of Z-numbers proposed in [12]. Let Zn1 ¼ A ~ ;R~ n ; Zn ¼ A ~ ;R
~ n ; n ¼ 1; N be
n n
ratings of the n-th object for the periods 1 and 2, an ; an ; #n ; #n ; n ¼ 1; N- weighted
1 2 1 2
points for the fuzzy numbers A ~1; A

~2; R ~ 1n ; R
~ 2n ; n ¼ 1; N. If a2n [ a1n , then the condition of
n n
the n-th object has improved, if an \a1n , then the condition of the n-th object has
2
become worse. If a2n ¼ a1n , then consider the relationship between #1n ; #2n . If #2n [ #1n ,
then the condition of the n-th object has improved, if #2n \#1n , then the condition of the
n-th object has become worse, if #2n ¼ #1n , then the condition of the n-th object has not
changed.
5 Conclusions
Rating evaluation model of objects within the framework of several qualitative char-
acteristics and model of monitoring of their functioning based on rating points under Z-
information are developed in this paper. The relevance of the studies is confirmed by
the need to develop such models in the conditions of their complete absence.
Ratings of objects are presented as Z-numbers. The first and second components of
the Z-numbers are values of Full Orthogonal Semantic Spaces. Operations with Z-
numbers are based on weighted points that aggregate the information of numbers over
all a-cuts of their components. Recognition of the state of objects is carried out based
on the choice of the minimum distance from the rating of the object (in the form of a Z-
number) to the levels of the linguistic scale used to evaluate the success of the func-
tioning of objects. The distance between the Z-number is determined on the basis of
weighted points. When determining the distance, both components of Z-numbers are
taken into account. Monitoring the functioning of objects is carried out on the basis of a
comparison of the first components of the Z-numbers, and in case of their equality on
the basis of the second components of the Z-numbers.
Thus, this paper makes it possible under Z-information to realize a rating evalua-
tion, analyze the results obtained and make appropriate decisions based on these
results. Studies further suggest the following: application of the developed model to
practical data on the assessment of green plant species in the urban conditions of the
large cities and the development of fuzzy arithmetic for Z-numbers and the distance
between them based on weighted segments.
References
1. Aliev, R.A., Zeinalova, L.M.: Decision-making under Z-information. In: Human-Centric
Decision-Making Models for Social Sciences, pp. 233–252 (2013)
2. Zadeh, L.A.: A note on Z-numbers. Inf. Sci. 14(181), 2923–2932 (2011)
3. Kang, B., Wei, D., Li, Y., Deng, Y.: Decision making using Z-numbers under uncertain
environment. J. Inf. Comput. Sci. 7(8), 2807–2814 (2012)
4. Aliev, R.A., Alizadeh, A.V., Huseynov, O.H.: The arithmetic of discrete Z-numbers. Inf.
Sci. 1(290), 134–155 (2015)
5. Aliev, R.A., Huseynov, O.H., Zeinalova, L.M.: The arithmetic of continuous Z-numbers.
Inf. Sci. 373, 441–460 (2016)
6. Aliev, R.К., Huseynov, O.H., Aliyeva, K.R.: Aggregation of an expert group opinion under
Z-information. In: Proceedings of the Eighth International Conference on Soft Computing,
Computing with Words and Perceptions in System Analysis, Decision and Control, pp. 115–
124 (2015)
7. Yager, R.R.: On Z-valuations using Zadeh’s Z-numbers. Int. J. Intell. Syst. 3(27), 259–278
(2012)
8. Sharghi, P., Jabbarova, K.: Hierarchical decision making on port selection in Z-environment.
In: Proceedings of the Eighth International Conference on Soft Computing, Computing with
Words and Perceptions in System Analysis, Decision and Control, pp. 93–104 (2015)
9. Wang, F., Mao, J.: Approach to multicriteria group decision making with Z-numbers based
on topsis and power aggregation operators. Math. Probl. Eng. 2019, 1–18 (2019)
10. Aliyev, R.R., Talal Mraizid, D.A., Huseynov, O.H.: Expected utility based decision making
under Z-information and its application. Comput. Intell. Neurosci. 3, 364512 (2015)
11. Dutta, P., Boruah, H., Ali, T.: Fuzzy arithmetic with and without a-cut method: a
comparative study. Int. J. Latest Trends Comput. 2(1), 99–107 (2011)
12. Kang, B., Wei, D., Li, Y., Deng, Y.: A method of converting Z-number to classical fuzzy
number. J. Inf. Comput. Sci. 9(3), 703–709 (2012)
13. Yager, R.R., Filev, D.P.: On the issue of defuzzification and selection based on a fuzzy set.
Fuzzy Sets Syst. 5(3), 255–272 (1993)
14. Aliyev, R.R.: Similarity based multi-attribute decision making under Z-information. In:
Proceedings of the Eighth International Conference on Soft Computing, Computing with
Words and Perceptions in System Analysis, Decision and Control, pp. 33–39 (2015)
15. Poleshchuk, O. M.: Novel approach to multicriteria decision making under Z-information.
In: Proceedings of the International Russian Automation Conference (RusAutoCon),
p. 8867607 (2019). https://doi.org/10.1109/rusautocon.2019.8867607
17. Ryjov, A.: Description of objects in human-machine information systems, application of
fuzzy systems. In: Proceedings of the International Conference on Application of Fuzzy
Systems, pp. 246–249 (1994)
18. Ryjov, A.P.: The concept of a full orthogonal semantic scope and the measuring of semantic
uncertainty. In: Proceedings of the Fifth International Conference Information Processing
and Management of Uncertainty in Knowledge-Based Systems, pp. 33–34 (1994)
19. Poleshchuk, O.M., Komarov, E.G., Darwish, A.: Comparative analysis of expert criteria on
the basis of complete orthogonal semantic spaces. In: Proceedings of the 19th International
Conference on Soft Computing and Measurements, SCM 2016, pp. 369–373 (2016)
20. Darwish, A., Poleshchuk, O.: New models for monitoring and clustering of the state of plant
species based on sematic spaces. J. Intell. Fuzzy Syst. 3(26), 1089–1094 (2014)
21. Poleshchuk, O., Komarov, E.: The determination of rating points of objects and groups of
objects with qualitative characteristics. In: Proceedings of the Annual Conference of the
North American Fuzzy Information Processing Society – NAFIPS 2009, p. 5156416 (2009)
22. Chang, Y.-H.: Hybrid fuzzy least- squares regression analysis and its reliability measures.
Fuzzy Sets Syst. 119, 225–246 (2001)
23. Poleshuk, O.M., Komarov, E.G.: Multiple hybrid regression for fuzzy observed data. In:
Proceedings of the Annual Conference of the North American Fuzzy Information Processing
Society – NAFIPS 2008, p. 4531224 (2008)
An Intelligent Decision Support System:
Application of Fuzzy Tools and System
Dynamics Modelling
Merve Kadan(&), Gökhan Özkan, and Mehmet Hilmi Özdemir
STM Savunma Teknolojileri Mühendislik ve Ticaret A.Ş., Ankara, Turkey

[email protected]
Abstract. This study aims to create a decision support model for a country to
improve its economic capacity. The Economic Complexity Index, a worldwide
index, was used for listing potential products that the country has not produced
so far, yet is deemed to be close to produce with the extant capabilities. Then,
using the Fuzzy Analytical Hierarchy Process, products that will provide the
highest added value among the potential products were selected and strategic
importance weights were calculated for. Subsequently, an economic model was
developed with system dynamics methodology where strategic importance
weights are used for each potential product. The model may provide a strategic
level foresight capability for the decision-makers in simulating the behavior
patterns and future value of a potential product on the economy of the country.
It is believed that, this study combining the fuzzy analytic hierarchy process
and system dynamics modelling methodology is an example of intelligent
decision support system and will be beneficial to the literature and the
application.
Keywords: Fuzzy Analytical Hierarchy Process System dynamics

modelling Intelligent decision support systems
1 Introduction
Adam Smith states in his book “The Wealth of Nations” that the wealth of a country is
related to its expertise [1]. In this context, the abilities of nations can be associated with
their economic capacities. Using that principle, this study aims to create a decision
support model for a country to improve its’s economic capacity. Many studies in the
literature provide concrete evidence that decisions made by human intuitive judgments
are far from optimal solution [2]. Therefore, data based decision support models are
created to provide decision support while making strategic decisions [3]. The country
subject to this study is one of the least developed countries according to the United
Nations [4]. It is a fact that the weakest point of the investment policies and strategies
in the least developed countries is mostly based on personal thoughts and feelings. We,
with this study, aim to show the subject country a systematic way to make more
rational economic investments on the right products. Potential products that will
contribute to economic capacity of the country, but are not yet produced by the country,
have been determined through the Economic Complexity Index [5]. Fuzzy Analytical
1200 M. Kadan et al.
Hierarchy Process (FAHP) was used to calculate strategic importance weights for these
products. As a final step a dynamic model of country’s economic structure is prepared
using system dynamics modelling methodology.
In future studies, it is planned to expand the scope of the model for more com-
prehensive decision support system covering national capacity in different domains
such as military, health, and energy. This study is believed to be an example of an
intelligent decision support system as it combines the system dynamics methodology
with fuzzy logic, and therefore will be useful for the literature and practice.
In the following section of the study, adopted methods, data sources and imple-
mentation are described and the final section outlines the results.
2 Methodology
Developing a decision support system for the country requires going through various
phases as seen in Fig. 1.
Determining Preparing the

SelecƟon of
strategic economic
potenƟal
importance model for
products via
weights of strategic level
Economic
potenƟal foresight using
Complexity
products using System
Index
FAHP Dynamics
Fig. 1. Phases of the study
At the outset of the study, the potential products for the country have been extracted
from the Atlas of Economic Complexity database, which provides data about global
trade flows through markets and new growth opportunities for each country [5]. Then,
FAHP used for determining the strategic importance weights of these potential prod-
ucts. As a final step the economic model of the country prepared using system
dynamics methodology.
2.1 Fuzzy Analytical Hierarchy Process (FAHP)

Fuzzy Analytical Hierarchy Process (FAHP) is the advanced version of the classical
AHP that is frequently used in solutions of multiple-criteria decision-making problems
[6]. The FAHP method offers such advantages as the ability to take the preferences of
decision-makers into account, the ability to evaluate qualitative and quantitative vari-
ables together, and possess a flexible structure [7]. FAHP has six steps as shown in
Fig. 2.
An Intelligent Decision Support System 1201
Fig. 2. FAHP steps [8]
Within the scope of FAHP, having defined the problem the next step is creating a
pairwise comparison matrix among all criteria [6]. Subsequently, each column of the
matrix is normalized.
aij
aij ¼ ð1Þ
maxaij
Then, consistency check is done by determining eigenvector of comparison matrix.
aij
wi ¼ ð2Þ
n
Afterwards, largest eigenvector is clarified.

X
kmax ¼ ð nÞ
GM1nni X ð3Þ
Then, Consistency Index (CI) can be calculated as shown in below formula. If

CI 0, than comparison matrix is consistent.
kmax n
CI = ð4Þ
n1
After checking the consistency of judgments for matrix in each level, the next step
is to calculate the weight vector of the alternatives [7]. To set up triangular fuzzy
number, comparison values are transformed to fuzzy numbers using FAHP scale,
which has three values, the lowest value (L), middle value (M), and highest value (U).

Mij ¼ lij ; mij ; uij ð5Þ
Within the scope of weight value calculation step, fuzzy synthesis value (Si) is
calculated.
X
n
1
Si ¼ Mkjj X hP Pn i ð6Þ
m
j¼1 Mij
1
j¼1 i¼1
The degree of possibility of M2 M1 is:

8
< 1; if M2 M1
VðM2 M1 Þ ¼ 0; if L1 U2 ð7Þ
: l1 u2
ðm2 u2 Þðm1 l1 Þ ; etc:
Having obtained Si, weight values of the fuzzy vector are calculated as d′.
d 0 ðAiÞ ¼ minVSi [ Sk ð8Þ
The weight vector is given by:
W ¼ ðd ðA1 Þ; d ðA2 Þ; . . .; d ðAn ÞT

0 0 0 0
ð9Þ
As a final step the normalized weight vectors are determined as:
W ¼ ðdðA1 Þ; dðA2 Þ; . . .; dðAn ÞT ð10Þ
2.2 System Dynamics Modelling

System dynamics is a mathematical modeling technique that is widely used to
understand the nonlinear behavior of complex problems [9]. This technique is based on
the systems thinking approach and uses computerized simulations to analyze depen-
dencies, relationships, or interactions among components composing a system. System
dynamics modeling technique operates continuously to achieve results with high pre-
cision [10]. Calculations are repeated at each unit of time, and each calculation depends
on the previous calculation. This causal character is one of the strengths of the system
dynamics method. Consequently, the system dynamics method creates a reliable visual
reality for critical and complex problems. Creating causal loop diagrams, building
stock and flows, and inserting input values via converters can be listed as the main
modelling actions in system dynamics.
In order to reflect the nonlinear and complex nature of economic structure of the
country - the subject of this study - system dynamics method was selected.
3 Implementation
Economic Complexity Index used for identifying potential products as seen in Table 1.
Fuzzy AHP method was used to convert these potential products into comparable
structure and determine strategic importance weights for each potential product. Then,
in order to create a decision support system for long term foresight, system dynamics
modelling methodology was used.
Table 1. Potential products.

Product number Product name
1 Calendaring or other rolling machines, other than for metals or glass
2 Compression-iginition intemal combusion piston engines
3 Self-propelled railway coaches
4 Appliances for thermostically controlled valves
5 Electric soldering machines
6 Lathes for removing metal
7 Parts and accessories for metal working
8 Knives and blades for machines
9 Interchangeable tools for machines
10 Industrial electric furnaces
3.1 Fuzzy AHP Implementation

The strategic importance weights of potential products were calculated with FAHP
using expert opinions. The final scores obtained with the FAHP are shown in Table 2.
Table 2. FAHP weights of potential products

Weight
Product 1 0.22
Product 2 0.24
Product 3 0.98
Product 4 0.74
Product 5 0.15
Product 6 0.48
Product 7 0.20
Product 8 0.95
Product 9 0.10
Product 10 0.09
As a result of FAHP, Product 3, Product 4 and Product 8 were identified as the

products having highest strategic importance weights.
3.2 System Dynamics Modelling Implementation

Firstly, a causal loop diagram was built at the outset of the modeling to reflect the
causality relations among parameters defining economic capacity of the country. Then,
a stock and flow diagram were built to model the behavior of the economic capacity.
The strategic importance weights and global trade size of the potential products are
used inserted into the model via converters. When we run the model, we get a foresight
about the behavior of the economic capacity of the country in a time scale. An interface
outlook is provided in Fig. 3, on which users can get the economic capacity trends
creating what-if scenarios by tuning the strategic weights of the products.
Fig. 3. System dynamics interface outlook
The developed system dynamics model shows that Product 3, Product 4, and
Product 8 have highest added-value potential to the country.
4 Conclusion
This study aims to create a decision support model to improve the economic capacity of
one of the least developed countries. The combination of the Economic Complexity
Index, Fuzzy Analytical Hierarchy Process, and System Dynamics modelling
methodologies have been used in the study, which reflect interdisciplinary feature.
With the developed model, strategic level foresight about the impacts of potential
products on the country’s economy can be seen.
In future studies, it is planned to expand the scope of the model to reflect different
domains of national capacity such as military, health and energy, and thus to establish a
more comprehensive decision support system. This study is believed to be an example
of a smart decision support system and will be useful for literature and practice.
References
1. Grieve, R.H.: Adam Smith’s ‘Wealth of Nations’: the Legacy of a Great Scottish Economist,
Understanding the Scottish Economy (1983)
2. United Nations: Development Programme Human Development Reports. http://hdr.undp.
org/en/content/2019-human-development-index-ranking. Accessed 20 Feb 2020
3. The Atlas of Economic Complexity: http://atlas.cid.harvard.edu/. Accessed 24 Feb 2020
4. Druzdzel, M.J., Flynn, R.R.: Decision Support Systems, Encyclopedia of Library and
Information Science (2002)
5. Marin, G.: Decision Support Systems, Faculty of Computer Science for Business
Management Romanian American University (2008)
6. Noor, A., Fauadi, M., Jafar, F.A.: Fuzzy analytic hierarchy process (FAHP) integration for
decision making purposes: a review. J. Adv. Manuf. Technol. 11, 139–154 (2017)
7. Chen, S.: A fuzzy AHP approach for evaluating customer value of B2C companies.
J. Comput. 6(2), 224–231 (2011)
8. Putra, M.S.D., Andryana, S., Fauziah, A.G.: Fuzzy analytical hierarchy process method to
determine the quality of gemstones. In: Advances in Fuzzy Systems (2018)
9. Sterman, J.: Business Dynamics, System Thinking and Modeling for a Complex World, vol.
19 (2000)
10. Forrester, J.: System dynamics, systems thinking, and soft OR. Syst. Dyn. Rev. 10, 245–256
(1994)
Fuzzy Uncertainty Modelling in Cost and Cash
Flow Forecasting in Project
Dorota Kuchta(&) and Adam Zabor
Department of Management Systems and Organisation Development,

Wrocław University of Science and Technology,
Wybrzeże Wyspańskiego 27, 50-370 Wrocław, Poland
{Dorota.kuchta,adam.zabor}@pwr.edu.pl
Abstract. Numerous projects (e.g. more than 50% of IT projects) are not fin-
ished within budget and cause serious financial problems to the organisations
implementing them. It is thus important to be able to predict the cost of projects
and cash flows related to them early enough, in order to be able to assess with
the necessary anticipation whether the necessary financial means will be
available on time and if not, to take in time the necessary measures to solve the
menacing problem. In the paper the sources of uncertainty with respect to
project cost and cash flows will be identified and their modelling by means of
fuzzy sets will be proposed. Such issues neglected in the project management
literature as various taxes, duties and impositions, which belong to the area in
which project managers are not experts and where they do not follow the
detailed regulations, will be discussed too. They are of high importance, because
legal changes in this area come often as a surprise for project managers and lead
to serious liquidity problems. The approach will be illustrated with a real world
case study, in which one of the authors was the member of the project team.
Keywords: Project cash flow Project finance Project budgetary risk
1 Introduction
It is well known that projects are often terminated at a much higher cost than their
initial budget foresaw. In Chaos Report (2015) we find the information that over 55%
of IT projects exceed their budgets. But the total project budget is only one side of the
problem. Project managers have to face also another one, not quite independent of the
budget problem but still of a different nature: that of project cash flows. During the
whole project course cash payments have to made and the project manager needs to
have enough cash inflows or resources for that. If the necessary payments are not made,
the project may fall into serious problems. Various penalties and interests or even
broken contracts may result in even higher project total cost or in the impossibility of
achieving certain project outcomes. On the other hand, often project cash flows are not
known exactly in advance.
That is why a lot of research has been made on project cash fuzzy modelling.
However, the existing approaches have two basic drawbacks:
Fuzzy Uncertainty Modelling in Cost and Cash Flow Forecasting in Project 1207
• the modelling is performed on a high level of generality, i.e. the project cash flows
are modelled on the level of compound magnitudes, without analyzing their indi-
vidual components and relations among them;
• the modelling is performed only once, in the project planning phase. The fuzzy
estimates of project cash flows are not updated later, which may result in their
significant obsolescence and inadequacy.
Thus, the objective of this paper is to propose a method of project cash flow fuzzy
modelling which would be a trial to make up for the above drawbacks. The method will
be described and illustrated by means of a case study.
The outline of the paper is thus as follows: In Sect. 2 main sources of uncertainty in
projects which may influence project cash flows are described. In Sect. 3 the existing
research on fuzzy modelling of project cash flows is summarized and evaluated. In
Sect. 4 we draw the reader’s attention to certain aspects of fuzzy modelling which are
not used in the existing literature on fuzzy modelling of project cash flows, but which
might help to overcome its above mentioned shortcomings. In Sect. 5 we describe our
proposal of the approach to fuzzy modelling of project cash flows. In Sect. 6 the
proposal is illustrated with a real world project. The paper terminates with some
conclusions.
2 Sources of Uncertainty in Project Finance
Project managers face many types of uncertainty at the planning stage of project. It
causes difficulties in preparing an optimal plan. The sources of these uncertainties vary.
Uncertainty related to financial categories present in the project plans or affecting the
values expressed in these plans are of key importance. Below we present a proposal to
divide the sources of uncertainty and extract from them some key financial factors.
There are many classifications of sources of uncertainty in the literature. It is worth
quoting the classification from Atkinson et al. (2006), which was created on the basis of
research carried out by the Rethinking Project Management Network. It is an organi-
zation of academics from fifteen UK universities and many senior practitioners from
private, public and voluntary sector organizations. That is mostly why we consider it as
worth mentioning.
Authors of the aforementioned paper divide uncertainties in projects into three
basic groups:
1. Uncertainty in estimates.
2. Uncertainty associated with project parties.
3. Uncertainty associated with stages in the project life cycle.
Uncertainty in estimates – the most obvious field of uncertainty regarding cost,
duration or quality forecasts. The sources of uncertainty in this case are e.g. lack of
data, insufficient detail. There are also factors related to the interpretation of these data,
such as lack of experience, lack of appropriate procedures or ignorance of the
requirements. Uncertainty associated with project parties – this category contains
sources as level of performance, quality, availability. It is possible that the parties will
1208 D. Kuchta and A. Zabor
perceive the project’s goals and their performance differently. Uncertainty associated
with stages in the project life cycle – these are typical issues depending on the stage of
project life cycle. They can vary according to project specification.
Uncertainty in estimates are key to this paper, from which we extract financial
sources of uncertainty. We propose a division of these factors into macroeconomic and
microeconomic ones:
1. On a macroeconomic scale there are:
1:1. Inflation,
1:2. interest rates,
1:3. taxes.
2. On a microeconomic scale:
2:1. costs/expenditures amounts,
2:2. terms of expenses and receipts,
2:3. conditions of financing.
The first group are factors that are not considered directly when planning a project
but will affect the quantities located in a project plan. It is worth bearing in mind their
impact on budgeted numbers.
Inflation is the first example of such factors. Inflation has an impact on all figures in
the project budget and thus on revenues and costs. Interest rates can affect the cost of
financing. If the project budget involves a loan with variable interest rates, its costs may
vary. This may have both positive and negative effects. Taxes are a category that is
usually not characterized by short-term volatility. At least when it comes to the scale of
taxation which can basically be described as certain data. However, there are some
situations that can be an exception, in case of changes in the law.
The second group are the financial factors on a microeconomic scale. Their possible
volatility should be referred directly to the shape of the budget. These factors partly
result from the previously described group of macroeconomic factors. The first factor is
costs – or to be more precise – expenditures. We propose to consider a category of
expenditures rather than costs as costs refer more to an accounting point of view. This
approach allows to broaden it for example by payment terms. It is obvious that the
prices of certain goods and services may change over time. This results either from
inflation, as previously described, or from changes in prices on the markets. Therefore,
possible changes in prices should be considered at the planning stage. The prices
volatility can be minimized by negotiating contracts for services possibly early.
Another factor connected with expenditures is payment terms for goods and services.
The liquidity of the entity running the project will depend on the payment deadlines
and their timing. In the case of earlier payments, there may be a need to finance them
with a loan, which will generate additional costs.
The above issues lead to the last category mentioned – the conditions for loans or,
more broadly, for obtaining financing. The exact terms of obtaining a loan may not be
known at the planning stage, when managers have not yet entered negotiations with
financial institutions. In the case of projects organized with the participation of grants
from public funds, which can often take place in the example considered here,
uncertainty to the amount of funding received may exist.
3 Project Cash Flows and Modelling of Their Uncertainty –

State of Art
In the previous section it was shown how many uncertainty factors influence project
finance. It is thus clear that if we want to analyse project cash flows, we have to identify
all the most important factors. But the papers in the scientific literature which refer to
modelling of uncertainty of project cash flows by means of fuzzy numbers (Cheng et al.
2010; Djatmiko et al. 2019; Etgar et al. 1997; Fathallahi and Najafi 2016; Maravas and
Pantouvakis 2019; Maravas and Pantouvakis 2012; Mohagheghi et al. 2017; Shavandi
et al. 2012; Tabei et al. 2019; Yu et al. 2017) use an expert base approach to express
cash flows by means of fuzzy numbers, asking the expert to give fuzzy estimates for
total cash flow planned for a certain moment or linked to a project activity. No types of
cash flow are distinguished and no factors influencing them. But if the factors
influencing cash flows are so numerous and diversified (on the microeconomic and
macroeconomic scale, referring to various project parties and to various project phases
– which we explained in the previous section), it cannot be possible for an expert, even
a very experienced one, to grasp all of them in his or her fuzzy estimation. Thus, in our
opinion the only way to model uncertainty in project cash flows in an efficient way is to
try to identify their types, their components and all the factors influencing them.
The other reproach we have to formulate with respect to the existing literature on
cash flows modeling in projects, is a complete disregard of dynamism. Fuzzy models of
cash flows are done once and for all and the fact that projects are element of a dynamic
world, where things do change almost continuously, is fully ignored. And the basic
rules of project risk management and project control (Project Management Institute
2008) require that risk and uncertainty have to be identified, quantified and then
managed and controlled throughout the whole project. In our opinion ignoring this rule
may seriously distort the whole cash flow modelling.
A typology of project cash flows is found e.g. in Gatti (2018). The most important
categories are: (a) costs of fuel and other consumables, (c) costs of obtaining any other
materials, supplies, utilities or services for the project, (d) franchise, licensing, property,
real estate franchise, licensing, property, real estate, sales and excise taxes (e) employee
salaries, wages and other employment-related costs (f) insurance cost and (g) capital
expenditure. Each type of project cash flow is subject to different rules, behaves in a
different way and is influenced by other factor. Thus, in our opinion they cannot be
treated as one total. An additional problem is that, within each category, we have to
adjust individual cash flows, subtracting the quantities unpaid in the given moment
(with delayed payment, e.g. in case of materials purchased on credit) and adding the
debts relating to the given category and paid in the given moment (e.g. payment of an
invoice for materials purchased some time ago).
In conclusion, we have to state that due to the complexity of components of project
cash flows and factors influencing them, among which the time is a very important
factor, it is in our opinion necessary to subject them to a much more thorough analysis
than it has been done in the literature so far. An approach to such an analysis is proposed
in the Sect. 5. In Sect. 4 we will draw the readers attention to the fact that certain
elements of fuzzy modelling should be used in the proposed approach, in order to widen
its possibilities of real world reflection. These useful fuzzy modelling theory elements
are not taken into account in the existing literature on project cash flow fuzzy modelling.
4 Non-classical Fuzzy Theory Elements to Be Used

in the Proposed Approach
In any modelling based on fuzzy notions, the known elements of the fuzzy modeling
theory should be used. Fuzzy modelling is a vast area with a lot of findings (see e.g.
(Bector and Chandra 2005); (Novák et al. 2016); (Bezdek et al. 1999)), but in most
applications of fuzzy sets to project cash flow modelling only its basic elements are
used, which makes fuzzy modelling much less powerful. Here we will mention two
example elements of fuzzy modelling, which go beyond the fuzzy sets theory used in
the existing literature on modelling project cash flows.
The first problem are arithmetic operations of fuzzy sets. In all the existing appli-
cations of fuzzy modelling to project cash flow modelling all arithmetic operations *

are defined for k-cuts of fuzzy numbers A ~ and B~ (Ak ¼ akl ; aku ; k ¼ ½0; 1, Bk
analogously) as leading to a fuzzy number defined through the k-cuts (1).
k k k k k
min al bl ; al bu ; au b; aku bku ; max akl bkl ; akl bku ; aku b; aku bku ð1Þ
It is thus assumed that all the couples

x y; x 2 Ak and x 2 Bk ð2Þ
belong to the k-cut of the result. But sometimes non-classical fuzzy operations should
be used, because in fact not all the couples (2) are possible. This is the case for example
when we are dealing with a situation which can be informally described as (big with
big, small with small): akl and bku occurring together is impossible – because the fuzzy
numbers A ~ and B
~ refer to two correlated magnitudes, where in both the pessimistic or in
both the optimistic scenario can occur. In such a case another fuzzy subtraction should
be used, defined in Gani and Assarudeen (2012), in order to render properly the fuzzy
result of the subtraction.
Another issue which should be taken into account are so called linguistic modifiers
(Bouchon-Meunier and Jia 1992). If we have a fuzzy number expressing e.g. the notion
“big”, we can use linguistic modifiers to transform this fuzzy number into one
expressing the notion “vey big”, “averagely big” etc. Similarly, if we have a fuzzy
number expressing the notion “around 10”, we can transform the membership function
to render e.g. the idea “still around 10 but bigger values should have a higher possi-
bility”. Such transformations allow us to modify the existing fuzzy evaluation on the
basis of new information.
In the next section we will show how the wide possibilities of fuzzy modelling can
be used in modelling project cash flows. The method will require a lot of effort, so –
according to the philosophy of project risk and uncertainty management – it should be
applied only to relatively high cash flows, which may have a considerable influence on
project liquidity, and thus its ultimate success.
5 Proposal of an Approach to Project Cash Flow Fuzzy

Modelling
We propose to use the following approach to the modelling of project cash flows:
Step I. Choose a moment “zero” before project start;
Step II: Identify all the types of cash flow, separating them both with respect to the
foreseen moment of occurrence (i.e. cash flows which are not connected formally
(e.g. by one single invoice) should be treated separately) and with respect to their
types (thus if two cash flows are expected to be linked by a single invoice, but are of
a different type – see Sect. 3 – they should also be separated).
Step III: For each individual cash flow CF defined in Step II identify all its com-
ponents which cannot be divided any more. For example, a cash flow “payment for
materials purchase” (PMP) expected to take place in a certain moment of time
should be expressed e.g. as formula (3) shows (no fuzziness is introduced yet):
PMP ¼ np p npl p þ npe pe ð3Þ
where np: number of units purchased in the considered moment, p: purchase price, npl:
number of units purchased in the given moment which will be payed later, npe: number
of units purchased earlier which will be paid in the considered moment, pe – earlier
purchase price.
Let denote the formula defining CF as CF cCF CF CF CF
1 ; . . .; cnCF , where c1 ; . . .; cnCF are
individual components of the formula (e.g. in case of (3) we have components
np; p; npl; npe; pe).
Step IV: For each CF define factors that influence the individual components
cCF CF
1 ; . . .; cnCF . The set of factors for each component will be denoted as F cCF i ;i ¼
n o
CF;i
1; . . .; nCF and will contain factors fj ; j ¼ 1; . . .; m. These factors should be
identified using risk identification methods from the project risk management area
(Kerzner 2017; PMI 2009), with focus on financial factors, and will comprise such
factors as inflation, current general economic situation, current situation of the
suppliers etc.
Step V: For each CF, taking into account F cCF i ; i ¼ 1; . . .; nCF and experts
opinions, define the best (according to the knowledge of the present moment) fuzzy
estimates of cCF CF
1 ; . . .; cnCF and the best arithmetic operations to be used in

CF c1 ; . . .; cnCF : Use the results of CF c1 ; . . .; cCF
CF CF CF
nCF to make managerial
decisions in the project.
Step VI: If the project is still in a phase where there is risk and uncertainty with
respect to its final outcome, choose a control moment t and go to Step VII.
Otherwise STOP.
Step VII: Take all the cash flows CF which in moment t are still
uncertain. For each
CF CF CF
of them reidentify F ci ; i ¼ 1; . . .; nCF and recalculate CF c1 ; . . .; cnCF . Make
respective decisions. Go to Step VI.
Such an approach, to be used for large project cash flows, allows to have control
over their fuzzy estimates – in the sense that they will be generated:
• using expert opinions not with respect to complex, compound magnitudes, where it
is doubtful if the expert can take into account all the complexity of factors
influencing the magnitudes, but with respect to individual components of cash flows
and arithmetic operations on them;
• not only in one moment before the project start, but on an ongoing basic, in regular
control moments, which is in line with the basic philosophy of project management,
and additionally, takes into account the uncertainty linked to the project life cycle
(see Sect. 3)
The proposed approach will be now illustrated with a real-world case study.
6 Case Study
The example project will be a 3 days long scientific conference, due to take place in
June 2020 in an European large city, organized jointly by two universities, for a
number of participants equal roughly to 40–50. The conference is a cyclical one – it
takes part every two years in another European city. In such a project there may occur
all the sources of uncertainty enumerated in Sect. 3. Uncertainty due to various parties
taking part in the project may result from a potential divergence between goals of
organizational committee and scientific committee. The organizational committee may
focus mainly on performance described in financial terms when the scientific com-
mittee may put emphasis on scientific level of the conference. As far as project phases
are concerned, uncertainty may be linked to: determining the number of potential
participants in the planning phase, effective communication between organizers in
preparation phase, effective control in implementation phase (during the conference),
capturing knowledge and experience in the summary phase after the conference.
Inflation may affect planning accuracy in some ways. A conference is usually planned
at least several months in advance. During this period, even in conditions of standard
inflation, consumer goods prices may increase which will entail an increase in the
prices of goods and services. In the conditions of higher inflation these problems will
obviously be widened. On the other hand, price uncertainty will cause a problem with
determining the amount of revenues understood as contributions from participants. The
registration fee set at the planning stage will have to cover the costs of organizing the
conference. Attention has to be paid to taxes. There was a case when a change in tax
regulations resulted in a change in VAT rates paid by the organization carrying out the
project. A university organizing a conference, that had not previously been a VAT
payer, was assigned to pay VAT by change in national law. That situation caused
significant discrepancies between previously budgeted numbers and real expenses and
led to serious problems with accounting for the conference. Thus, there are important
uncertainties and risks linked to a conference project which may seriously affect the
formal closure of the conference, especially considering that conference organisers are
usually not experts in finance.
We will illustrate the approach from Sect. 5, above all with respect to one of the
most important cash flows of the conference: participation fees cash inflows (PF).
In January 2020 the following evident formula was formulated:
PF ¼ np pf ð4Þ
where np stands for the participants number and pf for the participation fee. In the same
moment triangular fuzzy numbers for both np and pf were delivered by experts, on the
basis of the numbers known from the previous conferences. It was set: np ¼
ð30; 40; 50Þ and pf ¼ ð1200; 1400; 1600Þ (in local currency). The difference between
the two fuzzy sets was that the concrete realisation of np was dependent on both
exterior and interior factors (on the publicity, but also on the situation at other uni-
versities or on the offer of other conferences), but the concrete realisation of pf was the
independent decision of the organisers, which had to be taken on the basis of the
relation between (4) and the outflows, which we modelled using the same procedure.
We will not enter here into the modelling of the outflows. Let us only mention that their
two big components were preparation of the conference materials and the organisation
of the conference sessions, breaks and social events. While choosing the subtraction
operator for the inflows and outflows, the subtraction from (Gani and Assarudeen 2012)
had to be chosen, as more participants mean higher outflows for coffee breaks and
social events.
In February 2020 the organizers had access to more information about the possible
number of participants, as they had been sent positive signals from the participants
from previous conferences. So it was possible to apply a modifier “still about forty, but
more than it was thought before” and update np to the trapezoidal fuzzy number
ð35; 40; 45; 55Þ. At the same time, in the fuzzy numbers representing cash outflows the
supports were narrowed, as the uncertainty linked to them became lower (after some
negotiations with the future suppliers).
In March 2020 it became clear that the coronavirus pandemic had started. The
whole concept of the conference had to be changed. The organizers were forced to
change the conference into an on-line event. The number of participants became less
certain than it was in February, both in the negative and positive direction (the new
conference form could be more attractive for some potential participants, because it
would be cheaper, but for other potential participants, who like personal exchange, it
might be less attractive). Thus np was updated to the trapezoidal fuzzy number
ð30; 40; 45; 60Þ. Of course, the conference fee had to be updated. It was clear it was to
be lower, but on the other hand its uncertainty degree (the width of the fuzzy number
support) became bigger, as prices of the conference proceedings became more
uncertain – due to new circumstances – and a tool for remote conferences was still to be
sought for among different offers (the prices of such tool were bound to increase,
because of an increased demand). Thus, pf was updated to ð800; 1000; 1300Þ.
In the moment when this paper was finished, the story was still going on. But its
previous course shows clearly that estimating the highest cash flows in projects cannot
be performed on a high level of generality and once for all in a fixed moment. This
process has to be detailed and repetitive. Only than the risk of running out of cash in the
project can be properly managed.
7 Conclusions
In the paper we propose a new approach to fuzzy modelling of project cash flows,
much more detailed and much more attentive to changes in the project and its context
than the existing approaches. Its application is linked to much effort, that is why it is
dedicated to important cash flows, on which the project fate may depend. For such cash
flows the application of the proposed procedure may be of primordial importance,
because lack of cash in any project implementation moment may be among the causes
of project failure – or at least serious problems.
The proposal is still in its initial stage. It needs real world case studies, where
project managers and experts would cooperate on its practical verification. Ideally, it
should undergo a development similar to that of project risk management systems:
where the process of risk identification, evaluation, analysis, management and constant
updating is systematically embedded into (in case of project mature (Souza and Gomes
2015) organizations) the organizational project management system.
The proposal needs also intensive cooperation with fuzzy modelling experts. Fuzzy
modelling, including various forms of fuzzy sets, operations on them and relations
between them has today a huge potential, which is little used in project financial
management. Making use of this potential, in cooperation with experts, might lead to
the creation of an efficient project finance management system.
Acknowledgments. This research was supported by the National Science Centre (Poland),
under Grant 394311, 2017/27/B/HS4/01881: Selected methods supporting project management,
taking into consideration various stakeholder groups and using type-2 fuzzy numbers.
References
Atkinson, R., Crawford, L., Ward, S.: Fundamental uncertainties in projects and the scope of
project management. ERA Eng. Environ. Sci. 24 (2006). https://doi.org/10.1016/j.ijproman.
2006.09.011
Bector, C.R., Chandra, S.: Fuzzy numbers and fuzzy arithmetic. In: Fuzzy Mathematical
Programming and Fuzzy Matrix Games. Studies in Fuzziness and Soft Computing, vol. 169.
Springer, Heidelberg (2005). https://doi.org/10.1007/3-540-32371-6_3
Bezdek, J.C., Dubois, D., Prade, H.: Fuzzy Sets in Approximate Reasoning and Information
Systems. Springer, New York (1999)
Bouchon-Meunier, B., Jia, Y.: Linguistic modifiers and imprecise categories. Int. J. Intell. Syst.
(1992). https://doi.org/10.1002/int.4550070105
Chaos Report. https://www.standishgroup.com/sample_research_files/CHAOSReport2015-Final.
pdf
Cheng, M.Y., Tsai, H.C., Sudjono, E.: Evolutionary fuzzy hybrid neural network for project cash
flow control. Eng. Appl. Artif. Intell. 23(4), 604–613 (2010). https://doi.org/10.1016/j.
engappai.2009.10.003
Djatmiko, B., Yulistyorini, A., Sugandi, R.M., Setyawan, E.: The application of fuzzy logic for
profit optimization to contractor project cash flow. In: IOP Conference Series: Materials
Science and Engineering, pp. 669, no. 1 (2019). https://doi.org/10.1088/1757-899X/669/1/
012060
Etgar, R., Shtub, A., LeBlanc, L.J.: Scheduling projects to maximize net present value—the case
of time-dependent, contingent cash flows. Eur. J. Oper. Res. 96(1), 90–96 (1997). https://doi.
org/10.1016/0377-2217(95)00382-7
Fathallahi, F., Najafi, A.A.: A hybrid genetic algorithm to maximize net present value of project
cash flows in resource-constrained project scheduling problem with fuzzy parameters. Sci.
Iranica 23(4), 1893–1903 (2016)
Gani, A.N., Assarudeen, S.N.M.: A new operation on triangular fuzzy number for solving fuzzy
linear programming problem. Appl. Math. Sci. (2012). https://doi.org/10.13140/2.1.3405.
8881
Gatti, S.: 5 - valuing the project and project cash flow analysis. In: Gatti, S. (ed.) Project Finance in
Theory and Practice, 3rd edn. (2018). https://doi.org/10.1016/b978-0-12-811401-8.00005-2
Kerzner, H.: A Systems Approach to Planning Scheduling and Controlling. New York (2017).
https://doi.org/10.1016/j.drudis.2010.11.015
Maravas, A., Pantouvakis, J.P.: A study of cash flows in projects with fuzzy activity durations.
In: EG-ICE 2010 - 17th International Workshop on Intelligent Computing in Engineering
(2019)
Maravas, A., Pantouvakis, J.P.: Project cash flow analysis in the presence of uncertainty in
activity duration and cost. Int. J. Proj. Manage. 30(3), 374–384 (2012). https://doi.org/10.
1016/j.ijproman.2011.08.005
Mohagheghi, V., Mousavi, S.M., Vahdani, B.: Analyzing project cash flow by a new interval
type-2 fuzzy model with an application to construction industry. Neural Comput. Appl. 28
(11), 3393–3411 (2017)
Novák, V., Dvorák, A., Perfilieva, I.: Insight into Fuzzy Modelling. Wiley, Hoboken (2016)
PMI: practice standard for project risk management. In: Project Management Institute (2009)
Project Management Institute: Project Management Body of Knowledge (PMBOK® Guide). In:
A Guide to the Project Management Body of Knowledge (PMBOK® Guide) - Fourth Edition
(2008). https://doi.org/10.1007/s13398-014-0173-7.2
Shavandi, H., Najafi, A.A., Moslehirad, A.: Fuzzy project scheduling with discounted cash flows.
Econ. Comput. Econ. Cybern. Stud. Res. 46(1), 219–232 (2012)
Souza, T.F., Gomes, C.F.S.: Assessment of maturity in project management: a bibliometric study
of main models. Procedia Comput. Sci. 55, 92–101 (2015)
Tabei, S.M.A., Bagherpour, M., Mahmoudi, A.: Application of fuzzy modelling to predict
construction projects cash flow. Periodica Polytech. Civil Eng. 63(2), 647–659 (2019). https://
doi.org/10.3311/PPci.13402
Yu, M.C., Dang, V.L., Yeh, H.C.: Measuring cash flow and overdraft for fuzzy project networks
with overlapping activities. J. Civil Eng. Manage. 23(4), 487–498 (2017). https://doi.org/10.
3846/13923730.2016.1210215
Fuzzy System for Building Energy Flows
Management
Penka V. Georgieva(&)
Informatics Department in FCSE, Burgas Free University, Burgas, Bulgaria

[email protected]
Abstract. In this paper, a fuzzy inference system for managing an installed on

a building micro grid PV systems is presented. The software system manages a
PV/battery micro grid-connected system and uses fuzzy mapping of three input
variables: power produced by PV panels, power in the battery and consumed
power. The output of the FIS, obtained after applying the fuzzy rules, is a
decision for choosing which power source is to be used. This research is part of
a project for optimizing the energy consumption of a building with the use of
independent alternative renewable energy sources. The input test data for the
optimization are the amount of used energy for lighting, heating, computers
power supply and other needs of one particular building (the building of Burgas
Free University) and the corresponding price of each of the used energy sources
(This research is funded by National Research Fund of Bulgaria under Contract
No. KP-06-COST-8/06.08.2019 for providing national co-financing for the
participation of Bulgarian teams in approved actions under the European pro-
gram for cooperation in the field of research and technology COST under the
project “Characteristics prediction and optimization of a photovoltaic system
with artificial intelligence methods”.).
Keywords: Fuzzy inference system Micro-grid PV system Smart building

management
1 Introduction
This article presents a fuzzy system for managing a micro PV system installed on the
building of Burgas Free University in Bulgaria. The research aims at creating and
implementing autonomous software management of a PV/battery system. The pre-
sented in this paper system upgrades previously developed systems, because it incor-
porates the idea to enable electricity produced by PV to be used by the AC consumers
without going through the power distribution system (grid) and the power stored in the
batteries to be used by the DC consumers again avoiding the grid whenever possible.
The tree input variables: the quantity of electricity produced by the PV panels, the
quantity of electricity stored in the batteries and the power consumption (AC and DC)
are transformed into two input fuzzy variables. The output of the FIS is a decision for
action. This is part of a project to optimize the energy consumption of homes with
independent alternative energy source.
Fuzzy System for Building Energy Flows Management 1217
The paper is organized as follows: in the second section the current state of PV
systems and the potential for effective utility for Bulgaria are discussed in short; in the
third section a brief formal description of fuzzy rule-based inference systems is given;
the fourth section the fuzzy model for managing a micro build-on PV system is pre-
sented; the fifth section shows the Matlab simulation and the last one is the conclusion.
2 Development of PV Systems in Bulgaria
Photovoltaic systems (PV systems/solar energy systems/solar systems) are energy

systems designed to convert solar radiation into usable electricity through photovoltaic
elements. PV systems’ capacity ranges from several kilowatts to hundreds of mega-
watts. Currently, most of the PV systems are connected to the national electricity grids,
while stand-alone systems make up a small share worldwide. There are numerous
varieties of PV systems: equipped with tracking devices for detecting the sun’s tra-
jectory change for adjusting the position of the panels; equipped with batteries for
storing the generated energy; etc. [1].
The elements of a PV system are as follows: solar panels for absorbing solar radi-
ation and converting it into electricity; an inverter for converting electricity from DC to
AC; wiring, mounting elements and other elements needed for a working system.
The cost of photovoltaic panels is rapidly decreasing as a result of technology
development and increased production scale. The energy conversion efficiency of a
conventional solar panel has increased from 15% in 2004 to more than 22% in 2020
[2]. Financial incentives (e.g. preferential electricity prices generated by solar power)
largely support the creation of photovoltaic systems in different countries. The cost of
electricity from photovoltaic systems is competitive with the cost of electricity gen-
eration from conventional sources. Thus, at present, there are several million photo-
voltaic systems around the world.
By 2020 the rapidly expanding world photovoltaic market is estimated to reach the
capacity of 770 GWp which is about 77 times the installed capacity since 2006 (Fig. 1)
[3]. The largest photovoltaic system installers in terms of capacity are India, China,
European Union, United States, Japan, India, Australia, with the largest PV power
stations in the world build in India-Pavagada Solar Park (2050 GWp ) in 2019 and
Bhadla Solar Park (2245 GWp ) in 2020.
Light energy, measured in W=m2 , is the total energy obtained from the entire
spectrum of the solar radiation, but only a part of it is effective for photovoltaic systems.
These are rays with a length within the range of 300–1100 nm, including the visible light
(380–740 nm) and the infrared light (>740 nm). The sunlight in the visible spectrum
accounts for around 45% of the total solar energy, the invisible spectrum (ultraviolet
range) - 7% and the largest share - 48% - is due to the infrared light. Unfortunately,
infrared rays actually contribute to photovoltaic heating and only a small portion of it
can be transformed into electrical energy [4].
1218 P. V. Georgieva
Fig. 1. Worldwide growth of photovoltaics: cumulative capacity in GWp by regions
The effectiveness of a PV system depends mainly on the following three factors:

1. annual average solar radiation in hours, depending on the position of the Earth’s
surface against the Sun and the cloudiness over the area under consideration;
2. daily fluctuations of solar radiation, analogically depending on the position of the
Earth’s surface against the Sun and the cloudiness over the area under
consideration;
3. likelihood of sunlight, which directly depends on cloudiness and with likelihood
less than 0.2 it is economically unjustified to install PV systems [5, 6].
Bulgaria, with territory around 111000 km2 , has one of the most diverse climate in
Europe. Indeed, according to Köppen climate classification there are 7 climate types:
Cfa, Cfb, Csb, Dsa, Dfa, Dfc, ET (Fig. 2).
Fig. 2. Climate types in Bulgaria

The yearly average temperature in Bulgari is 12° (C) and the early global irradiation
- 1 517 kWh=m2 k. The annual electricity consumption in the country for 2019 is 32
340 000 MWh, while the cumulative installed PV capacity is 1 033 MWp.
The average annual total amount of solar radiation in Bulgaria is about 2 150 h and
the average annual total solar irradiation resource is 1 517 kWh/m2. In general, a total
amount of theoretical potential of solar energy falling on the territory of the country for
one year is approximately 13.103 ktoe and around 390 ktoe can be considered as an
available annual solar energy absorption potential.
Therefore, there is enough sun radiation on the territory of Bulgaria that can be used
to generate electricity from installed PV systems.
The fuzzy system, proposed in this paper, aims to support a more efficient man-
agement system of a micro grid PV system. The research is part of a project for
optimizing the energy consumption of a building with the use of independent alter-
native renewable energy sources. The input data for solar radiation (along with the
wind speed and the air temperature) are received from BFU METEO - a meteorology
station specially designed and installed on the building of Burgas Free University.
BFU METEO consists of sensors for monitoring the input variables, Fronius Sensor
Box, and FroniusDatalogger Web. Data is collected in a 5-min interval and stored in a
database in a server.
The micro grid PV system consists of:
• 20 polycrystalline PV panels, with max power approx. 235 Wp each, an invertor
SolarMax 13MT2 with the following technical specifications: Input values: MPP
voltage range 250 … 750 V, Minimum voltage for rated power 370 V, Maxi-
mum DC voltage 900 V, Maximum DC current 2 18 A, Number of MPP-
Trackers 2, Max. PV generator output per MPP tracker: 9 000 W, String connec-
tions 2 2, Connection type MC 4; Output values: Rated output power at cosu ¼
1 is 13 000 W, Maximum apparent output power 13 000 VA, Nominal mains
voltage 3 400 V, Maximum AC current 3 20, Mains nominal frequency/range
50 Hz/45 Hz …55 Hz; Grid connection Three-phase (3/N/PE);
• 10 monocrystalline PV panels, with max power approx. 150 Wp each; an invertor
IBC Serve Master 3300 MV with the following technical specifications: PV input
450 VDC, max 2 10 A; 100–350 VDC MPP; Output 230 VAC, 50 Hz, Class I,
3300 W/3600 W, 15.5 Amax.
The data are collected from the BFU METEO sensors every 5 min and then
automatically stored in the database. The data for the quantity of the electric power
produced by two types of installed PV panels (polycrystalline and monocrystalline) is
also collected and stored in the database at 5-min intervals [2].
3 Fuzzy Rule-Based Inference Systems
Following the general structure of a fuzzy system, a fuzzy rule-based inference system
consists of a knowledge base (rule base and database) and an inference machine [7–10].
Let N be the number of the input fuzzy variables Ki , i ¼ 1; 2; 3; . . .; N, and ni be the

number of terms Xij of Ki for each i with j ¼ 1; 2; 3; . . .; ni . Let S be the number of
output fuzzy variables Qs , s ¼ 1; 2; 3; . . .; S, and ps be the number of terms Ysp of Qs for
each s with p ¼ 1; 2; 3; . . .; ps .
Let lij ð xÞ be the membership function of the term Xij and lsp ð yÞ be the membership
function of Ysp . Then the overall number of the membership functions in the knowledge
base is
XN XS
N: n þ S:
i¼1 i s¼1
ps : ð1Þ

The crisp input values form a vector x ¼ x1 ; x2 ; . . .; xN . This vector is fuzzified

by calculating lij xi for each i and j. At this point there are
XN
N: i¼1
ni ð2Þ
membership values, stored in the database after that calculation.

The next step is to aggregate. For simplicity let min operator be used for the T-norm
and T-norm be used for the AND operator. Let M be the number of rules and the m-th
rule Rmnhas the form:o n o n o
if Km1 is Xm1 jm1 and Km2 is Xm2 jm2 and … and Kmk is Xmk jmk then
n o n o n o
Qm1 is Ym1 jm1 and Qm2 is Ym2 jm2 and … and Qml is Yml jml
and each rule has its weight wm , m ¼ 1; 2; 3; . . .; M.
Once the m-th rule is selected and put into the template, two consecutive calcu-
lations are made:
n o
IÞ Hm ¼ min lm1 jm x1 ; lm2 jm x2 ; . . .; lmk jm xk
1 2 k
and then
IIÞ Hom ¼ Hm :wm
After firing all the rules the corresponding values of the membership functions
lm
sp ¼ Hm for each term Ysp of the output variables are obtained. The number of these
o
values depends on the number of rules in which they are used.

The aggregation applies after calculating
n o
Psp ¼ max l1sp ; l2sp ; . . .; lM
sp ð3Þ
for each Ysp , s ¼ 1; 2; 3; . . .; S and p ¼ 1; 2; 3; . . .; ps .

The last step is defuzzification. For implementing any of the methods for
defuzzification, shown earlier, a numerical integration could be applied [11–13].
In Sugeno-type FIS each rule generates two values:
z i ¼ ai x þ bi y þ c i
and
wi ¼ WðF1 ð xÞ; F2 ð yÞÞ
where x and y are the values of Input1 and Input2, F1 ð xÞ and F2 ð yÞ are the values of
membership functions for Input1 and Input2.
The final output of the system is the weighted average over all rule outputs:
PN
wi :zi
OUTPUT ¼ Pi¼1
N ;
i¼1 wi
where N is the number of rules.
4 The Fuzzy Model for Managing a Micro Build-on PV

System
The quality of the produced by PV system electric power is out of the scope of this
paper and it is assumed that the frequency, amperage and voltage are with standard
values, thus usable by the consumers.
In advance, the power produced by PV panels, power in battery and consumed
power are measured and the percentage of consumption from the AC consumers and
from the DC consumers covered by each of the energy sources is calculated. These
percentages become the input variables in the fuzzy systems.
Thus, there are 2 input variables ðN ¼ 2Þ. The first one consists of five fuzzy terms
ðn1 ¼ 5Þ:
K 1 ¼ Input1 ¼ fConsumption of AC consumers covered by PV production in %g

¼ fX11 ; X12 ; X13 ; X14 ; X15 g ¼ fCritical low; Low; Enough; High; Excessg;
and the second – of three fuzzy terms ðn2 ¼ 3Þ:
K 2 ¼ Input2 ¼ fConsumption of DC consumers covered by electricity stored in

batteries in %g ¼ fX21 ; X22 ; X23 g ¼ fLow; Enough; Highg
The output is one ðS ¼ 1Þ with six terms ðp1 ¼ 5Þ:
Q1 ¼ Output ¼ fActiong ¼ fY11 ; Y12 ; Y13 ; Y14 ; Y15 g

( )
AC consumes from PV; DC consumers from batteries:
¼ :
Charge batteries from PV; PV to grid; Switch to grid
Thus, the total number of membership functions in the knowledge base is 13.
Three of the fuzzy terms (Low, Enough, High) of Input1 have Gaussian mem-
bership functions and the other two (Critical low and Excess) have z-shaped mem-
bership functions. And because this model aims at utilizing the power produced by the
PVs to the maximum possible extend, the priority is given to the variable
Input1 = {Consumption of AC consumers covered by PV production in %}.
The second energy source (in this model - the batteries) plays a reserve role, so the
fuzzy terms of the variable Input2 = {Consumption of DC consumers covered by
electricity stored in batteries in %} has only three terms: Low and High have z-shaped
membership function, while Enough has a Gaussian membership function with the
corresponding parameters.
The total number of rules that can be generated in this model is 75, but the proposed
rule-base consists of 20 rules and they have the form

if K1 isX1j1 and K2 isX2j2 then Q1 isY1j1 :
5 Matlab Simulation
After the implementation of the above model in MatLab, experimental tests have been
carried out. The tests aim at studying the behavior of the fuzzy system while changing
the values of the input variables. One such test is visualized on Fig. 3. The percentage
of consumption of the AC consumers covered by PV is 90% and the percentage of
Fig. 3. Firing the rules of the fuzzy system PVS3 for operating a PV/battery micro grid
connected system
consumption of the DC consumers covered by batteries is 85%, so there is a degree of

membership to terms X14 and X15 of K 1 ¼ Input1 and to terms X21 and X22 of K 2 ¼
Input2 and therefore the rules 6, 7, 8, 11, 12, 13, 15, 16 and 19 are fired. These rules
imply a defuzzified output with value 0.627 which means PV to grid.
6 Conclusion
In this paper a fuzzy model for supporting the process of managing a micro grid built
on PV system is presented.
The simulation tests are conducted in MatLab. This research is part of a project for
optimizing the energy consumption of a building with the use of independent alter-
native renewable energy sources. The input test data for the optimization are the
amount of used energy for lighting, heating, computers power supply and other needs
of one particular building (the building of Burgas Free University).
The main goal for future research is the full implementation of the proposed system
as an autonomous software system managing the power flows. In the process of ver-
ification and validation the system it is very important to find possible optimal
adjustments with respect to cost.
References
1. Georgieva, P.V., Dolchinkov, R.: Fuzzy models for managing a micro grid PV system. In:
42nd International Convention on Information and Communication Technology, Electronics
and Microelectronics (MIPRO), pp. 1066–1071 (2019)
2. What are the most efficient solar panels on the market? Solar panel cell efficiency explained.
https://news.energysage.com/what-are-the-most-efficient-solar-panels-on-the-market/.
3. https://en.wikipedia.org/wiki/Growth_of_photovoltaics
4. Marion, B., Adelstein, J.: Performance parameters for grid-connected PV systems. In: 31st
IEEE Photovoltaic Specialists Conference, Lake Buena Vista, FL, United States, Institute of
Electrical and Electronics Engineers Inc., Piscataway, NJ 08855-1331, USA (2005)
5. Dolchinkov, R., Georgieva, P.: Effectiveness of Solar Tracking Systems, Yearbook of
Burgas Free University, vol. XXVII (2012). (in Bulgarian)
6. Dolchinkov, R.: Efficiency of Using Photovoltaic Systems in Construction, Donetsk,
Ukraine, vol. XXXII (2015). ISSN: 1412-221-X
7. Georgieva, P.: Parameters of GFSSAM. In: ACM International Conference Proceeding
Series, pp. 85–92 (2018)
8. Georgieva, P.: A genetic fuzzy system for asset allocation. In: 19th International Symposium
on Electrical Apparatus and Technologies, SIELA (2016)
9. Georgieva, P.V.: Genetic fuzzy system for financial management. Cybern. Inf. Technol. 18
(2), 20–35 (2018)
10. Georgieva, P.V., Popchev, I.P., Stoyanov, S.N.: A multi-step procedure for asset allocation
in case of limited resources. Cybern. Inf. Technol. 15(3), 41–51 (2015)
11. Georgieva, P.V., Popchev, I.P.: Fuzzy logic Q-measure model for managing financial
investments. Comptes Rendus de L’Academie Bulgare des Sciences 66(5), 651–658 (2013)
12. Georgieva, P., Popchev, I.: Cardinality problem in portfolio selection. Lecture Notes in
Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture
Notes in Bioinformatics), LNCS, vol. 7824, pp. 208–217 (2013)
13. Georgieva, P.: FSSAM: A fuzzy rule-based system for financial decision making in real
time. In: Handbook of Fuzzy Sets Comparison-Theory, Algorithms and Applications,
pp. 121–148. Science Gate Publishing (2016)
Fuzzy RFM Analysis: An Application
in E-Commerce
Basar Oztaysi1(&) and Mert Kavi2

1
İstanbul Technical University, 34367 İstanbul, Turkey
[email protected]
2
Modanisa.com Head Office,
Altunizade, Kuşbakışı Cd. No:27/1, 34662 İstanbul, Turkey
Abstract. RFM (recency, frequency, monetary) analysis is an essential tool for

customer segmentation, which is very important for marketing, communication,
and even operations management activities. RFM is a widely adopted seg-
mentation tool since it can be accomplished by using purchase transactions. In
this study, we use the purchase transactions of Modanisa.com, which is a global
e-commerce website from Turkey, to build a fuzzy RFM module. In the first
step, the transaction data is converted to R, F, M data parameters, and then the
fuzzy c-means algorithm is used to build the RFM clusters.
Keywords: RFM analysis Fuzzy c-means E-commerce
1 Introduction
As an effect of globalization, there has been a massive increase in competition both for
traditional and e-commerce companies. Communications, campaigns, and other mar-
keting efforts have become more critical, and customer segmentation is a crucial tool for
success in these actions. Besides, determining the customer value and maintaining the
retention of profitable customers has become more important than the acquisition of new
customers. Customer segmentation approaches can be accomplished by using different
customer characteristics such as location, age, sex, income, lifestyle, and purchase
behavior. The results of customer segmentation can be used to guide the business decision
about building marketing plans, identify trends, new product development efforts, and
deliver relevant products. Customer segmentation can also be used to personalizes the
messages of individual customers to communicate with the target groups better.
RFM (recency, frequency, monetary) analysis is a basic tool for customer seg-
mentation, which is used for quantitatively evaluating the customers based on purchase
history. The techniques use three main dimensions. Recency shows how recently a
customer has purchased from the company; Frequency shows how often the customer
purchase and Monetary aspect shows how much the customer spends. In its classical
form, RFM technique first calculates the Recency, Frequency, and Monetary scores
from transactional data, and then by using a scoring method, these values are con-
solidated in the groups by using these three dimensions [1]. The groups are later used to
analyze and understand customer groups and predict future purchase activities.
1226 B. Oztaysi and M. Kavi
Electronic commerce is defined as the process of buying, selling, transferring, or

exchanging products, services, and/or information via computer networks [2]. As new
technologies emerge, electronic commerce has improved in terms of service quality and
transaction numbers. As new data storage and processing technologies improve, the e-
commerce companies had the chance to build the infrastructure required to capture and
store enormous amounts of customer transactions on the Internet. With the help of
these infrastructures, customers’ browsing histories and purchasing records can easily
be processed for further analysis. The aim of this study to provide information about
potential segmentation perspectives in e-commerce and propose a fuzzy RFM model,
which is a combination of traditional RFM analysis with a fuzzy c-means algorithm.
The real-world case study is from Modanisa.com, which is one of the top textile
products e-commerce websites.
The remaining of the paper is as follows, in the second section, a brief literature
review or RFM analysis and segmentation is given, then in the third section, RFM and
fuzzy clustering methods are introduced. Section four presents the real-world appli-
cation steps and the results of the study. Finally, the conclusion is given in the last
section.
2 Literature Review
In the literature, customer segmentation is accomplished by using different customer

data, so various segmentation type such as demographic-based segmentation, value-
based segmentation, propensity based segmentation, life-time period segmentation is
proposed [3]. The methods used for segmentation can be based on basic arithmetic
operations, expert experiences, or analytical methods. Traditional RFM analysis is in
the second group, where basic arithmetic operations are used. RFM is widely used in
the literature. Cluster analysis is a data mining technique that is highly adopted for
analytic segmentation [4].
Recently RFM has been used in the literature in combination with other techniques.
Cheng and Chen [5] integrate the rough set theory and RFM method in order to provide
a value-based segmentation approach. McCarty and Hastak [6] focus on comparing
major segmentation techniques. In their study, the authors compare RFM method with
decision tree and logistic regression models. In another study, Chen et al. [7] integrate
sequential pattern analysis with RFM by using the purchasing data of the customers.
Christy et al. [8] propose RFM ranking methodology as a novel customer segmentation
method. Anitha and Patil [9] integrate RFM method with the k-means algorithm to
build a segmentation method based on customer purchase data. Heldt et al. [28] focus
on predicting the customer value per product and propose a prediction method based on
RFM. Khajvand et al. [10] focus on finding the lifetime value of customers and propose
a method based on RFM method. Coussement et al. [11] focus on the quality of data
used for segmentation and benchmark the results of RFM method with different seg-
mentation techniques such as logistic regression, decision tree, and RFM.
In recent years, segmentation literature has started to use novel data for segmen-
tation purposes. In one of the studies, Ahani et al. [12] use social media data in the
segmentation process and propose a decision making prediction model for SPA hotels.
Fuzzy RFM Analysis: An Application in E-Commerce 1227
Similarly, Oztaysi and Cevik Onar [13] use data from twitter, a social media platform,
and use the fuzzy c-means method to segment Twitter users. With the emergence of
mobile devices, customers’ locations stated to be a part of segmentation. Oztaysi et al.
[14] use customer location data from a county-wide beacon network and develop a new
segmentation approach. Similarly, Dogan et al. [15] concentrate on a Mall and pro-
posed a customer segmentation approach. In the first step, the authors tried to determine
the shops’ user-visits and then used a clustering algorithm to segment mall customers.
Aliahmadipour et al. [16] focus on hesitant fuzzy clustering method. Oner and Oztaysi
[17] propose interval-valued hesitant fuzzy clustering for location and customer seg-
mentation. Dogan and Oztaysi [18] focused on indoor activities on customers. In their
study, the authors first clustered the customer visit paths and build a gender prediction
model by using fuzzy c-medoids clustering. In another study, Oner and Oztaysi [19]
focus segmenting retailers; to this end, the authors collected various data about retailer
shops, including subjective data, and proposed a hesitant fuzzy segmentation model. In
a similar way, Oztaysi and Isık [20] use fuzzy clustering to segment the suppliers based
on their performances.
3 Methodology
3.1 Recency, Frequency, Monetary (RFM) Analysis

In the domain of segmentation, one of the earliest segmentation techniques is called
RFM analysis, and it is based on three simple purchase-related attributes. These
attributes are Recency (R) of purchase, Frequency (F) of purchase, and Monetary (M)
value of the purchase. Recency refers to the interval between the time that the latest
consuming behavior happens, and the present. Many direct marketers believe that
most-recent purchasers are more likely to purchase again than less-recent purchasers.
Frequency is the number of transactions that a customer has made within a certain
period. This measure is used based on the assumption that customers with more pur-
chases are more likely to buy products than customers with fewer purchases. Monetary
refers to the cumulative total of money spent by a particular customer. The concept of
RFM was introduced by Bult and Wansbeek [21] and has proven very effective when
applied to marketing databases.
In traditional RFM analysis, the study begins with recency, then frequency, and
finally, monetary value. In the first step, customers are sorted based on recency in
descending order. The customers are then split into five groups and given the top 20% a
recency score of 5, the next 20% a score of 4 and so on. The same steps are applied for
frequency and the customers are sorted from most frequent to lowest one. The first 20%
is coded as 5 and the least frequent group is coded as 1. The same processes are then
undertaken for monetary. Finally, all customers are ranked by concatenating R, F, and
M values. At the end of the study, each customer has an RFM code varying from 111 to
555. The strengths of RFM method come from the application simplicity, it can be
applied by a simple transactions table, and the amount of data is not important.
RFM analysis has been extended by other researchers. In a recent study, Ravasan
and Mansouri [22] proposed Weighted RFM in which they integrate ANP method with
RFM. In Weighted RFM method each R, F, M value is multiplied by a weight value,
wR, wF and wM according to its relative importance, and the segmentation is
accomplished based on these weighted values. In another study Li et al. [23] propose
Timely RFM, which focuses on product periodicity i.e., to analyze different product
demands at different times. Another extension of RFM analysis is RFD (Recency,
Frequency, Duration), which adopts the original model to the website domain by
adding the Duration parameter. For subscription-based businesses, RML (Recency,
Monetary, and Loyalty) model is proposed [24].
3.2 Fuzzy C-Means Method

In the domain of segmentation, the clustering procedure is used to define subgroups of
customers who have common properties. In combination with RFM, the fuzzy c-means
algorithm uses three-dimensional data, namely R, F, and M. Fuzzy c-means algorithm
is based on distance measures between data points. The algorithm groups the multi-
dimensional universe into partitions, which are represented in a partition matrix.
A fuzzy partition matrix is defined by using the following equations [25].
lik 2 ½0; 1; 1 i c; 1 k N;
X
c
lik ¼ 1; 1 k N;
i¼1
X
N
0\ lik \N; 1 i c
k¼1
The problem is defined by using an objective function defined as follows:
X
c X
N 2
J ðZ; U; V Þ ¼ ðlij Þm zj vi
i¼1 j¼1
the Z parameter in the objective function shows the set of data elements to be
clustered, U parameter represents the fuzzy partition matrix, and V parameter repre-
sents the cluster centers vector. N represents the number of data elements to be par-
titioned, µ shows the related membership value, c is the predefined clusters numbers,
and m is the fuzzifier parameter, which shows the fuzziness degree of the final clusters.
The parameter must be set to 1 or higher values. Fuzzy c-means algorithm acts as
traditional k-means when the fuzzifier parameter is set to one.
The steps of fuzzy c-means (FcM) clustering algorithm is defined as in the fol-
lowing [26]:
1. Initialize U = [uij] matrix, U(0)
2. At k-step: calculate the centres vectors V(k) = [vi] with U(k)
PN
i¼1 lm
ij : zj
vi ¼ PN
i¼1 lm
ij
3. Update U(k), U(k+1)
1
lij ¼ m1
2
Pc kzj vi k
k¼1 kzj vk k

If U ðk þ 1Þ U ðkÞ \d then STOP; otherwise go to step 2.
In order to reach meaningful and actionable results, the cluster numbers should be
well determined. There are various indexes in the literature to numerically define the
appropriateness of the clusters. In this study, we use Xie-Beni index [27] to define the
cluster numbers.
Pc PN m 2
i¼1 ðlij Þ zj vj
j¼1
XðZ; U; VÞ ¼ 2
c minvi vj
i6¼j
4 Fuzzy RFM Application
Modanisa.com is one of the leading eCommerce companies, selling textile products

globally. Modanisa is located in Turkey and focuses on selling textile products to more
than 150 countries worldwide. The company sells its 70.000 products to more than 150
countries. The company wants to segment the global customers according to the RFM
values. To this end, the purchase history of the customers is preprocessed and prepared.
A sample RFM data set is composed on R-value which shows the number of days
from the customers last purchase, F value which is calculated by using the number of
order in the last three months, and M value which is calculated by the average of basket
size of each customer (Table 1).
Table 1. RFM sample dataset

Customer id # Recency (R) #Frequency (F) #Monetary (M)
999991 23 4 65
999992 5 6 16
… . . .
… . . .
999995 2 3 214
In the next step, outliers are eliminated from the table, and the remaining data are
used for fuzzy clustering. In order to find the exact number of clusters, test including
various c parameters and the results are compared according to Xei-Beni index. The
lower values of Xie-Beni index refer to better clustering results. Since the lowest Xie-
Beni index belongs to c = 6 the c value is selected as size. In order to understand the
clusters, the centroid table is given in Table 2 can be used.
Table 2. Cluster center points in terms of z values (c = 6)

Customer id Recency (R) Frequency (F) Monetary (M)
1 –0.41 0.25 2.51
2 0.12 1.98 –0.34
3 –2.5 0.23 0.54
4 –0.5. –1.75 2.07
5 1.75 0.3 0.5
6 1.6 2.21 –1.7
The results shown in Table 2 refers that six different customer groups exist in the
dataset. The first group can be called the big baskets since the distinct property they
have is the high monetary value of their purchases. They have an average frequency of
buying their recency is just below the average. The second group can be called the
Frequent Buyers, they buy with a high frequency, but the monetary value of each
basket is just below the average. The third group can be called the Inactives; their
recency value is very low; in other terms, they are not using the system for a long time.
The fourth group can be called GoodBuyers. Their basket average is below the average
basket size, but their frequency is low, which means when they buy, they buy a lot, but
they do not use the system frequently. The fifth group can be called the Actives, the
only distinct property they have is the recency of their purchase. The last group can be
called the DailyBuyers. They act similarly to the second group; their frequency is very
high, which means they buy nearly every day, but the monetary value of their basket is
very low. Xie-Beni index shows the mathematical performance of the segmentation. In
order to check the usefulness of the segmentation results, the clusters are presented to
market professionals and found to be meaningful and actionable.
5 Conclusion
In this study, a real-world segmentation problem is examined, and RFM methodology

and Fuzzy c-means algorithm are integrated with the application phase. From the
marketing perspective, different segmentation perspectives should also be applied,
including value-based clustering and intention-based clustering. On the other hand,
other fuzzy and crisp clustering methods can be used with the same dataset, and the
results can be compared with the result of this study.
References
1. Haiying, M., Yu, G.: Customer segmentation study of college students based on the RFM.
In: International Conference on E-Business and E-Government, Guangzhou, pp. 3860–3863
(2010)
2. Turban, E., King, D., Liang, P.L., Turban, D.: Electronic Commerce 2012: Managerial and
Social Networks Perspectives. Prentice Hall, Upper Saddle River (2012)
3. Tsiptsis, K., Chorianopoulos, A.: Data Mining Techniques in CRM: Inside Customer
Segmentation. Wiley, New York (2009)
4. Han, J., Kamber, M., Pei, J.: Data Mining: Concepts and Techniques, 3rd edn. Morgan
Kaufmann Publishers, San Francisco (2011)
5. Cheng, C.H., Chen, Y.S.: Classifying the segmentation of customer value via RFM model
and RS theory. Expert Syst. Appl. 36(3), 4176–4184 (2009)
6. McCarty, M., Hastak, H.: Segmentation approaches in data-mining: a comparison of RFM,
CHAID, and logistic regression. J. Bus. Res. 60(6), 656–662 (2007)
7. Chen, Y.L., Kuo, M.H., Wu, S.Y., Tang, K.: Discovering recency, frequency, and monetary
(RFM) sequential patterns from customers’ purchasing data. Electr. Commer. Res. Appl. 8
(5), 241–251 (2009)
8. Christy, A., Umamakeswari, A., Priyatharsini Neyaa A.: RFM ranking – an effective
approach to customer segmentation. J. King Saud Univ. Comput. Inf. Sci (2018)
9. Anitha P., Patil, M.M.: RFM model for customer purchase behavior using K-Means
algorithm. J. King Saud Univ. Comput. Inf. Sci. (2019)
10. Khajvand, M., Zolfaghar, K., Ashoori, S., Alizadeh, S.: Estimating customer lifetime value
based on RFM analysis of customer purchase behavior: case study. Procedia Comput. Sci.
32011, 57–63 (2011)
11. Coussement, K., Van den Bossche, F.A.M., De Bock, K.W.: Data accuracy’s impact on
segmentation performance: benchmarking RFM analysis, logistic regression, and decision
trees. J. Bus. Res. 67(1), 2751–2758 (2014)
12. Ahani, A., Nilashi, M., Ibrahim, O., Sanzogni, L., Weaven, S.: Market segmentation and
travel choice prediction in Spa hotels through TripAdvisor’s online reviews. Int.
J. Hosp. Manage. 80, 52–77 (2019)
13. Oztaysi, B., Cevik Onar, S.: User segmentation based on twitter data using fuzzy clustering.
In: Data Mining in Dynamic Social Networks and Fuzzy Systems, pp. 316–333 (2013)
14. Öztayşi, B., Gokdere, U., Simsek, E.N., Oner, C.S.: A novel approach to segmentation using
customer locations data and intelligent techniques. In: Handbook of Research on Intelligent
Techniques and Modeling Applications in Marketing Analytics (2017)
15. Dogan, O., Oztaysi, B., Fernandez-Llatas, C.: Segmentation of indoor customer paths using
intuitionistic fuzzy clustering: Process mining visualization. J. Intell. Fuzzy Syst. 38(1), 675–
684 (2020)
16. Aliahmadipour, L., Torra, V., Eslami, E.: On hesitant fuzzy clustering and clustering of
hesitant fuzzy data. In: Torra, V., Dahlbom, A., Narukawa, Y. (eds.) Fuzzy Sets, Rough Sets,
Multisets and Clustering. Studies in Computational Intelligence. Springer, Cham (2017)
17. Oner, S.C., Oztaysi, B.: An interval valued hesitant fuzzy clustering approach for location
clustering and customer segmentation. In: Advances in Fuzzy Logic and Technology,
pp. 56–70 (2017)
18. Dogan, O., Oztaysi, B.: Gender prediction from classified indoor customer paths by fuzzy C-
medoids clustering. In: International Conference on Intelligent and Fuzzy Systems, Istanbul,
pp. 160–169 (2019)
19. Oner, S.C., Oztaysi, B.: An interval type 2 hesitant fuzzy MCDM approach and a fuzzy c
means clustering for retailer clustering. Soft. Comput. 22(15), 4971–4987 (2018)
20. Oztaysi, B., Isik, M.: Supplier evaluation using fuzzy clustering. In: Supply Chain
Management Under Fuzziness, pp. 61–79 (2014)
21. Bult, J.R., Wansbeek, T.: Optimal selection for direct mail. Mark. Sci. 14(4), 378–394
(1995)
22. Ravasan, A.Z., Mansouri, T.: A dynamic ERP critical failure factors modelling with FCM
throughout project lifecycle phases. Prod. Plann. Control 27(2), 65–82 (2015)
23. Li, L.H., Lee, F.M., Liu, W.C.: The Timely Product Recommendation Based on RFM
Method (2006). Corpus ID: 16924645
24. Birant, D.: Data mining using RFM analysis. In: Knowledge-Oriented Applications in Data
Mining (2011)
25. Ruspini, E.H.: Numerical methods for fuzzy clustering. Inf. Sci. 2(3), 319–350 (1970).
https://doi.org/10.1016/S0020-0255(70)80056-1
26. Diaz-Perez, F.M., Bethencourt-Cejas, M.: CHAID algorithm as an appropriate analytical
method for tourism market segmentation. J. Destination Mark. Manage. 5(3), 275–282
(2016)
27. Xie, X.L., Beni, G.: A validity measure for fuzzy clustering. IEEE Trans. Pattern Anal.
Mach. Intell. 13(8), 841–847 (1991)
28. Heldt, R., Silveria, C.S., Luce, F.B.: Predicting customer value per product: from RFM to
RFM/P. J. Bus. Res. (2019)
A New Structure of Nullnorms
on Bounded Lattices
Gül Deniz Çaylı(B)
Department of Mathematics, Faculty of Science,

Karadeniz Technical University, 61080 Trabzon, Turkey
[email protected]
Abstract. As a generalization of the concepts of t-norm and t-conorm

on a bounded lattice, nullnorms are important from a theoretical view-
point but also because of their numerous potential applications, such as in
expert systems, fuzzy quantifiers and neural networks. Nullnorms have
a zero element selected from the whole domain, regardless of whether
t-norms and t-conorms have a zero element on the lower and upper ele-
ments of a bounded lattice. There are several attempts to investigate the
structure of nullnorms on bounded lattices in literature. In this paper,
unlike the researches in the literature, we use the closure operators and
interior operators on a bounded lattice to construct new classes of null-
norms with a zero element under some additional conditions. Further-
more, several interesting examples are included to have a better under-
standing of new structures of nullnorms on bounded lattices.
Keywords: Bounded lattice · Construction method · Closure

operator · Interior operator · Nullnorm
1 Introduction
Nullnorms as aggregation operators unifying and generalizing the notions of

t-norms and t-conorms were introduced by Calvo et al. [4], and have been proved
to be useful in several areas, such as in expert systems, fuzzy quantifiers, and
neural networks. Mas et al. [20] introduced the concept of t-operators on the unit
interval, and then they showed that nullnorms and t-operators are equivalent
since both of them have identical block structures on [0, 1]2 in [21]. Nullnorms
are interesting also from theoretical point of view due to the fact that in fuzzy
logic nullnorms as aggregation operators maintain as many logical properties
as possible. It can be found some theoretical discussions concerning t-norms,
t-conorms, nullnorms and related operators on the unit interval in [5,10,14,
17,19,22–24,26,27,29]. Recently, because of the effective application fields of
aggregation operators on a bounded lattice compared to their counterparts on
the real unit interval [0, 1], they have received increasing attention. Karaçal et al.
[18] studied nullnorms on a bounded lattice and identified the smallest nullnorm
1234 G. D. Çaylı
and the greatest nullnorm with a given zero element. Afterward, the presence
of idempotent nullnorms on bounded lattices was investigated in [6], and the
fact that such a nullnorm does not need to exist was showed. Some approaches
for idempotent nullnorms on bounded lattices were also proposed in [7,28] with
some constraints. Moreover, the structure of nullnorms on bounded lattices was
investigated widely in subsequent studies [8,9].
In [15,16], it was proposed that closure and interior operators can be defined
on a lattice (℘ (A) , ⊆) of all subsets of a set A with set union as the join and set
intersection as the meet. Moreover, the closure operator cl (resp. interior opera-
tor int) on ℘ (A) is extended to a general lattice L where the condition cl (∅) = ∅
(resp. int (A) = A) is omitted. In this paper, we aim to introduce some methods
for constructing nullnorms by means of closure operators and interior operators
defined on a bounded lattice. Since the existing studies do not consider the exis-
tence of closure operators or interior operators on bounded lattices to construct
nullnorms, this paper can be seen as the important and meaningful supplement
of this research topic by using closure and interior operators from the theoreti-
cal point of view. The remainder of this paper is organized as follows. In Sect. 2,
we give some preliminaries about bounded lattices, and nullnorms, interior and
closure operators on them. In Sect. 3, based on the closure operator cl and the
interior operator int on a bounded lattice L, we propose two methods to obtain
new classes of nullnorms Vcl and Vint on L with the zero element s ∈ L\{0, 1}
with some constraints. Furthermore, some illustrative examples are presented to
show that the nullnorms Vcl and Vint generate new types of nullnorms different
from those introduced in [28]. We observe that our constructions do not yield an
idempotent nullnorm on a bounded lattice L, where there exist at least two ele-
ments incomparable with theirs zero element s ∈ L\{0, 1}. In addition, we note
that the nullnorms Vcl and Vint do not have to coincide with another t-conorm on
2 2
[0, s] except the t-conorm S∨ : [0, s] → [0, s] defined by S∨ (x, y) = x ∨ y for all
2 2
x, y ∈ [0, s] and another t-norm on [s, 1] except the t-norm T∧ : [s, 1] → [s, 1]
defined by T∧ (x, y) = x ∧ y. Finally, some conclusions and future works are
included in Sect. 4.
2 Preliminaries
In this section, we recall some basic results with the respect to nullnorms, closure
and interior operators on bounded lattices.
A lattice (L, ≤) is bounded if it has the top and bottom elements, which are
written as 1 and 0, respectively. Throughout this article, unless stated otherwise,
we denote L as a bounded lattice [1].
Definition 1 [6–9,18]. A binary operation V : L × L → L is called a nullnorm

if it is commutative, associative, non-decreasing in each variable and there exists
an element s ∈ L called the zero element, such that we obtain V (x, 0) = x for
all x ≤ s and V (x, 1) = x for all x ≥ s.
A New Structure of Nullnorms on Bounded Lattices 1235
V is called an idempotent nullnorm on L whenever V (x, x) = x for all

x ∈ L. We note that a t-norm T (resp. t-conorm S) is a special case of nullnorm
with s = 0 (resp. s = 1) (see [3,13]). For a, b ∈ L, we use the notation a
b
to denote that a and b are incomparable. For s ∈ L\{0, 1}, we denote Ds =
[0, s] × [s, 1] ∪ [s, 1] × [0, s] and Is = {x ∈ L | x
s}.
Definition 2 [11,12,16,25]. Let L be a lattice. A mapping cl : L → L (resp.

int : L → L) is called a closure operator (resp. an interior operator) if, for any
x, y ∈ L, it satisfies the following properties:
(1) x ≤ cl(x) (resp. int(x) x).
(2) cl(x ∨ y) = cl(x) ∨ cl(y) (resp. int(x ∧ y) = int(x) ∧ int(y)).
(3) cl(cl(x)) = cl(x) (resp. int(int(x)) = int(x)).
3 Construction Methods for Nullnorms

In this section, we construct new families of nullnorms on bounded lattices with
the zero element based on closure and interior operators, where some additional
assumptions are required. Furthermore, we investigate some characteristics of
these constructions of nullnorms on bounded lattices and present some corre-
sponding examples in order to have a better understanding of their structures.
Theorem 1. Let s ∈ L\{0, 1} such that a ∧ s = b ∧ s and a ∨ s = b ∨ s for

all a, b ∈ Is . Given a closure operator cl : L → L such that cl(k ∧ s) ∈ Is for
all k ∈ Is , then the function Vcl : L × L → L is a nullnorm on L with the zero
element s, where
⎧ 2
⎪
⎪ x∨y if (x, y) ∈ [0, s] ,
⎪
⎪
⎪
⎪ x∧y if (x, y) ∈ [s, 1] ,
2
⎪
⎪
⎪
⎪ s if (x, y) ∈ Ds ,
⎨
y ∧ (x ∨ s) if (x, y) ∈ Is × [s, 1],
Vcl (x, y) = (1)
⎪
⎪ x ∧ (y ∨ s) if (x, y) ∈ [s, 1] × Is ,
⎪
⎪
⎪
⎪ y ∨ (x ∧ s) if (x, y) ∈ Is × [0, s],
⎪
⎪ x ∨ (y ∧ s)
⎪
⎪ if (x, y) ∈ [0, s] × Is ,
⎩
cl (x ∧ s) ∨ cl (y ∧ s) if (x, y) ∈ Is2 .
Theorem 2. Let s ∈ L\{0, 1} such that a ∧ s = b ∧ s and a ∨ s = b ∨ s for

all a, b ∈ Is . Given a closure operator cl : L → L. If the function Vcl defined
by the formula (1) is a nullnorm on L with the zero element s, then there holds
cl(k ∧ s) ∈ Is for all k ∈ Is .
Remark 1. In Theorem 1, we note that the condition a ∨ s = b ∨ s for all a, b ∈ Is

cannot be omitted, in general. Consider the lattice L1 = {0, m, n, s, r, 1} such
that m, n ∈ Is , n < r, s < r, m
n and m
r. Define the closure operator
cl : L1 → L1 by cl(0) = cl (m) = m and cl (s) = cl (n) = cl (r) = cl (1) = 1.
Then we have Vcl (Vcl (1, m) , n) = r and Vcl (1, Vcl (m, n)) = 1. So, Vcl is not a
nullnorm on L1 since the associativity is violated. Hence, we observe that the
1236 G. D. Çaylı
condition a∨s = b∨s for all a, b ∈ Is is sufficient in Theorem 1. However, it is not

a necessary condition in Theorem 1. To demonstrate this argument, we define
new closure operator cl : L1 → L1 by cl(0) = cl (n) = n, cl (s) = cl (r) = r and
cl (m) = cl (1) = 1. By the method in Theorem 1, the function Vcl : L1 ×L1 → L1
is given by Table 1. It is possible to check that Vcl is a nullnorm on L1 .
Remark 2. In Theorem 1, it should be pointed out that the condition a ∧ s =
b ∧ s for all a, b ∈ Is cannot be omitted, in general. Consider the lattice L2 =
{0, k, m, r, s, p, q, 1} such that p, q ∈ Is , k, m < r < s, k ∨ m = r, k < p, m < q,
p
r, q
r and p
q. Define the closure operator cl : L2 → L2 such that
cl(0) = 0, cl (m) = cl (q) = q, cl (k) = cl (p) = p and cl (s) = cl (r) = cl (1) = 1.
Then we have Vcl (Vcl (0, p) , q) = r and Vcl (0, Vcl (p, q)) = s. Therefore, Vcl is
not a nullnorm on L2 since it is not associative. So, the fact that a ∧ s = b ∧ s for
all a, b ∈ Is is a sufficient condition in Theorem 1 . However, it is not a necessary
condition in Theorem 1. To illustrative this observation, we take the lattice
L3 = {0, k, m, s, p, q, 1} such that p, q ∈ Is , k, m < s, k ∨ m = s, k < p, m < q
and p
q the closure operator cl : L3 → L3 as cl(0) = 0, cl (m) = cl (q) = q,
cl (k) = cl (p) = p and cl (s) = cl (1) = 1. By use of the method in Theorem 1,
the function Vcl : L3 × L3 → L3 defined by Table 2 is a nullnorm on L3 .
Table 1. Nullnorm Vcl on L1 Table 2. Nullnorm Vcl on L3
Vcl 0 m n s r 1 Vcl 0 k m s p q 1
0 00 0 sss 0 0 kmskms
m 0n nsr1 k k ks sks s
n 0n nsrr m ms mss ms
s ss s sss s s s s ss s s
r sr r srr p k ks sp 1 1
1 s1 r sr1 q ms ms1 q 1
1 s s s s1 1 1
Theorem 3. Let s ∈ L\{0, 1} such that a ∧ s = b ∧ s and a ∨ s = b ∨ s for all

a, b ∈ Is . Given an interior operator int : L → L such that int (k ∨ s) ∈ Is for
all k ∈ Is , then the function Vint : L × L → L is a nullnorm on L with the zero
element s, where
⎧
⎪
⎪ x∧y if (x, y) ∈ [s, 1]2 ,
⎪
⎪ 2
⎪
⎪ x∨y if (x, y) ∈ [0, s] ,
⎪
⎪
⎪
⎪ s if (x, y) ∈ Ds ,
⎨
y ∧ (x ∨ s) if (x, y) ∈ Is × [s, 1],
Vint (x, y) = (2)
⎪
⎪ x ∧ (y ∨ s) if (x, y) ∈ [s, 1] × Is ,
⎪
⎪
⎪
⎪ y ∨ (x ∧ s) if (x, y) ∈ Is × [0, s],
⎪
⎪
⎪
⎪ x ∨ (y ∧ s) if (x, y) ∈ [0, s] × Is ,
⎩
int (x ∨ s) ∧ int (y ∨ s) if (x, y) ∈ Is2 .
Theorem 4. Let s ∈ L\{0, 1} such that a ∧ s = b ∧ s and a ∨ s = b ∨ s for all

a, b ∈ Is . Given an interior operator int : L → L. If the function Vint defined
by the formula (2) is a nullnorm on L with the zero element s, then there holds
int (k ∨ s) ∈ Is for all k ∈ Is .
Remark 3. Let s ∈ L\{0, 1} such that a ∧ s = b ∧ s and a ∨ s = b ∨ s for all a, b ∈
Is . Given a closure operator cl : L → L and an interior operator int : L → L
such that cl (k ∧ s) ∈ Is and int (k ∨ s) ∈ Is for all k ∈ Is . On the domain Is ×Is ,
the nullnorms V∨ and V∧ introduced in [28, Theorem 2] differ from the nullnorm
Vcl and Vint defined in Theorems 1 and 3, respectively. For (x, y) ∈ Is × Is ,
the values of Vcl and Vint are cl (x ∧ s) ∨ cl (y ∧ s) and int (x ∨ s) ∧ int (y ∨ s) ,
respectively, whereas V∨ and V∧ have the values x ∨ y and x ∧ y, respectively.
These classes of nullnorms have the same value on all remainder domains. It can
be clearly seen that V∨ and V∧ are idempotent nullnorms. When there is only one
element incomparable with the zero element s, since Vcl and Vint are idempotent
nullnorms, the nullnorms V∨ , V∧ , Vcl and Vint coincide with each other. However,
the nullnorms Vcl and Vint do not need to coincide with the nullnorms V∨ and
V∧ . We provide the following example to illustrative this observation.
Fig. 1. Lattice L4
Example 1. Consider the lattice L4 characterized in Fig. 1. If we apply the con-

structions in [28, Theorem 2], then we give the nullnorms V∨ , V∧ : L4 × L4 → L4
by Tables 3 and 4, respectively.
If we define the closure operator cl : L4 → L4 by cl(0) = 0, cl (s) = cl (k) =
cl (n) = n, cl (m) = cl (p) = cl(q) = cl(t) = t and cl(1) = 1, then by using
the approach in Theorems 1, the nullnorm Vcl on L4 is given by Table 5. The
nullnorm Vcl is different from the nullnorms V∨ , V∧ since Vcl (p, q) = t = m =
V∧ (p, q) and Vcl (t, k) = t = k = V∨ (t, k) = t. If we define the interior operator
int : L4 → L4 by int(0) = 0, int (m) = int (p) = int (q) = int (s) = m,
int (t) = t, int (k) = int (n) = k and int(1) = 1, then by use of the approach
in Theorem 3, the nullnorm Vint on L4 is given by Table 6. The nullnorm Vint
is different from the nullnorms V∨ , V∧ since Vint (p, q) = k, V∨ (p, q) = t and
V∧ (p, q) = m.
1238 G. D. Çaylı
Table 3. Nullnorm V∨ on L4 Table 4. Nullnorm V∧ on L4
V∨ 0 m p q t k s n 1 V∧ 0 m p q t k s n 1
0 0 mmmmmss s 0 0 mmmmmss s
m mmmmmmss s m mmmmmmss s
p mmp t t k snn p mmp mp p snn
q mmt q t k snn q mmmq q q snn
t mmt t t k snn t mmp q t t snn
k mmk k k k snn k mmp q t k snn
s s s s s s s ss s s s s s s s s ss s
n s s n n n n snn n s s n n n n snn
1 s s n n n n sn1 1 s s n n n n sn1
Table 5. Nullnorm Vcl on L4 Table 6. Nullnorm Vint on L4
Vcl 0 m p q t k s n 1 Vint 0 m p q t k s n 1
0 0 mmmmmss s 0 0 mmmmmss s
m mmmmmmss s m mmmmmmss s
p mmt t t t snn p mmk k k k snn
q mmt t t t snn q mmk k k k snn
t mmt t t t snn t mmk k k k snn
k mmt t t t snn k mmk k k k snn
s s s s s s s ss s s s s s s s s ss s
n s s n n n n snn n s s n n n n snn
1 s s n n n n sn1 1 s s n n n n sn1
Remark 4. Let L be a bounded lattice, cl : L → L be an arbitrary closure

2
operator and s ∈ L\{0, 1}. It should be noted that the nullnorm Vcl , Vint | [0, s]
2
is the t-conorm S∨ : [0, s] → [0, s] defined by S∨ (x, y) = x ∨ y for all x, y ∈ [0, s] .
However, Vcl and Vint do not need to coincide with another predefined t-conorm
2
except for the t-conorm S∨ on [0, s] . To illustrate this argument, considering
the lattice L5 = {0, r, n, m, s, k, 1} such that 0 < r < n < m < s < 1, n < k < 1,
2 2
k
s and k
m, we assume that Vcl , Vint | [0, s] = S : [0, s] → [0, s] is the
t-conorm defined by Table 7. Then, by applying the construction approaches in
Theorems 1 and 3, we have Vcl (Vcl (r, r) , k) = n (Vint (Vint (r, r) , k) = n) and
Vcl (r, Vcl (r, k)) = m (Vint (r, Vint (r, k)) = m). Since Vcl and Vint do not satisfy
associativity property, we cannot force Vcl and Vint to coincide with another
2
prescribed t-conorm except for the t-conorm S∨ on [0, s] . Similarly, we note
2 2
that the nullnorm Vcl , Vint | [s, 1] is the t-norm T∧ : [s, 1] → [s, 1] defined
by T∧ (x, y) = x ∧ y for all x, y ∈ [s, 1] . However, Vcl and Vint do not need to
2
coincide with another predefined t-norm except for the t-norm T∧ on [s, 1] . To
illustrate this observation, we consider the lattice L6 = {0, p, q, t, k, s, 1} such
that 0 < s < k < p < t < 1, 0 < q < k, q
s and assume that Vcl , Vint |
2 2
[s, 1] = T : [s, 1] → [s, 1] is the t-norm defined by Table 8. Then, by means of
the construction approaches in Theorems 1 and 3, we have Vcl (Vcl (p, t) , q) = k
(Vint (Vint (p, t) , q) = k) and Vcl (p, Vcl (t, q)) = s (Vint (p, Vint (t, q)) = s). Since
Vcl and Vint do not satisfy associativity property, we cannot force Vcl and Vint
2
to coincide with another prescribed t-norm except for the t-norm T∧ on [s, 1] .
Table 7. T-conorm S on [0, s]2 Table 8. T-norm T on [s, 1]2
S 0 r n ms T skpt 1
0 0 r n ms s ss sss
r r n mms k ss sss
n n mmms p ss ppp
mmmmms t ss pt t
s s s s s s 1 ss pt 1
4 Concluding Remarks
After the structure of nullnorms on the real unit interval [0, 1] was determined,
the structure of nullnorms on bounded lattices has attracted researchers’ atten-
tion. In the recent times, the approaches for yielding nullnorms on bounded lat-
tices were studied in the papers [2,6–9,18,28]. In this paper, we introduced some
methods to obtain new classes of nullnorms with the zero element s by using
closure operators and interior operators on a bounded lattice L under some addi-
tional conditions. We exemplified that these conditions play an important role
in our constructions, and thus, they yield a nullnorm only in particular cases.
Notice that since our constructions do not generate idempotent nullnorms when
there exist at least two elements incomparable with the zero element, they can-
not be coincide with the ones introduced in [28, Theorem 2]. We also noted that
both constructions do not need to coincide with another predefined t-conorm
2
except for the supremum t-conorm S∨ on [0, s] and the infimum t-norm T∧ on
2
[s, 1] . As our future work, we plan to introduce new classes of nullnorms, in
particular idempotent nullnorms, with more general structures by use of interior
and closure operators on bounded lattices.
References
1. Birkhoff, G.: Lattice Theory. American Mathematical Society Colloquium Publish-
ers, Providence (1967)
1240 G. D. Çaylı
2. Bodjanova, S., Kalina, M.: Nullnorms and T-operators on bounded lattices: coinci-
dence and differences. In: Medina, J., et al. (eds.) IPMU 2018, Communications in
Computer and Information Science, vol. 853, pp. 160–170. Springer, Cham (2018)
3. Bodjanova, S., Kalina, M.: Uninorms on bounded lattices with given underlying
operations. In: Halaš, R., et al. (eds.) AGOP 2019, Advances in Intelligent Systems
and Computing, vol. 981, pp. 183–194. Springer, Cham (2019)
4. Calvo, T., De Baets, B., Fodor, J.: The functional equations of Frank and Alsina
for uninorms and nullnorms. Fuzzy Sets Syst. 120, 385–394 (2001)
5. Calvo, T., Kolesárová, A., Komornı́ková, M., Mesiar, R.: Aggregation operators:
properties, classes and construction methods. In: Calvo, T., et al. (eds.) Aggrega-
tion Operators, New Trends and Applications, pp. 3–104. Physica-Verlag, Heidel-
berg (2002)
6. Çaylı, G.D., Karaçal, F.: Idempotent nullnorms on bounded lattices. Inf. Sci. 425,
153–164 (2018)
7. Çaylı, G.D.: Construction methods for idempotent nullnorms on bounded lattices.
Appl. Math. Comput. 366, 1–14 (2020)
8. Çaylı, G.D.: Nullnorms on bounded lattices derived from t-norms and t-conorms.
Inf. Sci. 512, 1134–1154 (2020)
9. Çaylı, G.D.: Some results about nullnorms on bounded lattices. Fuzzy Sets Syst.
386, 105–131 (2020)
10. Drewniak, J., Drygaś, P., Rak, E.: Distributivity between uninorms and nullnorms.
Fuzzy Sets Syst. 159, 1646–1657 (2008)
11. Drossos, C.A.: Generalized t-norm structures. Fuzzy Sets Syst. 104, 53–59 (1999)
12. Drossos, C.A., Navara, M.: Generalized t-conorms and closure operators. In: Pro-
ceedings of EUFIT 1996, pp.22–26. Aachen (1996)
13. Drygaś, P.: Isotonic operations with zero element in bounded lattices. In:
Atanassov, K., et al. (eds.) Soft Computing Foundations and Theoretical Aspect,
pp. 181–190. EXIT Warszawa (2004)
14. Drygaś, P.: A characterization of idempotent nullnorms. Fuzzy Sets Syst. 145,
455–461 (2004)
15. Engelking, R.: General Topology. Heldermann Verlag, Berlin (1989)
16. Everett, C.J.: Closure operators and Galois theory in lattices. Trans. Am. Math.
Soc. 55, 514–525 (1944)
17. Grabisch, M., Marichal, J.L., Mesiar, R., Pap, E.: Aggregation Functions. Cam-
bridge University Press, Cambridge (2009)
18. Karaçal, F., İnce, M.A., Mesiar, R.: Nullnorms on bounded lattices. Inf. Sci. 325,
227–236 (2015)
19. Li, G., Liu, H.W., Su, Y.: On the conditional distributivity of nullnorms over
uninorms. Inf. Sci. 317, 157–169 (2015)
20. Mas, M., Mayor, G., Torrens, J.: T-operators. Int. J. Uncertain Fuzziness Knowl.
Based Syst. 7, 31–50 (1996)
21. Mas, M., Mayor, G., Torrens, J.: The distributivity condition for uninorms and
t-operators. Fuzzy Sets Syst. 128, 209–225 (2002)
22. Mas, M., Mayor, G., Torrens, J.: The modularity condition for uninorms and t-
operators. Fuzzy Sets Syst. 126, 207–218 (2002)
23. Mas, M., Mesiar, R., Monserrat, M., Torrens, J.: Aggregation operators with anni-
hilator. Int. J. Gen. Syst. 34, 1–22 (2005)
24. Mesiar, R., Kolesárová, A., Stupnanová, A.: Quo vadis aggregation. Int. J. Gen.
Syst. 47, 97–117 (2018)
25. Ouyang, Y., Zhang, H.P.: Constructing uninorms via closure operators on a
bounded lattice. Fuzzy Sets Syst. 395, 93–106 (2020). https://doi.org/10.1016/
j.fss.2019.05.006
26. Qin, F., Wang, Y.M.: Distributivity between semi-t-operators and Mayor’s aggre-
gation operators. Inf. Sci. 336–337, 6–16 (2016)
27. Xie, A., Liu, H.: On the distributivity of uninorms over nullnorms. Fuzzy Sets Syst.
211, 62–72 (2013)
28. Wang, Y.M., Zhan, H., Liu, H.W.: Uni-nullnorms on bounded lattices. Fuzzy Sets
Syst. 386, 132–144 (2020)
29. Zong, W., Su, Y., Liu, H.W., De Baets, B.: On the structure of 2-uninorms. Inf.
Sci. 467, 506–527 (2016)
A New Perspective to Design Phase of Water
Supply Systems from Aspect of Water Demand
Using Fuzzy Automation
Halid Akdemir1,2, Ayşegül Alaybeyoğlu1(&),

and Ali Danandeh Mehr2
1
Izmır Katip Çelebi University, 35620 İzmir, Turkey
2
Antalya Bilim University, 07190 Antalya, Turkey
[email protected]
Abstract. This paper aims to reduce the setup cost of water supply systems
(SCWSS) by decreasing the peak values at water consumptions Water supply
system (WSS) is designed by paying attention to worst scenario. It means that a
WSS must supply water demanded to people at each scenario. Therefore, there
is waste some part of volume needed for transporting water to certain regions
that isn’t utilized all the time. The use of water should be manipulated as
desired. If the peak values of water demand occurred at a specific time can be
dispersed to whole day, then SCWSS can be reduced due to diminished max-
imum discharge capacity needed as well. Lately researchers have done so many
things on water tank by using fuzzy logic but not considering stabilization of
fluctuations at water demands. In this paper, a new intelligent valve (NIV) which
aims to reduce the SCWSS is developed for stabilizing the demand of water
occurred all day long by using fuzzy logic automation. The NIV can get stabilize
water use for a region by optimizing discharge of water to water tank. The paper
introduces a new perspective to design phase by focusing water demand change
and trend.
Keywords: Setup cost of water supply system Fuzzy logic Intelligent valve
1 Introduction
Water is the most crucial need of human beings. Water becomes drinkable with
somehow by being exposed some treatment and transmission process and it is trans-
mitted with different designs to meet demands from aspect of quality and quantity [1].
So many optimization and evaluation methods that may contain some stochastic and
metaheuristic methods can be utilized to analyze a water supply system (WSS).
Storage tanks and reservoirs are commonly used in drinking water distribution
systems to equalize pumping requirements and operating pressures, and to provide
emergency water for fire-fighting and pumping outages [2].
Water supply systems is designed by engineers considering fluctuations at water
demand because the water demand undulates; it reaches its peak values sometimes [3].
When engineers design a water supply system, they use some experimental coefficient
A New Perspective to Design Phase of Water Supply Systems 1243
varying from region to region and regarding to a certain area for supplying the peak
demands. Lately, researchers have done lots of things about water tank optimization by
using fuzzy logic [4, 5] by considering the fact that making the water tank full faster,
regularly or with a way for some aims [6]. This paper aims to introduce a new idea
about stabilizing the fluctuations at water demands in some institutions having an own
water tank and working with same logic of universities so setup cost can be reduced. In
this paper, imaginary scenarios are studied. They are imaginary due to local constraints
not allowing giving real data. A region named Balatçık being in Izmir City of Turkey is
investigated and used for 2 scenarios as examples of WSS. Izmir Katip Çelebi
University is selected as a case study with estimated values. 2 WSS types are examined
aspect of their cost and working principles. One of them is a WSS working with new
intelligent valve (NIV). Organization of the paper is as follows: Methodology of the
paper is given in Sect. 2, The proposed system which includes Perspective to Design
Phase of WSS and Fuzzy Logic Automations subsections are discussed in Sect. 3,
results and conclusion are given in Sect. 4 and Sect. 5 respectively.
2 Methodology
2.1 WSS Design

First, a model of WSS will be designed to be able to calculate the cost of each scenario
for a region in Izmir City. Only tank model of the University in the WSS of the region
will be a variable, so 2 tank models in the same system can be investigated. These tank
models working different principles are stated in Table 1. Water tanks work with same
principle with water distribution tanks. They absorb peak values; therefore, they are
selected in WSS so setup cost of water supply system (SCWSS) can be reduced.
Conventional tank models work with simple valve named ballcock. It has own char-
acteristics on the other hand proposed new valve model named NIV in this paper is a
valve which works with complicated logic and considers variables in stated Table 4.
Ballcock works with some basic physics principles same as toilet flush tanks and they
are cheap and simple, so they are useful but not from every point of view. Although a
tank working with a ballcock is cheap and handy, not efficient as much as thought. The
University has a conventional water supply system consisting of a tank under the
ground level. The tank has a classical faucet not needing control and working with
basic physics principles and there is a pump to give energy to water in the same area.
To annihilate the uncertainties causing the setup cost to increase while designing
because of the uncertainties plays a crucial role in WSS designs.
2.2 Solution Model

There are uncertainties impossible to be known at water consumption in WSS mean-
while people have same water use trend which enhances to the SCWSS because of the
usage congestion, so a system is needed to be able to ensure stabilization at water
demand fluctuations. The point is embedding automation to conventional system from
aspect of water usage stabilization by optimizing discharge.
1244 H. Akdemir et al.
2 scenarios will be examined, and they are stated in Table 1.
Table 1. Scenarios for evaluation.

Scenarios
Conventional tank with ballcock
Tank with NIV
Hydrological Design. Local data will be evaluated and used for hydrological
parameters such as number of populations, elevation of region, necessary equipment
for an example of WSS design and so on in Table 2.
Table 2. Hydrological parameters according to 2019 local data.

Region Population Average water demand Peak factor
Household area 5000 70lt/cap/day 2,5
Izmir Katip Çelebi Uni. 13000 20lt/cap/day 4
Peak consumption of water in a day will be determined according to Eq. 1.
ðP:F:Þhour ¼ Dmh =Dah ð1Þ
Peak factors are used for calculation of maximum water consumption. Peak factors
change according to characteristics of region. Peak factors which is selected for the
areas are stated in above in Table 2.
Fuzzy Logic. Fuzzy logic is a form of many-valued logic; it deals with reasoning that
is approximate rather than fixed and exact. Compared to traditional binary sets (where
variables may take on true or false values), fuzzy logic variables may have a truth value
that ranges in degree between 0 and 1 [7].
In this paper, fuzzy logic will be selected to give desired logic to the system.
3 The Proposed System
3.1 A New Perspective to Design Phase of WSS

The logic of new system is to fill the tank of university as much as possible when water
demand of whole region is low or fill the tank slowly when water demand is high for
whole region. It can be said partially that the system works reversely thus a region more
stabilize and therefore, lower discharge needed is attained, lower SCWSS can be
obtained.
Fig. 1. An example of WSS for the region.
WSS was designed according to Turkish Standards and needs so zones for pressure
regulations were considered. The system was formed as loop pattern, but water con-
sumption of university will be investigated aspect of its tank model. Water demand of
household area is constant and determined.
Distribution of water demand per hour in a day for an area can be estimated so
many methods. [8–10] In this paper, local data was used for estimation of water
demand distribution of the household area as stated in Fig. 2.
35.00
30.00 Household Area Supply
25.00
20.00
15.00
10.00
5.00
0.00
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Fig. 2. Distribution of water consumption per hour in a day for the household area. (m3/h)
Figure 2 shows that the area has peak value at a point therefore, WSS which
belongs to it should be designed to meet 29,52 m3/h water. The household area is the
part of the whole area (see Fig. 1).
3.2 Fuzzy Logic Automation

The logic of the automation valve is to give water to the tank as much as need, to
control the fluctuations in consumption and when it is necessary in a particular area. It
means that the water is presented to a person without any lack of service quality.
Inputs and output were arranged and selected by considering the variables having
potential to be able to influence on the system. Table 3 given below shows inputs and
output of the fuzzy rules. They were chosen by considering the most determinant
factors that can impact on needs in water demand. Why these intervals of inputs were
chosen is for an example of Katip Çelebi University.
Table 3. Variables of fuzzy automation.

Variables Numeric values
Season From 1 to 12
People From 0 to 20,000
Time From 0 to 24
Fill rate of the tank From 0 to 1
Contribution of valve From 0 to 1
Season just was arranged from September to August and was separated 2 part as
Winter and Spring. Number of people coming to The University was separated 2 part
as Low and High. Time was divided as 0–7, 7–11, 11–13, 13–16, 16–19, 20–24. Fill
rate of the tank changes between from 0 to 1. It can be measured some electronical
devices sending message to fuzzy control system. Last of all, contribution of valve is
proportion of both, discharge calculated in design phase and discharge needed at that
moment. All the inputs can be got from the system itself so there is no need to human
for controlling the system.
4 Result
4.1 Comparison of Each Scenario

Water tank of The University helps to cost of WSSs and conjunction in water demand
by reducing peak values at water consumption. Generally, conventional tank systems
are used in universities like Izmir Katip Çelebi University. Universities have own
characteristic of water consumption, but water tanks absorb it. On the other hand,
proposed valve system can manipulate water consumption as desired. Characteristics of
water consumption for each valve system of tank is stated in Fig. 3.
As seen in Fig. 3. a water tank working with conventional ballcock system has own
uncontrollable characteristic water demand but a water tank working with NIV system
can be controlled as desired so it can be arranged according to benefit of the whole area.
Effect of both valve systems on whole area is graphed in Fig. 4.
45.00
40.00 Ballcock Supply
35.00 NIV Supply
30.00
25.00
20.00
15.00
10.00
5.00
0.00
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Fig. 3. Comparison of each valve system aspect of water discharge to the water tank. (m3/h)
70.00
65.00 Household+NIV Supply
60.00
55.00 Household+Ballcock Supply
50.00
45.00
40.00
35.00
30.00
25.00
20.00
15.00
10.00
5.00
0.00
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Fig. 4. Comparison of both scenario aspect of water consumption. (m3/h)
WSS with NIV absorbs peak values in water demands at region of Balatçık so
SCWSS can be decreased because of diminished peak water demand values.
Although NIV has advantages in regard of the other from aspect of the peak values,
NIV will cause to increase in water tank capacity.
Table 4. Comparison of effect of each scenario on peak values and tank capacity.
Scenario Peak water demand (m3/h) Required tank capacity (m3)
Conventional WSS 57,82 164,70
WSS with NIV 43,02 261,00
Since WSSs are modelled by considering peak water demand values, WSS with
NIV helps to reduce SCWSS because of decrease in max discharge needed according
to calculated results.
4.2 Integration of Fuzzy Logic to NIV

Variables were selected according to their impact on WSS and an example of case
study. Water discharge to the water tank of The University and its water consumption
distribution can be arranged thanks to the intelligent fuzzy logic system which is able to
consider variables. Other than simple design in new valve system helps to diminish
peak demands, integration of appropriate fuzzy logic rules mentioned in Sect. 3.2 help
on conjunction in water use thus getting more efficient water discharge. According to
the fuzzy rules given Sect. 3.2, sample simulation can be seen in Table 5.
Table 5. Sample simulation.

Inputs Output
Season Person Time Fill rate of the tank Contribution of valve
4.5 10,800 9.45 0,741 0,517
Contribution of valve named v was calculated as 0.517 according to given inputs.

v explains that what amount of water will discharge to the water tank of The University.
qc=qs ¼ v ð2Þ
qc = Calculated discharge before in design phase that needs to flow to water storage
tank
qs = Discharge that changes simultaneously as regards inputs.
5 Conclusion
This paper aims that stabilization of fluctuations at water demand. When water demand
is low then it is time to supply water to universities having an own water tank in a
household area having water consumption characterization like Fig. 2. Stabilization at
water demand can decrease the peak discharge at water demand, the worst scenario for
water demand, so maximum discharge that must be provided gets lower (becomes more
stable) thus lower SCWSS. This can be obtained by using the logic introduced in this
paper thanks to fuzzy logic automation that can arrange the stabilization of fluctuations
at water demands for reducing SCWSS.
Results show that the target of this research was attained. Peak values were decreased
thus lower SCWSS but increase in water tank capacity needed was seen in calculations
which has not impact on cost of the system as much as transmission line has.
The study in this paper can be extended by considering the tank capacity required.
Setup cost decreases with proposed idea from aspect of the pipe cost but tank capacity
required will increase so is setup cost proportionally. These economic aspects should
be investigated in more detail way. Cost efficiency should be determined by interre-
lating with them.
References
1. Marques, J., Cunha, M.C., Sousa, J., Savić, D.: Robust optimization methodologies for water
supply systems design. Drink. Water Eng. Sci. 5(1), 31–37 (2012)
2. Rossman, L.A., Grayman, W.M.: Scale-model studies of mixing in drinking water storage
tanks. J. Environ. Eng. 125(8), 755–761 (1999)
3. Awe, O.M., Okolie, S.T.A., Fayomi O.S.I.: Analysis and optimization of water distribution
systems: a case study of Kurudu post service housing estate, Abuja, Nigeria. Results Eng.
5 (2020)
4. Dey, N., Mandal, R., Subashini, M.M.: Design and implementation of a water level
controller using fuzzy logic. Int. J. Eng. Technol. 5(3), 2277–2285 (2013)
5. Aydogmus, Z.: Implementation of a fuzzy-based level control using SCADA. Expert Syst.
Appl. 36(3), 6593–6597 (2009)
6. Wei, J.: Intelligent building control of water tank based on fuzzy theory. In: 2010
International Conference on Intelligent Computation Technology and Automation ICICTA
2010, vol. 2, pp. 549–552 (2010)
7. Mamdani, E.H., Assilian, S.: An experiment in linguistic synthesis with a fuzzy logic
controller. Int. J. Man Mach. Stud. 7(1), 1–13 (1975)
8. Qin, T., Boccelli, D.L.: Estimating distribution system water demands using Markov Chain
Monte Carlo. J. Water Resour. Plan. Manag. 145(7) (2019)
9. Wang, H., Wang, W., Cui, Z., Zhou, X., Zhao, J., Li, Y.: A new dynamic firefly algorithm
for demand estimation of water resources. Inf. Sci. (Ny) 438, 95–106 (2018)
10. Navarrete-López, C., Herrera, M., Brentan, B.M., Luvizotto, E., Izquierdo, J.: Enhanced
water demand analysis via symbolic approximation within an epidemiology-based
forecasting framework. Water (Switzerland) 11(2), 1–17 (2019)
Fuzzy Cognitive Map
Based PI Controller Design
Aykut Denizci, Sinan Karadeniz, and Cenk Ulu(&)
Department of Mechatronics Engineering,

Yildiz Technical University, 34349 Istanbul, Turkey
[email protected],
Abstract. Fuzzy cognitive maps (FCMs) are often used in applications such as
politics, health, engineering and information technologies. In control applica-
tions, FCMs are generally used as supervisor controllers. In literature, there are
very few studies using the fuzzy cognitive map approach for the main controller
design. In this study, three different PI-type controller models based on FCM
approach are proposed. These models have 5, 4 and 3 tuning parameters,
respectively. In order to show the effectiveness of the proposed FCM based
controllers, simulation studies are performed on a nonlinear system. Perfor-
mances of the proposed three models are compared with performances of a
classical PI controller and a fuzzy PI controller with a 3 3 rule base. The
comparison results show that the proposed three controller models exhibit better
control performances than the classical PI and the fuzzy PI controller. The
proposed fuzzy cognitive map based controller models can effectively be used as
a main controller especially in nonlinear system control applications.
Keywords: Fuzzy cognitive map PI Nonlinear system Control Fuzzy

logic
1 Introduction
The derivation of mathematical models of nonlinear systems is a difficult task. Soft

computing techniques provide effective results for modeling and control of these
systems [1–6]. One of these techniques is a fuzzy cognitive map (FCM) method [7].
FCM was first proposed by Kosko [8] to represent and analyze the causal relationships
between different concepts. Initially, this method was used as a way to explain political
decision making processes. Today, it is used in many different areas such as health,
engineering, security, economy, trade and strategy development [9–13]. The reason of
using FCM method in the field of engineering is that it can easily express systems that
are difficult to express mathematically.
In control applications, the FCM method is mostly used in supervisor controller
design [14, 15]. There are very few studies in literature where FCM is used in the main
controller design. In [16], FCM based controllers are designed for translational motion
and yaw angle rotation control of an unmanned aerial vehicle, and these controllers are
used in the path tracking application of the unmanned aerial vehicle. In [17], a FCM
structure is used as a main controller, and applied on a heatex process.
Fuzzy Cognitive Map Based PI Controller Design 1251
In this study, three different FCM based controller models in PI structure are
proposed. In these controller models, one or more of error, rate of error change, rate of
change of the error membership value and integral of error are used as controller inputs.
The proposed Model I, Model II and Model III have 5, 4 and 3 tuning parameters,
respectively. Performances of the proposed models are compared with performances of
a classical PI controller and a simple fuzzy controller with a 3 3 rule base on a
nonlinear system. The settling time, overshoot, integral absolute error (IAE) and
integral time absolute error (ITAE) values of system responses and control signal
ranges are used as performance criteria in the comparisons. The comparison results
show that the proposed three controller models outperform the classical PI and the
fuzzy PI controller.
This study is organized as follows. In Sect. 2, the fuzzy cognitive map concept is
given. In Sect. 3, the proposed controller models are presented. In Sect. 4, simulation
studies are given. Finally, in Sect. 5, the results and future studies are provided.
2 Fuzzy Cognitive Map
Fuzzy cognitive maps are extended version of classical cognitive maps. In FCM,
weights representing a causal relationship between concepts are defined within the
interval of [–1 1]. Fuzzy cognitive maps have the advantages of both artificial neural
networks and fuzzy logic. Another advantage of this method is that it can use the
knowledge and human experience when creating a fuzzy cognitive map. When creating
fuzzy cognitive maps, the relationships between the concepts are determined by Wij
weights. According to the conditions of these weights, the relationship between two
concepts can be defined as follows
• Wij > 0 indicates direct (positive) causality between two concepts.
• Wij < 0 indicates inverse (negative) causality between two concepts
• Wij = 0 indicates no relationship between two concepts
The inference in FCM can be defined as
!
ðk þ 1Þ ðk Þ
XN ðk Þ
Ai ¼f Ai þ j ¼ 1 Aj Wji ð1Þ
j 6¼ i
Here, A(k+1)
i is the value of Ci concept at time k + 1, A(k)
j is value of Cj concept at time
k, Wji is the effect of Cj concept on Ci concept, N is number of concepts and f (.) is the
threshold function.
1252 A. Denizci et al.
3 Proposed FCM-Based PI Controller Models

3.1 Model I
The structure of the proposed Model I is shown in Fig. 1. In this model, the error
(e) and the rate of error change (de) are used as the controller inputs, while the integral
of the output signal of the FCM structure is used as the controller output. This model
includes totally 5 tuning parameters which are two input gains, P and D, one output
gain, I, and two causal relationship weights, Wp and Wd. The mathematical expression
of the control signal is given as follows
Z

u¼I f lðe PÞ sgnðe PÞ Wp þ lðde DÞ sgnðde DÞ Wd ð2Þ
where, l is the membership degree, f (.) is the activation function, and sgn(.) indicates
the sign function. The sign function is used to obtain negative control signals, since the
membership functions are only positive definite in [0 1].
Fig. 1. Structure of Model I
3.2 Model II
To reduce the number of tuning parameters, in Model II, the rate of error membership
change value is used instead of the rate of error change. Therefore, no gain parameter is
used for this signal. The structure of the proposed Model II is shown in Fig. 2. This
model includes totally 4 tuning parameters which are one input gain, P, one output
gain, I, and two relationship weights, Wp and Wd. The control signal is defined as
Z
d ðlðe PÞ sgnðe PÞÞ
u¼I f lðe PÞ sgnðe PÞ Wp þ Wd ð3Þ
dt
Fig. 2. Structure of Model II
3.3 Model III

In Model III, the error and the integral of the error signals are used as the controller
inputs without any gain parameters. The structure of Model III is shown in Fig. 3. This
model has totally 3 tuning parameters which are one output gain, K, and two rela-
tionship weights, Wp and Wi. The mathematical definition of the controller output
signal is as follows.
Z Z
u ¼ K f lðeÞ sgnðeÞ Wp þ l e sgn e Wi ð4Þ
Fig. 3. Structure of Model III
4 Simulation Studies
In this section, in order to show the effectiveness of the proposed FCM based PI
controllers, simulation studies are performed on a nonlinear system. In the simulation,
the sampling time is chosen as 0.025 s. The mathematical model of the nonlinear
system is
y_ ¼ y þ 7y2 þ u ð5Þ
Here u and y are the control input and the system output, respectively. The unit step
function is selected as the system reference input. The parameters of the proposed con-
trollers which are given in Table 1 are determined by using a genetic algorithm (GA). The
population size and the number of generation are chosen as 100. Linear function with a
slope value of 1 is used as an activation function in all controllers. Membership functions
used for all three proposed methods are selected as an inverted triangle as shown in Fig. 4.
The reason for selecting this membership function type is to generate the maximum
controller signal value when the error is high in either negative or positive regions.
Table 1. Determined parameter values of the proposed controller models

Model I Model II Model III
P = 2.65 P = 3.32 Wp = 0.41
D = 0.14 Wp = 0.99 Wi = 0.97
Wp = 0.95 Wd = 0.13 K = 306.87
Wd = 0.97 I = 197.25
I = 399.12
Fig. 4. Membership function used in the proposed controller models
Performances of the proposed three controller models are compared with perfor-
mances of a classical PI controller and a fuzzy PI controller. Design parameters of the
PI controller and the fuzzy controller are determined by GA. PI controller parameters
are determined as P = 86.90 and I = 234.68. The Sugeno type controller structure is
chosen for the fuzzy PI controller. In order to make a fair comparison in terms of the
number of tuning parameters, a 3 3 rule base is preferred in the design. Input
membership functions and the rule table are shown in Fig. 5a and Fig. 5b, respectively.
The singleton values of output fuzzy sets Negative, Zero and Positive are chosen as –1,
0 and 1, respectively. The input and output scaling factors are determined as follows;
P = 3.05, D = 0.21, and I = 360.04.
Fig. 5. a) Input membership functions b) Rule table of the fuzzy PI controller
The obtained system responses and control signals are shown in Fig. 6a and b,
respectively. The overshoot, settling time, IAE and ITAE values of system responses
and control signal ranges are used as performance criteria in the comparisons. The
performance comparison results are given in Table 2.
Fig. 6. a) System responses b) Control signals
Table 2. Performance comparison table

Overshoot [%] Settling time [s] (±2%) Control signal range IAE ITAE
Model I 1.6 0.15 [–6.4 7.3] 0.0783 0.0038
Model II 0.2 0.125 [–6.5 14.7] 0.0683 0.0027
Model III 0 0.075 [–4.6 125.4] 0.0162 0.0004
Classical PI 8.8 0.55 [–7.3 86.9] 0.0392 0.0065
Fuzzy-PI 0.4 0.175 [–6.6 5.4] 0.0954 0.0056
The comparison results show that the proposed three controller models show better
performance than the classical PI and the fuzzy PI controller in terms of the settling
time, and ITAE. Model III has the best control performance among the others while the
classical PI controller has poor control performance with 8.8% overshoot and 0.55 s
settling time. Moreover, the range of the control signal generated by Model I is almost
the lowest. Since there is a tradeoff between the rising time and the amplitude of the
control signal, the classical PI and Model III have control signals with higher
amplitudes.
5 Conclusion
In this study, three FCM based PI controller models are proposed. The effectiveness of
the proposed controller models are shown through simulation studies performed on
nonlinear system control. The simulation results show that the proposed fuzzy cog-
nitive map based controller models can effectively be used as a main controller
especially in nonlinear system control applications.
In order to increase control performance in applications, the proposed controller
models can directly be used instead of the existing classical PI controllers. Addition-
ally, the nonlinear characteristics of the proposed controllers can easily be increased by
using different types of membership functions. Thus, control performance can be
improved further especially in the nonlinear control applications. Moreover, by using
the proposed fuzzy cognitive map based design approach, the general PD and PID type
controller structures can also be derived.
In the future studies, controller models in general PID structure will be developed
by using the proposed FCM based approach. In addition, practical applications of
proposed controller models will be realized.
References
1. Er, M.J., Deng, C., Wang, N.: A novel fuzzy logic control method for multi-agent systems
with actuator faults. In: 2018 IEEE International Conference on Fuzzy Systems (FUZZ-
IEEE), pp. 1–7. IEEE, Rio de Janeiro (2018)
2. Amuthameena, S., Monisa, S.: Design of fuzzy logic controller for a non-linear system. In:
IEEE International Conference on Electrical, Instrumentation and Communication Engi-
neering (ICEICE), pp. 1–7. IEEE, Karur (2017)
3. Azizi, M., Ejlali, R.G., Ghasemi, S.A.M., Talatahari, S.: Upgraded whale optimization
algorithm for fuzzy logic based vibration control of nonlinear steel structure. Eng. Struct.
192, 53–70 (2019)
4. Zhang, X., Ma, X., Zhu, H., Liu, H.: Neural network controller design for uncertain
nonlinear systems based on backstepping control algorithm. In: 2017 29th Chinese Control
and Decision Conference (CCDC), pp. 1623–1627. IEEE, Chongqing (2017)
5. Zhang, Y., Tao, G., Chen, M.: Adaptive neural network based control of noncanonical
nonlinear systems. In: IEEE Transactions on Neural Networks and Learning Systems,
pp. 1864–1877. IEEE (2016)
6. Slama, S., Errachdi, A., Benrejeb, M.: Model reference adaptive control for MIMO nonlinear
systems using RBF neural networks. In: International Conference on Advanced Systems and
Electric Technologies (IC_ASET), pp. 346–351. IEEE, Hammamet (2018)
7. Puheim, M., Vascak, J., Madarasz, L.: A proposal for multi-purpose fuzzy cognitive maps
library for complex system modeling. In: 13th IEEE International Symposium on Applied
Machine Intelligence and Informatics (SAMI), pp. 175–180. IEEE, Herl’any (2015)
8. Kosko, B.: Fuzzy cognitive maps. Int. J. Man Mach. Stud. 24, 65–75 (1986)
9. Najafi, A., Amirkhani, A., Papageorgiou, E.I., Mosavi, M.R.: Medical decision making
based on fuzzy cognitive map and a generalization linguistic weighted power mean for
computing with words. In: IEEE International Conference on Fuzzy Systems (FUZZ-IEEE),
pp. 1–6. IEEE, Naples (2017)
10. Hajek, P., Prochazka, O.: Interval-valued fuzzy cognitive maps for supporting business
decisions. In: IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), Vancouver,
pp. 531–536 (2016)
11. Sudhagar, C.: Role of fuzzy cognitive maps in smart education system. In: 4th MEC
International Conference on Big Data and Smart City (ICBDSC), Muscat, pp. 1–6 (2019)
12. Arruda, L.V.R., Mendonca, M., Neves, F., Chrun, I.R., Papageorgiou, E.I.: Artificial life
environment modeled by dynamic fuzzy cognitive maps. IEEE Trans. Cogn. Dev. Syst. 10
(1), 88–101 (2016)
13. Papageorgiou, E.I.: Review study on fuzzy cognitive maps and their applications during the
last decade. In: IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), Taipei
(2011)
14. Groumpos, P.P., Stylios, C.D.: Modelling supervisory control systems using fuzzy cognitive
maps. Chaos, Solitons Fractals 11(1–3), 329–336 (2000)
15. Stylios, C.D., Groumpos, P.P.: Modeling complex systems using fuzzy cognitive maps.
IEEE Trans. Syst. Man Cybern. Part A Syst. Hum. 34(1), 155–162 (2004)
16. Amirkhani, A., Shirzadeh, M., Papageorgiou, E.I., Mosavi, M.R.: Visual-based quadrotor
control by means of fuzzy cognitive maps. ISA Trans. 60, 128–142 (2016)
17. de Souza, L.B., Soares, P.P., Mendonca, M., Mourhir, A., Papageorgiou, E.I.: Fuzzy
cognitive maps and fuzzy logic applied in industrial processes control. In: IEEE International
Conference on Fuzzy Systems (FUZZ-IEEE), Rio de Janeiro, pp. 1–8 (2018)
On 2-Fuzzy Metric Spaces
Elif Güner(B) and Halis Aygün
Department of Mathematics, Kocaeli University,

Umuttepe Campus, 41380 Kocaeli, Turkey
{elif.guner,halis}@kocaeli.edu.tr
Abstract. The target of this study is to describe the notion of 2-fuzzy

metric spaces which is the extension of 2-metric space in the setting of
that the area of triangle shaped by three distinct points is a convex, non-
negative, normal and upper semi-continuous fuzzy number. For this aim,
we first describe the notion of 2-fuzzy metric spaces, study some of their
properties and give some examples. Then we investigate the relationship
between 2-fuzzy metric spaces and 2-Menger spaces. Also, we discuss the
triangle inequalities and its level forms in 2-fuzzy metric spaces. Finally,
using the obtained properties and relationships we show that every 2-
fuzzy metric induces a Hausdorff topology which is metrizable.
Keywords: Fuzzy number · Triangle inequality · 2-metric · 2-fuzzy

metric · Hausdorff topology
1 Introduction
The notion of 2-metric spaces which is a nonlinear extension of an idea of ordi-
nary metric spaces was given by Gähler in the 1960s [4]. Geometrically a value
of three points under a 2-metric function represents the area of a triangular
region shaped by these three points as vertices. Since in the extensions of crisp
metric spaces only 2-metric space is not equivalent to a crisp metric space via
topologically, this notion is extensively studied by many researchers [1,4].
In 1987, Zeng [6] described the concept of probabilistic 2-metric which
expresses the situation where the area shaped by three points is not exactly
known, but the probabilities of the possible values of this area are known. Golet
[5] introduced the notion of 2-Menger space which is a special kind of proba-
bilistic 2-metric space where the triangular inequality is given with the help of
a triangular norm.
The main purpose of this paper is to introduce the notion of 2-fuzzy met-
ric spaces which is the expansion of 2-metric space considering the fuzziness.
This paper is contained the following sections: In Sect. 2, we recollect some basic
notions and relevant concepts that are used in the main section. In Sect. 3, we
define the notion of 2-fuzzy metric space in which we consider that the area of
the triangular region shaped by three distinct points is a convex, non-negative,
normal and upper semi-continuous fuzzy number. After we discuss the triangle
On 2-Fuzzy Metric Spaces 1259
inequalities and its level forms in 2-fuzzy metric spaces, we investigate the rela-
tionship between 2-fuzzy metric spaces and 2-Menger spaces. Using the obtained
properties and relationships we show that every 2-fuzzy metric induces a Haus-
dorff topology and the constructed topology is metrizable. Finally, we give a
brief summary in Sect. 4.
2 Preliminaries
In this part, we recollect some notions such that fuzzy number, 2-metric spaces
and 2-Menger spaces which are used in the main section.
Definition 1 [2].
(1) A mapping u : R → [0, 1] is said to be a fuzzy number (FN, shortly).
(2) A FN u is said to be convex if u(t1 ) ≥ min(u(t2 ), u(t3 )) where t2 ≤ t1 ≤ t3 .
(3) A FN u is said to be normal if there is a t ∈ R such that u(t) = 1.
(4) The set [u]α = {t|u(t) ≥ α}, α ∈ (0, 1], is called an α-level set of the FN
u. The α-level set of the fuzzy number u is a closed interval [λα , ρα ] for
all α ∈ (0, 1] where the values λα = −∞ and ρα = ∞ are admissible. If
λα = −∞, then [λα , ρα ] means (−∞, ρα ].
(5) A FN u is called non-negative if u(t) = 0 for each t < 0.
We will denote the set of all convex, normal and upper semi-continuous FNs by
G and the non-negative elements of G by E.
Each u ∈ R may be thought as a FN ū as follows:

1, t=u
ū(t) = . (1)
0, t = u
A sequence (un ) ⊆ G converges to u ∈ G if limn→∞ λα (un ) = λα (u) and

limn→∞ ρα (un ) = ρα (u) for each α ∈ (0, 1], where [u]α = [λα (u), ρα (u)] and
[un ]α = [λα (un ), ρα (un )]. If (un ) ⊆ E and u = 0̄, then for each α ∈ (0, 1], we
have the following implication
0 = λα (u) ≤ ρα (u) ≤ λα (un ) ≤ ρα (un ). (2)

Therefore, limn→∞ un = 0̄ if and only if limn→∞ ρα (un ) = 0. The readers can
found the other algebraic operations and properties of FNs in [2].
Definition 2 [4]. Let U be a non-empty set. A mapping σ defined on U × U × U

taking value in the set of real numbers is said to be a 2-metric on U provided
that the following conditions are satisfied:
(2M1) There is a point z ∈ U such that σ(x, u, z) = 0 for each pair of
different points x, u ∈ U ,
(2M2) σ(x, u, z) = 0 when at least two of x, u, z are equal,
(2M3) σ(x, u, z) = σ(x, z, u) = σ(z, x, u) = σ(z, u, x) = σ(u, x, z) =
σ(u, z, x) for each x, u, z ∈ U ,
1260 E. Güner and H. Aygün
(2M4) σ(x, u, z) ≤ σ(x, u, w) + σ(x, w, z) + σ(w, u, z) for each x, u, z, w ∈ U.

Provided that σ is a 2 − metric on U , the ordered pair (U, σ) is called a
2−metric space. The value σ(x, u, z) represents the area of the triangular region
shaped by x, u and z in U as its vertices (see [1]).
Proposition 1 [3]. Let (U, σ) be a 2-metric space. Then the collection B =

{Bσ (x, ε) : x ∈ U, ε > 0} is a base for a crisp topology on U where Bσ (x, ε) =
{u ∈ U : supz∈U σ(x, u, z) < ε } that is called the 2-metric topology induced by d
and denoted by τσ .
Definition 3 [3]. Let (U, τ ) be a topological space. If there exists a 2-metric σ

on U such that τ = τσ , then (U, τ ) is said to be a 2-metrizable topological space.
Definition 4 [5]. Let Δ+ be the set of all non-decreasing and left continuous
function F : R → [0, 1] such that F (0) = 0 and limt→∞ F (t) = 1. Let N ∈ Δ+
be defined as
0, t ≤ 0
N (t) = (3)
1, t > 0
Let U be a nonempty set, ∗ be a t-norm and F : U × U × U → Δ+ be a mapping.
The triplet (U, F, ∗) is said to be a 2-Menger space provided that the following
conditions are satisfied:
(2MG1) There is a point z ∈ U such that Fxuz = N for each pair of different
points x, u ∈ U ,
(2MG2) Fxuz = N when at least two of x, u, z are equal,
(2MG3) Fxuz = Fxzu = Fuxz = Fuzx = Fzxu = Fzux for each x, u, z ∈ U ,
(2MG4) Fxuz (t1 +t2 +t3 ) ≥ Fxuw (t1 )∗Fxwz (t2 )∗Fwuz (t3 ) for each x, u, z, w ∈
U and t1 , t2 , t3 ≥ 0.
3 2-Fuzzy Metric Spaces
In this part, we present the structure of 2-fuzzy metric spaces, obtain some
properties related to the triangle inequalities and its level forms in 2-fuzzy metric
spaces, present some examples and study the metrizability of this spaces.
Definition 5. Let U be a non-empty set and CL , CR : [0, 1] × [0, 1] × [0, 1] →

[0, 1] be two non-decreasing, symmetric mappings in all variables satisfying
CL (0, 0, 0) = 0 and CR (1, 1, 1) = 1. A mapping M : U × U × U → E is said to
be a 2-fuzzy metric provided that the following conditions are satisfied:
(2FM1) There is a point z ∈ U such that M (x, u, z) = 0̄ for each pair of
different points x, u ∈ U ,
(2FM2) M (x, u, z) = 0̄ when at least two of x, u, z are equal,
(2FM3) M (x, u, z) = M (x, z, u) = M (u, x, z) = M (u, z, x) = M (z, x, u) =
M (z, u, x) for each x, u, z ∈ U ,
(2FM4) For each x, u, z, w ∈ U ,
(i) M (x, u, z)(t1 + t2 + t3 ) ≥ CL (M (x, u, w)(t1 ), M (x, w, z)(t2 ), M (w, u, z)(t3 )

whenever t1 ≤ λ1 (x, u, w), t2 ≤ λ1 (x, w, z), t3 ≤ λ1 (w, u, z) and t1 +t2 +t3 ≤
λ1 (x, u, z)
(ii) M (x, u, z)(t1 + t2 + t3 ) ≤ CR (M (x, u, w)(t1 ), M (x, w, z)(t2 ), M (w, u, z)(t3 )
whenever t1 ≥ λ1 (x, u, w), t2 ≥ λ1 (x, w, z), t3 ≥ λ1 (w, u, z) and t1 + t2 + t3 ≥

λ1 (x, u, z)
where
[M (x, u, z)]α = [λα (x, u, z), ρα (x, u, z)]
for each α ∈ (0, 1]. The quadruple pair (U, M , CL , CR ) is called a 2-fuzzy metric
space. The value M (x, u, z)(t) can be interpreted as the possibility that the area
of the triangle shaped by x, u and z in U as its vertices is t.
Remark 1.(1) Note that 2-metric space is a particular case of the 2-fuzzy metric
space. If we chose

0, if at least two of e1 , e2 , e3 are equal to 0
CL (e1 , e2 , e3 ) = 0, CR (e1 , e2 , e3 ) =
1, otherwise
(4)
then the triangle inequality (2FM4) implies the 2-metric triangle inequality
(2M4).
(2) M (x, u, z) : (0, ∞) → R+ is non-decreasing on [0, λ1 (x, u, z)] and non-
increasing on [λ1 (x, u, z), ∞].
Lemma 1. Let (U, M , CL , CR ) be a 2-fuzzy metric space. If CL (e1 , e2 , e3 ) ≥

max(e1 , e2 , e3 ) for each e1 , e2 , e3 ∈ [0, 1], then λ1 (x, u, z) = 0 for each x, u, z ∈ U.
If CR (e1 , e2 , e3 ) ≤ min(e1 , e2 , e3 ) for e1 , e2 , e3 ∈ [0, 1], then M (x, u, z)(t) = 0
for each t > λ1 (x, u, z).
Proof. Let CL (e1 , e2 , e3 ) ≥ max{e1 , e2 , e3 } for each e1 , e2 , e3 ∈ [0, 1]. Assume

that λ1 (x, u, z) > 0 for some x, u, z ∈ U . Let 0 < t < λ1 (x, u, z). By (2FM4)(i),
we have the following inequality
M (x, u, z)(t) ≥ CL (M (x, u, z)(t), M (x, z, z)(0), M (z, u, z)(0))

≥ max(M (x, u, z)(t), M (x, z, z)(0), M (z, u, z)(0)) = 1 (5)
which is a contradiction since t < λ1 (x, u, z). Hence λ1 (x, u, z) = 0. Let

CR (e1 , e2 , e3 ) ≤ min(e1 , e2 , e3 ) for each e1 , e2 , e3 ∈ [0, 1]. Assume that
M (x, u, z)(t) > 0 for each t > λ1 (x, u, z). Similarly, by (2FM4)(ii), we have
the following inequality
M (x, u, z)(t) ≤ CR (M (x, u, z)(t), M (x, z, z)(0), M (z, u, z)(0))

≤ min(M (x, u, z)(0), M (x, z, z)(t), M (z, u, z)(0)) = 0 (6)
which is a contradiction since t > λ1 (x, u, z).

Lemma 2. CR (e1 , e2 , e3 ) = max(e1 , e2 , e3 ) for each e1 , e2 , e3 ∈ [0, 1] in the

triangle inequality (2FM4)(ii) is coincident with the triangle inequality
ρα (x, u, z) ≤ ρα (x, u, w) + ρα (x, w, z) + ρα (w, u, z) (7)
for each α ∈ (0, 1] and x, u, z, w ∈ U.

Proof. Let CR (e1 , e2 , e3 ) = max(e1 , e2 , e3 ) for each e1 , e2 , e3 ∈ [0, 1] in the tri-
angle inequality (2FM4)(ii), α ∈ (0, 1] and x, u, z, w ∈ U. We may suppose that
ρα (x, u, w) < ∞, ρα (x, w, z) < ∞ and ρα (w, u, z) < ∞. Otherwise, proof is
completed. Assume that
ρα (x, u, z) > ρα (x, u, w) + ρα (x, w, z) + ρα (w, u, z) (8)
for each α ∈ (0, 1] and x, u, z, w ∈ U. Then there are t1 > ρα (x, u, w) ≥

λ1 (x, u, w), t2 > ρα (x, w, z) ≥ λ1 (x, w, z) and t3 > ρα (w, u, z) ≥ λ1 (w, u, z) such
that t1 +t2 +t3 = ρα (x, u, z) ≥ λ1 (x, u, z). Let us take α = M (x, u, z)(t1 +t2 +t3 ).
By (2FM4)(ii) and the assumption, we have
M (x, u, z)(t1 + t2 + t3 ) ≤ max(M (x, u, w)(t1 ), M (x, w, z)(t2 ), M (w, u, z)(t3 )) < α
(9)
which is a contradiction. To prove the sufficient condition let x, u, z, w ∈ U ,
t1 ≥ λ1 (x, u, w), t2 ≥ λ1 (x, w, z) and t3 ≥ λ1 (w, u, z). Since α = M (x, u, z)(t1 +
t2 + t3 ), we have
t1 + t2 + t3 ≤ ρα (x, u, z) ≤ ρα (x, u, w) + ρα (x, w, z) + ρα (w, u, z). (10)
Hence, we obtain that t1 ≤ ρα (x, u, w), t2 ≤ ρα (x, w, z) or t3 ≤ ρα (w, u, z). This

implies that M (x, u, w)(t1 ) ≥ α, M (x, w, z)(t2 ) ≥ α or M (w, u, z)(t3 ) ≥ α.
Therefore, we have the following inequality
max(M (x, u, w)(t1 ), M (x, w, z)(t2 ), M (w, u, z)(t3 )) ≥ α = M (x, u, z)(t1 + t2 + t3 )
(11)
for each x, u, z ∈ U when t1 ≤ ρα (x, u, w), t2 ≤ ρα (x, w, z), t3 ≤ ρα (w, u, z) and
t1 + t2 + t3 ≥ λ1 (x, u, z).
Lemma 3. CL (e1 , e2 , e3 ) = min(e1 , e2 , e3 ) for each e1 , e2 , e3 ∈ [0, 1] in the tri-
angle inequality (2FM4)(i) is coincident with the triangle inequality
λα (x, u, z) ≤ λα (x, u, w) + λα (x, w, z) + λα (w, u, z) (12)
for each α ∈ (0, 1] and x, u, z, w ∈ U.

We will show in the next example that each 2-metric space induces a 2-fuzzy
metric space with the suitable CR and CL .
Example 1. Let (U, σ) be a 2-metric space, CL = min and CR = max. Define
Mσ : U × U × U → E by

1, t = σ(x, u, z)
Mσ (x, u, z)(t) = (13)
0, otherwise
for each x, u, z ∈ U . It is clear that the conditions (2FM1)–(2FM3) are satisfied.

Now, we show that the condition (2FM4)(i)–(ii) are hold. Since λα (x, u, z) =
σ(x, u, z) = ρα (x, u, z) for each α ∈ (0, 1] and x, u, z ∈ U , we have
λα (x, u, z) ≤ λα (x, u, w) + λα (x, w, z) + λα (w, u, z) (14)
ρα (x, u, z) ≤ ρα (x, u, w) + ρα (x, w, z) + ρα (w, u, z). (15)
From Lemma 2 and 3, we obtain that the condition (2FM4)(i)–(ii) holds. There-
fore, (U, Mσ , min, max) is a 2-fuzzy metric space. Hence, we can induce a 2-
fuzzy metric space when a 2-metric space is given. In this example, if we let that
CL (0, 0, 1) = 0 and CR is right continuous instead of CL = min and CR = max,
then we deduce that (U, Mσ , CL , CR ) is a 2-fuzzy metric space.
Now, we show in the next example that each 2-Menger space induces a 2-
fuzzy metric space.
Example 2. Let (U, F, ∗) be a 2-Menger space and the mapping M : U ×U ×U →
E be defined by

0, t < txuz = sup{t|Fxuz (t) = 0}
M (x, u, z)(t) = (16)
1 − Fxuz (t), t ≥ txuz
It is obvious that M (x, u, z) ∈ E from the conditions of Fxuz . Let CR , CL :
[0, 1] × [0, 1] × [0, 1] → [0, 1], CR (e1 , e2 , e3 ) = 1 − [(1 − e1 ) ∗ (1 − e2 ) ∗ (1 − e3 )]
and CL = 0. Then CR is non-decreasing mapping such that CR (1, 1, 1) = 1 and
also
M (x, u, z)(t1 + t2 + t3 ) = 1 − Fxuz (t1 + t2 + t3 )
≤ 1 − [Fxuw (t1 ) ∗ Fxwz (t2 ) ∗ Fwuz (t3 )]
= CR (1 − Fxuw (t1 ), 1 − Fxwz (t2 ), 1 − Fwuz (t3 ))(17)
which show that the condition (2FM4)(ii) is satisfied. Also, if CR = max, then
(U, M , 0, max) is a 2-fuzzy metric space when M is defined in the same way.
The next example illustrates that the converse statement is valid if the 2-
fuzzy metric space satisfies the condition
lim M (x, u, z)(t) = 0, ∀x, u, z ∈ U. (18)
t→∞
In this case, we can induce a 2-Menger space given a 2-fuzzy metric space
satisfying the condition (18). Let (U, M , CL , CR ) be a 2-fuzzy metric space such
that CR (e, e, 1) = CR (e, 1, e) = CR (1, e, e) = 1 for each e ∈ (0, 1] and define

0, t ≤ λ1 (x, u, z)
Fxuz (t) = . (19)
1 − M (x, u, z)(t), t ≥ λ1 (x, u, z)
Since M (x, u, z) is non-decreasing and left continuous on [λ1 (x, u, z), ∞),
Fxuz is non-decreasing and left continuous on [0, ∞). Obviously, since the con-
dition (18), limt→∞ Fxuz = 1. Also, Fxuz (0) = 0. If we take e1 ∗ e2 ∗ e3 =
1− CR (1− e1 , 1− e2 , 1− e3 ) for each e1 , e2 , e3 ∈ [0, 1], then the condition (2MG4)
is satisfied. Therefore, (U, F, ∗) is a 2-Menger space.
Definition 6. Let (U, M , CL , CR ) be a 2-fuzzy metric space. A sequence (un )

in U is said to be
(i) convergent to u ∈ U if limn→∞ M (un , u, z) = 0̄ for each z ∈ U .

(ii) a Cauchy sequence if limn→∞ M (un , um , z) = 0̄ for each z ∈ U .
(iii) (U, M , CL , CR ) is said to be complete if each Cauchy sequence in U is con-
vergent to some point u ∈ U .
Theorem 1. Let (U, M , CL , CR ) be a 2-fuzzy metric space such that

CR (e, e, e) → 0 as e → 0+ and K(α, ε) = {(x, u) ∈ U ×U |supz∈U ρα (x, u, z) < ε}.
Then the family {K(α, ε)|0 < α ≤ 1, ε > 0} forms a base for a Haus-
dorff uniformity U on U × U . If {(αn , εn )} is a convergent sequence to (0, 0),
then {U (αn , εn )} is a countable base for U. Furthermore, the family of sets
Kx (α, ε) = {u ∈ U |supz∈U ρα (x, u, z) < ε} forms a local base for a Hausdorff
topology τM on U , this topology is termed the topology induced by M and also
τM is metrizable.
Proof. We will show that the family {K(α, ε)|0 < α ≤ 1, ε > 0} satisfies the
axioms which are to be a base for a Hausdorff uniformity. Let K(α, ε) be given.
Since M (x, x, u) = 0̄ for any x, u ∈ U , we get that (x, x) ∈ K(α, ε). Thus, if
we take Δ = {(x, x)|x ∈ U }, then Δ ⊂ K(α, ε). It is obvious that K(α, ε) is
symmetric. And, it is simple to show that K(min(α, α ), min(ε, ε )) ⊂ K(α, ε) ∩
K(α , ε ) for each α, α ∈ (0, 1] and ε, ε > 0.Now, let ε = 3ε and α be too small
positive number such that CR (α , α , α ) < α. Suppose that (x, w), (w, u) ∈
K(α , ε ). Then, we get that ε > λ1 (x, w, z), M (x, w, z)(ε ) < α and ε >
λ1 (u, w, z), M (u, w, z)(ε ) < α for each z ∈ U . Therefore, since M (x, u, z) is
non-increasing on [λ1 , ∞), we get
ε ε ε ε
M (x, u, z)( ) ≤ CR (M (x, u, w)( ), M (x, w, z)( ), M (w, u, z)( )) < α(20)
2 3 3 3
for each t ≥ max(ε, λ1 (x, u, z)). Thus, we obtain ρα (x, u, z) < 2ε and this means
that (x, u) ∈ K(α, ε). Hence, K(α , ε ) ◦ K(α , ε ) ⊂ K(α, ε). As a result, the
family {K(α, ε)|0 < α ≤ 1, ε > 0} is a base for a Hausdorff Uniformity on
U × U.
Corollary 1. If (U, M , CL , CR ) is a 2-fuzzy metric space, then the topological

space (U, τm ) is first countable.
Corollary 2. If (U, Mσ , CL , CR ) is the 2-fuzzy metric space which is given in

Example 1 where CL (0, 0, 1) = 0 and CR is right continuous, then the topology
τσ induced by σ is coincident with the topology TM σ generated by Mσ .
Definition 7. A topological space (U, τ ) is said to admit a compatible 2-fuzzy

metric if the exists a 2-fuzzy metric M on U such that τ = τM .
Theorem 2. If a topological space (U, τ ) admits a compatible 2-fuzzy metric M ,

then (U, τ ) is metrizable.
Proof. Let M be a compatible 2-fuzzy metric for the topological space (U, τ ). i.e.
τ = τM . Then, the topology τM induced by the 2-fuzzy metric M is metrizable
by Theorem 1. Hence (U, τ ) is metrizable.
Theorem 3. If a topological space (U, τ ) is 2-metrizable,then (U, τ ) admits a

compatible 2-fuzzy metric M .
Proof. Let (U, τ ) be a 2-metrizable topological space. Then, there is a 2-metric

on U such that τ = τσ . If σ is a 2-metric on U , then we can induce a 2-fuzzy
metric space (U, Mσ , CL , CR ) in the form given in Example 1. Also, it is known
that τM σ = τσ from Theorem 1. Hence, (U, τ ) admits the compatible 2-fuzzy
metric Mσ .
Theorem 4. If a topological space (U, τ ) is compact, then each compatible 2-

fuzzy metric space (U, M , CL , max) is complete.
Proof. Assume that (U, τ ) is a compact topological space and (U, M , CL , CR ) be

a compatible 2-fuzzy metric space for (U, τ ). Let (un ) be a Cauchy sequence in
(U, M , CL , CR ). Since τ = τM , (U, τM ) is compact. So, (un ) has a subsequence
(unk ) converging to u ∈ U . Let ε > 0 be given. Since (un ) is a Cauchy sequence
and (unk ) converges to u ∈ U , there is n1 , n2 ∈ N such that ρα (un , unk , z) < 3ε
for each n ≥ n1 and ρα (unk , u, z) for each n ≥ n2 and z ∈ U . If we choose
n0 = min(n1 , n2 ), we have ρα (un , u, z) ≤ ρα (un , u, unk ) + ρα (un , unk , z) +
ρα (unk , u, z) < ε by Lemma 2. Hence, we obtain that (un ) converges to u in
(U, M , CL , max) and (U, M , CL , max) is complete.
4 Conclusion
In this paper, we define the notion of 2-fuzzy metric space as a generalization of
2-metric space in the fuzzy manner. In this view, we consider the area shaped
by three points as a fuzzy number instead of a single real number. Then we
study the relations between the other related concepts and induce a topology
from this space and research the metrizability of this topology. In future work,
we will research the compactness and completeness of this structure in detail.
Also, the fixed point theory could be studied in this space and some application
areas could be found in mathematics, engineering, economics and so on.
References
1. Aliouche, A., Simpson, C.: Fixed points and lines in 2-metric spaces. Adv. Math.
229(1), 668–690 (2012)
2. Chai, Y., Zhang, D.: A representation of fuzzy numbers. Fuzzy Sets Syst. 295, 1–18
(2016)
3. Farhangdoost, M.R.: Metrizable and 2-metrizable topological spaces. J. Dyn. Syst.
Geom. Theor. 10(1), 61–69 (2012)
4. Gähler, S.: 2-Metrische Räume und ihre topologische struktur. Mathe. Nachr. 26,
115–118 (1963)
5. Golet, I.: Fixed point theorems for multivalued mapping in probabilistic 2-metric
Spaces. An. St. Univ. Ovidius Constanta 3, 44–51 (1995)
6. Zeng, W.Z.: Probabilistic 2-metric spaces. J. Math. Res. Exposition 7, 241–245
(1987)
Some Properties of Partial Fuzzy
Metric Topology
Ebru Aydogdu(B) , Başak Aldemir , Elif Güner , and Halis Aygün
Department of Mathematics, Kocaeli University,

Umuttepe Campus, 41380 Kocaeli, Turkey
{ebru.diyarbakirli,elif.guner,halis}@kocaeli.edu.tr,
[email protected]
Abstract. The objective of this work is to study some topological prop-

erties of partial fuzzy metric spaces. For this purpose, we first define a
topology generated by the partial fuzzy metric in the sense of Sedghi
et al. [8] We give some relationships between the topologies induced by
partial metric and partial fuzzy metric. Then, we investigate various
properties including countability, completeness and separation axioms
and we give Baire’s theorem for partial fuzzy metric spaces. Finally, we
obtain that partial fuzzy metric topological spaces are metrizable under
some assumptions.
Keywords: Partial metric · Fuzzy metric · Partial fuzzy metric ·

Partial fuzzy metric topology · Completeness
1 Introduction
The definition of partial metric spaces (PMS) was given by Matthews [5] as
an extension of ordinary metric space. The difference of this structure from
ordinary metric space is that the distance of any point to itself may not be equal
to zero. Bukatin et al. [1] showed that this concept is introduced for some needs
in computer calculations.
In 1994, George and Veeramani [2] described a new notion called fuzzy met-
ric space (FMS) and after their work, many researchers [4,6,7] studied in this
structure and obtained a lot of important properties especially as a topological.
In the last years, Yue and Gu [9], Sedghi et al. [8] and Gregori et al. [3]
presented the concept of partial fuzzy metric spaces (PFMS) in the different
meanings. Yue and Gu [9] defined the notion of PFMS by using the continuous
minimum triangular norm. After Sedghi et al. [8] introduced the structure of
PFMS which generalizes the notion of non-archimedian (strong) FM and showed
its properties. Gregori et al. [3] approached the concept of FPMS as an extention
to the fuzzy settings in the meaning of George and Veeramani [2] and PMS due
to Matthews [5] by using residue operator which is associated to continuous
triangular norm.
1268 E. Aydogdu et al.
In this paper, we study in PFMS in the sense of Sedghi et al. [8]. This paper
is contained the following sections: Sect. 2 introduces fundamental definitions
and relevant concepts used in the main section. In Sect. 3, we first define two
topologies on X via open balls and s-open balls in (X, F, ∧). Then we investigate
some relation with the topology generated by a given partial metric p on U . We
also obtain some characterizations for partial fuzzy metric topological spaces to
be T2 spaces. We show that there are some necessary conditions that partial
fuzzy metric topological space is metrizable and Baire space. Finally, a brief
summary is given in Sect. 4.
2 Preliminaries
Some basic notions and facts about PMS, FMS and PFMS required in the sequel
are given in this part.
Definition 1 [5]. A partial metric space (in short PMS) is a pair (U, p) such
that U is a non-empty set and p : U × U → R+ is a mapping provided the listed
conditions for each u, v, w ∈ U :
(PM1) p(u, u) ≤ p(u, v),
(PM2) u = v iff p(u, u) = p(u, v) = p(v, v),
(PM3) p(u, v) = p(v, u),
(PM4) p(u, v) ≤ p(u, w) + p(w, v) − p(w, w).
Definition 2 [5]. Let (U, p) be a PMS. The open ball with radius ε > 0 and
centered at a point u in U is defined by Bp (u, ε) = {v ∈ U : p(u, v) < p(u, u) + ε}.
Theorem 1 [5]. In a PMS (U, p), the collection of each open balls is the basis of
a T0 topology on U , this topology is called the partial metric topology and denoted
by τp .
Definition 3 [5]. Let (U, p) be a PMS and (un ) be a sequence in U .
(i) (un ) is said to converge to u ∈ U if limn→∞ p(un , u) = p(u, u).

(ii) (un ) is said to be a Cauchy sequence if limn→∞ p(un , um ) exists.
(iii) (U, p) is said to be complete if there is a point u ∈ U such that
limn,m→∞ p(un , um ) = p(u, u) = p(un , u).
Definition 4 [2]. Let U be non-empty set, ∗ be a continuous t-norm and M :

U × U × (0, ∞) → [0, 1] be a mapping. If the listed conditions are satisfied for
each u, v, w ∈ U and t, s > 0, then the triplet (U, M, ∗) is said to be a fuzzy
metric space(in short FMS):
(1FM) M(u, v, t) > 0,
(2FM) M(u, v, t) = 1 iff u = v,
(3FM) M(u, v, t) = M(v, u, t),
(4FM) M(u, v, t + s) ≥ M(u, w, t) ∗ M(w, v, s),
(5FM) M(u, v, ·) is continuous on (0, ∞).
Some Properties of Partial Fuzzy Metric Topology 1269
If a FMS (U, M, ∗) provide the following condition for each u, v, w ∈ U and

s, t > 0, then (U, M, ∗) is said to be a non-Archimedean (strong) FMS [?]
M(u, v, max{t, s) ≥ M(u, w, t) ∗ M(w, v, s). (1)
Theorem 2 [4]. If (U, M, ∗) is a FMS, then
(1) (U, τM ) is a T2 -space,

(2) (U, τM ) is a metrizable topological space,
where τM =< {BM (u, r, t)|u ∈ U, r ∈ (0, 1), t > 0} > and BM (u, r, t) = {v ∈
U |M(u, v, t) > 1 − r}.
Theorem 3 [2]. If (U, M, ∗) is a FMS, then the topological space (U, τM ) is a

Baire space i.e., every intersection of a countable family of dense open sets in
(U, τM ) is also dense.
Definition 5 [8]. Let U be a nonempty set, ∗ be a cont. t-norm and F : U ×U ×

(0, ∞) → [0, 1] be a mapping. If the listed conditions are held for each u, v, w ∈ U
and t, s > 0, then the triplet (U, F, ∗) is said to be a partial fuzzy metric space
(in short PFMS):
(PFM1) u = v if and only if F(u, u, t) = F(u, v, t) = F(v, v, t),
(PFM2) F(u, u, t) ≥ F(u, v, t) > 0,
(PFM3) F(u, v, t) = F(v, u, t),
(PFM4) F(u, v, max{t, s}) ∗ F(w, w, max{t, s}) ≥ F(u, w, t) ∗ F(w, v, s),
(PFM5) F(u, v, ·) : (0, ∞) → [0, 1] is continuous.
If F(u, v, t) = 1, then from (PFM1) and (PFM2) u = v. But F(u, v, t) may

not be equal to 1 when u = v. From (PFM4) for each u, v, w ∈ U ad t > 0, we
have
F(u, v, t) ∗ F(w, w, t) ≥ F(u, w, t) ∗ F(w, v, t). (2)
In the next example, we show that there are PFMSs (U, F, ∗) in which
F(u, v, ·) : (0, ∞) → [0, 1] may not be non-decreasing function for each u, v ∈ U .
Example 1. Let U = R and ∗ be a cont. t-norm which is defined by a ∗ b =

min(a, b). Consider the mapping F : U × U × (0, ∞) → [0, 1] defined by

e−t , if u = v;
F(u, v, t) = (3)
1/2e−t , if u = v.
Then, it is simple to see that (U, F, ∗) is a PFMS such that F(u, v, t) is a decreas-
ing function with respect to t for each u, v ∈ U .
Definition 6 [8]. Let (U, F, ∗) be a PFMS and (un ) be a sequence in U .
(i) (un ) is said to converge to a point u ∈ U if F(u, u, t) = limn→∞ F(un , u, t)

for each t > 0.
(ii) (un ) is said to be a Cauchy sequence if limn,m→∞ F(un , um , t) exists for each
t > 0.
Definition 7. A PFMS (U, F, ∗) is said to be complete if each Cauchy sequence

(un ) ⊂ U converges to a point u ∈ U such that limn,m→∞ F(un , um , t) =
F(u, u, t).
3 Partial Fuzzy Metric Topology

In this part, we study in the PFMS (U, F, ∧) where a ∧ b = min(a, b) for each
a, b ∈ [0, 1].
Definition 8. Let (U, F, ∧) be a PFMS, u ∈ U , ε ∈ (0, 1) and t > 0. Then

the set BF (u, ε, t) = {v ∈ U : F(u, v, t) > F(u, u, t) − ε} is called an open ball
centered at u with radius ε.
Note that in particular that an open ball BF (u, ε, t) always includes u itself
for each ε ∈ (0, 1) and t > 0.
Theorem 4. Let (U, F, ∧) be a PFMS. Then, the family of each open balls is
the base of a topology τF which is T1 .
Proof. Since u ∈ BF (u, ε, t) for each ε ∈ (0, 1) and t > 0, we have U =

BF (u, ε, t). Let t > 0 and BF (u, ε, t), BF (v, δ, t) be two open balls centered at
u, v ∈ U with radius ε, δ ∈ (0, 1), respectively. Let w ∈ BF (u, ε, t) ∩ BF (v, δ, t).
Now, we will show that there is a η ∈ (0, 1) such that BF (w, η, t) ⊂ BF (u, ε, t) ∩
BF (v, δ, t). Since w ∈ BF (u, ε, t) ∩ BF (v, δ, t), we have F(u, w, t) > F(u, u, t) − ε
and F(v, w, t) > F(v, v, t) − δ.
Let η = min{F(w, w, t) − F(u, u, t) + ε, F(w, w, t) − F(v, v, t) + δ} and h ∈
BF (w, η, t). Then
F(u, h, t) ≥ F(u, h, t) ∧ F(w, w, t) ≥ F(u, w, t) ∧ F(h, w, t)

≥ (F(u, u, t) − ε) ∧ (F(w, w, t) − η). (4)
If F(w, w, t) − F(u, u, t) + ε < F(w, w, t) − F(v, v, t) + δ, then
F(u, h, t) ≥ (F(u, u, t) − ε) ∧ (F(w, w, t) − (F(w, w, t) − F(u, u, t) + ε)

≥ (F(u, u, t) − ε) ∧ (F(u, u, t) − ε) = F(u, u, t) − ε. (5)
If F(w, w, t) − F(v, v, t) + δ < F(w, w, t) − F(u, u, t) + ε, then
F(u, h, t) ≥ (F(u, u, t) − ε) ∧ F(w, w, t) − (F(w, w, t) − F(v, v, t) + δ)

≥ (F(u, u, t) − ε) ∧ (F(v, v, t) − δ)
≥ (F(u, u, t) − ε) ∧ (F(u, u, t) − ε) = F(u, u, t) − ε. (6)
This means that h ∈ BF (u, ε, t). Similarly, we have that h ∈ BF (v, δ, t). Thus,
h ∈ BF (u, ε, t) ∩ BF (v, δ, t) and BF (w, η, t) ⊂ BF (u, ε, t) ∩ BF (v, δ, t). This is
completed that the set of each open balls is a basis for a topology on U .
To show that (U, τF ) is a T1 -space, assume that u = v. Then F(u, u, t) >

F(u, v, t). We know that u ∈ BF (u, ε, t) for each ε ∈ (0, 1) and t > 0. If we take
ε = F(u, u, t) − F(u, v, t) > 0, then v ∈/ B(u, ε, t) . Similarly v ∈ BF (v, δ, t) and
u∈/ BF (v, δ, t) for δ = F(v, v, t) − F(u, v, t) > 0. Hence (U, τF ) is a T1 topology.
Corollary 1. If (U, F, ∧) is a PFMS, then the topological space (U, τF ) is first
countable.
Proposition 1. Let (U, F, ∧) be a PFMS such that F(u, v, t) is nondecreasing
for each t > 0. Let the mapping MF : U × U × (0, ∞) → [0, 1] be defined as

F(u, v, t), u = v
MF (u, v, t) = . (7)
1, u=v
Then (U, MF , ∧) is a FMS such that τF ⊆ τMF .

Proof. It is clear that (1F M ), (2F M ), (3F M ), (5F M ) are satisfied. Now, we
only show that (4F M ) is satisfied.
Let u, v, w ∈ U . It is obvious that (4F M ) is satisfied when u = v = w. Since
F(u, v, t) is nondecreasing with respect to t, we have the followings:
Let u = v = w, then
MF (u, v, t) ∧ MF (v, w, s) = F(u, v, t) ∧ F(v, w, s)

≤ F(u, w, max{t, s}) ∧ F(v, v, max{t, s})
≤ F(u, w, max{t, s}) ≤ F(u, w, t + s)
≤ MF (u, w, t + s). (8)
Let u = y and u = w, then
MF (u, v, t) ∧ M(v, w, s) = M(u, v, t) ∧ M(v, w, s)

≤ M(u, w, max{t, s}) ∧ M(v, v, max{t, s})
= F(u, u, max{t, s}) ∧ F(v, v, max{t, s})
≤ 1 = M(u, w, t + s). (9)
Let u = v and v = w, then
MF (u, v, t) ∧ MF (v, w, s) = F(u, v, t) ∧ 1 = F(u, v, t)

≤ F(u, w, t + s) = MF (u, w, t + s). (10)
Hence, (U, MF , ∧) is a FMS.

Example 2. Let (U, p) be a PMS and Fp : U × U × (0, ∞) → [0, 1] be a mapping
defined as
t
Fp (u, v, t) = . (11)
t + p(u, v)
Then, (U, Fp , ∧) is a partial fuzzy metric on U and (U, Fp , ∧) is called standard
partial fuzzy metric induced by p.
Proposition 2. Let (U, Fp , ∧) be the standard PFMS induced by a given partial

metric p on U . Then the topology τp induced by p is coincident with the topology
τFp induced by Fp i.e., τp = τFp
Proof. The proof is easily obtained by the definition of τp and τFp .
Lemma 1. Let (U, Fp , ∧) be a standard PFMS and (un ) be a sequence in U .
(i) (un ) converges to u ∈ U in (U, Fp , ∧) if and only if (un ) converges to u ∈ U
in (U, p).
(ii) (un ) is a Cauchy sequence in (U, Fp , ∧) if and only if (un ) is a Cauchy
sequence in (U, p).
(iii) (U, Fp , ∧) is complete if and only if (U, p) is complete. Furthermore,
limn→∞ p(un , u) = p(u, u) = limn,m→∞ p(un , um ) if and only if
limn→∞ Fp (un , u, t) = Fp (u, u, t) = limn,m→∞ Fp (un , um , t).
Proof. (i), (ii) are clear from Definition 3, Definition 6 and Example 2.
(iii) Let (U, Fp , ∧) be a complete PFMS and un ) be a Cauchy sequence in (U, p).
By (ii), (un ) is also a Cauchy sequence in (U, Fp , ∧). Since (U, Fp , ∧) is com-
plete, we have (un ) is convergent in (X, Fp , ∧). By (i), (un ) is also convergent in
(U, p). Thus (U, p) is a complete PMS. Other side is obtained with the similar
proceeding.
The equality of limits is obvious by the definition of Fp .
Theorem 5. Let (U, F, ∧) be a PFMS. Then (U, τF ) is a T2 space if and only
if inf w∈U {F(u, u, t) − F(u, w, t) + F(v, v, t) − F(v, w, t)} > 0 for each u, v ∈ U .
Proof. Assume that (U, τF ) is a T2 space, then for any u = v in U , there exists
some ε > 0 such that B(u, ε, t) ∩ B(v, ε, t) = ∅. Hence there is no w satisfying
both F(u, w, t) > F(u, u, t) − ε and F(v, w, t) > F(v, v, t) − ε. This implies that
inf w∈U {F(u, u, t) − F(u, w, t) + F(v, v, t) − F(v, w, t)} ≥ ε > 0 for each u, v ∈ U .
Conversely, if (U, τF ) is not a T2 space, then there exist u = v in
U such that BF (u, ε, t) ∩ BF (v, ε, t) = ∅ for any ε > 0. Thus,
there exists a sequence (wn ) ∈ BF (u, n1 , t) ∩ BF (v, n1 , t). Hence we have
infw∈U F(u, u, t) − F(u, w, t) + F(v, v, t) − F(v, w, t) < n2 . This means that the
value of infimum equal to zero.
Theorem 6. Let (U, F, ∧) be a PFMS.
(1) (U, τF ) is a T2 space if and only if limn→∞ F(un , u, t) = F(u, u, t) and
limn→∞ F(un , v, t) = F(v, v, t) implies u = v.
(2) If limn→∞ F(un , u, t) = F(u, u, t) implies limn→∞ F(un , v, t) = F(u, v, t) for
each v ∈ U then (U, τF ) is a T2 space.
Corollary 2. If (U, F, ∧) is a PFMS such that F(u, v,·) : (0, ∞) → [0, 1] is
nondecreasing function for each u, v ∈ U , then (U, τF ) is a T2 -space.
Theorem 7. Let (U, F, ∧) be a PFMS such that F(u, u, t) = F(v, v, t) for each
u, v ∈ U and t > 0. Then the second countability and separability are equivalent
in (U, τF ).
Proof. Since each second countable space is also separable, then it is enough to
show that if (U, τF ) is separable then it is second countable. Let A be a countable
dense subset of U , u ∈ U and B = {BF (a, ε, t)|a ∈ A, ε ∈ Q ∩ (0, 1), t ∈
Q+ }. Now, we will show that B is a basis of τF . Let v ∈ BF (u, ε, t). Then
we have F(u, v, t) > F(u, u, t) − ε. Let δ be a positive rational number such
that δ < ε + F(u, v, t) − F(u, u, t). If v ∈ A, then BF (v, δ, t) ⊂ BF (u, r, t). Let
v ∈/ A. Since A is dense in U , there exists a sequence (un ) ⊂ A such that
limn→∞ F(un , v, t) = F(v, v, t) for each t > 0. Therefore, limn→∞ F(un , v, t) =
limn→∞ F(un , un , t) = F(v, v, t) for each t > 0. Thus, there is a n0 ∈ N such that
F(un , v, t) > F(un , un , t) − 2δ for each n ≥ n0 . Hence, we have v ∈ BF (um , 2δ , t)
for each m ≥ n0 . Since
F(um , u, t) ≥ F(um , u, t) ∧ F(v, v, t) ∧ F(um , v, t) ∧ F(u, v, t)

δ
> (F(um , um , t) − ) ∧ (F(u, u, t) − ε)
2
≥ (F(u, u, t) − ε) ∧ (F(u, u, t) − ε), (12)
then um ∈ BF (u, r, t). Hence, we obtain BF (um , 2δ , t) ⊂ BF (u, r, t). Therefore, B

is a countable base for (U, τF ).
Definition 9. Let (U, F, ∧) be a PFMS, u ∈ U , ε ∈ (0, 1) and t > 0. Then the

∗
set BF (u, ε, t) = {v ∈ U : Fuuvt
> 1 − ε} is called a s-open ball centered at u
1, F(u, u, t) = F(u, v, t)
with radius ε, where Fuuvt = .
F(u, v, t), otherwise
Theorem 8. Let (U, F, ∧) be a PFMS. Then the family of each s-open balls is
the base of a topology τF∗ which is T2 .
∗
Proof. Since for each ε ∈ (0, 1) the s-open ball BF (u, ε, t) contains u, we have
∗ ∗ ∗
that U = BF (u, ε, t). Let t > 0 and BF (u, ε, t), BF (v, δ, t) be two open balls
∗
centered at u, v ∈ U with radius ε, δ ∈ (0, 1), respectively. Let w ∈ BF (u, ε, t) ∩
∗ ∗
BF (v, δ, t). Now, we will show that there is a η ∈ (0, 1) such that BF (w, η, t) ⊂
∗ ∗ ∗ ∗
BF (u, ε, t) ∩ BF (v, δ, t). Since w ∈ BF (u, ε, t) ∩ BF (v, δ, t), we have F(u, w, t) >
1 − ε and F(v, w, t) > 1 − δ.
∗
Let η = min{ε, δ} and h ∈ BF (w, η, t). Then we obtain
F(u, h, t) ≥ F(u, h, t) ∧ F(w, w, t) ≥ F(u, w, t) ∧ F(h, w, t)

≥ (1 − ε) ∧ (1 − η) ≥ (1 − ε) ∧ (1 − ε) = 1 − ε (13)
∗
This means that h ∈ BF (u, ε, t). Similarly, we have that h ∈ BF∗ (v, δ, t). Thus,
∗ ∗ ∗ ∗ ∗
h ∈ BF (u, ε, t) ∩ BF (v, δ, t) and BF (w, η, t) ⊂ BF (u, ε, t) ∩ BF (v, δ, t). This is
completed that the set of all s-open balls is a basis for a topology on U .
To show that (U, τF∗ ) is a T2 -space, assume that u = v. Then 1 ≥ F(u, u, t) >
∗
F(u, v, t) and 1 ≥ F(u, u, t) > F(u, v, t). We know that u ∈ BF (u, ε, t) for each
ε ∈ (0, 1) and t > 0. If we take ε = 1 − F(u, v, t) > 0, then u ∈ B(u, ε, t),
v ∈ B(v, ε, t) and B(u, ε, t) ∩ B(v, ε, t) = ∅. Hence (U, τF∗ ) is a T2 topology.
Remark 1. If (U, F, ∧) is a PFMS, then τF ⊂ τF∗ .
Remark 2. If (U, F, ∧) is a PFMS such that F(u, v, ·) : (0, ∞) → [0, 1] is nonde-

creasing function for each u, v ∈ U , then τF∗ = τMF .
Corollary 3. If (U, F, ∧) is a PFMS such that F(u, v, ·) : (0, ∞) → [0, 1] is

nondecreasing function for each u, v ∈ U , then (U, τF∗ ) is a metrizable topological
space.
Corollary 4. If (U, F, ∧) is a PFMS such that F(u, v, ·) : (0, ∞) → [0, 1] is

nondecreasing function for each u, v ∈ U , then the topological space (U, τF∗ ) is a
Baire Space.
Question 1. Let (U, F, ∧) be a PFMS. Now, we don’t know the topological space
(U, τF ) is metrizable or not. If the topological space (U, τF ) is metrizable, can
we find the necessary condition that the topological space (U, τF ) is metrizable?
Question 2. Let (U, F, ∧) be a PFMS. Now, we don’t know the topological space
(U, τF ) is a Baire space or not. If the topological space (U, τF ) is a Baire space,
can we find the necessary condition that the topological space (U, τF ) is Baire
space?
4 Conclusion
In this work, we have obtained different topologies via open balls in PFMS
(U, F, ∧) in the sense of Sedghi et al. [8]. We have studied some properties of one
of these topological spaces such as completeness, metrizability and to be a Baire
space. And we left some open questions about the other one. For future work,
we will continue to research to answer these questions and study compactness
properties of partial fuzzy metric topological spaces.
References
1. Bukatin, M., Kopperman, R., Matthews, S., Pajoohesh, H.: Partial metric space.
Am. Math. Monthly 116(8), 708–718 (2009)
2. George, A., Veeramani, P.: On some results in fuzzy metric spaces. Fuzzy Sets Syst.
64(3), 395–399 (1994)
3. Gregori, V., Minana, J.J., Miravet, D.: Fuzzy partial metric spaces. Int. J. Gen.
Syst. 48(3), 260–279 (2019)
4. Gregori, V., Romaguera, S.: Some properties of fuzzy metric spaces. Fuzzy Sets
Syst. 115(3), 485–489 (2000)
5. Matthews, S.: Partial metric topology. Ann. N. Y. Acad. Sci. 728, 183–197 (1994)
6. Piera, A.S.: A contribution to the study of fuzzy metric spaces. Appl. Gen. Topology
2(1), 63–75 (2001)
7. Rodriguez-López, J., Romaguera, S.: The Hausdorff fuzzy metric on compact sets.
Fuzzy Sets Syst. 147(2), 273–283 (2004)
8. Sedghi, S., Shobkolaei, N., Altun, I.: Partial fuzzy metric space and some fixed point
results. Commun. Math. 23(2), 131–142 (2015)
9. Yue, Y., Gu, M.: Fuzzy partial (pseudo-) metric space. J. Intell. Fuzzy Syst. 27(3),
1153–1159 (2014)
Fuzzy Classifier Based Colorimetric
Quantification Using a Smartphone
Öykü Berfin Mercan and Volkan Kılıç(B)
Electrical and Electronics Engineering Graduate Program,

Izmir Katip Celebi University, Izmir, Turkey
[email protected]
Abstract. The recent advances in smartphone technology have led to

the emergence of simple, low-cost and portable platforms that aim reli-
able qualitative and quantitative analysis of colored solutions. In this
paper, we present a new approach based on fuzzy classification for quan-
titative detection of peroxide relying on color change due to concentration
level. The proposed fuzzy classification algorithm was integrated with a
smartphone via a cloud system. Our custom-designed Android applica-
tion was developed with a user-friendly interface to allow the users to
capture the image of peroxide strips and crop the region of interest on the
image for quantification. Average RGB (Red-Green-Blue) values in the
cropped region were calculated to transfer these values to our proposed
fuzzy logic classifier via a cloud system to quantify the peroxide content.
The proposed fuzzy classifier evaluates the concentration value based on
RGB values and sends back to the Android application via the cloud
system. Experimental investigations show the advantage of the proposed
system for the quantification of peroxide.
Keywords: Fuzzy classification · Image processing · Android

application
1 Introduction
Smartphone-based colorimetric analysis has attracted an increasing amount of
attention in the last decade due to its potential applications in e.g. point-of-care
analyses [9,12], water monitoring [8,11,13] chemical sensing [3,10,19], glucose
sensing [5] and immunoassays [4]. Smartphone colorimeters have been found
to be promising for qualitative and quantitative analysis due to fact that they
offer rapid, sensitive, and user-friendly analysis with low sample consumption,
and provide mobility which enables to perform analysis in remote locations
without sophisticated laboratory infrastructure. Recent developments in smart-
phone technology with the increasing capability in camera, sensors, processing
power and wireless internet network connection have led to facilitate the com-
plex biomedical and chemical tests by integration with other platforms capable
of running advanced and complex software applications supported by artificial
intelligence.
Fuzzy Classifier Based Colorimetric Quantification Using a Smartphone 1277
Artificial intelligence is the computer science that enables a machine or com-

puter system to mimic human intelligence system, learn from experience, update
with new information and solve complex problems. It is based on mathemati-
cally elegant and consistent theories including artificial neural networks, expert
systems and genetic algorithms and fuzzy logic. Lotfi Zadeh introduced the fuzzy
logic [22] as a computing approach with any real number output in between 0
and 1 (both inclusive) rather than classical true or false logic. Applications of
fuzzy logic include medicine [15], engineering [16] and education [21]. Fuzzy logic
and its extensions offer flexibility on the approximation of output which leads
to be used in complicated systems. In medicine, fuzzy systems are used to diag-
nose and evaluate diseases such as cancer and tumor [6]. The fuzzy logic-based
decision support system is employed to classify raw milk with respect to quality
[1]. SPOT images are classified using fuzzy logic classification process [14].
In this study, we proposed fuzzy logic classifier to quantify hydrogen per-
oxide concentration level based on the color of test strips. We integrated the
proposed system with our Android application, named FuzzyChem, to calculate
the concentration level. FuzzyChem calculates mean RGB values of the region
of interest in peroxide strips and then transfers these values to our proposed
fuzzy logic classifier to quantify the peroxide content. After the proposed fuzzy
classifier evaluates the concentration value based on RGB values, it is sent back
to the FuzzyChem via the cloud and FuzzyChem displays concentration value on
the screen. To the best of our knowledge, peroxide quantification has not been
previously implemented with fuzzy logic as we do here.
The rest of this paper is organized as follows: Sect. 2 introduces fuzzy logic
foundations for colorimetric quantification. Section 3 explains the general struc-
ture of our proposed model and classification procedure. Section 4 presents the
dataset, Android application and discussion of the results. Closing remarks with
future directions are given in Sect. 5.
2 Fuzzy Logic
This section presents a brief review of the fuzzy logic approach which includes
membership function, if-then rules and fuzzy interference system.
In classical logic, an element is evaluated whether it belongs to the set or not.
Mathematically, membership value is assigned to an element that takes values 0
or 1, but dealing with only 0 or 1 values is no longer satisfactory enough because
there is no certain phenomenon in nature always. Contrarily, in fuzzy logic,
each element is valued in the range of 0 and 1. Defining problems by natural
linguistic terms is one of the advantages of fuzzy logic. Thus, verbal data in
natural language are used successfully in fuzzy logic.
Fuzzy logic has three main parts: fuzzification, fuzzy rule-based system, and
defuzzification [2]. First, input variables, crisp values, are determined. Crisp
inputs are converted to fuzzy values using membership functions. The process
of fuzzification converts a crisp input value to a fuzzy set. Each element of fuzzy
set has degree of membership. Membership values of the set members are not
1278 Ö. B. Mercan and V. Kılıç
limited to just 0 or 1. Instead, it can be any value in between. In this progress,

a fuzzy set is completely characterized by its membership function which deter-
mines the membership degree of the fuzzy set element. There are many mem-
bership functions developed including triangle, trapezoid, and Gaussian which
are commonly used. The fuzzy rule is simple “if-then” rules, and the degree
of membership coming from the fuzzification is evaluated using these rules to
obtain fuzzy results.
IF x IS A AND y IS B THEN z IS C (1)
where x and y input variables with values defined as A and B fuzzy sets. The out-
put variable, z is also defined by a C fuzzy set. Fuzzy inference system involves
if-then rules which can be determined with various inference methods. The best
known of these are Mamdani and Sugeno methods [17]. In the Mamdani fuzzy
inference system, membership values are calculated according to the rules trig-
gered by input values. Membership levels from fuzzification are associated with
the output membership function using fuzzy operators. In the Mamdani system,
output membership functions are fuzzy sets and need the defuzzification process
to convert fuzzy sets to crisp values. In the Takagi–Sugeno fuzzy inference sys-
tem, output membership functions are linear or constant. In the Sugeno method,
unlike the Mamdani method, polynomial functions are used for the output [7].
The output of if-then rule is determined by polynomial functions based on input
variables [18]. Thus, the output of each rule is defined as a crisp number. A
weighted average is used to achieve the final output.
A typical rule in a Sugeno fuzzy model has the following form:
If x = A and y = B then z = f (x, y) (2)
where x, y are input variables and z is output variable, A and B are fuzzy sets
and f (x, y) is a mathematical function [20].
3 Proposed Approach
This section describes the general structure of our proposed system based on
fuzzy logic.
3.1 General Structure of the Proposed Model

The proposed system is capable of quantifying peroxide content based on the
color of test strips with the fuzzy logic approach. Each test strip has a patch that
changes its color relying on the concentration level of the solution. The patch
image in the test strip contains the pixel information in terms of RGB values.
Mean RGB values of this patch are used as input variables in the fuzzy logic
system. These values are converted to fuzzy values using membership function
as RGB values are crisp inputs. Then, an inference is made based on defined rule
statements. In a rule-based fuzzy logic system, the output variable is obtained
Fig. 1. Proposed peroxide quantification model
based on the constructed rule statements which give the concentration level.
Figure 1 shows the general structure of our proposed model.
11 membership functions are defined for each input variable and triangle
(trimf) is used to define membership function in the quantification process as it
is found to be adequate based on extensive experimental studies. The proposed
fuzzy system needs three parameters (lower limit, mean and upper limit) for
the triangle membership function given in Table 1. In this table, mf represents
membership functions. Values of lower limit, mean and upper limit are obtained
using image processing algorithms. First, the peroxide dataset is transferred to
a computer to process in MATLAB environment. The region of interest which
shows the color change caused by the hydrogen peroxide concentration is cropped
and then the main RGB values of cropped regions are calculated. In addition
to mean RGB values, upper and lower limits are also determined based on the
changes in standard ambient conditions that may occur in experimental studies,
and the effect of experimental error margins on color change.
The hydrogen peroxide concentration level of the solutions is the constant val-
ues used for the output membership function definition which leads to employing
Sugeno inference to model the overall system. After the definition of the input
and output variables, the rule statements are constructed to control the out-
put variable. In Table 2, sample fuzzy rules for the proposed model are shown.
The fuzzy inference system has been defined based on variables, membership
functions and the rules which makes the proposed system capable of calculating
quantification of peroxide relying on color change due to the concentration level.
3.2 Android Application: FuzzyChem
The proposed fuzzy logic system was integrated with our custom-designed
Android application. We developed a simple and user-friendly smartphone appli-
cation for colorimetric quantification of peroxide content, named as FuzzyChem.
Table 1. Lower limit, mean and upper limit of training strips
Red Green Blue

mf Lower Mean Upper mf Lower Mean Upper mf Lower Mean Upper
limit limit limit limit limit limit
0 ppm R11 154 157.76 160 G11 149 152.89 156 B11 126 131.03 134
0.5 ppm R10 150 155.77 160 G10 154 154.82 160 B10 131 133.57 137
1 ppm R9 138 143.42 150 G9 145 148.02 152 B9 122 127.11 132
2 ppm R8 125 138.06 145 G8 149 151.26 154 B8 131 133.61 140
3 ppm R7 98 116.11 128 G7 130 140.13 150 B7 120 126.51 131
6.5 ppm R6 58 76.87 98 G6 101 115.58 129 B6 109 113.67 120
10 ppm R5 33 42.95 57 G5 88 94.75 100 B5 103 105.63 110
20 ppm R4 21 27.03 35 G4 66 76.78 87 B4 97 99.05 104
30 ppm R3 16 19.84 24 G3 54 67.77 74 B3 88 96.83 103
65 ppm R2 10 14.66 20 G2 39 43.37 53 B2 61 73.47 88
100 ppm R1 15 18.67 23 G1 34 38.88 45 B1 45 51.75 61
Screenshots of the FuzzyChem app given in Fig. 2 present the steps of the quan-
tification process. When the user runs the FuzzyChem, tips are displayed on
the opening page as shown in Fig. 2a. The user can select a test strip from the
gallery or capture a new image using the camera as in Fig. 2b. After the test
strip is selected, the region of the interest is cropped and average RGB values
are calculated (Figs. 2c and d). The user needs to tap the “calculate” button to
transfer these values to the remote server via the Firebase cloud system that
runs fuzzy logic classifier to quantify the peroxide content (Fig. 2e). The pro-
posed fuzzy classifier evaluates the concentration value based on RGB values
and the result comes back to the FuzzyChem via the Firebase. The FuzzyChem
displays peroxide content on the screen as shown in Fig. 2f.
Table 2. Sample fuzzy rules for the proposed system

Fig. 2. Steps of colorimetric peroxide quantification with FuzzyChem


Hydrogen peroxide (H2 O2 ) have 11 main color classes based on 0, 0.5, 1, 2, 3, 6.5,
10, 20, 30, 65 and 100 ppm concentration levels. Test strips have been dipped
into each different concentrations of hydrogen peroxide solutions. Color changes
were observed in test strips based on the concentration of solutions. These test
strips have been captured under the controlled illumination condition with the
LG G4 Android-based smartphone under sunlight. The reader is referred to [19]
for details of the dataset.
A custom-designed FuzzyChem app is capable of calculating the H2 O2 con-
centration. In this study, the proposed approach was tested on the different
peroxide test strips. Firstly, the patch is cropped and then mean RGB values
are calculated in FuzzyChem. These values are transferred to the fuzzy classifier
system running on the server to quantify the peroxide content. In the process
of quantifying H2 O2 concentration, membership functions have been defined by
using mean, upper and lower RGB values. The fuzzy classifier evaluates the con-
centration value based on RGB values and sends back to FuzzyChem. The result
shows that the proposed system is capable of calculating the concentration level
relying on color change.
5 Conclusions
In this paper, we presented fuzzy classifier-based colorimetric quantification for

the detection of H2 O2 using a smartphone. Our proposed approach was evalu-
ated on the peroxide dataset and then integrated with a simple and user-friendly
FuzzyChem Android app that can calculate H2 O2 content based on the color of
the test strips. It calculates RGB values in the cropped region and then transfers
to the remote server which runs our proposed fuzzy logic system to quantify the
H2 O2 concentration. The concentration value is calculated based on the RGB
values and returned to the FuzzyChem via the Firebase. Experiments on the
dataset show the advantage of our proposed system for colorimetric quantifi-
cation of H2 O2 with significant performance. This study has the potential for
colorimetric detection of different solutions which could be interesting directions
for future work. The proposed approach could be further improved by extending
the dataset with other smartphone brands under ambient light conditions.
References
1. Akıllı, A., Atıl, H., Kesenkaş, H.: Çiğ süt kalite değerlendirmesinde bulanık mantık
yaklaşımı. Kafkas Üniversitesi Veteriner Fakültesi Dergisi 20(2), 223–229 (2014)
2. Anand, M.S., Tyagi, B.: Design and implementation of fuzzy controller on FPGA.
Int. J. Intell. Syst. Appl. 4(10), 35–42 (2012)
3. Bayram, A., Horzum, N., Metin, A.U., Kılıç, V., Solmaz, M.E.: Colorimetric
Bisphenol-A detection with a portable smartphone-based spectrometer. IEEE Sens.
J. 18(14), 5948–5955 (2018)
4. Coskun, A.F., Nagi, R., Sadeghi, K., Phillips, S., Ozcan, A.: Albumin testing in
urine using a smart-phone. Lab Chip 13(21), 4231–4238 (2013)
5. Gölcez, T., Kiliç, V., Şen, M.: Integration of a smartphone application with a µ
pad for rapid colorimetric detection of glucose. In: Medical Technologies Congress
(TIPTEKNO), pp. 1–4. IEEE (2019)
6. Hassan, M.R., Hossain, M.M., Begg, R.K., Ramamohanarao, K., Morsi, Y.: Breast-
cancer identification using HMM-fuzzy approach. Comput. Biol. Med. 40(3), 240–
251 (2010)
7. Jang, J.S.R., Sun, C.T., Mizutani, E.: Neuro-fuzzy and soft computing-a compu-
tational approach to learning and machine intelligence [book review]. IEEE Trans.
Autom. Control 42(10), 1482–1484 (1997)
8. Kilic, V., Alankus, G., Horzum, N., Mutlu, A.Y., Bayram, A., Solmaz, M.E.: Single-
image-referenced colorimetric water quality detection using a smartphone. ACS
Omega 3(5), 5531–5536 (2018)
9. Kılıç, V., Horzum, N., Solmaz, M.E.: From sophisticated analysis to colorimetric
determination: smartphone spectrometers and colorimetry. In: Color Detection.
IntechOpen (2018)
10. Kiliç, V., Şen, M.: Smartphone-based colorimetric analysis for the detection of
H2O2 using a µ pad. In: Medical Technologies Congress (TIPTEKNO), pp. 1–4.
IEEE (2019)
11. Kocakusak, G., Bayram, A., Kilic, V., Horzum, N., Solmaz, M.: Smartphone-based
detection of dyes in water for environmental sustainability. Anal. Methods 9, 579–
585 (2017). https://doi.org/10.1039/C6AY03073D
12. Mutlu, A.Y., Kılıç, V., Özdemir, G.K., Bayram, A., Horzum, N., Solmaz, M.E.:
Smartphone-based colorimetric detection via machine learning. Analyst 142(13),
2434–2441 (2017)
13. Mutlu, A.Y., Kılıç, V.: Machine learning based smartphone spectrometer for harm-
ful dyes detection in water. In: 26th Signal Processing and Communications Appli-
cations Conference (SIU), pp. 1–4. IEEE (2018)
14. Nedeljkovic, I.: Image classification based on fuzzy logic. Int. Arch. Photogramme-
try Remote Sens. Spatial Inf. Sci. 34(30), 3–7 (2004)
15. Özdemir, A., Alaybeyoglu, A., Balbal, K.F.: Bulanık mantığın eğitim alanındaki
uygulamaları. Bilim Eğitim Sanat ve Teknoloji Dergisi 3(1), 45–50 (2019)
16. Phuong, N.H., Kreinovich, V.: Fuzzy logic and its applications in medicine. Int. J.
Med. Inf. 62(2–3), 165–173 (2001)
17. Ross, T.J.: Fuzzy Logic With Engineering Applications. Wiley, New York (2005)
18. Rout, S.S., Misra, B.B., Samanta, S.: Competency mapping with sugeno fuzzy
inference system for variable pay determination: a case study. Ain Shams Eng. J.
9(4), 2215–2226 (2018)
19. Solmaz, M.E., Mutlu, A.Y., Alankus, G., Kilic, V., Bayram, A., Horzum, N.: Quan-
tifying colorimetric tests using a smartphone app based on machine learning clas-
sifiers. Sens. Actuators B 255, 1967–1973 (2018)
20. Takagi, T., Sugeno, M.: Fuzzy identification of systems and its applications to mod-
eling and control. IEEE Trans. Syst. Man Cybern. SMC–15(1), 116–132 (1985)
21. Yen, J., Langari, R.: Fuzzy Logic: Intelligence, Control, and Information, vol. 1.
Prentice Hall, Upper Saddle River (1999)
LabHub: A New Generation Architecture
Proposal for Intelligent Healthcare Medical
Laboratories
Bengi Tugcu Idemen1(&), Emine Sezer2, and Murat Osman Unalir2

1
Panates Bilisim, Izmir, Turkey
[email protected]
2
Department of Computer Engineering, Ege University, Izmir, Turkey
{emine.sezer,murat.osman.unalir}@ege.edu.tr
Abstract. With the development of technologies such as Internet of Things,

Cloud Computing, Big Data and Machine Learning, created business models
and information systems based on these business models have the need for
restructuring. Medical Laboratories are environments where technical devices
are located. New generation devices also have the IoT infrastructure. Thus, these
devices support the transfer of device data to the relevant health information
system through cloud computing. Monitoring the operating conditions, esti-
mated maintenance periods and calibration settings for the specified devices on a
common platform will increase the quality and reliability of the measurement
results to be determined through the device. In this paper, a new generation LIS
architecture, named LabHub, is proposed. With the platform which will be
developed on LabHub architecture, traceability of the devices in all public and
private sector, domestic and international medical laboratories will be provided
with a cloud application that serves as a software principle. The data collected
through the device will be stored in the cloud database with a scalable big data
model in the cloud application by protecting the privacy of personal data and
even free from personal data. Any violations that may occur in this data will be
determined in a special way to the device and the operation performed on the
device. Methods based on machine learning algorithms will be used to identify
contradictory situations. Thus, the infrastructure for both the measurement of the
desired quality of the devices and the reliable recording of the measurement
results will be developed.
Keywords: Big data Laboratory Information System Anomaly detection

Preventive maintenance
1 Introduction
The healthcare domain is a geographically dispersed, quite huge and complex area
where different institutions and organizations both get service and also render service
[1]. It has a large number of health data throughout the health processes that begin
before the birth of the individual and continue throughout his life. Health records began
to be stored textually with the usage of computer technologies, and the concept of EHR
LabHub: A New Generation Architecture Proposal for Intelligent Healthcare 1285
(Electronic Health Records) emerged with the application of tables and databases in
information technologies.
The possibility of a new application in the healthcare field to cause permanent
disabilities or fatal consequences in the individual has delayed the effective use of
information and communication technologies in the healthcare field. Primarily, the
personal information, demographic features and payment based administrative in-
formation of the individual who applied to the health institution started to be record-ed
in the databases. However, factors such as the speed, efficiency and 24/7 access offered
by information and communication technologies have increased the trend towards new
applications in the healthcare field. Information technologies such as big data analysis,
artificial intelligence, machine learning and image processing promise to provide
effective solutions to healthcare providers in diagnosis and treatment.
When the individual’s first consultation to physician for any health problem, the
first medical method usually applied is the request for a medical test. This health
practice can be routinely requested not only in acute problems but also in the moni-
toring of chronic diseases. The rather complex and big healthcare field is divided into
subfields to be worked better. This has also been the case for medical tests, and medical
laboratories are divided into different subfields such as biochemistry, hematology, and
microbiology.
Laboratory Information Systems (LIS) have been developed in order to transmit the
results of the analysis made in medical laboratories to the Hospital Information
Management System and to be seen results by the doctor who made the request and by
the patient. It is seen that these systems have the features of recording results (‘‘by
transferring the result directly to the system by using the medical devices that offer IoT
feature in the field’’ or ‘‘for devices without IoT feature, a lab technician enters the
result to the system’’) and querying results.
Even if subdomains involve special cases, it is a obligatory that a medical test pro-
duces the right data in the fastest way. The ability of a medical test to produce the right
result involves different processes – that start with the right request of the medical staff
– as taking the sample correctly from the individual, transporting this sample to the
relevant laboratory in the right conditions, preparing the sample for testing, preparing
the adjuvants correctly, and calibrating the test device (accuracy of the calibration of
the test device).
LIS used in health institutions is not yet capable of monitoring all of these processes.
The main purpose of the LabHub architecture proposed in this paper is to intro-duce a new
generation laboratory information system approach that is aware of big data by adding to
LIS the ability to detect anomalies and the features to offer restorative maintenance
advice. Section 2 deals with LIS literature studies in the field. Section 3 describes the
LabHub architecture and its components. Finally, conclusions and future works are
presented.
1286 B. T. Idemen et al.
2 Literature Review
Laboratory Information Management Systems (LIMS) or LIS have been among the
study fields of Information and Communication Technologies for many years. Test
results that have been stored only on paper until recently have been moved to semi-
autonomous systems in recent years [2]. While the aim in first versions for these
information systems was only controlling the costs, the next developed systems began
to give services like recording the medical tests carried out, presenting these results to
patients or responsible health personnel, and querying all [3].
In recent years, the “Internet of Things Technology” which is expected to support
more autonomous systems in the industry has started to be applied in many different
fields such as health, factory automation and smart cities [4–6]. Increasingly, more
satisfaction is provided from the results achieved. Today, some of the devices used in
medical laboratories for testing purposes include IoT technologies [7]. The sensors in
the device only operate certain warnings and alarms on the device. What is the stim-
ulated problem should be analyzed by the human user and the problem should be
solved based on human interaction. Data transfer from these devices to software sys-
tems is usually done through the computer connection of the devices [8].
Different devices can be used for the same medical tests in different medical lab-
oratories. Different devices used for the same test may have different reference values.
In order for the data coming from different devices to have the same meaning in the
systems used in the institutions, the necessity of processing these data according to the
reference values, in other words, the contradiction in the results due to devices arises as
another problem. Identification of all information such as reference values belonging to
devices, conditions of use, kits used, calibration, quality control, test, sample on a data
model, also monitoring and processing of this data have an effect to increase the
accuracy of medical test results. If it is considered that the devices will be affected by
the factors such as light, sound, particles, temperature, humidity in the physical con-
ditions of the environment, it is seen that the data belonging to these factors are also
important in producing correct test result. As known, any medical test is related with a
person. Any test that is carried out in a laboratory brings quite huge volumetric data
about device and data about the environment continuously. These situations can be
explained with 3 V features of big data: velocity, volume and variety [9, 10].
To define and process continuously flowing big data that comes from the medical
test devices and laboratories causes scalability challenges [11]. It is not possible to
process the incoming data without defining it and to obtain new data from the incoming
data [12]. In addition to defining various data coming from different test devices, there
is another necessity that this data can be transmitted to the country’s health information
systems, such as SağlıkNet as in our country [13]. Health Information Standards have
been developed and used to create a common terminology in the health domain to
ensure interoperability between HIS. However, different medical information standards
have been developed for different purposes to serve in specific clinics. Besides the ICD
[14] codes used for diagnostic purposes, the HL7/FHIR [15] standard is used for data
transmission in worldwide. Even if the medical laboratory has its own information
system, it must use and support these standards in the data definition in order to
transmit the data it produces to the centres to ensure interoperability. In other side,
“Tıbbi Laboratuvar Yönetmeliği” [16] has been published in order to determine the
conditions that the laboratories must meet the minimum to work legally in Turkey.
Internal and external audits of laboratories should be carried out regularly. In these
audits, ISO Standards can be applied to accreditation or certification, or they may be
regulatory. Among the important examples of accreditation standards, international
standards commonly used are ISO 17025 [17] and ISO 15189 [18]. The main purpose
of ISO 15189 is the accreditation of medical laboratories, ensuring the test results are
secured by patients and healthcare professionals. Over time, some additional require-
ments were needed for medical laboratories, clinical researchers, and post-test appli-
cations. It is expected that the standards will also have sanctions including the use of
information and communication technologies in order to minimize the errors that may
arise from human impact in line with changing needs.
Information and Communication Technologies have been used in Health Infor-
mation Systems and Laboratory Information Systems for last decade. In recent years,
the trend in technologies such as big data and IoT in all information systems has
brought a new dimension to the healthcare field, also. Considering the studies con-
ducted in the literature, it is observed that there is no system yet that monitors the test
process including from taking a sample from the patient to producing the result, detects
anomaly and also provides preventive maintenance.
Considering all these problems, it is aimed to propose a solution for medical
laboratories to transfer environment sensor data and data that is produced from medical
devices to the cloud with the purpose of monitoring the analyze process from the
sample taking from the patient to the production of the result. LabHub, the new
generation architecture, proposed in this paper is described in details in the next
section.
3 New Generation LIS Architecture: LabHub
LIS has been the centre of attention for information technologies for a long time,
especially with the aim of transferring the medical test results carried out to the
Hospital Information Management System (HIMS) or to be displayed by the patient or
the related health personnel. However, a system for the laboratories does not exist yet
that can be work with the logic of plug/play to the existing one to handle the devices
meet optimal operating conditions independent of their brands, and where the cali-
bration, maintenance and quality control processes of the devices are followed. The test
results produced by medical test equipment cannot be related only monitoring and
processing, but also data related to the device requires big data analytics. Machine
learning technologies, which will ensure that the data collected from different devices
are used for possible device and test result anomaly determinations, will both speed up
the diagnosis process and support the effective, accurate and reliable monitoring of the
patient. Not only testing but also predicting anomaly of the device will have a cost-
reducing effect.
By considering the above requirements, LabHub, a new generation LIS architecture,
is proposed, which will add new capabilities to the information systems of medical
laboratories in the health domain. LabHub is designed as a platform that includes big
data aware processes, which are designed for anomaly detection services and preventive
maintenance services with supervised learning artificial algorithms, and IoT-based cloud
technologies. Figure 1 shows the basic components of LabHub architecture.
Fig. 1. LabHub: system architecture.
As seen in Fig. 1, the stakeholders of the system can be anybody or institutions like
healthcare organization, healthcare professionals or private medical laboratories. Lab-
Hub architecture will be developed on the cloud application structure. It will give
services to multiple users. The system to be designed as a cloud platform will direct the
requests from the client to the data and services reserved for them through a presen-
tation layer which will be implemented according to the multi-tenant application
model.
The Medical Laboratory Monitoring Service will collect any required data for the
devices and also from the environment where the corresponding device located to
produce an effective and accurate result through IoLT (Internet of Laboratory Thing)
plug-in. The most important feature of the IoLT plug-in, which is an IoT device, can be
added to any medical test device with a plug-and-play logic regardless of the which
laboratory used and the software on that device. The characteristics of the environment
where each device is located may change. For example, a medical test device may be
located under the air conditioner or next to a window or in the darkest corner. The
optimum operating conditions of each device can be different as mentioned in their user
manuals. Therefore, each device needs a special IoLT plug-in to sense its own envi-
ronmental conditions. This data will be transmitted to the Cloud Based Big Data
Management Platform, which will be also developed for LabHub platform. The data
will be described, defined and processed according to the Medical Test Device
(MDTO) Ontology to be processed by Anomaly Detection Service and Preventive

Maintenance Server.
The Medical Test Device Ontology (MTDO), which will be developed in the
LabHub design process, will support the identification of big data and the use of the
inference and reasoning capabilities provided by semantic web technologies. MTDO
will be developed in OWL language that is recommended and develop by W3 for
semantic web language [19]. The ontology development methodologies given in [20]
and [21] will be used to define MTDO. Considering the objectives of reusability and
interoperability in the process of developing ontology, concepts that can be added to
ontology from Healthcare Information Standards will be selected. Contradictions
between the concepts will be determined, and it will be possible to add new concepts
by making inferences. It will be tested that the newly added concepts do not cause any
conflict in their relations with the existing concepts. By using MDTO ontology to
describe the information of the systems, thanks to the capabilities of semantic web
technologies, the system will support SQL and NoSQL databases.
In addition to the managing the big data of the medical laboratories, machine
learning techniques will be used for anomaly detection and device/environment pre-
ventive maintenance recommendation. Machine learning technologies will be used in
the development of these services. Anomaly Detection Service will use the supervised
classification method which will developed by the kNN technique to detect unwanted
anomalies. It will activate the necessary warning mechanism in any violation. Although
the system is capable of stopping the test process, the current legislation does not allow
the test period to be stopped autonomously. Therefore, it is necessary to use warning
mechanism. Preventive Maintenance Service will also use supervised learning method
for classification will be carried out by LSTM technique. The classification will include
four levels: emergency, high, and medium and low. This service will give suggestions
to the healthcare institution for preventive maintenance and calibration about test
devices in the laboratory across the LabHub platform.
4 Conclusions and Future Work
In this paper, LabHub, an intelligent new generation laboratory information system

architecture is proposed. The main purpose of LabHub is to develop a big data man-
agement platform that will collect, transfer, manage and process the data of the medical
devices in the clinical laboratories and the environment in which these devices are
located, and provide the services to analyze these data for anomaly detection and
predictive maintenance through the cloud.
It is seen that information technologies have been used in HIMS and LIMS for
many years when the studies and systems are examined. However, there is not any
system in laboratories that manage the test analyzing process starting from where the
sample is taken from the patient through in which the test result is produced with the
environmental variables of this process are monitored and the data related to the device
are collected.
To collect data from the medical test devices and the environment where these
devices are located, a plug and play plugin named IoLT will be developed. Sensors will
be added to an embedded computer to detect the ambient physical conditions required

for the device to operate in the most accurate way. Thereby with IoLT plugin ensures
continuous data flow to the LabHub platform.
Data coming to the cloud through the plug-in conforms to the definition of big data
in speed, variety and volume aspects. Therefore, the information model of the platform
will be developed as a big data model. Medical Test Device Ontology (MTDO) will be
developed in order to identify the contexts and features between data in big data.
The fact that the data transferred to the cloud is completely about private health data
entails the obligation to provide privacy. Confidentiality and reliability will be ensured
in data transfer since the platform will be used by multiple laboratories. Therefore,
requirement for a multi-tenant application architecture arises. While applications and
data are isolated from other tenants, role-based access models will be considered for the
institution. For anomaly detection and preventive maintenance, supervised learning
algorithms of artificial intelligence techniques will be used to give proper services.
When the proposed architecture is equipped on the LabHub platform, first of all, the
workload of the health personnel will be reduced. In addition, smart services such as
calibration, preventive maintenance recommendations, and anomaly detection will
benefit from time and cost, while improving the reliability and accuracy of test results.
This will provide faster diagnosis and treatment.
References
1. Sezer, E., Can, Ö., Ünalır, M.O., Bursa, O.: An ontology-based integration approach for
medical information standards. In: 25th Signal Processing and Communications Applica-
tions Conference (SIU), Antalya, Türkiye (2017)
2. Cucoranu, I.C.: Laboratory information systems management and operations. Surg. Pathol.
Clin. 8(2), 153–157 (2015)
3. Plebani, M.: Harmonization in laboratory medicine: requests, samples, measurements and
reports. Crit. Rev. Clin. Lab. Sci. 53(3), 184–196 (2016)
4. Sezer, E., Ünalır, M.O., Yıldız, F., Gümüşkavak, A., Akçay, N.: Sağlık Alanı için
Kişiselleştirilmiş Nesnelerin İnterneti Platformu. Acad. Perspect. Procedia 1(1), 311–320
(2018)
5. Islam, S.M., Kwak, D., Kabir, M.H., Hossain, M., Kwak, K.: The internet of things for
health care: a comprehensive survey. IEEE Access 3, 678–708 (2015)
6. Li, S., Da Xu, L., Zhao, S.: The internet of things: a survey. Inf. Syst. Front. 17(2), 243–259
(2015)
7. Gope, P., Hwang, T.: BSN-Care: a secure IoT-based modern healthcare system using body
sensor network. IEEE Sens. J. 16(5), 1368–1376 (2015)
8. Darwish, A., Hassanien, A.E., Elhoseny, M., Sangaiah, A.K., Muhammad, K.: The impact of
the hybrid platform of internet of things and cloud computing on healthcare systems:
opportunities, challenges, and open problems. J. Ambient Intell. Human. Comput. 10(10),
4151–4166 (2019)
9. Andreu-Perez, J., Poon, C.C., Merrifield, R.D., Wong, S.T., Yang, G.Z.: Big data for health.
IEEE J. Biomed. Health Inf. 19(4), 1193–1208 (2015)
10. Belle, A., Thiagarajan, R., Soroushmehr, S.M., Navidi, F., Beard, D.A., Najarian, K.: Big
data analytics in healthcare. BioMed. Res. Int. 2015, 1–16 (2015)
11. Archenaa, J., Anita, E.M.: A survey of big data analytics in healthcare and government.
Procedia Comput. Sci. 50, 408–413 (2015)
12. Luo, J., Wu, M., Gopukumar, D., Zhao, Y.: Big data application in biomedical research and
health care: a literature review. Biomed. Inform. Insights 8, 1–10 (2016)
13. Sebetci, O., Aksel, M.: An overview of the studies of health ınformation systems in Turkey.
Int. J. Comput. Sci. Eng. 4(8), 100–106 (2016)
14. ICD. https://www.who.int/classifications/icd/en/
15. HL7/FHIR. https://www.hl7.org/fhir/
16. T.C. Sağlık Bakanlığı.: Tıbbi Laboratuvarlar Yönetmeliği. Resmi Gazate (2013)
17. ISO 17025. https://www.iso.org/ISO-IEC-17025-testing-and-calibration-laboratories.html
18. ISO 15189. https://www.iso.org/standard/56115.html
19. OWL. https://www.w3.org/OWL/
20. Noy, N.F., McGuinness, D.L.: Ontology Development 101: A Guide to Creating Your First
Ontology (2001)
21. Sezer, E., Ünalır, M.O., Can, Ö., Bursa, O.: A methodology for defining working healthcare
ontologies. Int. J. Sci. Technol. Res. 5(15), 146–156 (2019)
Center-of-Gravity Real Options Method
Based on Interval-Valued Fuzzy Numbers
Jani Kinnunen1(&), Irina Georgescu2, and Mikael Collan3

1
Åbo Akademi University, Turku, Finland
[email protected]
2
Bucharest University of Economics, Bucharest, Romania
[email protected]
3
Lappeenranta-Lahti University of Technology, Lappeenranta, Finland
[email protected]
Abstract. This paper introduces an extension to the center-of-gravity fuzzy

pay-off method, CoG-FPOM by Borges and others from 2018 meant for the
valuation of real options by using interval-valued fuzzy numbers. The CoG-
FPOM was developed as a response and a remedy to the identified inconsistency
with financial theory in the original fuzzy pay-off method for real option val-
uation (FPOM). The use of interval-valued fuzzy numbers introduced in this
paper allows taking into account a higher level of uncertainty and imprecision
than is possible with the original CoG-FPOM model. This higher level of
uncertainty can be encountered in many application areas. The proposed
approach builds on using triangular upper and lower membership functions of
fuzzy numbers, which represent expected pay-off distributions. The applicability
of the CoG-FPOM in high uncertainty situations is improved by allowing the
use of ranges instead of single numbers when inputting values for scenarios. An
illustrative numerical application is presented in the context of mergers and
acquisitions, where an acquiring company may receive options to abandon that
is, rights to divest non-core business units acquired together with the desired
parts of a target company. The model outcomes will be discussed and compared
with the results from other pay-off-based fuzzy real option valuation models
including the original FPOM.
Keywords: Fuzzy pay-off method Real option Center of gravity Interval-

Valued fuzzy numbers Possibility theory Mergers and acquisitions
1 Introduction
Borges and others [1] introduced a center-of-gravity fuzzy pay-off method, CoG-FPOM,
for real option valuation as a response to a found inconsistency in the original fuzzy
pay-off method (FPOM) [3]. They showed an example that illustrated a triangular fuzzy
net present value (NPV) based case, where the original model was found to be arguably
inconsistent with financial theory in the way that the real option value given by the
model from the triangular NPV of the case was lower than the value of the possibilistic
mean for the case, ceteris paribus. This is theoretically incorrect, as the real option
value should under these said circumstances be equal or higher than the possibilistic
Center-of-Gravity Real Options Method 1293
mean. The center-of-gravity approach was able to overcome the problem and in this
paper an interval-valued fuzzy number (IVFN) based variant of the method is pre-
sented. The new variant allows analysts to use ranges for inputting scenario informa-
tion in the valuation with the model. IVFNs were introduced by Zadeh [2]. The
valuation logic underlying the fuzzy pay-off models is akin to the one presented in the
probabilistic simulation-based Datar-Mathews method (DMM) [4]. The FPOM of
Collan et al. [3] uses the same components of valuation logic but operates in a partially
possibilistic setting with fuzzy numbers. The method can operate under higher
imprecision and uncertainty that does not allow for a model to be built (due to
structural uncertainty). The method computes the real option value (ROV) as the
positive area of a fuzzy NPV over the whole area of the same fuzzy NPV times the
possibilistic mean of the positive area of the fuzzy NPV (A+). The possibilistic mean is
used to derive what can be called a possibilistic expected value for the positive side of
the fuzzy NPV. The ROV calculated for an investment project with fuzzy NPV A can
be defined as:
R1
Að xÞdx
ROV ¼ R 10 E ðA þ Þ: ð1Þ
1 Að xÞdx
1 1
The “weight” in Eq. (1), R
Að xÞdx= R
Að xÞdx, is the ratio of the area of the positive
0 1
fuzzy NPV divided by the total area under what is typically a triangular distribution that
is defined by three single (crisp) numbers. This is where the difference with the
interval-valued models is apparent – they allow the use of ranges to represent the
uncertainty found in the tails of the fuzzy NPV – this is likely to be valuable for
practitioners of strategic investments in high-uncertainty environments. Mezei et al. [5]
introduced a possibilistic interval-valued pay-off model based on [3] and later a
credibilistic interval-valued model [6] based on [7] was published. Kinnunen and
Georgescu [8] used IVFNs with trapezoidal fuzzy distributions building on the earlier
three types of models: the possibilistic FPOM from [3] and [4], the credibilistic model,
Cred-POM [7], as well as, the CoG-FPOM [1].
The FPOM has been widely applied in many industries and investment domains
including mergers and acquisitions (M&As) [9–13]. A credibilistic scenario-based
model was applied to a portfolio of startup companies [12] and later the Cred-POM [7]
was applied to corporate acquisition targets in its triangular form [8]. The discussed
three types of models have been compared [13] with a corporate acquisition applica-
tion, where revenue and cost synergies are created during an M&A process. The
credibilistic model has shown to typically be stable in valuing investments between the
values obtained from FPOM and the CoG-FPOM, while the latter leads to higher
values than the other two modeling options, cf. [6, 8, 13].
This paper is structured as follows. Section 2 presents the required preliminary
notations, Sect. 3.1 recalls the original center-of-gravity model, and Sect. 3.2 extends it
to the use of interval-valued fuzzy numbers. Section 4 presents a numerical M&A
example and Sect. 5 concludes the paper.
1294 J. Kinnunen et al.
2 Preliminary Notations
Let A be a fuzzy number, whose level sets are ½ Ac ¼ ½a1 ðcÞ; a2 ðcÞ, c 2 ½0; 1 [13, 14].
Carlsson and Fullér [15] defined the possibilistic expected value Eð AÞ of the fuzzy
number A:
Z1
E ð AÞ ¼ c½a1 ðcÞ þ a2 ðcÞdc: ð2Þ
0
The center of gravity ECoG ð AÞ of the fuzzy number A is defined by [1]:

R1
1 xAð xÞdx
ECoG ð AÞ ¼ R1
: ð3Þ
1 Að xÞdx
A triangular fuzzy number A = (a, a, b) with a 2 R and a; b 0 has the form

[16, 17]:
8
< 1 ax
a ; if a a x a
Að xÞ ¼ 1 xa
b ; if a x a þ b ð4Þ
:
0; otherwise:
Proposition 1 [1]. If A is a triangular fuzzy number (a, a, b), then E ð AÞ ¼ a þ ba 6 and
3aa þ b
ECoG ð AÞ ¼ 3 .
An interval-valued fuzzy set (IVFS) is a function A from a universe X to the set of
closed intervals of [0, 1]. Thus, for all u 2 X, A(u) will be the interval ½AL ðuÞ; AU ðuÞ. In
this way, one obtains two fuzzy sets AL : X ! ½0; 1 and AU : X ! ½0; 1. We remark
that AL ðuÞ AU ðuÞ, for any u 2 X.
If AU and AL are fuzzy numbers, then A is an interval-valued fuzzy number (IVFN).
For each IVFN A, the possibilistic expected value of A is introduced [18]:
E ðAL Þ þ E ðAU Þ
EIV ð AÞ ¼ : ð5Þ
2
A triangular interval-valued fuzzy number A will be defined by the triangular fuzzy

numbers AU ¼ ða; a1 ; b1 Þ and AL ¼ ða; a2 ; b2 Þ. According to Proposition 1, E ðAU Þ ¼
b1 a1 b2 a2
aþ 2 and E ðAL Þ ¼ a þ 2 , therefore by (5) one gets:
b1 þ b2 a1 þ a2
EIV ð AÞ ¼ a þ : ð6Þ
4 4
The center of gravity EIVCoG ð AÞ of an interval valued fuzzy number is defined by:
ECoG ðAL Þ þ ECoG ðAU Þ

EIVCoG ðA Þ ¼ : ð7Þ
2
If A is a triangular interval-valued fuzzy number with AU ¼ ða; a1 ; b1 Þ and

AL ¼ ða; a2 ; b2 Þ, then ECoG ðAU Þ ¼ 3a þ a31 þ b1 and ECoG ðAL Þ ¼ 3a þ a32 þ b2 , hence by (7):
a1 þ a2 þ b 1 þ b 2
EIVCoG ðA Þ ¼ a þ : ð8Þ
6
3 Center-of-Gravity Real Options Method
In this section we, firstly, recall the original center-of-gravity fuzzy pay-off method,
CoG-FPOM, for real option valuation using triangular fuzzy numbers in Sect. 3.1 and,
secondly, derive the interval-valued model, IV-CoG-FPOM, using the interval-valued
triangular fuzzy numbers in Sect. 3.2.
3.1 The Original Center-of-Gravity Method

Borges et al. [1] introduced the CoG-FPOM for a triangular fuzzy number. The
membership function of the triangular fuzzy number A = (a, a, b) was defined in
Eq. (4). The center-of-gravity real option value is (cf. [1]):
R1 R1
Að xÞdx Að xÞdx R 1 0 xAð xÞdx
ROVCoG ¼ R 10 ECoG ðA þ Þ ¼ R 10 R1 ; ð9Þ
1 Að x Þdx 1 A ð x Þdx 1 Að xÞdx
where the center-of-gravity expected value, ECoG(A+), for the positive side of the
triangular fuzzy number A, replaces the E(A+) in formula (1), while the weight com-
ponent is the same as with the original CoG-FPOM (as well as with the seminal FPOM
[3]). We have the following four cases depending on the location of origin, which
represents NPV = 0 (cf. [1]).
Case 1, 0 a /: In case 1, the whole NPV distribution is above zero. In this case,
the ROVCoG value is simply the expected value of the whole distribution (Proposition
1) with full weight:
3a a þ b
ECoG ð AÞ ¼ ; ð10Þ
3
Weight1 ¼ 1; ð11Þ
which makes also the ROVCoG ð AÞ ¼ Weight1 ECoG ð AÞ ¼ 3aa3 þ b.

Case 2, a / 0 a: In case 2, part of the NPV distribution is negative with a

positive peak a. The expected value of the positive side and the weight [1]:
/ ða þ bÞ3 a3 ða þ bÞ
ECoG ðA þ Þ ¼ h i; ð12Þ
3 aða þ bÞ2 a2 ða þ bÞ
2
2a a/ þ b
Weight2 ¼ ; ð13Þ
/ þb
Case 3, a 0 a þ b: In case 3, the NPV distribution has negative peak, but a

positive b + b. Now, the expected value and the weight [1]:
aþb
ECoG ðA þ Þ ¼ ; ð14Þ
3
ða þ bÞ2
Weight3 ¼ : ð15Þ
bð/ þ bÞ
Case 4, a þ b 0: In case 4, the whole NPV distribution is negative. In this case:
ECoG ðA þ Þ ¼ 0; ð16Þ
Weight4 ¼ 0: ð17Þ
Thus, also the real option value ROVCoG ðA þ Þ ¼ Weight4 ECoG ðA þ Þ ¼ 0.
3.2 The Center-of-Gravity Method with IVFNs

For interval-valued fuzzy numbers, there are six cases instead of the above four. We
present analytical solutions for the expected values of normalized fuzzy variables and
the weights, which allow their direct implementations in any spreadsheet software. The
expected value of the positive part of an IVFN A defined in Eq. (7) is used.
Case 1, 0 \ a a1 : The fuzzy numbers, AU and AL are both fully positive. Thus,
the expected values of the positive NPV are obtained from Eq. (10) plugged into
Eq. (7):

1 3a a1 þ b1 3a a2 þ b2 a1 þ a2 b1 þ b2
EIVCoG A þ ¼ þ ¼a þ ð18Þ
2 3 3 6 6
and as the total areas under AU and AL are positive, the weight is, cf. Equation (11):
Weight1 ¼ 1: ð19Þ
Thus, also the real option value, Weight1 EIVCoG ðA þ Þ, becomes:
a1 þ a2 b1 þ b2
IVROVCoG ðA þ Þ ¼ a þ : ð20Þ
6 6
Case 2, a a1 \ 0 \ a a2 : Accordingly, in case 2, the AL is fully positive, i.e. it is

obtained from Eq. (10) like in case 1, but AU is partly negative and obtained plugging
Eq. (12) into Eq. (7):
2 3
1 4 /1 ða þ b1 Þ3 a3 ða1 þ b1 Þ 3a a2 þ b2 5
EIVCoG A þ ¼ h iþ : ð21Þ
2 3 a1 ða þ b Þ2 a2 ða1 þ b Þ 3
1 1
The weight is, accordingly, from Eq. (11) and Eq. (13):
2
2a /a 1 þ b1 þ a2 þ b2
Weight2 ¼ ; ð22Þ
/1 þ b1 þ /2 þ b2
and the real option value:

2
2a /a 1 þ b1 þ a2 þ b2
IVROVCoG ðA þ Þ ¼
/12 þ b1 þ /2 þ b2 3
1 4 /1 ða þ b1 Þ3 a3 ða1 þ b1 Þ 3a a2 þ b2 5
h iþ : ð23Þ
2 3 a1 ða þ b Þ2 a2 ða1 þ b Þ 3
1 1
Case 3, a a2 \ 0 \ a: In case 3, both fuzzy numbers are partly negative, the

peak is positive, and their average is obtained by plugging Eq. (12) into Eq. (7) and the
weight is obtained from Eq. (13):
2 3
1 4 /1 ða þ b1 Þ3 a3 ða1 þ b1 Þ /2 ða þ b2 Þ3 a3 ða2 þ b2 Þ 5
EIVCoG A þ ¼ h iþ h i ;
2 3 a1 ða þ b Þ2 a2 ða1 þ b Þ 3 a2 ða þ b2 Þ2 a2 ða2 þ b2 Þ
1 1
ð24Þ
2 2
4a /a 1 þ b1 aa2 þ b2
Weight3 ¼ ; ð25Þ
/1 þ b1 þ /2 þ b2
2 2
4a /a 1 þ b1 aa2 þ b2
IVROVCoG ðA þ Þ ¼ ð26Þ
/1 þ b1 þ /2 þ b2
2 3
1 4 /1 ða þ b1 Þ3 a3 ða1 þ b1 Þ /2 ða þ b2 Þ3 a3 ða2 þ b2 Þ 5
h iþ h i
2 3 a1 ða þ b Þ2 a2 ða1 þ b Þ 3 a2 ða þ b2 Þ2 a2 ða2 þ b2 Þ
1 1
Case 4, a \ 0 \ a þ b2 : In case 4, the peak is negative and the expected values of

both fuzzy numbers to be averaged are obtained from Eq. (14) plugged into Eq. (7) and
the weight from Eq. (15):

1 a þ b1 a þ b2
EIVCoG A þ ¼ þ ; ð27Þ
2 3 3
a2 a2
4a þ b1 þ b1 þ b2 þ b2
Weight4 ¼ ; ð28Þ
/1 þ b1 þ /2 þ b2
IVROVCoG ðA þ Þ ¼ Weight4 EIVCoG ðA þ Þ

2 2
4a þ ba þ b1 þ ba þ b2 1 a þ b1 a þ b2
¼ 1 2
þ ð29Þ
/1 þ b1 þ /2 þ b2 2 3 3
Case 5, a þ b2 \ 0 \ a þ b1 : In case 5, the lower fuzzy number AL is fully

negative (cf. Equation (16)), and the upper fuzzy number AL is obtained from Eq. (14)
like above, while the weigh from Eq. (15):

1 a þ b1
EIVCoG A þ ¼ ; ð30Þ
2 3
a2
2a þ b1 þ b1
Weight5 ¼ ; ð31Þ
/1 þ b1 þ /2 þ b2
a2
2a þ b1 þ 1 a þ b1
b1
IVROVCoG ðA þ Þ ¼ ð32Þ
/1 þ b 1 þ /2 þ b 2 2 3
Case 6, a þ b1 \ 0: Cf. Equation (16) and Eq. (17): In the trivial case 6,
EIVCoG A þ ; Weight6 , as well as, IVROVCoG ðA þ Þ are 0.
4 Numerical Example: Option to Abandon in M&A
A merger and acquisition process provides different real options for an acquiring
company that include, e.g., revenue and cost synergies, cross-sales potential, and
divestment/abandon options, which have been framed and valued as real options, cf.
[9, 10, 13, 19, 20]. A divestment option means having the right to sell/abandon, e.g.,
non-core business parts received together with “desired” parts of an acquisition target.
Table 1 presents cumulative NPV cash flows at t = 0, for five years received from
selling of a non-core production unit. Managers want to value the option to sell it.
Table 1. Discounted (NPV) cash flows (USD ´000) from a corporate divestiture
Year 1 2 3 4 5
Optimistic scenario 87–96 179–197 266–294 355–392 457–506
Base scenario 27 54 79 103 130
Pessimistic scenario –31 –60 –85 –105 –135
The NPV scenarios have been created for the optimistic, base, and pessimistic
cases. It is seen in Table 1 that the downside risk is considered to be less imprecise than
the upside potential - a range for cash-flows is used for the optimistic scenario, while
single number values are used for the other two scenarios. The pessimistic scenario has
the NPV of $-135,000 (a /), the base-case scenario $130,000 (a), and the optimistic
scenario between $457,000 (a þ b2 ) and $506,000 (a þ b1 ). The situation represents
Case 3 and IV-ROVCoG is obtained using Eq. (26) with 88.9% weight and the expected
positive side NPV of $184,350; IV-ROVFPOM is obtained from [5] and IV-ROVCred
from [6] to get: IV-ROVCoG = $163,777; IV-ROVFPOM = $133,450; IV-
ROVCred = $146,496. We note that IV-ROVCoG > IV-ROVCred > IV-ROVFPOM
meaning that the CoG model suggests the M&A target with the abandonment option to
be more valuable with the IV-ROVCoG than would be using the other two methods,
which is in line with non-interval comparisons in [6, 8], and in [13].
5 Conclusions
This paper introduced a new interval-valued method for real option valuation based on
recent center-of-gravity model, which has proven to be consistent with financial theory
[1]. The fuzzy pay-off models have been demonstrated intuitive and easy to implement
[21], which, together with the provided analytical solutions, allow easy applications
using any spreadsheet software by practitioners or in future research.
References
1. Borges, R.E.P., Dias, M.A.G., Dória Neto, A.D., Meier, A.: Fuzzy pay-off method for real
options: the center of gravity approach with application in oilfield abandonment. Fuzzy Sets
Syst. 353, 111–123 (2018)
reasoning-I. Inf. Sci. 8(3), 199–249 (1975)
3. Collan, M., Fullér, R., Mézei, J.: Fuzzy pay-off method for real option valuation. J. Appl.
Math. Decis. Syst. 2009, 14 (2009)
4. Datar, V., Mathews, S.: A practical method for valuing real options: the Boeing approach.
J. Appl. Corp. Finan. 19, 95–104 (2007)
5. Mezei, J., Collan, M., Luukka, P.: Real option analysis with interval-valued fuzzy numbers
and the fuzzy pay-off method. In: Kacprzyk, J., et al. (eds.) Advances in Fuzzy Logic and
Technology, vol. 641, Springer, Cham (2018)
6. Kinnunen, J., Georgescu, I.: Credibilistic real options analysis using interval-valued
triangular fuzzy numbers. In: Proceedings of the IRES-WMFARE International Conference,
Jakarta, Indonesia, pp. 32–37 (2020)
7. Collan, M., Fullér, R., Mezei, J.: Credibilistic approach to the fuzzy pay-off method for real
option analysis. J. Appl. Oper. Res. 4(4), 174–182 (2012)
8. Kinnunen, J., Georgescu, I.: Fuzzy real options analysis based on interval-valued scenarios
with a corporate acquisition application. Nordic J. Bus. (2020). Association of Business
Schools Finland
9. Collan, M., Kinnunen, J.: A procedure for the rapid pre-acquisition screening of target
companies using the pay-off method for real option valuation. J. Real Options Strategy 4(1),
117–141 (2011)
10. Collan, M., Kinnunen, J.: Acquisition strategy and real options. IUP J. Bus. Strategy 4, 45–
65 (2009)
11. Kinnunen, J., Collan, M.: Supporting the screening of corporate acquisition targets. In:
Proceedings of the 42nd International Conference on System Sciences (HICSS), Waikoloa,
Hawaii, pp. 1–8 (2009)
12. Georgescu, I., Kinnunen, J.: Credibility measures in portfolio analysis: from possibilistic to
probabilistic models. J. Appl. Oper. Res. 3(2), 91–102 (2011)
13. Kinnunen, J., Georgescu, I.: Support system for evaluating synergy real options in M&A. In:
Proceedings of the International Conference on Management and Information Systems
(ICMIS), 29–30 September, Bangkok, Thailand, pp. 408–418 (2019)
14. Georgescu, I.: Possibility Theory and the Risk. Springer, Heidelberg (2012)
15. Carlsson, C., Fullér, R.: On possibilistic mean value and variance of fuzzy numbers. Fuzzy
Sets Syst. 122(2), 315–326 (2001)
16. Dubois, D., Prade, H.: Fuzzy Sets and Systems: Theory and Applications. Academic Press,
New York (1980)
17. Dubois, D., Prade, H.: Possibility Theory: An Approach to Computerized Processing of
Uncertainty. Plenum Press, New York (1988)
18. Carlsson C., Fullér R., Mezei J.: Project selection with interval-valued fuzzy numbers. In:
Proceedings of the 12th IEEE International Symposium on Computational Intelligence and
Informatics (CINTI), Budapest, Hungary, pp. 23–26 (2011)
19. Bruner, R.: Applied Mergers and Acquisitions. Wiley, New York (2004)
20. Alvarez, L., Stenbacka, R.: Takeover timing, implementation uncertainty, and embedded
divestment options. Rev. Financ. 10, 1–25 (2006)
21. Kinnunen, J., Georgescu, I.: Intuitive fuzzy real options in digital coaching for strategic
investment decisions. In: Proceedings of International Conference on Business Excellence
(ICBE), 2020, Bucharest, Romania (2020)
Fusing Fuzzy Preferences in Contexts
of Social Choice: Towards a Topological
Structure
Marı́a Jesús Campión1 , Esteban Induráin2 ,

and Armajac Raventós-Pujol1(B)
1
Institute for Advanced Research in Business and Economics (INARBE),
Universidad Pública de Navarra, 31006 Pamplona, Spain
[email protected]
2
Institute for Advanced Materials (InaMat), Universidad Pública de Navarra,
31006 Pamplona, Spain
Abstract. We discuss the difference between ordinal and cardinal

axioms in the fuzzy Arrovian models. We expose how to add a topological
structure in the fuzzy Arrovian models as well as why topological struc-
ture and continuity is a reasonable hypothesis for these models. Finally,
we show how with a topological structure, a cardinal-like Arrovian model
could be handled using differential equations and inequations.
Keywords: Fuzzy Social Choice · Arrow’s theorem · Topology ·

Aggregation of preferences · Differential topology
1 Introduction
A key problem in Mathematical Social Choice is that of fusing individual pref-

erences –as inputs– into a social one –as the final output– in a way that the
resulting social preference reflects in a way as many of the features of the indi-
vidual preferences as possible. To guarantee that, in most typical models in
Social Choice literature, the rules to aggregate the individual preferences should
accomplish a list of axioms or restrictions that our common sense, so-to-say,
would impose to the fusion method.
However, in Social Choice the famous and well-known Arrow’s impossibility
theorem arises (see e.g. [2,12,13]). It clearly states that under an apparently
mild set of conditions that our common sense would naturally impose to any
aggregation rule that fuses individual preferences into a social one, no such
rule is possible. In other words, the restrictions of “common sense” become
incompatible.
Due to the impossibility result stated by Arrow’s theorem, we see that any
rule to aggregate rankings in that model and context will have, a fortiori, some
kind of bad behavior or undesired properties. That is, it will not accomplish
some of the desired conditions for a potential rule.
1302 M. J. Campión et al.
Nevertheless, all this happens in the crisp setting, where the preferences
are dychotomic, only taking values in the set {0, 1}. If X is a nonempty set
of alternatives, and we consider that an element x ∈ X is related to another
element y ∈ X, it happens that the relationship is either void (=0) or total
(=1.) That is, either the elements are not related at all, or they are fully related.
No intermediate situation is allowed.
Unlike the crisp classical approach, in several generalizations of the Arrovian
model that deal with fuzzy preferences to formalize uncertainty, there is still
room for some possibility results and, consequently, for the existence of suitable
aggregation rules. The generalizations from the crisp to the fuzzy setting are not
unique. Here, it is crucial to establish well a suitable definition of the concept
of a fuzzy preference. It is also crucial to set well the restrictions that will be
required to the aggregation rules. Needless to say, they should extend in a sense
those considered in the crisp models. But, once again, the generalizations to the
fuzzy setting are not unique. As a matter of fact, in this literature about pos-
sible extensions to the fuzzy setting of the classical Arrovian model, one may
encounter both impossibility and possibility theorems, depending on the defini-
tions considered, the extensions made, their terms and the conditions imposed
(see e.g. [7,9,11]). Sometimes the conditions imposed to the fusion rules are still
compatible, and the aggregation of fuzzy individual preferences is feasible, and
in other situations they are incompatible, and an impossibility result appears.
Among the fuzzy Arrovian models that could generate those possibility and
impossibility results, some authors [3,4] have distinguished two approaches to
perform the fuzzy generalizations of the crisp Arrovian context, namely the
ordinal approach (e.g. [9,14]) and the cardinal approach (e.g. [7,8,10]).
On the one hand, in the ordinal approach, given two alternatives x and y, the
theory focus on the comparison of the degree of preference of x over y (R(x, y))
when compared to the degree of y over x (R(y, x)) by means of the usual order of
the unit interval [0, 1]. In other words, the stress is put on which of the following
situations occurs: R(x, y) > R(y, x), R(x, y) = R(y, x) or R(x, y) < R(y, x).
Here, the degree of the preference itself, as number in the unit interval, (R(x, y))
is not so relevant. We just pay attention to the fact of R(x, y) being equal, greater
or smaller than R(y, x). On the other hand, the cardinal approach focus on the
degree of the preference, and the value of R(x, y) as a number in [0, 1] plays a
decisive role.
On the ordinal approaches, a binary relation R can be defined for every
preference R as follows: We declare that x R y ⇔ R(x, y) ≥ R(y, x). In some
cases, by means of that binary relation R , the fuzzy model can be analyzed by
means of a suitable associated crisp model. Because of the impossibility theorems
arising in the Arrovian crisp setting, it can be proved that in relevant fuzzy
ordinal approaches, an impossibility result arises, too.
For this reason, we opt for focusing on cardinal approaches, since they could
still give rise to some possibility results.
Moreover, most of the fuzzy aggregation rules with cardinal approach have
an algebraic basis (e.g. weighted means [10], WOWA [1] or medians [7]).
Fusing Fuzzy Preferences in Contexts of Social Choice 1303
Unfortunately, algebraic expressions can not describe in an easy manner the

whole set of fuzzy aggregation rules.
Bearing this in mind, we propose the exploration of other methods and tools
from analysis, topology and differential equations as well as the algebraic ones. In
crisp contexts arising in Social Choice, topological methods were already intro-
duced in the 1980’s by Chichilnisky and Heal among others, as an alternative
to the Arrovian models. And these new methods were proved to produce some
possibility results, depending on particular aspects of the topology considered on
the space of individual preferences (see [6]). Moreover, our topological approach
is different from the one introduced by Chichilinsky and Heal in [6]. Indeed in
that model, the starting point is an abstract space of preferences X, endowed
with a topology, such that an aggregation rule here is a continuous and symmet-
ric (i.e.: invariant through rearrangements of n-tuples) map from X n into X.
Being that set X totally abstract, the Chichilnisky-Heal setting is also valid if X
is a set of fuzzy preferences endowed with some suitable topology. However, in
our study we deal with fuzzy models already settled in the literature, and trying
to generalize the Arrovian ones to a fuzzy approach. Therefore, the models that
we consider here differ from those analyzed in [6]. As far as we know, the consid-
eration of topological tools on these fuzzy Arrovian models models constitutes
a novelty.
Our aim in the present work, is to introduce and analyze how the problem of
social aggregation of fuzzy preferences can be translated into some differential
and topological framework.
The structure of the present work goes as follows: After the Introduction,
where we also motivate the work to be done, in the second section we will fur-
nish a general framework for fuzzy Arrovian models. In the third section, we
will describe how the sets of fuzzy preferences can be understood as topological
Euclidean subspaces. In the fourth section, we will translate some fuzzy Arrovian
axioms to the differential language. These translations will be possible because
the axioms follow the cardinal approach. Indeed, dealing with the ordinal app-
roach, the analogous translations become troublesome. In the last section, we
will discuss the advantages and the scope of studying the problem of aggrega-
tion of fuzzy preferences from this new perspective, and we will give some clues
for their resolution, that is, for the obtention of new possibility results.
2 Fuzzy Preferences and Arrovian Models
This section is devoted to exposing the fuzzy Arrovian problem. There are many
fuzzy Arrovian models in the literature, each of them is an extension of the
classical one. Here, we will provide a few examples of extensions and we will use
them to illustrate the differences between ordinality and cardinality.
Let X stand for the set of preferences and set a t-norm T and a t-conorm S.
Definition 1. A fuzzy preference on X is a function R : X × X → [0, 1] reflex-

ive (R(x, x) = 1 for all x ∈ X), T -transitive (R(x, y) ≥ T (R(x, z), R(z, y)) for
all x, y, z ∈ X) and S-connected (S(x, y) = 1 for all x, y, ∈ X). The set of all
fuzzy preferences on X is denoted by FP X .
Other authors use other transitivities instead of T -transitivities. A good

example is the weak transitivity used in [4,14].
Definition 2. A function R : X × X → [0, 1] is weak transitive if for all x, y, z ∈

X the following implication holds:
[R(x, y) ≥ R(y, x) and R(y, z) ≥ R(z, y)] ⇒ R(x, z) ≥ R(z, x). (1)
In Definition 1, T -transitivity is used. It takes into account the degree of the

preference. Therefore, it follows a cardinal approach. Unlike this, weak transi-
tivity follows an ordinal approach, because it does not consider the degree of
preference itself. It only considers the order relation between the degrees.
A method of decomposition derives a relation PR : X × X → [0, 1] (called
strict preference) from any preference R. Various methods have been used in the
literature. Our analysis does not need any particular method of decomposition.
For this reason, we assume that we have set a decomposition method D that
provides a strict preference PR for every preference R (PR = D(R)). We also
assume that the following conditions are satisfied by D for every x, y ∈ X and
R ∈ FP X : (i) PR (x, y) > 0 ⇒ PR (y, x) = 0, (ii) PR (x, y) > 0 ⇔ R(x, y) >
R(y, x), (iii) PR (x, y) ≤ R(x, y). In [11], many methods of decomposition have
been described.
We denote a society of n individuals by N = {1, . . . n}. A profile of prefer-
ences for the society N is an n-tuple of preferences R = (R1 , . . . , Rn ) ∈ FP n
representing the preferences of all members of the society.
Definition 3. A n-aggregation rule on the set of fuzzy preferences on X is a

function f : (FP X )n → FP X .
The axioms of the Arrovian model can be extended into the fuzzy setting in
different ways. Here we will expose a few extensions. Even though in the following
section we will work exclusively with the cardinal extensions, we will also expose
ordinal axioms to make clear the difference between both approaches.
The Paretian Property (see [13]) can be extended cardinally as follows:
Definition 4. A n-aggregation rule f : (FP X )n → FP X is Strong Paretian

if for every profile R and every pair of alternatives x, y ∈ X, Pf (R) (x, y) ≥
mini∈N {PRi (x, y)} holds.
In addition, it can be extended using an ordinal approach as:
Definition 5. A n-aggregation rule f : (FP X )n → FP X is Weak Paretian if

for every profile R and every pair of alternatives x, y ∈ X if PRi (x, y) > 0 for
all i ∈ N then Pf (R) (x, y) > 0.
Whereas on the Strong Paretian the degrees of the strict preferences are
used, on the Weak Paretian they are not important. On the second definition,
it is only taken into account the fact of the degrees being positive.
The extension of Independence of Irrelevant Alternatives (see [13])
can be extended as follows:
Definition 6. A n-aggregation rule f : (FP X )n → FP X satisfies indepen-

dence of irrelevant alternatives if for every pair of profiles R, R and every pair
of alternatives x, y ∈ X, if Ri (x, y) = Ri (x, y) and Ri (y, x) = Ri (y, x), then
f (R)(x, y) = f (R )(x, y) and f (R)(y, x) = f (R )(y, x) hold true.
Even though Definition 6 is the most used in the literature (see [7,9]), other
ordinal fuzzy extensions of Independence of Irrelevant Alternatives are also used
(see [11, Definition 4.9], [11, Definition 4.12] or [4, Definition ]).
Non-dictatorship (see [13]) can be extended as follows:
Definition 7. A n-aggregation rule f : (FP X )n → FP X is non-dictatorial if
there is no k ∈ N such that for every profile R and every pair of alternatives
x, y ∈ X, if PRi (x, y) > 0, then Pf (R) (x, y) > 0 holds.
Definition 8. A n-aggregation rule f : (FP X )n → FP X fails to be strongly

dictatorial if there is no k ∈ N such that for every profile R, every pair of
alternatives x, y ∈ X and every α ∈ [0, 1], if PRi (x, y) > α, then Pf (R) (x, y) > α
holds.
The Fuzzy Arrovian problem consists of studying the existence of aggregation

functions for each possible extension of the classical Arrovian model. There is a
huge literature studying for which combinations of axioms suitable aggregation
functions do actually exist. However, there is no general result classifying Fuzzy
Arrovian models according with the existence of aggregation functions. In [7,9,
14] specific models are studied, and [11] is a complete survey about Fuzzy Social
Choice.
The topological framework on the next section is inspired on cardinal exten-
sions, i.e. Definitions 4, 6 and 8.
3 Topological Spaces of Fuzzy Preferences

In this section, we will endow the Arrovian model with a topological structure.
First, we will give a topology to the set of preferences. Second, we will expose
how cardinal fuzzy axioms can be dealt with under a topological framework.
When the set of alternatives is finite, fuzzy preferences can be seen as vectors
on hyperdimensional cubes. Formally, if we denote X2 = {(x, y) ∈ X 2 : x = y},
we can associate every fuzzy preference R, to the vector vR = (vi )i∈X2 ∈ [0, 1]X2
with v(x,y) = R(x, y)1 . So, we can see the set of fuzzy preferences on X as a subset
of [0, 1]X2 (see [5] for an study of the geometrical aggregation of finite fuzzy sets).
1
X2 is used as index set instead of X 2 because R(x, x) = 1 for all x ∈ X (reflexibility).
We can go further and give more mathematical structure to FRX . Since

[0, 1]X2 is a topological space with the Euclidean topology, we can provide FRX
with the topology inherited from [0, 1]X2 .
It is natural to endow the set of fuzzy preferences with a topology. It implies
that preferences are related to each other and are not independent entities. In our
case, the Euclidean topology has an associated metric. Therefore, the relation
induced by the topology between preferences can be interpreted as follows: if
two preferences are close, they represent similar states of opinion.
With the topological structure on preference spaces, we request aggregation
functions to be continuous. Continuous aggregation functions will imply that
close profiles will be aggregated similarly.
Topological structure and continuity are natural structures on fuzzy settings.
There are fuzzy opinions in which the degree of preference express a perception
of intensity and they can not be interpreted as an objective qualitative scale.
For example, the hot/cold perception. In such cases, when agents express their
preferences with a degree d ∈ [0, 1], the value d is vague in the sense that agents
could have reported a slightly higher or lower value for the same state of opinion.
For that reason, when we work with fuzzy preferences, it is important to consider
that agents can not translate their opinions into exact values.
In case we work with continuous aggregation functions, these slight differ-
ences on individual preferences can be translated to big differences on the aggre-
gation. Continuous functions guarantee that slight differences in individual pref-
erences will cause small differences on the aggregation.
If preferences are vectors on [0, 1]X2 , profiles are vectors on ([0, 1]X2 )n . A
profile R = (R1 , . . . , Rn ) is denoted by v R = (vik )i∈X2 with v(x,y)
k
= Rk (x, y).
k∈N
As a consequence, we can use the same coordinate for any aggregation function
f : (FP X )n → FP X : f = (fi )i∈X2 with each fi is a real function with variables
(vik )i∈X2 (the coefficients of the profile induced by the agent k ∈ N and the pair
k∈N
of alternatives (x, y) ∈ X2 ).
Once we have a topological structure on the set of preferences, adding a dif-
ferential structure is a reasonable next step. Since every continuous function can
be approximated by differential functions, we can study differential aggregation
functions first, and extend a fortiori our conclusions on continuous functions.
Next, we will provide conditions for differential aggregation functions to sat-
isfy the cardinal Arrovian axioms from the previous section.
Proposition 1. Let f : (FRX )n → FRX a differential aggregation func-
tion. f satisfies independence of irrelevant alternatives if, and only if, for every
x, y, a, b ∈ X and k ∈ N the following equations hold:
∂f(x,y)
k
≡0 if {x, y} = {a, b} (2)
∂v(a,b)
Sketch of proof. Set a k ∈ N , a, b, x, y ∈ X with {a, b} = {x, y} and a profile R.

For every t in a neighbourhood of 0 consider the profile Rt obtained by adding t
to the (a, b)-coordinate of v R . Applying independence of irrelevant alternatives

we obtain that f(x,y) (v Rt ) = f(x,y) (v R ) for all t. The result follows from the
application of the definition of a partial derivative. The converse implication
follows from the the fact that f(x,y) is constant with respect (a, b)-coordinate.
For the Strong Pareto axiom, we require to the decomposition method D to
be differentiable. The same argument used for aggregation functions can be used
here.
Proposition 2. Let f : (FRX )n → FRX a differentiable aggregation function.

A sufficient condition for f to be Strong Pareto is the satisfaction of the following
inequalities for every profile R, agent k ∈ N and pair of alternatives x, y ∈ X:
∂(D ◦ f(x,y) ) R ∂D k
k
(v ) ≥ (v R ) (3)
∂v(x,y) ∂v(x,y)
Sketch of proof. Consider the following

1 equality for a differentiable function
F : Rn → R: F (x) − F (x0 ) = 0 Dtx+(1−t)x0 F (x − x0 )dt (where Dp F is the
differential of F in p ∈ Rn ). If the previous formula is applied to D ◦ f(x,y) and
D, we prove that f is Strong Pareto.
This proposition states that if the aggregated strict preference increases faster
than the increasing of the strict preferences of all agents (while we modify the
profile), then the aggregation function is Strong Pareto.
The system of partial differential inequalities will depend on D. On the sim-
plest situation, for every preference R, the strict preference of x over y D(v R )(x,y)
depends only on the preferences between x and y (v(x,y) and v(y,x) ). However,
on other situations, D(v R )(x,y) depends on more variables and it increases the
complexity of the system of partial inequalities.
Finally, concerning the non-dictatorship condition, we expose the follow-
ing consideration: Whereas the independence of irrelevant alternatives and the
Pareto conditions concern the behaviour of the aggregation function along with
all preference set, non-dictatorship does not.
Typically, to assure the non-dictatorship condition, we only need to set the
aggregation of two profiles. With just setting the image of two profiles it can be
assured that no agent impose their opinion to the society.
Therefore, whereas in Propositions 1 and 2 we have exposed conditions which
constrain the behaviour of aggregation functions by means of differential equa-
tions and inequalities, the version of non-dictatorship is setting the image of at
least a pair of profiles, that is, additional boundary conditions to the differential
equations.
On this article, we have proposed how to add a topological structure to the

problem of setting and analyzing Arrovian fuzzy models (see also [15]). By doing
this, it is possible to use new techniques to find aggregation functions. We are

talking, in particular, about partial differential equations and inequations.
On the other hand, we have also opened the door to the consideration of
differential topology techniques. We leave for future work the study of whether
the topological space FP X is a differential manifold (with corners). If so, we
could study if the existence theorems of functions in differential topology can
prove the existence of aggregation functions.
It is also important to emphasize that introducing the use of differential
equations makes us think that the behaviour of aggregation functions can be
studied qualitatively. For example, quantifying how an agent’s opinion affects
social opinion. This can be useful in order to connect axiomatic studies with
applied modelling.
Acknowledgements. Thanks are given to the editor Prof. Dr. Kahraman and two
anonymous reviewers, for their valuable comments.
References
1. Alcantud, J.C.R., de Andrés Calle, R.: The problem of collective identity in a fuzzy
environment. Fuzzy Set. Syst. 315, 57–75 (2017)
2. Arrow, K.J.: Social Choice and Individual Values, 2nd edn. Yale University Press,
London (1963)
3. Basu, K.: Fuzzy revealed preference theory. J. Econ. Theory 32(2), 212–227 (1984)
4. Billot, A.: Economic Theory of Fuzzy Equilibria: An Axiomatic Analysis, 2nd edn.
5. Campión, M.J., Catalán, R.G., Induráin, E., Lizasoain, I., Raventós-Pujol, A.,
Valero, Ó.: Geometrical aggregation of finite fuzzy sets. Int. J. Approx. Reason.
103, 248–266 (2018)
6. Chichilnisky, G., Heal, G.: Necessary and sufficient conditions for a resolution of
the social choice paradox. J. Econ. Theory 31(1), 68–87 (1983)
7. Duddy, C., Perote-Peña, J., Piggins, A.: Arrow’s theorem and max-star transitivity.
Soc. Choice Welf. 36(1), 25–34 (2011)
8. Duddy, C., Piggins, A.: On some oligarchy results when social preference is fuzzy.
Soc. Choice Welf. 51, 717–735 (2018)
9. Dutta, B.: Fuzzy preferences and social choice. Math. Soc. Sci. 13(3), 215–229
(1987)
10. Garcı́a-Lapresta, J.L., Llamazares, B.: Aggregation of fuzzy preferences: some rules
of the mean. Soc. Choice Welf. 17(4), 673–690 (2000)
11. Gibilisco, M.B., Gowen, A.M., Albert, K.E., Mordeson, J.N., Wierman, M.J.,
Clark, T.D.: Fuzzy Social Choice Theory. Springer, Cham (2014)
12. Kelly, J.S.: Arrow Impossibility Theorems. Academic Press, New York (1978)
13. Kelly, J.S.: Social Choice Theory: An Introduction. Springer, Heidelberg (1988)
14. Mordeson, J.N., Gibilisco, M.B., Clark, T.D.: Independence of irrelevant alterna-
tives and fuzzy Arrow’s theorem. New Math. Nat. Comput. 8(2), 219–237 (2012)
15. Raventós-Pujol, A., Campión, M.J., Induráin, E.: Decomposition and arrow-like
aggregation of fuzzy preferences. Mathematics 8(3), 436 (2020)
Understanding the Behavior of Zadeh’s
Extension Principle for One-to-One Functions
by R Programming Language
Abbas Parchami(&) and Parisa Khalilpoor
Department of Statistics, Faculty of Mathematics and Computer,

Shahid Bahonar University of Kerman, Kerman, Iran
Abstract. The extension principle is one of the most fundamental principles in

the fuzzy set theory. It provides a powerful technique in order to extend the
domain of a classical function from real numbers to fuzzy sets. In the present
paper, we propose a new program in R software to graphically show the
behavior of the simplest type of the extension principle, where the function is
one-to-one. Numerical examples are given to show the results of the proposed
program by import-export figures and animations.
Keywords: Extension principle Fuzzy number One-to-one function
1 Introduction
Fuzzy set theory was formalized by Zadeh in his paper work entitled “Fuzzy sets” [4],
since its introduction in 1965, the theory of fuzzy sets has been deeply developed and it
has influenced in many applied fields. The extension principle of Zadeh is an effective
tool to develop the fuzzy arithmetic which its simplest version is quoted in bellow for
one variable function [2, 5, 6].
Definition 1.1 (Zadeh’s Extension Principle). Let f be a function such that f : X ! Y
and let A be a fuzzy subset of X: Then B ¼ f ðAÞ is a fuzzy subset of Y with the
membership function
(
sup AðxÞ if f 1 ðyÞ 6¼ ;
BðyÞ ¼ x2X: y¼f ðxÞ ð1Þ
0 if f 1 ðyÞ ¼ ;
in which f 1 ðyÞ ¼ fx : f ðxÞ ¼ yg.
1310 A. Parchami and P. Khalilpoor
Due to the Zadeh’s extension principle, every one-to-one correspondence, and more
generally every one-to-many correspondence, can be extended to fuzzy sets, i.e. sets
without precise boundaries. The complexities of the membership function of A and/or
the complexity of f function can cause some computational problems for the applicants
in various sciences. In this paper, we have studied about the extension principle as an
important concept in fuzzy mathematics and fuzzy statistics. In other words, we pro-
vide a computer program in R software which easily able to depict three following
graphs in one figure for the user: (1) the membership function of the imported fuzzy
number A, (2) the function f and (3) the membership function of the exported fuzzy set
B ¼ f ðAÞ. In other words, the objective of this research is visualization of the imported
fuzzy number and exported fuzzy set from the extension principle accompanied by the
extended function in one figure. We believe that the applied results of this program can
be useful for better understanding the behavior of Zadeh’s extension principle.
The organization of this paper is as follows. A computer program is presented in
Sect. 2 using “Fuzzy Numbers” package in R to show the behavior of the extension
principle. Several numerical examples are given in Sect. 3 to show the performance of
the proposed program. Conclusion part is given in the final section.
2 A Computer Program to Show the Behavior

of the Extension Principle
In this section, a computer program proposed to graphically show the behavior of the
extension principle by simultaneously plotting three following graphs in one figure:
(1) the membership function of the inputted fuzzy number in extension principle,
(2) the membership function of the output fuzzy set from the extension principle, and
(3) the extended one-to-one function.
The bellow program can be applied for teaching the graduate students in R software
after loud “FuzzyNumbers” package [1–3].
Understanding the Behavior of Zadeh’s Extension Principle 1311
library(FuzzyNumbers)
EX = function(A, f, knot.n=20, xlim=supp(A), ylim = supp(fA),
core.line=FALSE, grid=TRUE, ...){
fA = fapply(A, f)
alphas <- seq(0,1,len=knot.n+2)
PLFN.to.cuts = function(P, knot.n)
alphacut(P, round(( knot.n+1):0/(knot.n+1),4))
par(mar=c(.8,.8,1.8,1.8))
layout(mat = matrix(c(3, 0, 2, 1), nrow = 2, ncol = 2),
heights=c(2,1), widths=c(1,2))
x.vec = c( rev(PLFN.to.cuts(A,knot.n)[,1]),
PLFN.to.cuts(A,knot.n)[,2] )
y.vec = c(alphas, rev(alphas))
plot(x.vec, -y.vec, type='l', axes=FALSE,
frame.plot=TRUE, xlim=xlim, ylim=c(-1,0), ...)
Axis(side=2, labels=FALSE)
if(grid==TRUE) grid()
if(core.line==TRUE) abline(v=core(A), col=4, lty=3)
curve(f, col=2, xlab=NA,ylab=NA, xlim=xlim,ylim=ylim,...)
if(core.line==TRUE) abline(v=core(A), h=core(fA),
col=4, lty=3)
x.fA.vec=c(rev(PLFN.to.cuts(fA,knot.n)[,1]),
PLFN.to.cuts(fA,knot.n)[,2] )
plot(-y.vec, x.fA.vec, type='l', axes=FALSE,
frame.plot=TRUE, ylim=ylim, xlim=c(-1,0), ...)
if(core.line==TRUE) abline(h=core(fA), col=4, lty=3)
par(mfcol=c(1,1), mar=c(2,2,1,1))
}
It must be mentioned that the usage of the introduced EX function is as follows:

EX(A, f, knot.n=20, xlim=supp(A), ylim = supp(fA),
core.line=FALSE, grid=TRUE, ...)
in which A is the inputted fuzzy number in the extension principle and f is the one
variable function. The user able to control on the ranges X and Y axes by xlim and
ylim parameters, respectively. The user can adjust the accuracy of calculations in
plotting the membership functions by integer positive parameter knot.n with default
20. Meanwhile, the logical parameter grid adds calibration to the figure by
rectangular grid (with True default), and also the logical parameter core.line is
considered in the program for user’s control on drawing the core limits lines (with False
default value); see two examples of Sect. 3.
3 Numerical Examples
In this section, we provide two numerical examples from Zadeh’s extension principle
using the introduced EX function in province section. It must be emphasized that all the
programs and plots in this paper have been carried out using “FuzzyNumbers” package
in R software.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Example 3.1. In this example, the method of applying function f ðxÞ ¼ logðx þ 1Þ to
the trapezoidal fuzzy number A ¼ Tð0; 1; 2; 3Þ is graphically shown in Fig. 1. First, the
trapezoidal fuzzy number A is introduced by “FuzzyNumbers” package and then we
use the introduced function EX to plot the membership function of f ðAÞ (in the left side
of Fig. 1) as follows:
A <- as.PiecewiseLinearFuzzyNumber(
TrapezoidalFuzzyNumber(0,1,2,3), knot.n=100)
EX( A, function(x) sqrt(log(x+1)) )
Fig. 1. The result of EX function in Example 3.1 to compute the membership function of f ðAÞ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where f ðxÞ ¼ logðx þ 1Þ and A is a trapezoidal fuzzy number.
Fig. 2. The result of EX function in Example 3.1 to compute the membership function of f ðA2 Þ
pffiffiffiffiffi
where f ðxÞ ¼ x2 :
Moreover in the continue of Example 3.1, one can easily plots Fig. 2 for imply
the square root function on A using the bellow code. Note that the exported plot at the
left side of Fig. 2 is exactly the imported fuzzy number A into the extension principle,
pffiffiffiffiffi
since A2 ¼ A.
EX( A^2, function(x)sqrt(x) )
Example 3.2 In this example, we are going to show another ability of EX function to
create an animation in R software. This animation can be helpful for teaching graduate
students as well as understanding a part of the extension principle complexities in
practical problems. Let the trapezoidal fuzzy number A, as the input of the extension
principle, moves/shift from Tð0; 1; 2; 3Þ to Tð0; 1; 2; 3Þ 2 ¼ Tð2; 1; 0; 1Þ by 25
steps/frames. Under this input transformation, we are going to show the transformation
of the membership function of the output fuzzy number f ðAÞ in this example where
f ðxÞ ¼ x3 is the used function in extension principle. Due to the fact that we are not
able to print all the animation frames in this article, so we only printed the first and last
frames in Fig. 3 and the user can see the proposed animation with 25 frames by running
the bellow suggested program in R software.
for(i in seq(0,2,len=25)){
EX(A-i, function(x)x^3,xlim=c(-2,3),ylim=c(-10,15),core=T)
Sys.sleep(.2)
}
Fig. 3. The first (left figure) and the last (right figure) frames of the presented animation in
Example 3.2 to display the transfer of the membership function f ðAÞ in the extension principle
when f ðxÞ ¼ x3 and the fuzzy number A moves/shifts from Tð0; 1; 2; 3Þ to Tð2; 1; 0; 1Þ
4 Conclusions and Future Works
Many natural phenomena involve the ambiguity which can be formulated/recorded by

fuzzy numbers. The extension principle is a useful applied formula to extend any usual
function such that able to work with fuzzy numbers. This paper provides a computer
program in R software to graphically show the behavior of the extension principle by
simultaneously plotting (1) the membership function of input fuzzy number, (2) the
membership function of output fuzzy set from the extension principle, and (3) the
extended one-to-one function. Two numerical examples are presented for better
understanding of the proposed program.
The Zadeh’s extension principle has been designed for the extension of any arbi-
trary function, and not only for the extension of one-to-one functions. Hence, the
proposed program in this research can be generalized for the extension of any arbitrary
function in future works (e.g., for the extension of Normal probability density function,
sin function and etc.).
References
1. Gagolewski, M., Caha, J.: FuzzyNumbers Package: Tools to deal with fuzzy numbers in R.
R package version 0.4-1 (2015). https://cran.r-project.org/web/packages=FuzzyNumbers
2. Gagolewski, M., Caha, J.: A guide to the FuzzyNumbers package for R (FuzzyNumbers
version 0.4-1) (2015). http://FuzzyNumbers.rexamine.com
3. Parchami, A., Calculator.LR.FNs: Calculator for LR Fuzzy Numbers, R package version 1.1
(2016). http://CRAN.R-project.org/package=Calculator.L.R.FNs
5. Zadeh, L.A.: The concept of a linguistic variable and its application to approximate reasoning
I. Inf. Sci. 8, 199–251 (1975)
6. Zadeh, L.A.: The concept of a linguistic variable and its application to approximate reasoning
II. Inf. Sci. 8, 301–357 (1975)
Intuitionistic Fuzzy Z-numbers
Irem Ucal Sari(&) and Cengiz Kahraman

[email protected]
Abstract. Fuzzy sets and its extensions are widely used to determine mem-
bership of an element using different functions to decrease the effect of uncer-
tainty. In this paper, intuitionistic fuzzy Z-numbers that are a new extension of
fuzzy sets are proposed for the first time. First, fuzzy Z-numbers and intu-
itionistic fuzzy numbers are determined as preliminaries. Then definitions and
defuzzification of intuitionistic fuzzy Z-numbers are proposed.
Keywords: Fuzzy set extensions Intuitionistic fuzzy sets Fuzzy Z-numbers

Defuzzification
1 Introduction
Fuzzy sets have been first proposed by Lotfi. A. Zadeh in 1965 to handle uncertainty.
These fuzzy sets that are also named as “ordinary fuzzy sets” or “type-1 fuzzy sets” are
defined as a class of objects which are characterized by a membership degree assigned
to each object in the set ranging between 0 and 1 [1]. A non-membership function is the
complement of membership function that is 1 − l. Since then several extensions of
fuzzy sets have been proposed by various authors. Figure 1 shows the timeline of the
fuzzy extensions and their founders.
Fuzzy set extensions differ from each other based on the definitions of membership
functions. In type 2 fuzzy sets, membership functions are represented by fuzzy num-
bers. In intuitionistic fuzzy sets, non-membership degree is involved in addition to
membership degree in the representation of membership functions. The sum of these
degrees may be less than 1 indicating the hesitancy of the experts. In Hesitant fuzzy
sets, several possible membership degrees for a certain value are used. Pythagorean
fuzzy sets and fermatean fuzzy sets are extensions of intuitionistic fuzzy sets which
offer larger domains for membership and non-membership degrees. In fuzzy
Z-numbers, fuzzy restriction and fuzzy reliability to that restriction are defined.
The purpose of this paper is to propose some definitions for intuitionistic fuzzy
Z-numbers that is a new extension of fuzzy sets in which fuzzy restriction and certainty
of the fuzzy restriction are denoted by intuitionistic fuzzy numbers.
The rest of the paper is organized as follows: First, preliminaries of fuzzy
Z-numbers and intuitionistic fuzzy numbers are determined in Sects. 2 and 3,
respectively. Then, in Sect. 4, some definitions of intuitionistic fuzzy Z-numbers are
proposed. In Sect. 5, paper is concluded with future research suggestions.
Intuitionistic Fuzzy Z-numbers 1317
Ordinary
• Zadeh - 1965 [1]
Fuzzy Sets
Type-2 Fuzzy
• Zadeh - 1975 [2]
Sets
Interval • Zadeh - 1975 [2]

Valued Fuzzy • Sambuc - 1975 [3]
Sets • Jahn - 1975 [4]
• Grattan-Guiness - 1976 [5]
Intuitionistic
• Atanassov - 1986 [6]
Fuzzy Sets
Fuzzy
• Yager - 1986 [7]
Multisets
Netrosophic
• Smarandache - 1998 [8]
Sets
Nonstationary
• Garibaldi &Ozen - 2007 [9]
Fuzzy Sets
Hesitant
• Torra - 2010 [10]
Fuzzy Sets
Fuzzy Z
• Zadeh - 2011 [11]
Numbers
Pythagorean • Yager& Abbasov - 2013 [12]
Fuzzy Sets • Yager - 2013 [13]
Picture fuzzy
• Cuong - 2014 [14]
Sets
Orthopair
• Yager - 2016 [15]
Fuzzy Sets
Spherical
• Kutlu Gundogdu&Kahraman - 2019 [16]
Fuzzy Sets
Fermatean
• Senapati & Yager - 2019 [17]
Fuzzy Sets
Fig. 1. Timeline of fuzzy set extensions
2 Fuzzy Z-numbers

Zadeh [11] introduced Z-numbers that is an ordered pair of fuzzy sets; Z A; ~ B
~ in

which first component A ~ is a fuzzy restriction of the values of X variable, and the

second component B ~ is referred to as certainty of the first component.
The restriction Rð X Þ : X is A is referred to a possibilistic restriction shown in Eq. (1)
where lA~ ð xÞ is the membership function of A: ~
1318 I. U. Sari and C. Kahraman
Rð X Þ : X is A ! PossðX ¼ xÞ ¼ lA~ ð xÞ ð1Þ

Figure 2 represents a simple fuzzy Z-number; Z ¼ A;~ B
~ which has a trapezoidal
membership function for fuzzy restriction and a triangular membership function for
fuzzy reliability.
Fig. 2. A simple fuzzy Z number

~ ¼ hx; l ~ ð xÞijlð xÞ 2 ½0; 1 and B
Let A ~ ¼ hx; lB~ ð xÞijlð xÞ 2 ½0; 1 where l ~ ð xÞ is a
A A
trapezoidal membership function and lB~ ð xÞ is a triangular membership function. To
convert a Z number into a regular fuzzy number, Eqs. 2–4 could be used [18].
Equation 2 converts the reliability into a crisp number:
R
xl ~ ð xÞdx
a¼ R R ð2Þ
lR~ ð xÞdx
R
where denotes an algebraic integration.
The weighted Z-number can be denoted as Z~ a by adding the weight of the relia-
bility to the restriction:

Z~ a ¼ hx; lA~ a ð xÞijlA~ a ð xÞ ¼ alA~ ð xÞ; lð xÞ 2 ½0; 1 ð3Þ
The weighted Z number, in other words weighted restriction, can be converted to an

ordinary fuzzy number Z~ 0 which is shown in Fig. 3 using Eq. 4:

x
Z~ ¼
0
hx; lZ~ 0 ð xÞijlZ~ 0 ð xÞ ¼ lA~ pffiffiffi ; lð xÞ 2 ½0; 1 ð4Þ
a
Fig. 3. Ordinary fuzzy number transformed from Z-number
Atanassov [6] introduced intuitionistic fuzzy sets (IFS). An IFS in E is defined as;

~ ¼ hx; l ~ ð xÞ; v ~ ð xÞix 2 E
A ð5Þ
A A
where lA~ : E ! ½0; 1 defines the degree of membership and vA~ : E ! ½0; 1 defines the
degree of non-membership of the element x 2 E. A simple intuitionistic fuzzy number
which has triangular fuzzy membership and non-membership functions is shown in
Fig. 4.
Fig. 4. A simple intuitionistic fuzzy number
For every x 2 E, the relation between membership and non-membership functions

should satisfy the following equation:
0 lA~ ð xÞ þ vA~ ð xÞ 1 ð6Þ
Beside that, the degree of non-determinacy or hesitancy is determined as follows:

pA~ ð xÞ ¼ 1 lA~ ð xÞ vA~ ð xÞ ð7Þ
The main operations of the intuitionistic fuzzy sets are defined as follows:

A ~ ¼ x; l ~ ð xÞ þ lB~ ð xÞ l ~ ð xÞ:lB~ ð xÞ; v ~ ð xÞ:vB~ ð xÞx 2 E
~þB ð8Þ
A A A

~B
A: ~ ¼ x; l ~ ð xÞ:lB~ ð xÞ; v ~ ð xÞ þ vB~ ð xÞ v ~ ð xÞ:vB~ ð xÞx 2 E ð9Þ
A A A
Defuzzification of Intuitionistic Fuzzy Numbers: In the following, we propose two

defuzzification methods: one for triangular IFNs and one for trapezoidal IFNs:
Let Ii ¼ ðaL ; aM ; aU ; bL ; bM ; bU Þ be a triangular intuitionistic fuzzy number. Then,
the defuzzification is realized by using function defined in Eq. (10).
aL þ 2aM þ aU bL þ 2bM þ bU
df ¼ þ ð10Þ
4 s
where s is a very large number, such as 100, indicating the effect of non-membership
function on the IFN.
Let Ii ¼ ðaL ; aM1 ; aM2 ; aU ; bL ; bM1 ; bM2 ; bU Þ be a trapezoidal intuitionistic fuzzy
number. Then, the defuzzification is realized by using function defined in Eq. (11).
aL þ 2ðaM1 þ aM2 Þ þ aU bL þ 2ðbM1 þ bM2 Þ þ bU

df ¼ þ ð11Þ
4 s
4 Intuitionistic Fuzzy Z-numbers

An intuitionistic fuzzy Z-number is an ordered pair of fuzzy sets Z A ~I ; B
~ I in which
fuzzy restriction and certainty of the fuzzy restriction are denoted by intuitionistic fuzzy
numbers. Figure 5 shows the membership functions of an intuitionistic fuzzy
Z-number.
Fig. 5. A simple intuitionistic fuzzy Z-number


Let A ~ I ¼ hx; l ~ ð xÞ; v ~ ð xÞix 2 E ¼ ða2 ; a3 ; a4 ; a5 ; a1 ; a3 ; a4 ; a6 Þ and B
~I ¼
A A

hx; lB~ ð xÞ; vB~ ð xÞi x 2 E ¼ ðb2 ; b3 ; b4 ; b1 ; a3 ; b5 Þ where lA~ ð xÞ is a trapezoidal mem-
bership function and lB~ ð xÞ is a triangular non-membership function. The restriction
and reliability functions of A ~ I and B ~ I are defined as in Eqs. 12–15, respectively.
8 xa2
>
> a3 a2 ; if a2 x a3
>
< 1; if a3 x a4
lA~ I ð xÞ ¼ x ð12Þ
> aa5a
> ; if a4 x a5
>
: 5 4
0; Otherwise
8 a3 x
>
> a3 a1 ; if a1 x a3
>
< 1; if a3 x a4
vA~ I ð xÞ ¼ ð13Þ
>
>
xa4
; if a4 x a6
>
: a6 a4
0; Otherwise
8 xb2
< b3 b2 ;
> if b2 x b3
lB~ I ð xÞ ¼ b4 x
; if b3 x b4 ð14Þ
>
: b4 b3
0; Otherwise
8 b3 x
< b3 b1 ; if b1 x b3
>
vB~ I ð xÞ ¼ xb3 ; if b3 x b5 ð15Þ
>
: b5 b3
0; Otherwise
b2 þ 2b3 þ b4 b1 þ 2b3 þ b5
aI ¼ þ ð16Þ
4 s
The weighted Z-number can be denoted as Z~ a by adding the weight of the relia-
bility to the restriction:
nD E o
Z~ aI ¼ x; l e aI ð xÞ jl e aI ð xÞ ¼ aI l e ðxÞ ; lð xÞ 2 ½0; 1 ð17Þ
AI AI AI
The weighted Z number in other words weighted restriction can be converted to an

ordinary fuzzy number Z~ using Eq. 18:
0

x
Z~ ¼
0
x; lZ~ 0 ð xÞ jlZ~ 0 ð xÞ ¼ lA~ I pffiffiffiffi ; lð xÞ 2 ½0; 1 ð18Þ
aI
If the restriction function and reliability function are defined as in Fig. 6, the
calculations are modified as follows:

~ d1 ;d2 number, Z~d1 ;d2 ¼ A
Fig. 6. A simple Z ~ d1 ; R
~ d2

Let A~ I ðg ;d Þ ¼ x; l ~ ð xÞ; g1 ; v ~ ð xÞ; d1 x 2 E ¼ ða2 ; a4 ; a5 ; a7 ; g1 ; a1 ; a3 ; a6 ; a7 ; d1 Þ
1 1
A A
~ I ðg ;d Þ ¼ x; lB~ ð xÞ; g2 ; vB~ ð xÞ; d2 x 2 E ¼ ðb2 ; b3 ; b4 ; g2 ; b1 ; b3 ; b5 ; d2 Þ where
and B 2 2
ðd Þ ðd Þ
lA~ 1 ð xÞ is a trapezoidal membership function and lB~ 2 ð xÞ is a triangular membership
function.
In this case, restriction and reliability functions are defined as in Eqs. 19–22,
respectively.
8 xa2
>
> a4 a2 d1 ; if a2 x a4
>
<d ;
ðd Þ 1 if a4 x a5
lA~ 1 ð xÞ ¼ x ð19Þ
> a a d1 ; if a5 x a7
>
a 7
>
: 7 5
0; Otherwise
8
> g1 1
þ a3 g1 a1
>
>
> a3 a1 x a3 a1 ; if a1 x a3
>
<g ;
ðg Þ if a3 x a6
mA~ 1 ¼ 1 ð20Þ
> 1g1 x þ g1 a8 a6 ; if a x a
>
>
> a8 a6 a8 a6 6 8
>
:
0; Otherwise
8 xb2
< b3 b2 d2 ; if b2 x b3
>
ðd2 Þ
lB~ ð xÞ ¼ b4 x
d2 ; if b3 x b4 ð21Þ
>
: b4 b3
0; Otherwise
8
>
>
g2 1
x þ b3 g2 b1
; if b1 x b3
>
< 3 1
b b b 3 b 1
ðg2 Þ
mB~ ð xÞ ¼ g2 b5 b3 ð22Þ
b5 b3 x þ
1g2
>
> b5 b3 ; if b3 x b5
>
:
0; Otherwise

b2 þ 2 b3 þ b4 b1 þ 2 b3 þ b5
aI ¼ d2 þ ð1 g2 Þ ð23Þ
4 s
Then, the weighted Z~da1 ;d2 number can be denoted as in Eq. 24:

Z~da1 ;d2 ¼ d1
hx; l ð xÞijl d1 d1
ð x Þ ¼ aI l ; lð xÞ 2 ½0; 1 ð24Þ
eI aI
A eI aI
A eI aI
A
The ordinary fuzzy number converted from Z-fuzzy number can be given as in
Eq. 25:

b2 þ 2b3 þ b4 b1 þ 2b3 þ b5
Z~d1 ;d2 ¼
0
hx; l~dz01 ð xÞijl~zd01 ð xÞ ¼ lAd~1 x d2 þ ; x 2 ½0; 1
4 s
ð25Þ
5 Conclusion
Fuzzy extensions offer decision makers flexibility on the determination of fuzzy

membership functions. In this paper, intuitionistic fuzzy Z-numbers where fuzzy
restriction and fuzzy reliability of a Z-number are determined using intuitionistic fuzzy
numbers are defined for the first time and its basic characteristics are identified.
For further research, it is suggested to extend the intuitionistic fuzzy Z-numbers
using interval intuitionistic numbers for determination of fuzzy restriction and fuzzy
reliability of fuzzy Z-numbers. Besides, applications of intuitionistic fuzzy Z-numbers
on several decision making methods such as AHP, TOPSIS, WASPAS etc. could be
made to increase the information involved in the analysis.
References
reasoning—I. Inf. Sci. 8(3), 199–249 (1975)
3. Sambuc, R.: Function U-Flous. Application a l’aide au Diagnostic en Pathologie Thyroi-
dienne. These de Doctorat en Medicine. University of Marseille (1975)
4. Jahn, K.U.: Intervallwertige Mengen. Math Nach 68(1), 115–132 (1975)
5. Grattan-Guinness, I.: Fuzzy membership mapped onto intervals and many-valued quantities.
Math. Logic Q. 22(1), 149–160 (1976)
7. Yager, R.R.: On the theory of bags. Int. J. General Syst. 13(1), 23–37 (1986)
8. Smarandache, F.: Neutrosophy. Neutrosophic Probability, set, and logic, ProQuest
information & learning. Ann Arbor, Michigan, USA, vol. 105, pp. 118–123 (1998)
11. Zadeh, L.A.: A note on Z-numbers. Inf. Sci. 181(14), 2923–2932 (2011)
12. Yager, R.R., Abbasov, A.M.: Pythagorean membership grades, complex numbers, and
(2016)
17. Senapati, T., Yager, R.R.: Fermatean fuzzy sets. J. Ambient Intell. Human. Comput. 1–12
(2019)
18. Yildiz, N., Kahraman, C.: Evaluation of social sustainable development factors using
Buckley’s fuzzy AHP based on Z-numbers. In International Conference on Intelligent and
Fuzzy Systems, pp. 770–778. Springer, Cham, July 2019
Risk Assessment
Risk Assessment of R&D Activities Using
Intuitionistic Fuzzy AHP and FIS
Esra Ilbahar1,2(B) , Selcuk Cebi1 , and Cengiz Kahraman2

1
{eilbahar,scebi}@yildiz.edu.tr
2
Macka, 34367 Istanbul, Turkey
[email protected]
Abstract. National and international supports for research and develop-

ment projects have been increasing over the world. Risk management has a
vital impact on the success of these projects. In risk management, all of the
potential actions that may negatively affect the processes or outputs of a
project should be identified and their negative effects should be minimized
before starting the project by taking a set of prevention. Therefore, in this
study, risks in research and development projects are identified and evalu-
ated by using interval-valued intuitionistic fuzzy AHP and fuzzy inference
system. The most critical risks are determined as insufficient planning and
abnormal changes in costs.
Keywords: Research & Development · Risk assessment · Intuitionistic

fuzzy sets · Fuzzy inference system
1 Introduction
National and international supports for research and development (R&D) projects
of the entrepreneurs have been increasing day by day over the world. Although
some qualified business ideas benefit from these supports and become inspiring
success stories, many of them cannot achieve the desired success. The main rea-
son for this is not only that the idea of the project is worthless, but also the unex-
pected events causing huge costs. Moreover, if the entrepreneurs do not have expe-
rience in project and risk management, they may not be able to manage project
risks and make the project a success. For this reason, effective risk assessment in
R&D processes is crucial for the success of the project. Project risk management
includes planning process, identifying and analyzing the risks, implementing the
necessary measures, monitoring and control of the risks in a project [1]. In other
words, risk management is an effort to identify and manage potential problems
that may arise while implementing a project. For an effective risk management,
all potential actions that may have a negative influence on process or outcome of a
1328 E. Ilbahar et al.
project should be identified and their negative effects should be minimized before
starting the project by taking the necessary measures.
Since it contains many uncertainties and it is required to predict future
events, risk management is a difficult process. The success of risk manage-
ment depends on expertise and experience. In other words, the uncertainty in
R&D projects is higher because many phases of R&D projects have not yet
been repeated. Therefore, it is hard to use past experiences directly. Nowadays,
software-based applications and technological products have been increasingly
popular, thus risk analysis of software R&D projects has become more important
[2]. In this study, interval-valued intuitionistic fuzzy Analytic Hierarchy Process
(IVIF-AHP) is employed to identify the probability and severity weights of risks
in software R&D projects. Then, these weights are used with a fuzzy inference
system (FIS) to complete the risk assessment procedure. In this study, fuzzy
logic is needed for the risk evaluation, because the assessment process relies on
the expert opinions and the data collected from the experts for risk severity and
probability of occurrence are in linguistic form. The reason of preferring intu-
itionistic fuzzy sets is their ability to better represent the expert judgments by
considering both membership and non-membership values together with their
hesitancies.
The rest of this paper is organized as follows: Sect. 2 presents the literature on
risk management of R&D projects. The structure of the approach and an illus-
trative example are given in Sect. 3 and Sect. 4, respectively. Finally, concluding
remarks are presented in Sect. 5.
2 Literature Review on Risk Assessment of R&D Project

There have been several studies related to risk assessment in R&D projects and
product development process, since effective strategies for eliminating risks are
required to successfully develop a new product [3]. Adler et al. [4] investigated
strategic projects to identify whether risk management strategies within the con-
tract are related to main outcomes of project or not. In particular, the issue of
managing financial risks associated with unplanned changes in project outputs
is examined. Cooper [3] proposed the utilization of knowledge-based tools to
balance benefits and risks. In the study, it was indicated that failures in R&D
projects or product development process are caused by intrinsic and extrinsic
problems. Some intrinsic problems causing product failures are not meeting per-
formance, reliability, safety, or other requirements whereas extrinsic problems
are unfavorable reception in market and regulatory changes. Intrinsic problems
causing project failures are identified as violating resource constraints whereas
extrinsic problem is that the product can be previously developed by competi-
tors [3]. Yang et al. [5] used the belief rule-based (BRB) inference method for
the risk assessment of R&D projects. In the study, a new predictive evalua-
tion framework was introduced to eliminate the problem of oversized rule base
caused by too many risk factors. This new evaluation framework, BRB with
random subspaces, was validated by using the data obtained from R&D projects
Risk Assessment of R&D Activities 1329
in Chinese industries. Andre and Miguel [6] utilized Failure Mode and Effects
Analysis (FMEA) to systematize risk management in product development. In
the study of Wu et al. [7], an integration of FMEA and graphical evaluation
and review technique was used to manage risks related to product development.
Keizer et al. [8] examined the applicability of a risk reference framework to iden-
tify risks in R&D projects. Choi and Ahn [9] employed fuzzy theory and Markov
processes to calculate the risk degrees of processes in product development.
3 Methodology
In this section, we will describe an approach consisting of interval-valued intu-
itionistic fuzzy AHP and FIS to evaluate the risks in R&D projects [10]. The
framework of the approach is presented in Fig. 1.
Fig. 1. Framework of the study
In the first step, potential events that might have a negative impact on the
project should be defined through comprehensive literature review and expert
opinions. After the identification of all risks, pairwise comparisons of them with
respect to severity and likelihood of occurrence (probability) are carried out.
Pairwise comparisons of risks are obtained from the experts by using linguistic
terms and these terms are converted to the corresponding interval-valued intu-
itionistic fuzzy numbers as given in Table 1. Then, the probability and severity of
the risks are calculated by using interval-valued intuitionistic fuzzy AHP intro-
duced by Wu et al. [11].
Table 1. Linguistic scale for evaluation [10]
Linguistic terms [μL , μU ] [νL , νU ]

Very High (VH) [0.60, 0.80] [0.00, 0.20]
High (H) [0.45, 0.65] [0.15, 0.35]
Equal (E) [0.30, 0.50] [0.30, 0.50]
Low (L) [0.15, 0.35] [0.45, 0.65]
Very Low (VL) [0.00, 0.20] [0.60, 0.80]
In the second step, the rule based system presented in Table 2 is utilized to
identify the category of a risk by using the weights of probability and severity
obtained in the previous step. The explanations of the categories given in Table 2
are as follows:
– Category I: The risks in this category are qualified as catastrophic risk imply-
ing that the risk has a huge potential to cause severe damage or failure of the
project.
– Category II: The risks in this category are marginal risks which can pose a
major threat to the success of the project.
– Category III: The risks in this category are insignificant. In other words, the
risk might occur but it probably will not cause any damage [10].
Table 2. Rule based system for risk assessment [10]
Severity ratio for a risk: WRS

Likelihood ratio
WRS≥0.08 0.08>WRS≥0.06 0.06>WRS≥0.04 0.04>WRS≥0.02 0.02>WRS≥0
for a risk:WRP
0.08≤WRP Category I Category I Category I Category II Category III
0.06≤WRP<0.08 Category I Category I Category II Category II Category III
0.04≤WRP<0.06 Category I Category II Category II Category III Category III
0.02≤WRP<0.04 Category II Category II Category III Category III Category III
0.0≤WRP<0.02 Category III Category III Category III Category III Category III
4 Application
Many software projects fail due to various reasons. The main reasons may be
classified as fallows:
– R1. Integration
• R11. Insufficient planning
• R12. Insufficient resource allocation
• R13. Insufficient integration management
– R2. Scope
• R21. Inadequate identification of work packages
• R22. Inaccurate determination of requirements’ quantities
• R23. Inadequate scope control
– R3. Schedule
• R31. Inaccurate determination of work packages’ duration
• R32. Inaccurate determination of project duration
• R33. Inaccurate determination of the critical path
– R4. Cost
• R41. Insufficient cost estimates
• R42. Barriers to accessing finance
• R43. Abnormal changes in costs
– R5. Quality
• R51. Design that does not meet the requirements
• R52. Violation of customer-oriented design
• R53. Inadequate quality assurance program
– R6. Human Resources
• R61. Insufficient human resources
• R62. Lack of qualified personnel
• R63. Insufficiency in organization and defining responsibilities
– R7. Procurement
• R71. Deficiencies in contract articles
• R72. Competitiveness issues
• R73. Problems in procurement processes.
Table 3. Decision matrix for probability Table 4. Decision matrix for severity
comparison of main risks comparison of main risks
Risks 1 2 3 4 5 6 7 Risks 1 2 3 4 5 6 7
1 E H VL VL E H E 1 E H L VL E H E
2 L E VL VL L L E 2 L E VL VL L L H
3 VH VH E E VH VH VH 3 H VH E VL VH H H
4 VH VH E E VH VH VH 4 VH VH VH E VH VH VH
5 E H VL VL E H H 5 E H VL VL E H H
6 L E VL VL E E E 6 L E L VL E E H
7 VL L VL VL L E E 7 VL L L VL L L E
CR: 0.02 CR: 0.09
Table 5. Probability comparison of inte- Table 6. Severity comparison of integra-

gration risks tion risks
R1 R11 R12 R13 R1 R11 R12 R13

R11 E VH VH R11 E H VH
R12 VL E E R12 L E H
R13 VL E E R13 VL L E
CR: 0.000 CR: 0.033
Pairwise comparisons of the risks obtained in the linguistic form for an

R&D process, and consistency ratios of these pairwise comparisons are given
in Tables 3, 4, 5, 6,7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18. Consistency ratios of
these pairwise comparison matrices are computed as in the classical AHP method
by considering the crisp numbers corresponding to these linguistic terms.
Table 7. Probability comparison of scope Table 8. Severity comparison of scope

risks risks
R2 R21 R22 R23 R2 R21 R22 R23

R21 E H VH R21 E L H
R22 L E E R22 H E VH
R23 VL E E R23 L VL E
CR: 0.025 CR: 0.033
Table 9. Probability comparison of Table 10. Severity comparison of sched-

schedule risks ule risks
R3 R31 R32 R33 R3 R31 R32 R33

R31 E VL E R31 E H E
R32 VH E H R32 L E VL
R33 E L E R33 E VH E
CR: 0.025 CR: 0.025
Table 11. Probability comparison of cost Table 12. Severity comparison of cost
risks risks
R4 R41 R42 R43 R4 R41 R42 R43

R41 E VH E R41 E E L
R42 VL E VL R42 E E VL
R43 E VH E R43 H VH E
CR: 0.000 CR: 0.025
Table 13. Probability comparison of Table 14. Severity comparison of quality

quality risks risks
R5 R51 R52 R53 R5 R51 R52 R53

R51 E VH E R51 E E VH
R52 VL E VL R52 E E H
R53 E VH E R53 VL L E
CR: 0.000 CR: 0.025
Table 15. Probability comparison of Table 16. Severity comparison of human

human resources risks resources risks
R6 R61 R62 R63 R6 R61 R62 R63

R61 E L E R61 E E VL
R62 H E H R62 E E VL
R63 E L E R63 VH VH E
CR: 0.000 CR: 0.000
Table 17. Probability comparison of pro- Table 18. Severity comparison of pro-
curement risks curement risks
R7 R71 R72 R73 R7 R71 R72 R73

R71 E VL E R71 E VL VL
R72 VH E VH R72 VH E E
R73 E VL E R73 VH E E
CR: 0.000 CR: 0.000
Table 19. Overall weights of the probability and severity of the risks
Probability
R1 0.142 R2 0.110 R3 0.178 R4 0.195 R5 0.148 R6 0.121 R7 0.107
R11 0.067 R21 0.049 R31 0.048 R41 0.078 R51 0.060 R61 0.036 R71 0.028
R12 0.038 R22 0.031 R32 0.079 R42 0.038 R52 0.029 R62 0.050 R72 0.050
R13 0.038 R23 0.029 R33 0.051 R43 0.078 R53 0.060 R63 0.036 R73 0.028
Severity
R1 0.142 R2 0.119 R3 0.163 R4 0.199 R5 0.146 R6 0.130 R7 0.102
R11 0.062 R21 0.040 R31 0.058 R41 0.057 R51 0.061 R61 0.034 R71 0.020
R12 0.047 R22 0.052 R32 0.037 R42 0.053 R52 0.052 R62 0.034 R72 0.041
R13 0.032 R23 0.027 R33 0.068 R43 0.089 R53 0.033 R63 0.061 R73 0.041
The overall probability weights and severity weights of these risks are pro-
vided in Table 19. Then, risk analysis process is completed by using Table 2 and
categories of risks are provided in Table 20.
As a result of the proposed assessment, the most critical risks are identi-
fied as R11. Insufficient planning and R43. Abnormal changes in costs, followed
by R31. Inaccurate determination of work packages’ duration, R32. Inaccurate
determination of project duration, R33. Inaccurate determination of the criti-
cal path, R41. Insufficient cost estimates, R51. Design that does not meet the
requirements, R63. Insufficiency in organization and defining responsibilities,
and R72. Competitiveness issues. Since these risks have a potential to constitute
a problem in R&D processes, activities related to these risks must be constantly
monitored and the necessary measures should be taken. The rest of the risks are
determined as Category III, indicating that they can be ignored.
Table 20. Risk categories obtained by the approach
Risks Category
R11. Insufficient planning Category I
R12. Insufficient resource allocation Category III
R13. Insufficient integration management Category III
R21. Inadequate identification of work packages Category III
R22. Inaccurate determination of requirements quantities Category III
R23. Inadequate scope control Category III
R31. Inaccurate determination of work packages duration Category II
R32. Inaccurate determination of project duration Category II
R33. Inaccurate determination of the critical path Category II
R41. Insufficient cost estimates Category II
R42. Disabled in accessing finance Category III
R43. Abnormal changes in costs Category I
R51. Design that does not meet the requirements Category II
R52. Violation of Customer-oriented design Category III
R53. Inadequate quality assurance program Category III
R61. Insufficient human resources Category III
R62. Lack of qualified personnel Category III
R63. Inadequate organization and defining responsibility Category II
R71. Deficiencies in contract articles Category III
R72. Competitive problems Category II
R73. Problems in procurement processes Category III
5 Conclusions
R&D projects are vital for companies to compete in the global market and
there are several risks to be managed in these projects. Risk assessment of R&D
projects essentially relies on prediction of the likelihood of project success and
effective mitigation of risks [5]. However, there is uncertainty in both the possible
consequences and the probability of their realization [3]. Therefore, in this paper,
a combination of interval-valued intuitionistic fuzzy AHP and FIS which is able
to cope with uncertainty better is preferred for the risk assessment of R&D
projects. The utilized evaluation method showed that R11. Insufficient planning
and R43. Abnormal changes in costs are the most critical risks in R&D projects
and managers should take the necessary precautions for these risks. For further
research, this framework can be expanded and combined with other risk analysis
techniques and compared with the results obtained in this study.
References
1. Rose, K.H.: A guide to the project management body of knowledge (PMBOK R
Guide)—Sixth Edition, Newtown Square, PA: Project Management Institute,
ISBN 9781628253924 (2017)
2. Lu, X.N., Ma, Q.G.: Risk analysis in software development project with owners
and contractors. In: 2004 IEEE International Engineering Management Conference,
18–21 October 2004 (2004). https://doi.org/10.1109/IEMC.2004.1407488
3. Cooper, L.P.: A research agenda to reduce risk in new product development
through knowledge management: a practitioner perspective. J. Eng. Technol.
Manag. 20(1–2), 117–140 (2003)
4. Adler, T.R., Pittz, T.G., Meredith, J.: An analysis of risk sharing in strategic
R&D and new product development projects. Int. J. Project Manag. 34(6), 914–
922 (2016)
5. Yang, Y., Wang, J., Wang, G., Chen, Y.W.: Research and development project
risk assessment using a belief rule-based system with random subspaces. Knowl.
-Based Syst. 178, 51–60 (2019)
6. Andre, S., Miguel, P.A.C.: Failure mode and effects analysis (FMEA) in the context
of risk management in new product development. Int. J. Qual. Reliab. Manag.
25(9), 899–912 (2008)
7. Wu, D.D., Kefan, X., Gang, C., Ping, G.: A risk analysis model in concurrent
engineering product development. Risk Anal. Int. J. 30(9), 1440–1453 (2010)
8. Keizer, J.A., Vos, J.P., Halman, J.I.: Risks in new product development: devising
a reference tool. R&D Manag. 35(3), 297–309 (2005)
9. Choi, H.G., Ahn, J.: Risk analysis models and risk degree determination in new
product development: a case study. J. Eng. Technol. Manag. 27(1–2), 110–124
(2010)
10. Cebi, S., Ilbahar, E.: Warehouse risk assessment using interval valued intuitionistic
fuzzy AHP. Int. J. Anal. Hierarchy Process 10(2), 243–253 (2018)
11. Wu, J., Huang, H.B., Cao, Q.W.: Research on AHP with interval-valued intuition-
istic fuzzy sets and its application in multi-criteria decision making problems. Appl.
Math. Model. 37(24), 9898–9906 (2013)
Risk Matrix Technique Based Fuzzy MCDM
Risk Assessment Method and Application
in Glass Industry
Sukran Seker(&)
Department of Industrial Engineering, Yildiz Technical University, Istanbul,

Turkey
[email protected]
Abstract. Glass industry is known among the most dangerous sectors. Multi-
Criteria Decision Making (MCDM) methods are used widely for risk assessment
in various industry sectors. In this study, Risk Matrix applied for risk assessment
is integrated with fuzzy The Technique for Order Preferences by Similarity to an
Ideal Solution (TOPSIS) method. Fuzzy method is employed to deal with
uncertain information derived from the nature of industry and subjective eval-
uation of decision makers (DMs). Once the required risk parameters for Risk
Matrix’ are determined, the hazards are ranked using fuzzy TOPSIS method in
glass industry. Consequently, satisfactory results are obtained as a result of the
Risk Matrix based fuzzy MCDM risk assessment approach.
Keywords: Risk assessment Risk Matrix Fuzzy TOPSIS
1 Introduction
Glass is an important material used in a broad variety of the fields such as in the
science, industry and society. Glass is used in many different forms since it is inex-
pensive and has many desirable properties [1]. Furthermore, the 80 to 85% of the mass
production in the world’ glass industry is used as bottles for the food, beverage and
pharmaceutical industries, and flat glass for construction or motor vehicle production
[2]. Glass production includes three-part processes: the batch processing, forming
process, and the cold-end. The batch processing process the raw materials; forming
process conducts the manufacturing process and the cold end process consists of the
product-quality control and packaging [1].
In recent years, the glass industry like many industries, has revealed that providing
occupational health and safety (OHS) is as important as the production requirement.
Since glass production is delicate and fragile due to its structure, it contains variety
complex processes. This situation create many risks that threaten worker health and
safety in the workplace throughout the glass production process. The Bureau of Labor
Statistics reports that the glass manufacturing industry shows greater than average
accident rates for workplace. The common health hazards with respect to glass man-
ufacturing processes are exposure to chemical substances such as respirable airborne
particulates, ergonomic hazards, physical hazards such as noise exposures owing to
Risk Matrix Technique Based Fuzzy MCDM Risk Assessment Method 1337
heavy machinery, material and equipment, radiant energy, heat stress disorders and
infrared radiation [1]. As the part of risk management, risk assessment is a process in
which identified hazards are analyzed to determine the possible cause’s accidents and
also impacts are reduced or eliminated to protect the health and safety of the employees
[3].
Since the risk assessment can be implemented using qualitative and quantitative
information, it requires different risk assessment techniques. In this paper, traditional
risk assessment matrix approach based on Risk Matrix is integrated with Multi Criteria
Decision Making (MCDM) method to rank risky factors in the glass industry. Thus,
TOPSIS introduced by [4] is integrated with Risk Matrix technique under fuzzy
environment. MCDM method is selected since the application of MCDM methods are
widely used to solve risk management problems and efficient results are obtained. For
example, [5] presented a fuzzy entropy-weight MCDM method and applied to risk
assessment of hydropower stations. [3] presented fuzzy MCDM method based on the
risk matrix technique as risk assessment method and implement them on an aluminum
industry’s plant. [6] proposed a risk assessment method based on Pythagorean fuzzy
analytic hierarchy process (PFAHP) method using cosine similarity and neutrosophic
fuzzy AHP. [7] defined and prioritized 86 hazards at the Kerman coal deposit, Iran by
performing fuzzy TOPSIS. Failure Mode and Effects Analysis (FMEA) and
Fuzzy AHP are used for risk assessment of green elements to dangerous substances by
[8].
The study is organized as: the applied methodology is presented in Sect. 2. The
application of Fuzzy-MCDM based risk assessment method in glass industry is
introduced in Sect. 3. The results with preventive measures are presented in the same
section. Lastly, conclusion section is conducted in Sect. 4.
2 Methodology
As a part of risk management, risk assessment is the systematic process include

identifying hazards, evaluating risks, mitigating the risks and reporting the results. This
paper presents traditional Risk Matrix based fuzzy TOPSIS method for risk assessment
in glass industry. The methods are used throughout the paper are introduced in this
section.
2.1 Risk Assessment Matrix

The Risk Matrix technique is a logical technique which is extensively applied in the
field of OHS as risk assessment technique. The technique is based on risk value
includes severity (S) and likelihood (P). Risk value is obtained as [9]:
R:PS ð1Þ
Using Table 1 and Table 2 the P and S measurements of risk value are determined.
Once the decision matrix is constructed according to P and S of the risks, the tolerable
level of the risks is evaluated using Table 3.
1338 S. Seker
Table 1. Likelihood ratings (P) for risk matrix

Rating category Description
Rare (1) Very rarely
Unlikely (2) Remote (Once a year)
Possible (3) Occasional (A few events in a year)
Likely (4) Often (Monthly)
Almost certain (5) Very Often (Once a week, every day)
Table 2. Severity ratings (S) for risk matrix

Rating category Description
Insignificant (1) Loss of working hours and situations that do not require first aid
Minor (2) The situations that requires outpatient or first aid
Moderate (3) Minor injury situations requiring hospitalization
Major (4) Serious injuries and occupational disease requiring treatment
Catastrophic (5) Death or permanent disability
Table 3. The evaluation scale

Intolerable risks (25) Work should not be conducted, If the identified risks aren’t reduced
an acceptable level
Significant risks (15, Work should be stopped immediately until emergency precautions
16, 20) are taken
Intermediate risks (8, Work should be investigated to minimize the identified risks and to
9, 10, 12) take prevention measures
Acceptable risks (2, 3, The preventive measures should be applied continuously and these
4, 5, 6) controls should be observed
Insignificant risks (1) To conduct control processes isn’t necessary
2.2 Fuzzy Set Theory and Trapezoidal Fuzzy Numbers

Fuzzy Set is presented by [10] to handle uncertain information and extended versions
are widely used in the literature. Some definitions for generalized Trapezoidal fuzzy
numbers (TFNs) are illustrated in the following. A generalized TFNs can be illustrated
~ ¼ ðm1 ; m2 ; m3 ; m4 ; w ~ Þ and the membership function að xÞ : R ! ½0; 1 is
as a vector A A
shown as follows [11].
8 xm1
> m2 m1 xw^ x 2 ðm 1 ; m 2 Þ
>
> A;
>
< w^ x 2 ðm 2 ; m 3 Þ
að x Þ ¼ A; ð2Þ
>
> xm4
xw^ x 2 ðm 3 ; m 4 Þ
>
> m3 m4
: A;
0 x 2 ð1; a1 Þ [ ða4 ; þ 1Þ
where m1 m2 m3 m4 and w^ 2 ½0; 1.

A
~ ¼ ðm1; m2 ; m3; m4 Þ and B

The arithmetical operations of two positive TFNs A ~¼
ðn1; n2 ; n3; n4 Þ where m1 0 and n1 0 are defined as follows:
~B
A ~ ¼ ðm1 þ n1; m2 þ n2; m3 þ n3; m4 þ n4 Þ ð3Þ
~ k ¼ ðm1 þ k; m2 þ k; m3 þ k; m4 þ kÞ
A ð4Þ
~ B
A ~ ¼ ðm1 n1; C; m3 n3; m4 n4Þ ð5Þ
~
B
A ~ ðm1 * n1; m2 * n2; m3 * n3; m4 * n4Þ ð6Þ
~
k ¼ ðm1 k; m2 k; m3 k; m4 k Þ
A ð7Þ
^
The defuzzified (crisp) value of a fuzzy number A ¼ ðm1; m2 ; m3; m4 Þ is defined
as follows:
1 m3m4 m1m2
¼ ðm1 þ m2 þ m3 þ m4
D A Þ ð8Þ
3 ðm3 þ m4Þ ðm1 þ m2Þ
~ BÞ;
Euclidian distance dE ðA; ~ BÞ;
~ dH ðA; ~ ¼ ðm1; m2; m3; m4 Þ
~ between two TFNs A
~
and B ¼ ðn1; n2; n3; n4 Þ is computed as follows:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2 2 2
dE ðA; ~ ¼ ðm1 n1Þ þ 2ðm2 n2Þ þ 2ðm3 n3Þ þ ðm4 n4Þ
~ BÞ ð9Þ
6
2.3 Risk Matrix Based Fuzzy TOPSIS Risk Assessment Methodology

A MCDM problem consists of set of K (DMs) shown DM ¼ fDM1 ; DM2 . . .:; DMk g,
set of m possible hazards shown H ¼ fH1 ; H2 . . .:; Hm g, set of n criteria (risk factors)
shown C ¼ fC1 ; C2 . . .:; Cn g a set of performance ratings of hazards considering cri-
teria shown X ¼ fxij ; i ¼ 1; 2; . . .; m; j ¼ 1; 2; . . .; ng.
The Risk Matrix based on Fuzzy TOPSIS approach motivated from [12] is as
follows:
Step 1. The decision matrix is constructed. Once the DMs determine the importance
weights vector (W) of the risk factors, DM evaluate, DMs evaluate hazards with respect
to P and S using linguistic terms shown in Table 4.
2 C1 Cn 3
H1 ~x11 . . .: ~x1n
D ¼ . . . 6 .. .. .. 7 ð10Þ
4 . . . 5
Hm ~x ~xmn
m1
1340 S. Seker
Table 4. Linguistic terms and corresponding TFNs for weighting criteria and rating alternatives
Linguistic Term Trapezoidal fuzzy Number (TFN) for TFN for rating
weighting criteria alternatives
Very Low (VL) (0, 0,0.1, 0.2) (0, 0,1, 2)
Low (L) (0.1,0.2,0.2,0.3) (1, 2, 2, 3)
Medium Low (0.2, 0.3, 0.4, 0.5) (2, 3, 4, 5)
(ML)
Medium (M) (0.4, 0.5, 0.5, 0.6) (4, 5, 5, 6)
Medium High (0.5, 0.6, 0.7, 0.8) (5, 6, 7, 8)
(MH)
High (H) (0.7,0.8,0.8,0.9) (7, 8, 8, 9)
Very High (VH) (0.8, 0.9, 1, 1) (8, 9, 10, 10)
~ ¼ ½w
W ~ 1; w
~ 2 ; . . .; w
~ n ð11Þ

where ~xij ¼ aij ; bij ; cij ; dij and w ~ j1 ; w
~j ¼ w ~ j2 ; w
~ j3 ; w
~ j4 ; i ¼ 1; 2; . . .; m j ¼ 1; 2; . . .; n
be positive TFNs.
Step 2. Aggregate decision matrices and determine the weights of P and S. Using
Aggregation operator in Eq. (13) individual evaluations are aggregated to obtain
aggregated decision matrix and to determine the weights of the P and S of the hazards.
Let K DMs evaluate m hazards with respect to n criteria using positive trapezoidal
fuzzy numbers R ~ k ¼ ðak ; bk ; ck ; dk Þ. The aggregated fuzzy ratings can be defined as
~ ¼ ða; b; c; d Þk ¼ 1; 2; . . .; K
R
XK 1 XK
~ ¼ min fak g ; b ¼ 1
R b k ; c ¼ c; d¼
maxfdk g
ð12Þ
k K k¼1 K k¼1 k k

min aijk P P max dijk
where a ¼ ;b ¼ K1 Kk¼1k¼1 bijk , c ¼ K1 Kk¼1 cijk , d ¼ .
k k
Assume the importance weight of the kth DM for n criteria

~ jk ¼ w
w ~ jk1 ; w
~ jk2 ; w
~ jk3 ; w
~ jk4 ; i ¼ 1; 2; . . .; m j ¼ 1; 2; . . .; n ð13Þ
Step 3. Normalize decision matrix. The decision matrix is normalized considering cost
and benefit criteria as

~ ¼ ~rij
N ð14Þ
mxn
where B and C are the sets of benefit and cost criteria respectively
!
aij bij cij dij
~rij ¼ ; ; ; ; j2B ð15Þ
dj dj dj dj

aj aj aj a j
~rij ¼ ; ; ; ; j2C ð16Þ
dij cij bij aij
dj ¼ max dij ; j 2 B ð17Þ
a
j ¼ min aij ; j 2 C ð18Þ
Step 4. Obtain the weighted normalized decision matrix. The weighted normalized
matrix is obtained as:

~ ¼ ~vij
V i ¼ 1; 2; . . .; m j ¼ 1; 2; . . .; n ð19Þ
mxn
where ~vij ¼ ~rij w

e j.
Step 5. Determine positive ideal solution (PIS, A ) and negative ideal solution (NIS,
A ). PIS and NIS can be defined as:

A ¼ ~v1 ; ~v2 . . .:; ~vn ð20Þ

A ¼ ~v v
1 ;~ v
2 . . .:; ~n ð21Þ

where ~vj ¼ max vij4 and ~v
j ¼ min vij1 .
Step 6. Calculate the distance from PIS and NIS: The distance from PIS and NIS is
calculated as:
Xn
di ¼ j¼1
dv ð~vij ; ~vj Þ; i ¼ 1; 2; . . .; m ð22Þ
Xn
di ¼ j¼1
dv ð~vij ; ~v
j Þ; i ¼ 1; 2; . . .; m ð23Þ
Step 7. Obtain Closeness coefficient (CC): Closeness coefficient of each hazard ðCCi Þ
is calculated as:
di
CCi ¼ i ¼ 1; 2; . . .; m ð24Þ
di þ di
Step 8. Rank the hazards. Descending order of CCi is determined to rank the alter-
natives. The hazard has the highest CC is selected as the most dangerous hazard.
3 Case Study
In this section fundamental hazards encountered in glass industry are ranked and after
prioritization of the hazards required preventive measures are recommended. The glass
industry is known as risky industry since the produced products are delicate and fragile,
1342 S. Seker
and requires complex process. Glass production includes three-part operations: the
batch house, the hot end, and the cold end. The fundamental hazards are identified as
Cutting due to splattered broken glass (H1), Respiratory problems due to hazardous
chemicals (H2), Slips, Trips and Falls due to scattered broken glass (H3), Popping out
materials (H4), Noise exposure due to machinery (H5), Ergonomic hazards due to
manual handling (H6), heat stress due to Temperature Exposure (H7), Contamination
risk due to hazardous materials (H8), Machine and Electrical hazards with respect to
interact with machinery or equipment (H9) by five experts (DMs) who have at least five
year experience in glass industry. The stepwise of the risk assessment method is
explained as follows:
Step 1. Determine the hazards with P and S value. Fundamental risk factors associated
with glass industry are determined by five DMs. The DMs determine P value of
occurring these hazards and S value of the hazards using linguistic terms in Table 4.
The related TFNs are translated for weighting of P and S and rating of hazards with
respect to P and S factors.
Step 2. Aggregate decision matrix: The aggregated weights of P and S are determined
using Eq. (13). The aggregated decision matrix is obtained using the same equation.
Aggregated decision matrix is represented in Table 5.
Step 3. Establish the weighted decision matrix. The weighted decision matrix is
obtained using Eq. (19). The weighted decision matrix is shown in Table 6.
Step 4. Determine the PIS and NIS. The PIS and NIS are determined as given in
Eqs. (20–21).
Step 5. Calculate the distance from PIS and NIS: The distance from PIS and NIS are
calculated using Eqs. (22–23).
Step 6. Obtain CC of each hazard. CC of each hazard is obtained and the results are
shown in Table 7.
Step 7. Rank hazards and suggest required measures: The ranking order of hazards is
obtained and shown in Table 7.
Table 5. Fuzzy decision matrix

Risk Probability Severity (S)
(P)
H1 5 6.8 7.4 9 2 5.4 6.4 8
H2 4 6.2 6.8 9 5 6.4 7.2 9
H3 2 5.2 6 8 4 5.2 5.4 8
H4 2 5.2 5.4 9 4 5.6 5.6 9
H5 4 5.6 6.2 8 4 5.8 6.6 8
H6 5 6.4 7.2 9 4 6.6 7 9
H7 5 6.8 7.4 9 7 8 8 9
H8 4 6.2 6.8 9 2 5.8 6.6 9
H9 5 7.2 7.6 9 5 7.6 7.8 9
Table 6. Weighted fuzzy decision matrix

Risk Probability (P) Severity (S)
H1 0.101 0.123 0.134 0.182 0.136 0.17 0.202 0.545
H2 0.101 0.134 0.147 0.228 0.121 0.151 0.17 0.218
H3 0.114 0.152 0.175 0.455 0.136 0.202 0.21 0.273
H4 0.101 0.169 0.175 0.455 0.121 0.195 0.195 0.273
H5 0.114 0.147 0.163 0.228 0.136 0.165 0.188 0.273
H6 0.101 0.126 0.142 0.182 0.121 0.156 0.165 0.273
H7 0.101 0.123 0.134 0.182 0.121 0.136 0.136 0.156
H8 0.101 0.134 0.147 0.228 0.121 0.165 0.188 0.545
H9 0.101 0.12 0.126 0.182 0.121 0.14 0.143 0.218
Table 7. Distances from PIS and NIS and ranking of hazards.

Risk A* A− Ci* Rank
H1 0.007 0.109 0.063 4
H2 0.003 0.118 0.026 6
H3 0.014 0.094 0.131 1
H4 0.013 0.096 0.122 2
H5 0.005 0.107 0.047 5
H6 0.003 0.120 0.025 7
H7 0.001 0.132 0.006 9
H8 0.011 0.107 0.092 3
H9 0.002 0.129 0.012 8
Accordingly, the most dangerous risks are obtained as: Slips, Trips and Falls due to
scattered broken glass (H3), Popping out materials (H4) and Contamination risk due to
hazardous materials (H8).
Many serious cuts and eye injuries occur due to breaking and flying of the glass.
Accidents can be prevented by keeping walking and working surfaces clean and dry
and by giving non-slip shoes to employees. Floors and surfaces should be kept clean to
prevent accidents due to slips, rips and falls due to scattered broken glass. Since glass
production requires a large number of processes, the wastes must be removed from the
environment as they are in a large number of quantities.
4 Conclusion
In this paper, traditional Risk Matrix based Fuzzy MCDM method is introduced for
evaluating hazards encountered during glass manufacturing. Accordingly, the weights
of Risk Matrix parameters are obtained using Fuzzy evaluation. Fuzzy sets is used to
avoid limitation of a crisp values in risk assessment approach and inconsistency in
decision making process. Fuzzy TOPSIS is used to rank hazards encountered in glass
1344 S. Seker
industry. The results are satisfactory for glass manufacturing processes which have
many hazardous situations. The method can be applied for other industries such as
maritime, construction etc. with different fuzzy extensions.
References
1. Abdel-Rasoul, G.M., Al-Batanony, M.A., Abu-Salem, M.E., Taha, A.A., Faten, U.: Some
health disorders among workers in a glass factory. Occup. Med. Health Aff. 1, 2 (2013)
2. John, H., Hooper, P: Glass Waste, Waste, A Handbook for Management, pp. 151–165
(2011)
3. Gul, M., Guneri, A.F.: A fuzzy multi criteria risk assessment based on decision matrix
technique: a case study for aluminum industry. J. Loss Prevent. Process Industrie 40, 89–100
(2016)
4. Hwang, C.L., Yoon, K.: Multiple Attribute Decision Making: Methods and Applications, A
State-of-the-Art Survey. Springer, New York (1981)
5. Ji, Y., Huang, G.H., Sun, W.: Risk assessment of hydropower stations through an integrated
fuzzy entropy-weight multiple criteria decision-making method: a case study of the Xiangxi
River. Expert Syst. Appl. 42(12), 5380–5389 (2015)
6. Tepe, S., Kaya, İ.: A fuzzy-based risk assessment model for evaluations of hazards with a
real-case study. Hum. Ecol. Risk Assess. Int. J. 26(2), 512–537 (2020)
7. Mahdevari, S., Shahriar, K., Esfahanipour, A.: Human health and safety risks management
in underground coal mines using fuzzy TOPSIS. Sci. Total Environ. 488, 85–99 (2014)
8. Hu, A.H., Hsu, C., Kuo, T., Wu, W.C.: Risk evaluation of green components to hazardous
substance using FMEA and FAHP. Expert Syst. Appl. 36(3), 7142–7714 (2009)
9. Main, B.W.: Risk assessment: a review of the fundamental principles. Prof. Saf. 49(12), 37–
47 (2012)
11. Ghorabaee, M.K., Amiri, M., Zavadskas, E.K., Hooshmand, R., Antuchevičienė, J.: Fuzzy
extension of the CODAS method for multi-criteria market segment evaluation. J. Bus. Econ.
Manag. 18(1), 1–19 (2017)
12. Chen, C.T., Lin, C.T., Huang, S.F.: A fuzzy approach for supplier evaluation and selection in
supply chain management. Int. J. Prod. Econ. 102(2), 289–301 (2006)
A New Risk Assessment
Approach for Occupational Health
and Safety Applications
Selcuk Cebi(&)

Yildiz Technical University, Besiktas, Istanbul, Turkey
[email protected]
Abstract. Risk assessment is a critical issue in terms of Occupational Health

and Safety around the world. The major responsibility of the employers is to
protect their employees from the possible hazardous in the working environment.
For this, risk assessment in the working environment should be executed as a
proactive policy to manage occupational risks. In general, the risk is obtained by
using severity, possibility, and frequency parameters in the industry. However, it
is impossible to measure risk severity by quantitative tools. Besides, sometimes,
possibility and frequency have to be evaluated by using subjective evaluation
since there are not any statistical data for these parameters in the working
environment. Furthermore, although there are various outcomes of any risk, one
of the most possible outcomes are considered during the evaluation. However,
the effects of other consequences of any risk should be considered. For this, in
this study, a new approach for the risk assessment has been proposed to literature
to overcome these difficulties. In the proposed approach, the fuzzy inference
system and Multi fuzzy sets have been integrated to consider the difficulties given
above. The proposed approach is applied to public transportation.
Keywords: Risk assessment Multi fuzzy sets Fuzzy inference

Occupational health and safety
1 Introduction
One of the most important proactive measures that should be applied in order to
provide the health and safety of employees in the working environment is risk man-
agement. This proactive procedure is vital for companies to protect their employees and
services for various potential risks before they occur since it provides a set of proce-
dures to avoid potential hazards, to minimize their impacts, and to cope with the
undesired results. In terms of occupational health and safety, risk management pro-
cedure consists of four main phases. These are (i) Identification, (ii) Assessment,
(iii) Control, and (iv) Monitor and Review [1]. The identification phase includes
identification of hazards and then identifications of risk or risks caused by hazard
sources. The assessment phase involves the definition of risk analysis method, col-
lection data for the related risk analysis, and finally scoring the identified risks. The
assessment phase involves the definition of risk analysis method, collection data for the
1346 S. Cebi
related risk analysis method, and finally scoring the identified risks. Usually, the risk
analysis method is based on the prediction of probability and the prediction of severity
when any potential risk occurred. The control phase is based on the planning of control
measures and eliminating the undesired effects of analyzed risks. The control measures
are planned to apply from the highest-ranked risk to the lowest one. There are four
strategies for this phase [1]; (i) Risk avoidance (elimination of the risk source), (ii) Risk
sharing/transferring (shifting of risk from one section/company to another), (iii) Risk
reduction (applying engineering measurements at hazard source, applying collective
measures protect and organizational measurements, and at last providing personal
protective equipment) (iv) Risk acceptance (not taking any actions to the risk). Finally,
the monitor and review phase are to observe whether the determined control measures
are fit or not and recover the process. Risk management process is illustrated in Fig. 1.
Fig. 1. Risk management process
The success of the risk management process is based on experience and consul-
tation. Therefore, a risk assessment team is required for the implementation of the risk
management process. In the evaluation, the risk assessment team generally used a
linguistic scale for the risk likelihood and risk severity to analyze risk magnitude.
Sometimes, a pairwise comparison for risk likelihood and risk severity may be utilized
[2–4]. These evaluation procedures are categorical and therefore they include imprecise
information. Therefore, there are various studies considering these relations in Fuzzy
[5, 6]. Furthermore, the potential consequences of an accident may be various such as
simple injuries, loss of working days, loss of limbs, occupational disease, and death [7].
The occurrence likelihood of risk may also have a wide range of scales. For instance,
some accidents may not occur although it has an 80% possibility. Moreover, there are
several occupational health and safety specialists in the risk assessment team who
presents his/her prediction for a risk. Another problem is the aggregation of their
judgments on the risks. In this paper, a new approach for occupational risk analysis will
be proposed to cope with the obstacles given above.
A New Risk Assessment Approach for Occupational Health and Safety Applications 1347
The rest of this paper is organized as follows. A literature review is given in Sect. 2.
The proposed risk evaluation approach is given in Sect. 3. Section 4 provides an
illustrative example of the proposed approach. Finally, concluding remarks are given in
Sect. 5.
2 Literature Review
In the literature, various new methods or extensions of the current risk assessment
methods were introduced. For instance, Uzun and Cebi [8] proposed a new perspective
of view including Fuzzy Set Theory and Kano Model to analyze the perceived effi-
ciency of protective and preventive occupational health and safety measures imple-
mented in the construction sector. Kosztyan et al. [9] proposed a multi-level and
flexible risk assessment approach to provide customized risk evaluation to allow
integrating successfully the elements of risk evaluation and the method was applied to a
manufacturing company producing an electric motor. Yazdi et al. [10] proposed a
novel integrated multiple criteria decision-making approach to improve the efficiency
of conventional FMEA. Mete et al. [11] developed a decision support system based
including Pythagorean fuzzy VIKOR and applied it to the assembly of pipeline auger
process to illustrate the effectiveness of the proposed method. Dabbagh and Yousefi
[12] proposed a hybrid approach including the fuzzy cognitive map (FCM), multi-
objective optimization on the basis of ratio analysis (MOORA) and failure mode and
effect analysis (FMEA) to overcome the drawbacks in the conventional FMEA while
calculating risk priority number (RPN). Haghighi and Torabi [13] utilized a fuzzy
multi-criteria decision-making method and a fuzzy inference system to cope with the
uncertainty in the risk assessment process. Karasan et al. [3] proposed a new risk
assessment method which is called Safety-Critical Effect Analysis including severity,
probability, frequency and detectability parameters. Ilbahar et al. [4] integrated
Pythagorean fuzzy sets and fuzzy Mamdani methods for the risk analysis. Yousefi et al.
[14] introduced a new risk assessment approach including data envelop analysis
(DEA) and FMEA methods. Panchal and Kumar [15] developed a fuzzy decision
support system including grey relation analysis (GRA) and FMEA approaches to Risk
analysis of the compressor house unit. Lolli et al. [16] proposed a group decision
support system including FMEA and PROMETHEE methods. The effectiveness of the
proposed method was illustrated by applying it to a case study. Rao et al. [17] proposed
a new integrated method including dynamic fault tree analysis (FTA) and Monte Carlo
simulation. The developed method has been applied to an electrical power supply
system of a nuclear power plant.
3 Proposed Method
As it is mentioned in the Introduction Section, the risk assessment is one of the main
phases of the risk management process. In this step, a new approach for the occupa-
tional risk assessment method will be proposed. The steps of the proposed method are
as follows:
1348 S. Cebi
Step 1. Identification: In this step, various definitions for the analysis such as risk
assessment team, the linguistic scales used for the evaluation, potential risks have been
conducted.
Step 2. Prediction Likelihood of the Potential Risk: Each occupational health and
safety specialist presents his/her judgments on the likelihood of each risk defined in
Step 1. For this definition, Multi-Fuzzy Set (FMS) is used. The FMS is defined by
Sebastian and Ramakrishnan [18] as follows:
Definition: “Let X be a nonempty set, N the set of all natural numbers and
fLi : i 2 N g a family of complete lattices. A multi-fuzzy set A in X is a set of ordered
sequences:
A ¼ fhx; l1 ð xÞ; l2 ð xÞ; . . .li ð xÞ; . . .i : x 2 X g; ð1Þ
where li 2 LXi ; for i 2 N and the function lA= ⟨l1, l2,…⟩ is called a multi-member-
ship function of multi-fuzzy set A.”
Each occupational health and safety (OHS) specialist present his opinion on like-
lihood (RL) as follows;

RLm
i ¼ x; lm m m
1 ð xÞ; l2 ð xÞ; . . .li ð xÞ; . . . : x 2 X ; ð2Þ
where RLi, m and li represents the likelihood occurrence of the risk, the occupational
health and safety specialist, and the membership function value of the term of the used
linguistic scale, respectively. The union operation of fuzzy multi sets firstly discussed
by Yager [19] is used for the aggregation of the OHS specialists’ predictions as follow:
RLi ¼ RL1i [ RL2i [ RL2i [ . . . [ RLm

i ð3Þ
1
RLi ¼ x; l1 ð xÞ _ l21 ð xÞ _ . . . _ lm m
1 ð xÞ; l2 ð xÞ _ l2 ð xÞ _ . . . _ l2 ð xÞ; . . . : x 2 X
1 2
ð4Þ
Step 3. Prediction Severity of the Potential Risk: Similar to Step 2, Each occupational
health and safety specialist presents his/her judgements on the potential severity (S) of
each risk defined in Step 1.
m
RSm
i ¼ x; l1 ð xÞ; lm m
2 ð xÞ; . . .li ð xÞ; . . . : x 2 X ; ð5Þ
where Si, m and li represents the potential severity of the risk, the occupational
health and safety specialist, and the membership function value of the term of the used
linguistic scale, respectively. The union operation is used for the aggregation of the
OHS specialists’ predictions on potential severity as follow:
RSi ¼ RS1i [ RS2i [ RS2i [ . . . [ RSm

i ð6Þ
1
RLi ¼ x; l1 ð xÞ _ l21 ð xÞ _ . . . _ lm m
1 ð xÞ; l2 ð xÞ _ l2 ð xÞ _ . . . _ l2 ð xÞ; . . . : x 2 X
1 2
ð7Þ
Step 4. Calculation Risk Magnitude of the Potential Risk: In this step, risk magnitude
is calculated by using Fuzzy Mamdani [20]. The fuzzy Mamdani is widely used while
calculating risk magnitude [3–6] since it provides an effective tool to cope with
imprecise and vague information [21]. In this paper, it is the first time, the application
steps of Mamdani Fuzzy Inference technique are interpreted by using Multi-Fuzzy
Sets.
The fuzzy inference system requires a rule base defined by experts including
several rules relationships among risk magnitude (RM), risk severity (RS), and risk
likelihood (RL) parameters. To illustrate if-then rule type for fuzzy multi sets, Eq. 8 is
given.

Rk :IF RL is lkRL1 ð xÞ; lkRL2 ð xÞ; . . . lkRLi ð xÞ ; RS is lkRS1 ð xÞ; lkRS2 ð xÞ; . . . lkRSi ð xÞ ;

THEN RM is lkRM1 ð xÞ; lkRM2 ð xÞ; . . . lkRMi ð xÞ ; k ¼ 1; 2; ::; K ð8Þ
where lkRLi ; lkRSi ; and lkRMi presents membership value of multi fuzzy sets
RLi; RSi; and RMi, respectively. By using max-min operation (Eq. 9), the value of
RM is obtained.

lkRMi ð yÞ: _Kk¼1 lkRL1 ð xÞ; lkRL2 ð xÞ; . . .lkRLi ð xÞ ^ RS is lkRS1 ð xÞ; lkRS2 ð xÞ; . . .lkRSi ð xÞ
ð9Þ
where y 2 Y; x1 2 X1 ; and x2 2 X2 represents universe of RM; RL; and RS: Since

the obtained output from fuzzy inference system is a fuzzy set, it is required to
defuzzify output into a crisp value. For the defuzzification process, center-average
method given by Eq. 10 is used.
P
z i l ð yÞ
RM ¼ Pi¼1 RM ð10Þ
i¼1 lRM ð yÞ
4 Application
In this application, the method proposed in the study will be applied to the analysis of
the occupational risks faced by public bus drivers. Public bus drivers drive a fixed-route
transit bus and they face various occupational risks. The steps of the application are as
follows:
1350 S. Cebi
For the evaluation process, a risk assessment team consists of 3 occupational health
and safety specialists (OHS-Si) is created. In the evaluation, the linguistic scales given
in Table 1 [5] and Table 2 are selected for RS and RL, respectively.
Table 1. Linguistic scale used in the evaluation of RS

Type Risk severity Linguistic term
RS No loss of working time Very Low (VL)
No loss of working days Low (L)
Loss of working days Medium (M)
Loss of working weeks High (H)
Permanent unfitness/Occupational disease/Death/ Very High (VH)
Table 2. Linguistic scale used in the evaluation of RL

Type Risk likelihood Linguistic term
RL Rare (Once ever) Very Low (VL)
Seldom (Once or twice in a month) Low (L)
Several (Once or twice in a week) Medium (M)
Usually (Once or twice in a day) High (H)
Always Very High (VH)
Then, occupational risks are defined by the team. Actually, these risks cause long-
term health effects of public bus drivers. The main risk can be classified as follows:
• R1. Exposure to physical work factors: Public bus drivers are constantly exposed to
vibrations, noise, vapors, and fumes including sulfur dioxide, carbon monoxide,
nitrous oxides, and diesel particulates, etc.
• R2. Exposure to violence: Conflicting with customers due to their insecure behavior
causes an intentional use of power, threatened or action against public bus drivers.
• R3. Wrong sitting positions: Public bus drivers have to sit for a long period on the
seat of the bus; and consequently, sitting in a wrong sitting position causes joint and
muscle pain.
• R4. Stress: Shift work, traffic, noise, and some customers’ insecure behaviors cause
stress on public bus drivers (Table 3).
The likelihood and severity of the potential risks given above are predicted by the
team as given in Table 3. Then, co-decision matrix is constructed by using Eq. 4 and
Eq. 7 (Table 4)
Then, the risk magnitude is obtained by using the fuzzy inference system. The
fuzzy inference system requires a rule base including several rules relationships among
risk magnitude (RM), risk severity (RS), and risk likelihood (RL) parameters. These
rules are defined by experts as given in Table 5.
Table 3. Risk Likelihood and risk severity values

RL RS
VL L M H VH VL L M H VH
R1 OHS_S1 0.2 0.3 0.5 0.4 0.3 0.3
OHS_S2 0.1 0.5 0.4 0.2 0.7 0.1
OHS_S3 0.3 0.7 0.1 0.5 0.4
R2 OHS_S1 0.3 0.4 0.3 0.3 0.5 0.2
OHS_S2 0.1 0.2 0.7 0.5 0.5
OHS_S3 0.4 0.5 0.1 0.7 0.2 0.1
R3 OHS_S1 0.4 0.5 0.1 0.3 0.7
OHS_S2 0.2 0.6 0.3 0.1 0.4 0.5
OHS_S3 0.3 0.4 0.3 0.2 0.2 0.6
R4 OHS_S1 0.6 0.3 0.2 0.3 0.5 0.2
OHS_S2 0.3 0.7 0.2 0.7 0.1
OHS_S3 0.4 0.5 0.1 0.4 0.3 0.3
Table 4. The aggregated values

RL RS
VL L M H VH VL L M H VH
R1 0.3 0.7 0.5 0.4 0.7 0.4
R2 0.3 0.4 0.7 0.1 0.7 0.5 0.2
R3 0.4 0.6 0.3 0.2 0.4 0.7
R4 0.6 0.7 0.2 0.4 0.7 0.5 0.2
Table 5. The rule bases
Likelihood of occurrence (RL)

VL L M H VH N: Negligible
VL N N N N N Mi: Minor
Severity (RS)
L N N N Mi Mi Ma: Major
M N N Mi Ma Ma C: Catastrophic
H N Mi Ma Ma C
VH N Mi Ma C C
1352 S. Cebi
Risk magnitudes are obtained by using Eq. 8–9 and the rule based in Table 5 as
follows:

0 0:3 0:7 0:5 0:7 0:2 0:1 0
R1 ¼ ; ; ; ; R2 ¼ ; ; ;
N Mi Ma C N Mi Ma C

0:2 0:4 0:6 0:3 0:6 0:7 0:5 0:2
R3 ¼ ; ; ; ; R4 ¼ ; ; ;
N Mi Ma C N Mi Ma C
The risk values are defuzzified as R1 (9.4), R3 (8.0), R4 (7.9), and R2 (2.2) by using
Eq. 10 and the linguistic scale given in Fig. 2 [21].
Fig. 2. Rectangular fuzzy numbers for risk magnitudes
5 Conclusion
Risk management is a proactive method to prevent employees from working accidents.

The main objective of this paper is to propose a new approach to literature. The
proposed approach includes a fuzzy inference system and multi fuzzy sets. Then, the
method is applied to the analysis of occupational risks faced by public bus drivers.
According to the outcomes of the application, “exposure to physical work factors” is
found as the riskiest factor. The proposed approach provides the following advantages:
• The method considers all of the potential consequences of an accident such as
simple injuries, loss of working days, loss of limbs, occupational disease, and death,
• The method considers all possibilities of the likelihood of risk occurrence.
• The method considers imprecise information in the categorical evaluation procedure
of occupational health and safety studies.
• The risk magnitude is obtained by using a rule-based in parallel to the human
thinking process.
In this study, each risk is evaluated as an independent. In fact, each risky factor may
affect another one. Therefore, for further study, the proposed approach may be
extended to include the interactions among the risk factors. Furthermore, the proposed
method may be extended to intuitionistic fuzzy sets.
References
1. ISO 45001 Occupational Health and Safety Standards
2. Merrick, J.R.W., Van Dorp, J.R., Singh, A.: Analysis of correlated expert judgments from
extended pairwise comparisons. Decis. Anal. 2(1), 17–29 (2005)
3. Karasan, A., Ilbahar, E., Cebi, S., Kahraman, C.: A new risk assessment approach: safety
and critical effect analysis (SCEA) and its extension with pythagorean fuzzy sets. Saf. Sci.
108, 173–187 (2018)
4. Ilbahar, E., Karasan, A., Cebi, S., Kahraman, C.: A novel approach to risk assessment for
occupational health and safety using Pythagorean fuzzy AHP & fuzzy inference system. Saf.
Sci. 103, 124–136 (2018)
5. Acuner, O., Cebi, S.: An effective risk-preventive model proposal for occupational accidents
at shipyards. Brodogradnja 67(1), 67–84 (2016)
6. Cebi, S., Akyuz, E., Sahin, Y.: Developing web based decision support system for
evaluation occupational risks at shipyards. Brodogradnja 68(1), 17–30 (2017)
7. Cinar, U., Cebi, S.: A hybrid risk assessment method for mining sector based on QFD, Fuzzy
Logic and AHP. In: International Conference on Intelligent and Fuzzy Systems, pp. 1198–
1207 (2019)
8. Uzun, M., Cebi, S.: A novel approach for classification of occupational health and safety
measures based on their effectiveness by using fuzzy kano model. J. Intell. Fuzzy Syst. 38
(1), 589–600 (2020). https://doi.org/10.3233/JIFS-179432
9. Kosztyan, Z.T., Csizmadia, T., Kovacs, Z., Mihalcz, I.: Total risk evaluation framework. Int.
J. Qual. Reliab. Manage. (2020). https://doi.org/10.1108/IJQRM-05-2019-0167
10. Yazdi, M., Nedjati, A., Zarei, E., Abbassi, R.: A reliable risk analysis approach using an
extension of best-worst method based on democratic-autocratic decision-making style.
J. Clean. Prod. 256, 120418 (2020)
11. Mete, S., Serin, F., Oz, N.E., Gul, M.: A decision-support system based on Pythagorean
fuzzy VIKOR for occupational risk assessment of a natural gas pipeline construction. J. Nat.
Gas Sci. Eng. 71, 102979 (2019)
12. Dabbagh, R., Yousefi, S.: A hybrid decision-making approach based on FCM and MOORA
for occupational health and safety risk analysis. J. Saf. Res. 71, 111–123 (2019)
13. Haghighi, M.S., Torabi, S.A.: Business continuity-inspired fuzzy risk assessment framework
for hospital information systems. Enterp. Inf. Syst. (2019). https://doi.org/10.1080/
17517575.2019.1686657
14. Yousefi, S., Alizadeh, A., Hayati, J., Baghery, M.: HSE risk prioritization using robust DEA-
FMEA approach with undesirable outputs: a study of automotive parts industry in Iran. Saf.
Sci. 102, 144–158 (2018)
15. Lolli, F., Ishizaka, A., Gamberini, R., Rimini, B., Messori, M.: FlowSort-GDSS – a novel
group multi-criteria decision support system for sorting problems with application to FMEA.
Expert Syst. Appl. 42(17–18), 6342–6349 (2015)
16. Panchal, D., Kumar, D.: Risk analysis of compressor house unit in thermal power plant using
integrated fuzzy FMEA and GRA approach. Int. J. Ind. Syst. Eng. 25(2), 228–250 (2017)
1354 S. Cebi
17. Rao, D.K., Gopika, V., SanyasiRao, V.V.S., Kushwaha, H.S., Verma, A.K., Srividya, A.:
Dynamic fault tree analysis using Monte Carlo simulation in probabilistic safety assessment.
Reliab. Eng. Syst. Saf. 94, 872–883 (2009)
18. Sebastian, S., Ramakrishnan, T.V.: Multi-fuzzy sets: an extension of fuzzy sets. Fuzzy Inf.
Eng. 1, 35–43 (2011)
19. Yager, R.R.: On the theory of bags. Int. J. General Syst. 13, 23–37 (1986)
20. Mamdani, E.H.: Application of fuzzy logic to approximate reasoning using linguistic
synthesis. IEEE Trans. Comp. 26, 1182–1191 (1977)
21. Zeng, J., An, M., Smith, N.J.: Application of a fuzzy based decision making methodology to
construction project risk assessment. Int. J. Project Manag. 25, 589–600 (2007)
Intuitionistic Fuzzy Group Decision Making
on Risk Assessment and Control Efficiency
of Accounting Information Systems
with TOPSIS
Aygülen Kayahan Karakul(&), Rozi Mizrahi, and Berna Tektaş

İzmir Kâtip Celebi University, Cigli, Izmir, Turkey
{aygulen.kayahan,rozi.mizrahi,
berna.tektas}@ikcu.edu.tr
Abstract. Accounting information systems (AIS) play a key role in risk

assessment and control in modern enterprise management. Risk management is
the process of planning, organizing, managing and controlling the necessary
activities to control unexpected losses that may occur in the enterprises. The
importance of risk assessment and control requires to use advanced fuzzy Multi
Criteria Decision Making techniques. The aim of this paper is to apply the
Intuitionistic Fuzzy TOPSIS method to evaluate the risk assessment and control
efficiency of AIS of firms in three different sector: production, trade and service
sectors. The risk factors are determined as financial risks, legal risks, reputation
risks, strategic risks, operational risks and external risks. The prioritization of
risks has been made by academic experts and evaluation of organizations’ AIS
has been made by risk-oriented audit professionals. As a result, the ranking of
firms according to their ability of coping with the risks in their AIS is obtained
as production firm, service firm and trade firm respectively.
Keywords: Intuitionistic Fuzzy TOPSIS Risk assessment and control

Accounting Information Systems (AIS)
1 Introduction
Enterprises have to establish strategies with flexibility for adaptation to ever-changing

conditions. Knowledge is one of the essential sources of strategy which allow com-
panies to be successful in competition as well as the effective use of information
systems [1]. Accounting information systems (AIS) is focus point for businesses and it
generates new data for other information systems. AIS play an important key role for
risk assessment and control in contemporary enterprise management. This importance
indicates to use advanced quantitative techniques to solve problems in AIS of enter-
prises. So fuzzy Multi Criteria Decision Making (MCDM) techniques will be helpful to
use Decision Making (DM) problems in this field by their ability to handle vagueness.
Intuitionistic fuzzy sets are widely used in MCDM problems. Bolturk, Gulbay and
Kahraman [2] used intuitionistic and neutrosophic fuzzy sets for facility location
1356 A. K. Karakul et al.
selection. Torun and Gördebil [3] used Intuitionistic Fuzzy TOPSIS and Pic-
ture Fuzzy TOPSIS to rank the alternatives in their research; Yalcin, Kılıc and Guler [4]
used Intuitionistic Fuzzy DEMATEL and Intuitionistic Fuzzy TOPSIS in the selection
of projects.
In this research the Intuitionistic Fuzzy TOPSIS method has been used to evaluate
the risk assessment and control efficiency of enterprises in their AIS. Intuitionistic
fuzzy TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution)
consider the hesitations of decision makers which is a natural part of process of
decision making. The research can contribute to literature since it serves as a sample of
using a fuzzy MCDM technique to solving a real life problem in risk assessment of
AIS.
The rest of paper is organized as follows: In Sect. 2. Theoretical information about
risk assessment and control is presented, In Sect. 3. Intuitionistic fuzzy TOPSIS
method is introduced. In Sect. 4. the application of a real life problem has been
implemented. The conclusion is given the last section.
2 Risk Assessment and Control
The common risk factors in literature are determined as financial risks, legal risks,
dignity risks, strategic risks, operational risks and external risks [5]. Financial risks can
be grouped as credit risk, interest rate risk, currency risk, market risk and liquidity risk.
Legal risks occur as a result of misinterpretation or unawareness of changes in laws.
Reputation risk means the decline of dignity of the enterprise and it effects the customer
loyalty, market share, and employee motivation of company. Strategic risks represents
the disharmony of enterprise to changing market conditions in the long term. Insuffi-
ciency of the working staff, technological deficiencies and organizational structure
problems causes the operational risks. External risks are the risks encountered due to
external factors and risks of natural disasters or cyber-attacks are in this group [6].
Risk management process contains the stages of identifying risks, developing risk
correction tools, and lastly assessment and control the risk [7]. Enterprises benefit from
AIS in the risk assessment and control process. In this context, the effective and
efficient use of AIS systems is of great importance in the risk management process. The
development of AIS programs to include the risk assessment and control function has
brought the “Risk Based Control” phenomenon to the business literature. Risk-oriented
auditing is a part of internal audit and is an intuitive approach that focuses on more
risky areas than low-risk areas. It is based on the principle of determining risk profiles
and allocating resources on this basis [5]. In order to identify and control the risks the
enterprises can be guided by before the risks occur.
Intuitionistic Fuzzy Group Decision Making on Risk Assessment 1357
3 Method
3.1 Intuitionistic Fuzzy Sets
Intuitionistic Fuzzy Sets are firstly introduced by Atanassov as the following definitions
and theorems [8]: Let, X is a non-empty set, then intuitionistic fuzzy set A is defined as
in the Eq. (1) as the membership and non-membership functions lA ð xÞ and mA ðxÞ
respectively.
A ¼ fhx; lA ð xÞ; mA ðxÞijx 2 Xg ð1Þ
The sum of membership and non-membership degree does not need to be equal to
1, while it lies between 0 and 1 as shown in the Eq. (2)
0 lA ð xÞ þ mA ðxÞ 1 ð2Þ
The degree of hesitancy pA defined as given in the Eq. (3)
pA ¼ 1 lA ð xÞ mA ðxÞ ð3Þ
3.2 Intuitionistic Fuzzy TOPSIS

TOPSIS which is introduced by Chen and Hwang [9] and Hwang and Yoon [10] is
based on the ranking of alternatives which is nearest to positive ideal solution and is
farthest from negative ideal solution simultaneously.
Intuitionistic fuzzy TOPSIS method allows to consider the importance degree of
decisions of different DMs who has different degree of expertise on the subject. Table 1
shows the Linguistic terms and their corresponding Intuitionistic Fuzzy Numbers (IFNs)
for rating the DMs significance of Criteria and rating of the alternatives. Many DM
problems has been solved by using this method.
Table 1. Linguistic terms and IFNs for rating the DMs, significance of criteria and the
alternatives
For rating of DMs and criteria significance For rating the alternatives
Very Important (VI) (0.90, 0.10, 0.00) Extremely High (EH) (0.75, 0.10, 0.15)
Important (I) (0.75, 0.20, 0.05) High (H) (0.60, 0.25,0.15)
Medium (M) (0.50, 0.45, 0.05) Medium (M) (0.50, 0.50, 0.00)
Unimportant (UI) (0.35, 0.60, 0,05) Low (L) (0.25, 0.60, 0.15)
Very Unimportant (VU) (0.10, 0.90, 0.00) Extremely Low (EL) (0.10, 0.75, 0.15)
Source: Yalcın, Kılıc and Gular [4]
The importance degree of kth DM is given in the Eq. (4) where there are l number
of DMs.

lk þ pk ðl lþk mk Þ Xl
kk ¼ P ; k ¼1 ð4Þ
k
l lk k¼1 k
k¼1 lk þ p k ð l þ mk Þ
k
The evaluations of DMs are calculated by the Eqs. (5) and (6).

ð1Þ ð2Þ ðlÞ ð1Þ ð2Þ ð3Þ ðlÞ
wj ¼ IFWAk wj ; wj ; . . .; wj ¼ k1 wj k2 wj k3 wj . . . kl wj ð5Þ
Yk¼1 k Yk¼1 kk Yk¼1 k Yk¼1 k
ðk Þ k ðk Þ ðk Þ k ðk Þ k
wj ¼ 1 l 1 lj ; l mj ; l 1 lj l 1 mj
ð6Þ

W ¼ w1 ; w2 ; w3 ; . . .; wj ð7Þ
The scores of DMs are represented as matrix by the Eq. (8) and Eq. (9).

ð1Þ ð2Þ ðlÞ ð1Þ ð2Þ ð3Þ ðlÞ
rij ¼ IFWAk rij ; rij ; . . .; rij ¼ k1 rij k2 rij k3 rij . . . kl rij ð8Þ
Yl Yl kk Yl Yl
ðk Þ kk ðk Þ ðk Þ k k ðkÞ kk
rij ¼ 1 k¼1
1 lj ; k¼1 mj ; k¼1 1 lj k¼1
1 mj
ð9Þ
Aggregated weighted intuitionistic fuzzy decision matrix offered by Xu [11] can be

defined by the Eq. (10) and Eq. (11)

RW ¼ x; lAi ð xÞ lW ð xÞ; mAi ð xÞ þ mW ð xÞ mAi ðxÞ mW ð xÞ x 2 X ð10Þ
pAi W ð xÞ ¼ 1 mAi ð xÞ mW ð xÞ lAi ð xÞ lW ð xÞ þ mAi ðxÞ mW ð xÞ ð11Þ
Intuitionistic fuzzy positive and negative ideal solutions A and A are obtained by
Eqs. (12)–(16) where J1 & J2 are benefit and cost criteria respectively.

A ¼ lA W xj ; mA W xj A ¼ lA W xj ; mA W xj ð12Þ

lA W xj ¼ maxi lAi W xj j 2 J1 ; mini lAi W xj j 2 J2 ð13Þ

mA W xj ¼ mini mAi W xj j 2 J1 ; maxi mAi W xj j 2 J2 ð14Þ

lA W xj ¼ mini lAi W xj j 2 J1 ; maxi lAi W xj j 2 J2 ð15Þ

mA W xj ¼ maxi mAi W xj j 2 J1 ; mini mAi W xj j 2 J2 ð16Þ
For the calculation of distances to positive and negative ideal solutions, the nor-
malized Euclidian distance offered by Schmidt and Kacprzyk [12] with Eqs in (17)–
(18) as separation measures.
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
1 Xn h 2 2 2 i
S ¼ lAi W xj lA W xj þ mAi W xj mA W xj þ pAi W xj pA W xj
2n j¼1
ð17Þ
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
1 Xn h 2 2 2 i
S ¼ lAi W xj lA W xj þ mAi W xj mA W xj þ pAi W xj pA W xj
2n j¼1
ð18Þ
Relative closeness coefficient Ci is calculated by Eq. (19) which enables to rank
the alternatives in the last stage.
Si
Ci ¼ ; 0 Ci 1 ð19Þ
Si þ Si
4 Application
The real life problem of ranking three enterprises, operating in three different sectors,
according to their control ability of risks in their AIS is handled in this section. Three
DMs, (DM1, DM2 & DM3) who rated the weights of criteria, are academic experts of
titles professor, associate professor and assistant professor in the field of Accounting
and Finance. Three other experts (DM4, DM5 & DM6) who rate the alternatives are
working in the auditing of firms with different seniorities. The alternatives of the
research are three firms which are operating in the production (A1), trade (A2) and
service sector (A3). The importance degree of DMs are handled as linguistic variables,
very important, important and medium and their weights are calculated as 40%; 36%
and 24% respectively by using Eq. (4). Ratings of DMs on criteria and the weights of
those criteria obtained by Eq. (5) and Eq. (6) has been given in Table 2.
Table 2. Ratings of academic DMs on criteria and calculated weights of criteria

DM1 (0.40) DM2 (0.36) DM3 (0.24) Weights of criteria
C1: Financial risks VI VI I (0.88, 0.12, 0.01)
C2: Legal risks I I VI (0.80, 0.17, 0.03)
C3: Reputation risks M M UI (0.47, 0.48, 0.05)
C4: Strategic risks M I UI (0.59, 0.36, 0.05)
C5: Operational risks UI VI I (0.74, 0.24, 0.02)
C6: External risks VU VI M (0.65, 0.35, 0.01)
The ratings of Accountant Profession DMs on alternative firms are given in the
Table 3.
Table 3. Ratings of accountant profession DMs on alteratives

Firms Criteria
C1 C2 C3 C4 C5 C6
DM4 (0.40) A1: Production sector H EH M H H H
A2: Trade sector H M L H L M
A3: Service sector M H H H EH M
DM5 (0.36) A1: Production sector H H H H EH M
A2: Trade sector M M L M L M
A3: Service sector L M H H H H
DM6 (0.24) A1: Production sector EH H H M EH H
A2: Trade sector H H M L H M
A3: Service sector H H H EH H H
The opinions of all DMs are aggregated by using the IFWA operator defined in the
Eq. (8) and Eq. (9). In the next step the criteria weights are applied to Aggregated
Decision Matrix by using Eqs. (10) and (11). The result matrix is given in Table 4.
Table 4. Aggregated weighted decision matrix

C1 C2 C3 C4 C5 C6
A1 0.72 0.26 0.03 0.66 0.3 0.04 0.31 0.63 0.06 0.4 0.53 0.07 0.63 0.34 0.03 0.46 0.5 0.04
A2 0.53 0.41 0.05 0.56 0.37 0.07 0.19 0.76 0.05 0.34 0.59 0.07 0.4 0.55 0.05 0.32 0.64 0.04
A3 0.51 0.44 0.05 0.56 0.37 0.07 0.31 0.63 0.06 0.48 0.46 0.06 0.61 0.36 0.03 0.43 0.53 0.04
Positive and negative ideal solution sets are determined as in the Table 5 by using
Eqs. (12)–(16).
Table 5. Positive and negative ideal solution sets

C1 C2 C3 C4 C5 C6
A 0.72 0.26 0.03 0.66 0.3 0.04 0.31 0.63 0.06 0.48 0.46 0.06 0.63 0.34 0.03 0.46 0.5
A 0.51 0.44 0.05 0.56 0.37 0.07 0.19 0.76 0.05 0.34 0.59 0.07 0.4 0.55 0.05 0.32 0.64
Lastly the alternatives are ranked after the calculation of distances to positive ideal
and negative ideal solutions. The separation measures and closeness coefficients in
Eqs. (17)–(19). The results are given in the Table 6.
Table 6. Separation measures, closeness coefficients and rankings of alternatives

S S C i Ranking
A1 in production sector 0.03 0.14 0.82 1
A2 in trade sector 0.14 0.01 0.05 3
A3: service sector 0.08 0.11 0.58 2
According to Table 6, by calculating the Separation Measures and Closeness

Coefficients the ranking of firms is obtained as A1-firm in production sector; A3-firm in
Service Sector and A2-firm in Trade Sector respectively.
5 Conclusion
The contribution of this research to literature will be its guidance to implementers and
researchers to use fuzzy MCDM techniques for their financial or auditing decisions
about risk assessment of AIS of the enterprises. The advantage of valuing the different
experts’ experiences with different weights of the Intuitionistic fuzzy TOPSIS method
has been successfully used as well as the positive effect of fuzzy environment of
method to decision making process. As a result of the research the risk valuation and
control efficiency of the AIS of the production enterprise has been found higher than
the enterprises in other sectors, which shows that those enterprises are more powerful
and competitive in the market according to others. This success is about their ability of
technological intensive and qualified labor investments of the production sector in
Turkey which is a prerequisite for coping with the risks of AIS. The varied and
diversified risks that businesses face in the capital markets highlights the importance of
risk-based auditing. In this context, the efficient use of AIS will increase the effec-
tiveness of risk management and risk-oriented audit activities, and will help businesses
survive in competitive market conditions. In this sense it is expected that this research
will enlighten to enterprises with its way of using advanced qualitative DM techniques.
For further research it is recommended to apply extension of Intuitionistic Fuzzy
Sets to Pythagorean fuzzy sets and Neutrosophic sets as well as to apply all other
MCDM methods.
Acknowledgements. Authors would like to deep gratitude to academic experts as also to

Accountant Professionals for their contributions.
References
1. Mizrahi, R.: Accounting and business information systems in SMEs. In: Çelik, Ş. (ed.)
SMEs in Turkey Managerial Perspective IBSA. Studies in Management and Innovation, vol.
5, pp. 118–138. Kellner Publishing House, Bremen (2020)
2. Bolturk, E., Gülbay, M., Kahraman, C.: Location selection by intuitionistic fuzzy and
neutrosophic aggregation operators. In: Kahraman, C., Çebi, S., Onar, S.Ç., Öztayşi, B.,
Tolga, A.Ç., Sarı, İ.U. (eds.) Intelligent and Fuzzy Techniques in Big Data Analytics and
Decision Making, pp. 527–536. Springer, Cham (2020)
3. Torun, H., Gördebil, M.: Multi criteria decision making based on TOPSIS method with
extended fuzzy sets. In: Kahraman, C., Çebi, S., Onar, S.Ç., Öztayşi, B., Tolga, A.Ç., Sarı, İ.
U. (eds.) Intelligent and Fuzzy Techniques in Big Data Analytics and Decision Making,
4. Yalçın, A.S., Kiliç, H.S., Guler, E.: Research and development project selection via IF-
DEMATEL and IF-TOPSIS. In: Kahraman, C., Çebi, S., Onar, S.Ç., Öztayşi, B., Tolga, A.
Ç., Sarı, İ.U. (eds.) Intelligent and Fuzzy Techniques in Big Data Analytics and Decision
Making, pp. 625–633. Springer, Cham (2020)
5. Usul, H., Mizrahi, R.: Risk Odaklı Denetim. Detay Yayıncılık, Ankara (2016)
6. Thomas, E.M.: The Seven Step Process to Risk Based Auditing, vol. 2, no. 4 (2003)
7. Emhan, A.: Risk Yönetim Süreci ve Risk Yönetmekte Kullanılan Teknikler, Atatürk U.
IIBF, vol. 23, no. 3 (2009)
8. Atanassov, K.T.: Intuitionistic Fuzzy Sets: Theory and Applications. Studies in Fuzziness
and Soft Computing. Springer, Heidelberg (1999)
9. Chen, S.J., Hwang, C.L.: Fuzzy multiple attribute decision making. Methods and
applications. In: Lecture Notes in Economics and Mathematical Systems, vol. 375.
pp. 298–486 (1992)
10. Hwang, C.L., Young, J.L., Ting, Y.L.: A new approach for multiple objective decision
making. Comput. Oper. Res. 20(8), 889–899 (1993)
11. Xu, Z.: Intuitionistic fuzzy aggregation operators. IEEE Trans. Fuzzy Syst. 15(6), 1179–
1187 (2007)
12. Szmidt, E., Kacprzyk, J.: Distances between intuitionistic fuzzy sets. Fuzzy Sets Syst. 114,
505–518 (2000)
Evaluation of Occupational Safety Risk
in Underground Mining Using Fuzzy Bayesian
Network
Fatma Yaşlı1(&) and Bersam Bolat2

1
Anadolu University, Eskisehir, Turkey
[email protected]
2
Istanbul Technical University, Istanbul, Turkey
Abstract. Underground mining includes multi processes that are performed in

labor-intensive way. This study analyses the occupational safety issues sur-
rounding these processes using a fuzzy set approach and Bayesian Network. In
order to reduce safety risk, it is necessary to specify, investigate, and measure
the risk comprehensively. In the absence of recorded data about the events, fault
tree approach has played an effective role to reveal the causes of the events and
Fuzzy Bayesian Network provides a causative and probabilistic approach for the
events with their importance. The analysis is performed on 13 undesired events
that stemmed from equipment injury to explosion occurred during the processes
of the mining. Results show that the main factors of occupational safety are not
employee error, on the contrary, the managerial approach to the education,
planning, and inspection related to occupational health and safety. We believe
that this study could be helpful for evaluating the safety risk of the multi-process
systems comprehensively and proposing strategic planning for mitigating the
risks.
Keywords: Fuzzy Bayesian Network Occupational safety Risk analysis

Underground mining
1 Introduction
Including developed countries, the number of serious injuries and fatalities still remains
high in the mining industries despite substantial efforts the industry has put and the
technological development over the past twenty years [1–3]. In countries, particularly
where labor costs are low and creating less incentives to increase efficiency, it still
requires a large amount of labor force for mining activities [4]. Mining sector employees
are exposed to significant occupational safety risks all over the world due to these labor-
intensive processes. Dust diseases, material dropping on them, hitting an object, colli-
sion, or overexertion are also very common events related to occupational health and
safety (OHS) in under-ground mining [5–7] in addition to the collapsed and explosion
accidents which are widely considered within the coal-related literature [8–12].
In this study, undesired OHS events occurring in underground mining, which have
not been extensively discussed in the literature, are discussed from the risk analysis
1364 F. Yaşlı and B. Bolat
perspective. An underground chrome mine located in Turkey is considered for the

study. When the historical OHS events of the mine are inquired, it is seen that the data
contain do not contain certainty and detailed for the statistical risk analysis. Using the
Fuzzy Bayesian Network (FBN) is a very appropriate approach to perform risk analysis
in areas where historical data is not sufficient or easily accessible. Because Bayesian
Network (BN) allows causal analysis without reliable statistical data, while the fuzzy
set approach allows reaching the probability data related to the events practically.
This study firstly relates to present a methodology that can be used in any multi-
process system to evaluate the risk. After giving the details about it, through the case
study of an underground mining located in Turkey, it is demonstrated that the proposed
methodology provides a feasible method for accurately identifying and evaluating the
occupational safety risk. In the last section of the study, recommendations and future
suggestions are provided about the using BN and fuzzy approach in risk studies.
2 Methodology
The methodology firstly considers the qualitative configuration of the network which
reveals the undesired events with their causal relationships by using fault tree (FT) with
the BN approach and the quantitative configuration of the network which provides the
probabilistic data about the events by using the fuzzy approach. Then, it provides many
comprehensive inferences about the events using the Bayes Theory.
2.1 Construction of the BN for Multi-top Event Systems: Using Fault

Tree
It is necessary to determine the sub-processes and define the undesired events which may
harm the employees for constructing the BN. The causal structures related to the occur-
rence of the events can be elicited by considering and questioning each undesired event as
a top event within the scope of the FT. Specifying of the network’s top events which
represent the undesired events should be made in consensus with analysts and experts
during the network configuration, due to the fact that many events are handled in the
paper, unlike other BN studies. In FTA, the center is the top event [13] and all different
combinations of sub-events that are effective in the occurrence of the top event are
revealed. This study with the difference of literature, the causative analysis of the unde-
sired occupational events in a multi-process system is performed based on FTs and then all
proposed FTs have been integrated into a single BN. The conversion of the FT into the BN
is carried out through to the mapping relationship that both methods have [14].
2.2 Quantitative Structure of the BN: Using Fuzzy Approach

BN focuses on our thoughts on the precise events that we have preliminary information
and if there are other events related to this event, depending on the conditional
probabilities, our thoughts change in relation [15]. All nodes in the network are vari-
ables of the BN model [16]. Event probability values should be determined regarding
the causal relationships of the events as the network parameters. To determine the
Evaluation of Occupational Safety Risk in Underground Mining 1365
parameters, the fuzzy approach is used. It allows that the network probability values are
considered within a range using the linguistic expressions. Linguistic terms are rep-
resented in the form of fuzzy numbers. In this study, triangular fuzzy numbers (TFNs)
(l, m, u) are used for the elicitation of the probabilities due to they have a convenient
reference function processing and presents a quite easiness for algebraic operation [8].
In the uncertain environment of the risk analysis, fuzzy logic provides a quite rea-
sonable approach. It is based on the Fuzzy set theory which is introduced by Zadeh in
1965 [17].
The fact that the probabilities will be inferred by the experts reveals the importance
of the experts’ knowledge and experience levels both on the field and the concept of
probability. So, the judgments of a group of experts with different knowledge and
experience level are gathered for the analysis and they are rearranged according to each
one’s importance weight. Where P ~ si indicates the fuzzy numbers of the linguistic
expression of experts j for the probability of node i under the state s combinations, wi
denotes the importance of the expert and P ~ s
i presents the aggregated occurrence
probability value of the event i under related condition (see in Eq. 1).
s s s Xm
~ s
P i ¼ li ; mi ui ¼ ~ sij ðlsi ; msi ; usi Þ
wj P ð1Þ
j¼1
After aggregation the subjective possibility judgments of different experts as the

form of TFNs, the defuzzification process is provided to have one exact probability
value correspond to each fuzzy value. “Mean area method” is used for determining the
crisp probability of the nodes (see in Eq. 2).
i þ 2mi þ ui
ls s s
i ¼
Ps ð2Þ
4
2.3 Bayesian Inferences

BN modeling is expressed as a simple mathematical formula that calculates the con-
ditional and marginal probabilities of a random event [18]. Therefore, the Bayes
Theory provides a conditional investigation of the events. If “variable a” is occurred in
the network, the occurrence probability of the “variable b” which caused by “a”, is
shown as PðbjaÞ, and the determination formulation of the occurrence probability of
“b” is as follows (see in Eq. 3). This is also called forward inquiry for BN.
Pðb; aÞ
PðbjaÞ ¼ ð3Þ
PðaÞ
In BN, backward queries are also made to estimate the cause of an event whose
result is known. It is expressed as “if the result “b” has given in the network, the
probability that “a” may be caused” and it is calculated as in Eq. 4. It enables the
uncertainty in the relationship between the parent and the event to be reduced and the
characteristic associated with the occurrence of the event to be defined more

specifically.
Pða; bÞ
PðajbÞ ¼ ð4Þ
PðbÞ
The joint probability of the occurrence of “a” and “b” events, which are in a cause
and effect relationship with each other, is presented as P (a, b) or P (b, a). Due to the
equity of them in both two-way conditional probability inquiries, the following for-
mulation that explains the Bayes Theorem is reached (see in Eq. 5).
PðajbÞ PðbÞ
PðbjaÞ ¼ ð5Þ
PðaÞ
The relationship between forward and backward inquiries is explored by the

statistician ‘Thomas Bayes’ in the 18th century.
3 Safety Risk Evaluation of Underground Mining
In this section, a case study of the occupational safety risk among the processes of an
underground mine is addressed to illustrate the application of the proposed risk eval-
uation model. It has been observed that the accidents and injuries experienced by the
employees during the multi-processes of the underground mine have spread to a wide
range such as falling, physical strain, falling into the cage cavity, or lung disease. For
this reason, all of these accidents and injuries have been called as undesired events.
3.1 Construction of the BN

With the collaboration of the analyst and the four experts, subprocesses and the related
undesired events are revealed. It is determined that 13 undesired events occur during
mining processes and the causal structures of the events are revealed by using FT (see
Fig. 1). And all the FTs are converted to a BN (see Fig. 2).
3.2 Determining the Network Parameters Using Fuzzy Logic

For determining the network parameters, the linguistic probability statements presented
to the experts should provide sufficient assessment about the events. Preliminary work
is carried out with experts to understand their experience on probability and to establish
the appropriate probability scale. The TFNs, as described in Table 1, are proposed to
represent ambiguous data. On the other hand, Table 2 shows the information about the
participants of the study and their importance scores.
All the network variables are determined by the expert as the form of linguistic
expressions. Then aggregated TFNs and crisp probability values are calculated. The
probability values related to the node of “Lung-Dust Disease” are shown in Table 3.
Fig. 1. 13 fault trees which represent the causal structure of the undesired events.
Fig. 2. Proposed BN.
Prior probabilities of the network are the first trigger values required to determine
event probabilities within the scope of the analysis and they are given in Table 4. After
the prior probabilities and the all CPT values that show the causal relationships
between the events within the network are elicited and calculated according to the
fuzzy approach, Bayesian inference about the undesired events can be performed.
Table 1. Linguistic expressions and their corresponding fuzzy numbers.

Linguistic expression Triangular fuzzy number
Extremely rare (ER) (0, 0, 0.0001)
Rare (R) (0.0001, 0.001, 0.01)
Low (L) (0.01, 0.05, 0.20)
Slightly low (SL) (0.15, 0.30, 0.45)
Medium (M) (0.40, 0.50, 0.60)
Slightly high (SH) (0.55, 0.70, 0.85)
High (H) (0.80, 0.95, 0.999)
Almost always (AA) (0.99, 0.999, 0.9999)
Almost (A) (0.9999, 1, 1)
Table 2. Participated experts and their related importance scores.

Expert Profe. Experience Graduation Probability knowledge Imp. score
position
E1 Manager 15–20 Master High OHS knowledge 0.2982
E2 Chief 10–15 Bachelor High OHS know. and 0.2632
engineer risk an. exp.
E3 Engineer 10–15 Master High OHS knowledge 0.2456
E4 Sergenat >20 High Intermediate OHS 0.1930
major school knowledge
Table 3. Expert Judgments for the node of Lung-Dust disease.

Parent States of the parent nodes
nodes
Dust Yes No
system
Using PPE Yes No Yes No
Expert 1 (0.0001, 0.001, (0.15, 0.30, (0.15, 0.30, (0.55, 0.70,
0.01) 0.45) 0.45) 0.85)
Expert 2 (0.0001, 0.001, (0.15, 0.30, (0.40, 0.50, (0.80, 0.95,
0.01) 0.45) 0.60) 0.999)
Expert 3 (0.0001, 0.001, (0.55, 0.70, (0.40, 0.50, (0.80, 0.95,
0.01) 0.85) 0.60) 0.999)
Expert 4 (0.01, 0.05, 0.20) (0.55, 0.70, (0.55, 0.70, (0.99, 0.999,
0.85) 0.85) 0.9999)
Aggregated (0.00027, 0.01, (0.325, 0.475, (0.354, 0.479, (0.762, 0.885,
0.047) 0.625) 0.604) 0.955)
Probability 0.017 0.4754 0.4789 0.8717
Table 4. Fuzzy prior probabilities of the root causes.

Root causes States and prior probabilities
Geological structure Competent rock Faulted rock
(0.339, 0.539, 0.739) 0.538 (0.261, 0.461, 0.661) 0.462
OHS education Adequate Inadequate
(0.304, 0.504, 0.704) 0.503 (0.296, 0.496, 0.696) 0.497
Ground condition Straight Rough
(0.3, 0.5, 0.7) 0.5 (0.3, 0.5, 0.7) 0.5
OHS inspection Adequate Inadequate
(0.702, 0.819, 0.917) 0.813 (0.083, 0.183, 0.298) 0.187
Employee attention Adequate Inadequate
(0.339, 0.539, 0.739) 0.538 (0.261, 0.461, 0.661) 0.462
OHS planning Adequate Inadequate
(0.825, 0.89, 0.942) 0.886 (0.058, 0.11, 0.175) 0.114
3.3 Bayesian Inferences

According to the prior probabilities and CPTs determined by the experts, Bayesian
inferences related to the network are performed. Inferences include the forward and
backward inquiries through the network. The analysis is performed by Genie 2.4
software.
Forward Inquiries. Obtaining the occurrence probabilities of the top events in the
network are performed by the forward inquiries. The results are given in Fig. 3.
Fall of the cage cavity 0.0081

Entrapment due to the wagon 0.1525
Fall of materials on employee 0.1311
Vehicle's hit to employee 0.1451
Roolover of digger 0.0701
Roof falling 0.0974
Falling 0.3563
Overexterion 0.3563
Fall of scales 0.3719
Gas leakage- flooding 0.0089
Equipment injury 0.245
Exposure to explosion 0.0857
Lung-dust disease 0.2568
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Fig. 3. Occurrence probabilities of undesired events in BN.
Backward Inquiries. When new information included to the BN, the posterior
probabilities by updating the prior probabilities of the events are determined and the
events which are most likely to cause the undesired events are specified. The results
related to the posterior probabilities of the root causes are shown in Table 5, show the
causes that particular attention should be paid for undesired events. It can be seen easily
that the causes “OHS education” and “Employee attention” give more contribution to
the occurrence of the undesired events.
Table 5. Posterior probabilities of the root causes.

Undesired Faulted Inadequate Rough Inadequate Inadequate Inadequate
events rock OHS educ. work. OHS insp. empl. Attent. OHS plan.
cond.
Lung-dust 0.78 0.38
disease
Exposure to 0.94 0.41 0.48
explosion
Equipment 0.8 0.55 0.58
injury
Gas leakage- 0.83 0.68 0.47
flooding
Fall of scales 0.92 0.56 0.19 0.49
Overexertion 0.79 0.57 0.61
Falling 0.77 0.59 0.59
Roof falling 0.85 0.69 0.28
Rollover of 0.87 0.61 0.69
digger
Vehicle’s hit 0.82 0.56 0.6 0.13
to employee
Fall of 0.84 0.35 0.6
materials on
emp.
Entrapment 0.82 0.62
due to wagon
Fall of the 0.98 0.83
cage cavity
4 Conclusion and Results
The fuzzy approach presents a quite easiness to gather the data using linguistic
information from the experts for inferencing the reasoning relationships between the
events. Fuzzy logic handle with insufficient knowledge about the criticality of the
events to occur to each other. FT, BN, and fuzzy logic provide quiet useful methods
together for investigating the undesired events in any process. When FT and BN
present structural perspective to analyze the events qualitatively, on the other hand
fuzzy logic and BN present probabilistic and reliable perspective them quantitatively.
The events whose occurrence probabilities are less than 5% are accepted as small
probability events (8). But it is seen that the occurrence probabilities of many events in
this study are quite higher than 5% due to the occurring conditions of the occupational
accidents are related to the mining environment. So, the reasons of the accidents need
to be roughly considered for reducing and mitigating the risks.
Human factors are accepted as the main accident cause in many risky environ-
ments. But according to the result, in the underground mining sector, human factors are
related to the managemental issues and just accepted that they can be associated with
and unpreventable lack of attention of human. So according to the results of the study,
to prevent occupational accidents, related OHS education of the workers should be
considered as the issue that needs attention. The proposed methodology to analyze the
risk can be used by analysts and decision-makers about OHS in any process industry. It
provides a comprehensive evaluation of the occupational safety risk to develop the
required risk mitigation and reducing strategies.
References
1. Komljenovic, D., Loiselle, G., Kumral, M.: Organization: a new focus on mine safety
improvement in a complex operational and business environment. Int. J. Min. Sci. Technol.
27(4), 617–625 (2017)
2. Samantra, C., Datta, S., Mahapatra, S.S.: Analysis of occupational health hazards and
associated risks in fuzzy environment: a case research in an Indian underground coalmine.
Int. J. Injury Control Saf. Promot. 24(3), 311–327 (2016)
3. Boudreau, J., Ehrlich, J., Sanders, S., Winn, A.: Social choice violations in rank sum scoring:
a formalization of conditions and corrective probability computations. Math. Soc. Sci. 71,
20–29 (2014)
4. Ghodrati, B., Hadi Hoseinie, S., Garmabaki, A.H.S.: Reliability considerations in automated
mining systems. Int. J. Min. Reclam. Environ. 29(5), 404–418 (2015)
5. Sanmiquel, L., Rossell, J.M., Vintró, C.: Study of Spanish mining accidents using data
mining techniques. Saf. Sci. 75, 49–55 (2015)
6. Kissell, F. N.: Handbook for Dust Control in Mining (No. 9465). US Department of Health
and Human Services, Public Health Service, Centers for Disease Control and Prevention.
National Institute for Occupational Safety and Health. Pittsburgh Research Laboratory
(2003)
7. Horberry, T., Burgess-Limerick, R., Steiner, L.: Human centred design of equipment:
lessons from the mining industry (2016)
8. Li, M., Wang, D., Shan, H.: Risk assessment of mine ignition sources using fuzzy Bayesian
network. Process Saf. Environ. Prot. 125, 297–306 (2019)
9. Tong, X., Fang, W., Yuan, S., Ma, J., Bai, Y.: Application of Bayesian approach to the
assessment of mine gas explosion. J. Loss Prev. Process Ind. 54, 238–245 (2018)
10. Jiang, W., Qu, F., Zhang, L.: Quantitative identification and analysis on hazard sources of
roof fall accident in coal mine. Procedia Eng. 45, 83–88 (2012)
11. Maiti, J., Khanzode, V.V.: Development of a relative risk model for roof and side fall fatal
accidents in underground coal mines in India. Saf. Sci. 47(8), 1068–1076 (2009)
12. Palei, S.K., Das, S.K.: Sensitivity analysis of support safety factor for predicting the effects
of contributing parameters on roof falls in underground coal mines. Int. J. Coal Geol. 75(4),
241–247 (2008)
13. Bobbio, A., Portinale, L., Minichino, M., Ciancamerla, E.: Improving the analysis of
dependable systems by mapping fault trees into Bayesian networks. Reliab. Eng. Syst. Saf.
71(3), 249–260 (2001)
14. Xue, S. S., Li, X. C., Xu, X. Y.: Fault tree and Bayesian network based scraper conveyer
fault diagnosis. In: Proceedings of the 22nd International Conference on Industrial
Engineering and Engineering Management, pp. 783–795. Atlantis Press, Paris (2015)
15. Garvey, M.D., Carnovale, S., Yeniyurt, S.: An analytical framework for supply network risk
propagation: a Bayesian network approach. Eur. J. Oper. Res. 243(2), 618–627 (2015)
16. Hänninen, M., Kujala, P.: Influences of variables on ship collision probability in a Bayesian
belief network model. Reliab. Eng. Syst. Saf. 102, 27–40 (2012)
17. Zadeh, L.A.: Fuzzy Sets. Inf. Control 8, 338–353 (1965)
18. Jones, B., Jenkinson, I., Yang, Z., Wang, J.: The use of Bayesian network modelling for
maintenance planning in a manufacturing industry. Reliab. Eng. Syst. Saf. 95(3), 267–277
(2010)
Risk Analysis for Digitalization Oriented
Sustainable Supply Chain Using Interval-
Valued Pythagorean Fuzzy AHP
Nurşah Alkan(&)

[email protected]
Abstract. Digitalization arising with advancements in technologies offers

immense opportunities for the supply chain. The digitalization provides raised
value creation and productivity advantages in terms of optimized processes and
systems, better resource management, and smarter supply chains. Therefore,
with digitalization, organizations consider the adoption of sustainable practices
in their supply chains as both a competitive advantage and a necessity for long-
term survival. However, the development and implementation of sustainable
supply chains include many risks, largely due to the existence of several risks
such as economic, technological, political, social, and the lack of resources.
Sustainable supply chains in a digital environment are an important issue that
needs to be addressed due to these risks. This study is to examine the risks
towards the advancement of technology more closely on sustainable supply
chain environments. The aim is to create a model that evaluates the risks that can
be encountered under the digitalization perspective of the sustainable supply
chain. Firstly, in this study, the main and sub-criteria of the risk assessment
model are determined by a comprehensive literature review and experts’ opin-
ion. Subsequently, the interval-valued Pythagorean fuzzy analytic hierarchy
process (IVPF-AHP) method is used to evaluate the related risks arising in the
sustainable supply chain with digitalization.
Keywords: Digital transformation Sustainable supply chain Risk

assessment Pythagorean fuzzy sets AHP
1 Introduction
In recent years, many businesses are increasingly extending their commitment to

sustainable business practices towards their supply chains since sustainable supply
chain management is an important provider of significant improvements for the
community in the long term [1, 2]. Businesses struggle to adopt sustainable supply
chain practices in their business activities because of competition increasing. Also, they
struggle to fulfill the continuously changing needs of customers along with providing a
sustainable relation in business. Therefore, industrial managers aim to create an
innovative business environment in their businesses by adopting advanced technolo-
gies like 3D printing, Internet of Things, Cloud Computing, Data Analytics to meet
1374 N. Alkan
these struggles [2]. These technologies are the enablers of digitalization importantly
transforming the behavior of supply chain management in terms of sustainability since
they can communicate with each other by reconfiguring themselves according to
customer requirements, production parameters, and environmental performance [2, 3].
Digitization also facilitates highly organized interconnections between materials,
goods, and equipment in addition to meeting customer requirements in terms of supply
chain [3]. Also, digitalization-based sustainability in the supply chain helps industrial
managers in issues such as resource efficiency, equipment effectiveness, environmental
protection and control initiatives, employee and community welfare, smarter, agile and
flexible processes in their supply chains. However, it poses several risks for businesses
in issues such as ensuring efficient utilization of natural resources, data quality, and
credibility, technological innovation to better develop the process and quality, com-
plexity issues. Therefore, businesses need to recognize and to analyze accurately the
several risks related to the adoption and effective implementation of the digital-oriented
sustainable supply chain. For this, a set of risks should be evaluated with a method-
ological approach to manage and ensure the sustainability of the implementation of
digitalization trends in various business operations and the supply chain activities [4].
The evaluation and prioritization of the risks constitute a decision problem involving
many criteria and sub-criteria. Therefore, this study aims to analyze the risks identified
by an extensive literature review and expert opinions by addressing the sustainability
effects in the digitization of the supply chain. These risks can be evaluated by using a
multi-criteria decision making (MCDM) methodology. In the literature, there are many
MCDM methods such as AHP. The AHP method introduced by Saaty [1, 5, 6] based
on pairwise comparisons is one of the most used and most popular methods among
MCDM problems. The AHP method allows the evaluation of expert judgments by
using pairwise comparisons to obtain the weight of the criteria in the decision-making
process. AHP, which shows the relationship between criteria with a hierarchical
structure, provides an easy solution by dividing a large and complex problem into
smaller and easier problems [2].
In this study, the AHP method based on PFSs is applied to provide more freedom to
by evaluating the vagueness and impreciseness of the risk evaluation process for
decision-makers. To the best of our knowledge, this study is the first decision-making
procedure that performs digital transformation-based risk analysis in the sustainable
supply chain. Thus, this study helps for a successful digital transformation imple-
mentation in businesses’ sustainable supply chains by determining the most important
risks. In this way, this study is a guiding qualification both for businesses and for
academics who want to research this subject.
The rest of the paper has been organized as follows: Sect. 2 gives preliminaries of
PFSs and IVPFSs. The proposed methodology is presented in Sect. 3. IVFF-AHP
method is applied to analyze the risks that can be faced in digitalization-oriented
sustainable supply chains of businesses in Sect. 4. Finally, the obtained results and
future research directions are discussed in Sect. 5.
Risk Analysis for Digitalization Oriented Sustainable Supply Chain 1375
2 Pythagorean Fuzzy Sets (PFSs)
In this section, the basic concepts and the mathematical operations of PFSs and IVPFSs
have been briefly introduced.
2.1 Preliminaries: Pythagorean Fuzzy Sets

PFSs, an extension of intuitionistic fuzzy sets by Atannasov [7] have been introduced
by Yager [8]. PFSs are characterized by the membership and non-membership degrees.
In PFSs, the sum of membership degree and non-membership degree assigned by
experts can be larger than one, but their squared sum must be less than or equal to 1.
PFSs are defined in Definition 1 as follows:
Definition 1. Let X be a fixed set. A Pythagorean fuzzy sets P ~ in X is an object having
the form [9]:

~ ¼ ðx; lP~ ð xÞ; mP~ ð xÞÞjx 2 X
P ð1Þ
where the function lP~ : X ! ½0; 1 defines and tP~ : X ! ½0; 1 the degree of membership
~ respectively. For every x 2 X, it holds that:
and non-membership of the element x 2 X to P,
0 ðlP~ ð xÞÞ2 þ ðmP~ ð xÞÞ2 1 ð2Þ
The degree of hesitancy is calculated as follows:

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pP ð X Þ ¼ 1 lP~ ð xÞ2 mP~ ð xÞ2 ð3Þ
2.2 Preliminaries: Interval - Valued Pythagorean Fuzzy Sets (IVPFSs)

IVPFSs introduced by Zhang [9] are defined in Definition 2 as follows:
Definition 2. Let X be a fixed set. An IVPFS P ~ is denoted over X is given as [10]:

P~ ¼ ðx; lP~ ð xÞ; mP~ ð xÞÞjx 2 X ð4Þ
where lP~ ð xÞ ½0; 1 and tP~ ð xÞ ½0; 1 denote the membership degree and non-
membership degree of the element x 2 X to the set P, ~ respectively. For each x 2 X,
lF~ ð X Þ and tF~ ð X Þ are interval values, lP~ ð xÞ and mP~ ð xÞ having the lower and upper
interval value are defined as follows:

lP~ ð xÞ ¼ lLP~ ð xÞ; lU
~ ð xÞ ½0; 1
P
ð5Þ

mP~ ð xÞ ¼ mLP~ ð xÞ; mU
~ ð xÞ ½0; 1
P
ð6Þ
2
2
where the expression is subject to condition 0 ðlU
~
P
ð x ÞÞ + t U
~
P
ð x Þ 1.
1376 N. Alkan
h i
For every x 2 X, pP~ ð xÞ ¼ pLP~ ð xÞ; pU
~
P
ð x Þ is called as the degree of
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2
hesitancy in IVPFSs, where pLP~ ð xÞ ¼ 1 lU ~ ð xÞ mP
P ~ ð xÞ
U
and pU
~ ð xÞ ¼
P
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
2 2
1 lLP~ ð xÞ mLP~ ð xÞ .
L U
L U
~ 1 ¼ lL1 ; lU
Definition 3. Let P 1 ; m1 ; m1 and P~ 2 ¼ lL2 ; lU
2 ; m2 ; m2 be two
IVPFNs and k [ 0. The arithmetical operations of these two IVPFNs are defined as
follows:
02 r
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2 2 2 3 1
lP~ þ lP~ lP~ lP~ 7 h L L U U iC
L L L L
B6
~ 2 ¼ B6 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~1 P 7 C
@4 2 2 2 2 5; tP~ 1 tP~ 2 ; tP~ 1 tP~ 2 A ð7Þ
1 2 1 2
P
lP~U
þ lP~ lP~
U U
lP~
U
1 2 1 2
0 2 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2 2 2 31
B h i 6 tLP~ þ tLP~ tLP~ tLP~ 7C
P ~ 2 ¼ B lL~ lL~ ; lU~ lU~ ; 6 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
~1 P 1 2 1 2
7C ð8Þ
@ P1 P2 P1 P2 4 2 2 2 2 5A
tP~
U
þ tP~ tP~
U U
tP~
U
1 2 1 2
02sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3 1
2 k 2 k k k
~ 1 ¼ @4 1 1 lL~
kP ; 1 1 lU ~
5; tL~ ; tU~ A ð9Þ
P 1 P 1 P1 P1
0 2sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi31
2 k 2 k
k k
~ k1
P ¼ @ lP~ 1 ; lP~ 1
L U
; 4 1 1 tP~ L ; 1 1 tU 5A ð10Þ
1
~
P 1
h
i h i
~i ¼
Definition 4. Let P lLP~ ; lU
~i ; t L
;
~i Pt U
~i ði ¼ 1; 2; . . .::; nÞ be a set of IVPFSs and
i P P
Pn
w ¼ ðw1 ; w2 ; . . .:; wn ÞT be weight vector of P ~ i with wi ¼ 1, then an interval-valued
i¼1
Pythagorean fuzzy weighted geometric ðIVPFWGÞ operator of dimension n is a
mapping IVPFWG: P ~n ! P~ that [11]:
0 hQ
wi Q wi i 1
k
lLP~ ; ki¼1 lU ~i ;

B "s
i¼1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
i
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
#C
IVPFWG P~1; P ~k ¼ B
~ 2 ; . . .; P
@ Qk 2 w i Qk
2 wi C A
1 i¼1 1 tP~ L ; 1 i¼1 1 tP~ U
i i
ð11Þ
In this section, the AHP method based on PFSs has been suggested to evaluate and
prioritize the risks of digitalization oriented sustainable supply chains. The steps for the
proposed methodology are presented as follows [12].

Step 1. Construct the compromised pairwise comparison matrix R ¼ rij mxm with
respect to experts’ opinions given in Table 1 which is constituted by interval-valued
Pythagorean fuzzy information.
Table 1. Linguistic scale and Pythagorean fuzzy number equivalents

Linguistic terms IVPF numbers
Certainly Low Importance (CLI) <[0,0], [0.9,1]>
Very Low Importance (VLI) <[0.1,0.2], [0.8,0.9]>
Low Importance (LI) <[0.2,0.35], [0.65,0.8]>
Below Average Importance (BAI) <[0.35,0.45], [0.55,0.65]>
Average Importance (AI) <[0.45,0.55], [0.45,0.55]>
Above Average Importance (AAI) <[0.55,0.65], [0.35,0.45]>
High Importance (HI) <[0.65,0.8], [0.2,0.35]>
Very High Importance (VHI) <[0.8,0.9], [0.1,0.2]>
Certainly High Importance (CHI) <[0.9,1], [0,0]>
Equal Importance (EI) <[0.5,0.5], [0.5,0.5]>
Step 2. Check the consistency of pairwise comparison matrixes ðRÞ. Here, to check
consistency the experts’ judgments, match with the crisp numbers corresponding in
Saaty’s scale the linguistic terms given in Table 1.
Step 3. Aggregate the individual judgments of each expert by using IVPFWG aggre-
gation operator given in Eq. (11).
Step 4. Find the differences matrix D ¼ dij mxm by using Eqs. (12) and (13):
2 2
dijL ¼ lLij mU
ij ð12Þ
2 2
dijU ¼ lU
ij mLij ð13Þ

Step 5. Calculate the interval multiplicative matrix S ¼ sij mxm by using Eqs. (14) and
(15):
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
L
sLij ¼ 1000dij ð14Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
U
ij ¼
sU 1000dij ð15Þ
1378 N. Alkan
Step 6. Obtain the indeterminacy value H ¼ hij mxm of the rij by using Eq. (16):

hij ¼ 1 l3ijU l3ijL m3ijU m3ijL ð16Þ

Step 7. Calculate unnormalized weights T ¼ tij mxm by multiplying the indetermi-

nancy values H ¼ hij mxm with the interval multiplicative matrix S ¼ sij mxm by
using Eq. (17):
!
sLij þ sU
ij
tij ¼ hij ð17Þ
2
Step 8. Find the normalized priority weights wi of criteria and sub-criteria by using
Eq. (18).
Pm
j¼1 tij
wi ¼ m Pm
P ð18Þ
i¼1 j¼1 tij
4 Application
With advances in technology, digital transformation has affected all sectors including
supply chain and logistics and has become an important topic by academics and
business nowadays. Businesses have been focused on sustainable practices and
information and communication technology tools in their supply chains to both sustain
their supply chains and benefit from the opportunities brought by digitalization. This
has led the academy and business to address digitalization and sustainable supply
chains in different ways.
Thanks to digitalization, businesses will enable resources such as materials, energy,
water, and products are allocated more efficiently using real-time data from their pro-
duction systems and supply chain partners. However, businesses are to face many risks
to achieve digital transformation in their sustainable supply chains. Therefore, they want
to identify and prioritize the risks they face in adopting and implementing both the
sustainable supply chain and digitalization. In this context, it is very important to
conduct a risk analysis to guide the businesses. In this paper, a risk analysis is conducted
using an MCDM methodology based on PF-AHP. Four main criteria and seventeen sub-
criteria for risk analysis are identified through literature review [2–4, 13–15] and
experts’ opinions as given in Table 2. Then, the pairwise comparison matrices are
constituted to evaluate the risk factors by three decision-makers. Table 3 represents
pairwise comparisons for the main risk factors of decision-makers according to lin-
guistic terms given in Table 1. After applying the steps of the Pythagorean fuzzy AHP
presenting in Sect. 3, the overall and local weights of all main and sub- risk factors are
shown in Table 4.
Table 2. Main and sub-risk factors for digitalization sustainable supply chain
Organizational (O) O1: Lack of top management support and commitment
O2: Lack of digital culture
O3: Reluctant behavior for digital transformation
O4: Financial constraint
O5: Lack of know-how and skills to use technologies and technical
expertise
Technological (T) T1: Lack of standardization and data sharing protocols
T2: Security challenges
T3: Low technological advancements and progress in business
T4: Lack of infrastructure and internet-based networks
Strategic (S) S1: Lack of R&D activities
S2: Lack of collaboration and coordination
S3: Lack of tools, methods, techniques, and indicators for digital
transformation in the sustainability supply chain
S4: Poor innovation
S5: Uncertain economic benefit of digital investment
External Factors (E) E1: Cultural differences in supply chain partners
E2: Lack of governmental supported policies
E3: Lack of customer’s awareness and tendency for digitalization and
sustainability
According to the results obtained, the most important main risk factor is determined
as Organizational risk with 0.44° as shown in Table 4. Other risk factors are ranked as
technological risk, strategic risk, and external factors risk, respectively. Considering
the overall weights of sub-risk factors for each main risk factor, “O1: Financial con-
straints” places in the first rank as the most important risk within the sub-risk factors.
“T3: Low technological advancements and progress in business” and “O5: Lack of
know-how and skills to use technologies and technical expertise” follow this rank as
second and third, respectively.
Table 3. Main criteria pairwise comparisons according to linguistic terms

DM1 DM2 DM3
O T S E O T S E O T S E
O EI AAI HI VHI EI AAI AAI HI EI EI AAI HI
T BAI EI AAI HI BAI EI AAI HI EI EI AAI HI
S LI BAI EI AAI BAI BAI EI AAI BAI BAI EI AAI
E VLI LI BAI EI LI LI BAI EI LI LI BAI EI
CR 0.044 0.076 0.016
1380 N. Alkan
Table 4. Weights of main and sub-risk factors

Main criteria O T
Weights 0.44 0.323
Sub-criteria O1 O2 O3 O4 O5 T1 T2 T3 T4
Local weights 0.059 0.099 0.03 0.478 0.335 0.114 0.306 0.508 0.072
Overall weights 0.026 0.044 0.013 0.21 0.147 0.037 0.099 0.164 0.023
Main criteria S E
Weights 0.158 0.079
Sub-criteria S1 S2 S3 S4 S5 E1 E2 E3
Local weights 0.067 0.343 0.496 0.039 0.06 0.248 0.695 0.058
Overall weights 0.01 0.054 0.078 0.006 0.009 0.02 0.055 0.005
5 Conclusions
In today’s competitive conditions, businesses need to both ensure sustainability in their

supply chains and benefit from the opportunities brought by digitalization. Businesses’
adoption of a digital-based sustainable supply chain brings with it many risks as well as
offers several opportunities. Therefore, it is very crucial for businesses to create a road
map by addressing these risk factors while adopting digitalization. This work aims to
analyze the risks faced with digitalization initiatives for accomplishing sustainability
orientation in supply chains. The analysis of the risks constitutes an MCDM problem
involving many criteria and sub-criteria. Several risks related to the adoption and
implementation of digitalization trends in sustainability activities of the supply chain
are prioritized using Pythagorean fuzzy AHP that provides more flexibility and
reflecting uncertainty better than ordinary fuzzy AHP in linguistic assessments of
criteria. After applied the proposed methodology, the most important main risk factor is
determined as Organizational risk. Considering the sub-risk factors, “Financial con-
straints”, “Low technological advancements and progress in business” and “Lack of
know-how and skills to use technologies and technical expertise” are ranked as first,
second, and third important sub-risk factors. These results show that “Financial con-
straints” is the most crucial risk factor that businesses need to first address. For further
research, different MCDM methodologies can be applied for this addressed problem
and can be evaluated different risk factors. Besides, different extensions of ordinary
fuzzy sets such as intuitionistic fuzzy sets, hesitant fuzzy sets, spherical fuzzy sets can
be applied in this problem framework.
References
1. Rasmi, S., Kazan, C., Türkay, M.: A multi-criteria decision analysis to include environ-
mental, social, and cultural issues in the sustainable aggregate production plans. Comput.
Ind. Eng. 132, 348–360 (2019)
2. Luthra, S., Mangla, S.: Evaluating challenges to Industry 4.0 initiatives for supply chain
sustainability in emerging economies. Process Safety and Environmental Protection 117,
168–179 (2018)
3. Jabbour, A.L., Jabbour, C., Foropon, C., Filho, M.: When titans meet – can industry 4.0
revolutionise the environmentally sustainable manufacturing wave? The role of critical
success factors. Tech. Forecast. Soc. Change 132, 18–25 (2018)
4. Mangla, S., Govindan, K., Luthra, S.: Prioritizing the barriers to achieve sustainable
consumption and production trends in supply chains using fuzzy analytical hierarchy
process. J. Clean. Prod. 151, 509–525 (2017)
5. Saaty, T.L.: The Analytic Hierarchy Process: Planning, Priority Setting, Resource
Allocation. MacGraw-Hill, New York (1980)
6. Saaty, T.L.: Decision making with the analytic hierarchy process. Int. J. Serv. Sci. 1(1), 83–
98 (2008)
8. Yager, R.: Pythagorean fuzzy subsets. In: IFSA World Congress and NAFIPS Annual
Meeting, IFSA/NAFIPS 2013, pp. 57–61. IEEE (2013)
pythagorean fuzzy sets. Int. J. Intell. Syst. 29(12), 1061–1078 (2014)
10. Peng, X., Yang, Y.: Fundamental properties of interval-valued pythagorean fuzzy
aggregation operators. Int. J. Intell. Syst. 31(5), 444–487 (2016)
11. Garg, H.: New exponential operational laws and their aggregation operators for interval-
valued pythagorean fuzzy multicriteria decision-making. Appl. Soft Comput. J. 38, 988–999
(2016)
12. Karasan, A., İlbahar, E., Kahraman, C.: A novel pythagorean fuzzy AHP and its application
to landfill site selection problem. Soft. Comput. 23(21), 10953–10968 (2018)
13. Saberi, S., Kouhizadeh, M., Sarkis, J., Shen, L.: Blockchain technology and its relationships
to sustainable supply chain management. Int. J. Prod. Res. 57(7), 2117–2135 (2019)
14. Muñoz-Villamizar, A., Solano, E., Quintero-Araujo, C., Santos, J.: Sustainability and
digitalization in supply chains: a bibliometric analysis. Uncertain Supply Chain Manage. 7
(4), 703–712 (2019)
15. Sanders, N., Boone, T., Ganeshan, R., Wood, J.: Sustainable supply chains in the age of AI
and digitization: research challenges and opportunities. J. Bus. Logist. 40(3), 229–240
(2019)
Intelligent Quality
Design for Six Sigma and Process Capability
Using Penthagorean Fuzzy Sets
Elif Haktanır1,2(&) and Cengiz Kahraman1

1
Istanbul Technical University, 34367 Besiktas, Istanbul, Turkey
2
Altınbas University, 34217 Bagcilar, Istanbul, Turkey
[email protected]
Abstract. Design for six sigma (DFSS) is a structured approach for designing
new products, processes and/or services or redesigning the existing ones. It
builds quality into the product during the development and ensures the customer
expectations are understood, met, or exceed. In DFSS, variation is aimed to be
minimized by applying the techniques of the six sigma approach. Some of these
techniques are process control and process capability analysis. Vagueness and
impreciseness in these analyses such as the statistical distribution of the con-
sidered process can be handled by the fuzzy set theory. Recently, ordinary fuzzy
sets have been extended to several new types of fuzzy sets such as Pythagorean
fuzzy sets, spherical fuzzy sets, picture fuzzy sets, and intuitionistic fuzzy sets. In
this paper, we propose a new extension of intuitionistic fuzzy sets, which is called
Penthagorean fuzzy sets and employ it in the process capability analysis of a
production process. Penthagorean fuzzy process capability indices are developed
and applied to an Autonomous Underwater Vehicle (AUV) design example.
Keywords: Design for six sigma Process capability Penthagorean fuzzy

sets Autonomous Underwater Vehicle
1 Introduction
In recent years, the dynamics of the market and the speed of technological changes
have forced companies to succeed in new product development. In this challenging
competitive environment, the key of growth for companies is being skilled and suc-
cessful in developing new products. The new product development activity includes
both existing products and new product development projects in the company.
Improvement of the existing products that already produced within the company is
evaluated within this scope. Quality improvement of existing products, special addi-
tions or modifications constitute existing product improvement projects.
Design for Six Sigma (DFSS) is a step-by-step business process management
method for designing and developing new products or services based on some statis-
tical tools such as linear regression. DFSS is a proactive approach that can be utilized to
understand the customers’ needs by not only meeting but exceeding them. Different
phases in the DFSS process are named in many various ways in the literature such as
DMADV (Design, Measure, Analyze, Design, Verify) or IDOV (Identify, Design,
Optimize, Verify).
1386 E. Haktanır and C. Kahraman
Since DFSS employs other widely used design tools like QFD, FMEA, Design of
Experiments (DOE), Robust Design, it is one of the most comprehensive design tools
which allows a lower NPD cost, shorter time to market, lower production costs,
reduced defects and waste but increased quality, reliability and robustness comparing to
the other methods.
DFSS has been utilized by many researchers in the literature for product design and
development such as electric kettle [1], spiral bevel gears [2], eco-motorbike [3],
portable water filter [4], an automotive component [5].
The inputs of the DFSS approach can be customer needs, business needs, raw
materials, constraints and the outputs can be quality products, processes or services
designed to reach six sigma levels [6]. Since the inputs contain vague and imprecise
information by its nature, fuzzy logic can be successfully applied in such procedures to
deal with it. Since the introduction of ordinary fuzzy sets to the literature by Zadeh [7],
they have been extended to many new forms such as intuitionistic fuzzy sets, neu-
trosophic sets, hesitant fuzzy sets, Pythagorean fuzzy sets, spherical fuzzy sets, and q-
rung orthopair fuzzy sets.
In this study, a new extension of fuzzy sets, penthagorean fuzzy sets, has been
developed and used with the aim of design for six sigma and process capability. The
proposed method has been illustrated with an Autonomous Underwater Vehicle
(AUV) design example.
The rest of the paper is organized as follows: Sect. 2 briefly explains process
capability and six sigma approach. Section 3 introduces penthagorean fuzzy sets.
Section 4 gives the steps of the proposed penthagorean fuzzy six sigma approach.
Section 5 demonstrates the application of the proposed method for AUV design. The
last section concludes the paper with future directions.
2 Process Capability Analysis and Six Sigma Approach
Reducing the process variability is the main purpose of quality improvement efforts.
Process capability analysis (PCA) and six sigma approach (SSA) are the two well-
known tools for quality improvement. PCA compares the specification limits and the
distribution of the considered process, which is generally normal but sometimes non-
normal. Based on process capability indices, it decides whether the process is capable
or not. SSA tries to reduce process variability so that 12 sigma can be included between
lower and upper specifications. SSA allows the mean of process distribution to shift 1.5
sigma to left or right from the center of specification limits. This yields maximum 3.4
defects per one million products.
Process capability indices Cp and Cpk are given by Eqs. (1) and (2), respectively.
A capable process is presented in Fig. 1.
USL LSL
Cp ¼ ð1Þ
6r

lpop LSL USL lpop
Cpk ¼ min ; ð2Þ
3r 3r
Design for Six Sigma and Process Capability 1387
Fig. 1. A capable process.
pop : population
where USL: upper specification limit, LSL: lower specification limit, l
mean, and r: process standard deviation.
Fig. 2. Process capability in SSA.
Figure 2 illustrates the position of a process distribution in SSA where the non-
conforming percentage is P(z < −4.5) + P(z > 7) = 3.4 10−6 or P(z < −7) + P
(z > 4.5) = 3.4 10−6.
Figure 3 presents the roadmap of DMADV which is one of the DFSS frameworks
including the phases define, measure, analyze, design, and verify. In this paper we deal
with the phase analyze through PCA and SSA.
D M A D V
Define Measure Analyse Design Verify
• Define the project • Determine • Identfiy functions. • Develop design, • Verify design
customer Generate and test/optimise performance.
requirements. select concepts. design Implement design.
components and
complete design.
Fig. 3. Roadmap for DFSS.

3 Penthagorean Fuzzy Sets
Intuitionistic fuzzy sets (IFS) are the main source of other extensions of fuzzy sets such
as Pythagorean fuzzy sets and fermatean fuzzy sets. Yager [8] has proposed the general
definition of power intuitionistic fuzzy sets and called them as q-rung orthopair fuzzy
sets (QROFS). Based on QROFS, IFS can be abbreviated as P1FS whereas Pytha-
gorean fuzzy sets as P2FSS and fermatean fuzzy sets as P3FS. Our proposed fuzzy sets
Penthagorean fuzzy sets are abbreviated as P5FS thereafter. The definition of P5FSs is
given in Eq. (3). Score and accuracy functions for P5FS are given in Eqs. (5) and (6).
Some arithmetic operations with P5FSs are given in Eqs. (7–12). Aggregation opera-
tors for P5FSs are given in Eqs. (13) and (14).
~ can be given as:
Definition 1. Let T be a fixed set. A P5FS M

~ ¼
M T; lM~ ðtÞ; mM~ ðtÞ jt 2 T ; ð3Þ
~ ! ½0; 1 is the membership degree and the function

where the function lM~ ðtÞ : M
~
mM~ ðtÞ : M ! ½0; 1 is the non-membership degree, moreover, for every t 2 T, it
satisfies:
ðlM~ ðtÞÞ5 þ ðmM~ ðtÞÞ5 1 ð4Þ

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

ffi
In addition, the degree of hesitancy is defined as pM~ ðtÞ ¼ 1 ðlM~ ðtÞÞ5 ðmM~ ðtÞÞ5 .
5
~ ¼ ðlM~ ; mM~ Þ be a P5F number, a score value S and an accuracy

Definition 2. Let M
value H are:

S M ~ 2 ½1; 1
~ ¼ ðlM~ ðtÞÞ5 ðmM~ ðtÞÞ5 ; S M ð5Þ

~ ¼ ðlM~ ðtÞÞ5 þ ðmM~ ðtÞÞ5 ; H M
H M ~ 2 ½1; 1 ð6Þ

~ 1 ¼ lM~ ; mM~ and M
Definition 3. Let be M ~ 2 ¼ lM~ ; mM~ two P5F numbers. Then,
1 1 2 2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
5 5 5 5 5
~1 M
M ~2 ¼ lM~ 1 þ lM~ 2 lM~ 1 lM~ 2 ; mM~ 1 mM~ 2 ð7Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
5 5 5 5 5
~1 M
M ~2 ¼ lM~ 1 lM~ 2 ; mM~ 1 þ mM~ 2 mM~ 1 mM~ 2 ð8Þ
0v ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u 5 1 ( )
u lM~ l5M~ mM~ mM~ 2 pM~ 1
M ~2 ¼ @
~1 M t
5
1 2
; 1 A if l ~
l ~ ; m ~ min m ~ ; ð9Þ
1 l5M~ mM~ 2 M1 M2 M1 M2
pM~ 2
2
0 vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 ( )
u 5
lM~ 1 u mM~ m5M~ l p
~1
M ø ~2 ¼ @
M
lM~ 2 1 mM~
M~
pM~ 2
M~
; t 1 5 2 ;A if lM~ 1 min lM~ 2 ; 2 1 ; mM~ 1
mM~ 2
5
ð10Þ
2
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi !
5
k k
~1 ¼
kM
k
1 1 lM~ 1 ; mM~ 1 ð11Þ
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi !
k 5
k
~ 1k ¼
M lM~ 1
5
; 1 1 mM~ 1 ð12Þ
where k > 0.

~ l ¼ lM~ ; mM~ ðl ¼ 1; 2; . . .; nÞ be a set of P5F numbers, and
Definition 4. Let M l l
P5FWA is defined as:

P5FWA# M~ 1; M~ 2 ; . . .; M ~ n ¼ #1 M ~ 1 #2 M ~ 2 . . . #n M
~n
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi! !
5
Yn 5
#l Yn #l ð13Þ
¼ 1 l¼1 1 lM~ l ; l¼1 mM~ l
where # ¼ ð#1 ; #2 ; . . .; #P nÞ
T
is the weight of M ~ l ðl ¼ 1; 2; . . .; nÞ with M
~l 2
n
½0; 1ðl ¼ 1; 2; . . .; nÞ and l¼1 #l ¼ 1.

Definition 5. Let M ~ l ¼ lM~ ; mM~ ðl ¼ 1; 2; . . .; nÞ be a set of P5F numbers, and
l l
P5FWG is defined as in Eq. (14):

P5FWG# M~ 1; M
~ 2 ; . . .; M
~n ¼ M ~ 1#1 M ~ 2#2 . . . M ~ n#n
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi!!
Yn #l 5
Yn 5
#l ð14Þ
¼ l¼1
l ~l
M ; 1 l¼1
vM~ l
T ~ ~
Pn # ¼ ð#1 ; #2 ; . . .; #n Þ is the weight of Ml ðl ¼ 1; 2; . . .; nÞ with Ml 2 ½0; 1 and
where
l¼1 #l ¼ 1.
4 Penthagorean Fuzzy Six Sigma Approach
P5F process capability indices can be calculated by following the proposed equations
given below.
The possible USL and LSL values in Table 1 are aggregated by using Eqs. (13) and
(14) and the aggregated USL and LSL values are obtained as USLagg ;

lUSLagg ; vUSLagg and LSLagg ; lLSLagg ; vLSLagg , respectively.

Table 1. Possible P5F USL and LSL values.

Possible values of USL P5FS Possible values of LSL P5FS

USLl lUSLl ; vUSLl LSLl lLSLl ; vLSLl

USLm lUSLm ; vUSLm LSLm lLSLm ; vLSLm

USLu lUSLu ; vUSLu LSLu lLSLu ; vLSLu
The possible r values in Table 2 are aggregated

by

using Eqs. (13) and (14) and the

aggregated r value is obtained as ragg ; lragg ; vragg .
Table 2. Possible r values.

Possible r values P5FS

rl lr ; vr
l l
rm lr ; vrm
m
ru lru ; vru
pop values in Table 3 are aggregated by using

The possible population mean l

Eqs. (13) and (14) and the aggregated l pop value is obtained as l popagg ;

llpopagg ; vlpopagg .
Table 3. Possible l
pop values.
Possible lpop values P5FS

popl
l llpop ; vlpopl
l

popm
l llpopm ; vlpopm

popu
l llpopu ; vlpopu
~ p and P5F C
P5F C ~ pk are obtained by using Eqs. (15) and (16).
~ ~ USL ; l ; v LSL ; l ; v
~ p ¼ U SL LSL ¼
agg USLagg USL agg agg LSLagg LSL agg
C
r
6~ 6 ragg ; lragg ; vragg
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
l 5 l 5 v
USLagg LSLagg ; 5 USL1agg l LSL5agg ; vUSL agg
LSLagg
LSLagg
¼
ð15Þ
6 ragg ; lragg ; vragg
0 0rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi11
lUSLagg 5 lLSLagg 5 u
5 u vUSLagg 5
B USLagg LSLagg B 1 l u
5 vragg 5 CC
t vLSLagg
5
¼B B CC
LSLagg
@ ; @ ;
6 ragg l ragg 1 v 5 AA
ragg
ð16Þ
The minimum of C ~ pl and C

~ pu can be determined by using the score function in
Eq. (5) or the accuracy function in Eq. (6). If these indices are equal or greater than 1,
the process is capable otherwise it is not capable. In six sigma approach, the width
between USL and LSL is required to be 12r. When the process mean is centered, both
Cp and Cpk are equal to 2. When a shift of 1.5r from the center occurs, Cp still equals to
2 whereas Cpk becomes 1.5.
After the calculation of C ~ p and C

~ pk , Eqs. (17) and (18) are used for the final
decision on process capability, respectively:
pffiffiffiffiffiffi
Cp 5 SF ; if SF [ 0
Cpc ¼ ð17Þ
Cp ð1 þ SF Þ; if SF\0
pffiffiffiffiffiffi
Cpk 5 SF ; if SF [ 0
c
Cpk ¼ ð18Þ
Cpk ð1 þ SF Þ; if SF\0
where SF is the result of score function; Cpc and Cpk

c
are conditional process capability
indices.
5 Application
Autonomous Underwater Vehicles (AUV) are the vehicles that can operate underwater
without a human occupant. AUVs are being used for many tasks such as commercial,
research, hobby, air crash investigations, and military applications. Figure 4 and Fig. 5
show an AUV and its acoustic hull window (AHW), respectively. An AHW is used for
oceanographic research and their length can be up to 4 ft long.
Fig. 4. An AUV.
Fig. 5. An AHW.
Two possible lathes may process AHW parts of AUVs, which are presented in
Fig. 6. The mean and standard deviation of the AHW width differ with respect to the
lathes and these parameters are not exact values as well as specification limits. Table 4
gives the three possible values of these parameters as P5F numbers. These values are
equally weighted parameters, i.e. #li ¼ 1=3, i = 1, 2, 3. P5F specification limits are
(3.97; 0.85, 0.20) for LSL and (4.01; 0.95, 0.10) for USL.
(a) Lathe A (b) Lathe B
Fig. 6. Two possible lathes.
Table 4. Values of AHW parameters as P5F numbers.

Lathe A Lathe B
Mean of AHW width (feet) (3.970; 0.90, 0.20) (3.960; 0.90, 0.10)
(3.975; 0.95, 0.10) (3.968; 0.94, 0.08)
(3.980; 0.95, 0.20) (3.976; 0.92, 0.15)
Standard deviation of AHW width (feet) (0.013; 0.90, 0.20) (0.008; 0.90, 0.15)
(0.014; 0.95, 0.10) (0.009; 0.96, 0.10)
(0.015; 0.85, 0.15) (0.010; 0.93, 0.20)
Using Eqs. (15) and (16), process capability indices are obtained as in Table 5.
Table 5. P5F process capability indices.

Lathe A Lathe B
P5FWA P5FWG P5FWA P5FWG
~p
C (0.9524; 0.9345, (0.9540; 0.9471, (1.4815; 0.9104, (1.4876; 0.9159,
0.4998) 0.4996) 0.4998) 0.4996)
~ pk
C (0.8333; 0.7784, (0.8348; 0.8273, (1.4074; 0.7631, (1.4130; 0.7513,
0.6299) 0.5402) 0.5312) 0.6216)
Using the score function given in Eq. (5), the crisp values of crisp process capa-
bility indices are calculated as in Table 6.
Table 6. Crisp process capability indices.

Lathe A Lathe B
P5FWA P5FWG P5FWA P5FWG
~p
C 0.8821 0.8960 1.3350 1.2948
~ pk 0.5957
C 0.6734 1.0364 0.9624
Based on the process capability indices given in Table 6, it is seen that the process
with Lathe B is a capable process whereas the process with Lathe A is incapable. There
is also a slight difference between aggregation operators. P5FWA indicates that process
B is certainly capable while P5FWG indicates that process B is almost capable. For
process A, both aggregation operators indicate that it is an incapable process.
6 Conclusions
Process capability indices are the indicators of a process how well it is located between
specification limits. When incomplete data exist in this analysis, intuitionistic fuzzy
approaches can be used. In this paper, an extension of intuitionistic fuzzy sets, namely
penthagorean fuzzy sets, has been employed in the calculation of process capability
indices under vague and imprecise data. Our analysis indicates that the result of a crisp
analysis can change when uncertainty conditions are considered. Empirical equations
have been proposed for the expected process capability values. An application for
process capability analysis of an AUV component has been given. For further research,
we suggest q-rung orthopair fuzzy sets to be used in this analysis in order to generalize
our study.
References
1. Hassan, M.F., Chandra Mohan, T.R., Ismail, A.E., Taib, I., Mahmood, S., Kafuku, J.M.:
Application of design for six sigma (DFSS) in sustainable product design: a case study on an
electric kettle. In: IOP Conference Series: Materials Science and Engineering vol. 607, no.
1 (2019)
2. Peng, S., Ding, H., Tang, J., Tang, Y.: A novel collaborative manufacturing model requiring
both geometric and physical evaluations of spiral bevel gears by design for six sigma. Mech.
Mach. Theory 133, 625–645 (2019)
3. Frizziero, L., Liverani, A., Nannini, L.: Design for six sigma (DFSS) applied to a new eco-
motorbike. Machines 7(3), 52 (2019)
4. Fahrul Hassan, M., Jusoh, S., Zaini Yunos, M., Arifin, A.M.T., Ismail, A.E., Rasidi Ibrahim,
M., Zulafif Rahim, M.: Application of design for six sigma methodology on portable water
filter that uses membrane filtration system: a preliminary study. In: IOP Conference Series:
Materials Science and Engineering, vol. 243, no. 1 (2017)
5. Suresh, K.M., Asokan, P., Vinodh, S.: Application of design for six sigma methodology to an
automotive component. Int. J. Six Sigma Compet. Adv. 10(1), 1–23 (2016)
6. Amer, Y., Luong, L., Lee, S.: Case study: optimizing order fulfillment in a global retail supply
chain. Int. J. Prod. Econ. 127(2), 278–291 (2010)
8. Yager, R.: Generalized orthopair fuzzy sets. IEEE Trans. Fuzzy Syst. 25, 1222–1230 (2017)
Defects Control Charts Using Interval-Valued
Penthagorean Fuzzy Sets
Elif Haktanır1,2(&) and Cengiz Kahraman1

1
Istanbul Technical University, Besiktas, 34367 Istanbul, Turkey
2
Altınbas University, Bagcilar, 34217 Istanbul, Turkey
[email protected]
Abstract. Defects control charts are one of the attributes control charts which
are constructed based on the number of defects on a product or in an inspection
unit. Since the term ‘defect’ can be vague due to the subjective decisions of the
quality control operator, it cannot be precisely defined. In such cases, fuzzy set
theory can be used to add more information and flexibility to attribute based-
process control. After the introduction of the ordinary fuzzy sets to the literature,
it has been extended to many new types of fuzzy sets. One of the recent
extensions, q-rung orthopair fuzzy sets, which is a generalization of intuitionistic
fuzzy sets and Pythagorean fuzzy sets, provides a broader space to express
ambiguous information, For the q-rung orthopair fuzzy sets the sum of the qth
power of the membership degree and the qth power of the non-membership
degree is equal to or less than one. In this study, the fifth power of the q-rung
orthopair fuzzy sets is used to represent the impreciseness in the construction of
the defects control charts and names as penthagorean fuzzy sets. An application
in a TV production system is presented as an illustrative example.
Keywords: Defects control charts Attributes control charts Interval-valued

penthagorean fuzzy sets Intuitionistic fuzzy sets q-rung orthopair fuzzy sets
1 Introduction
Control charts are one of the seven basic quality control tools (histogram, check sheet,
flow chart, cause and effect diagram, Pareto chart, scatter diagram, and control chart)
being used by a variety of industries. Control charts are the graphs used for analyzing
how the processes change over time. They include a central line for the average, an
upper line for the upper control limit (UCL), and a lower line for the lower control limit
(LCL) that are determined from historical data. Decisions can be made based on the
process variation by comparing current data with respect to these limits. Process
variation can be either consistent (under control) or it can be unpredictable (out of
control). Out of control process indicates that there is an assignable cause which must
be corrected whereas an under-control process indicates that there are random causes of
variation, which are not important for the process.
Depending on the data being monitored, there are two broad categories of control

charts: variable and attribute control charts. Variables control charts (X&R
and X&S)
plot continuous measurement process data in a time-ordered sequence while attribute
Defects Control Charts Using Interval-Valued Penthagorean Fuzzy Sets 1397
control charts (u, c, p, and np) are used to monitor discrete data for nonconforming
units and nonconformities per inspection unit.
Due to the ambiguity of the attribute data, traditional control charts may be inad-
equate for defects control charts. In such cases, fuzzy control charts can be successfully
used to capture and handle the vagueness. In this study, interval-valued penthagorean
fuzzy (IVP5F) sets are developed and applied into defects control charts.
In one of the earliest studies on fuzzy control charts, Bradshaw [1] presented fuzzy
control charts’ advantages over traditional control charts depending on their ability to
represent reality more accurately. Wang and Raz [2] constructed control charts and
determined the center line and the control limits using linguistic variables and fuzzy
logic. Taleb and Limam [3] discussed different procedures of constructing control
charts for linguistic data, based on fuzzy and probability theory. Gülbay et al. [4]
developed a-cut fuzzy control charts for linguistic data. Gülbay and Kahraman [5]
developed direct fuzzy approach as an alternative to fuzzy control charts. Sentürk et al.
[6] introduced the control charts of “fuzzy nonconformities per unit with a-cut”. In
some of the recent studies, Sakthivel et al. [7] proposed a new method based on fuzzy
logic for monitoring attribute quality characteristics in statistical control charts. Tekşen
and Anagün [8] developed interval type-2 fuzzy c-control charts using ranking meth-
ods. Aslam et al. [9] designed a new attribute control chart under neutrosophic
statistics. Ercan-Teksen [10] proposed intuitionistic fuzzy c-control charts using fuzzy
comparison methods.
The rest of the paper is organized as follows: Sect. 2 briefly explains classical and
fuzzy defects control charts. Section 3 introduces IVP5F sets. Section 4 gives the steps
of the proposed IVP5F control charts for the defects approach. Section 5 demonstrates
the application of the proposed method for a TV production system. The last section
concludes the paper with future directions.
2 Defects Control Charts
2.1 Classical c Control Chart

In the crisp case, control limits for number of nonconformities are calculated by:
CL ¼ c ð1Þ
pffiffiffi
LCL ¼ c 3 c ð2Þ
pffiffiffi
UCL ¼ c þ 3 c ð3Þ
where c is the mean of the nonconformities.
2.2 Fuzzy ~c Control Chart

In the ordinary fuzzy case, each sample, or subgroup, can be represented by a trape-
zoidal fuzzy number ða; b; c; d Þ or a triangular fuzzy number ða; b; b; d Þ. Center line,
a;
f is the mean of fuzzy samples, and it is represented by ð
CL b; c; dÞ where
a;
b; c and d
are the arithmetic means of the values a; b; c and d, respectively. The center line is
obtained as in Eq. (4).
Pn Pn Pn Pn
aj bj j¼1 cj dj
f ¼ a;
b; c; dÞ
j¼1 j¼1 j¼1
CL ; ; ; ¼ ð ð4Þ
n n n n
and
f ¼ ða; b; c; dÞ ¼ ðCL1 ; CL2 ; CL3 ; CL4 Þ

CL ð5Þ
Lower and upper control limits are as given in Eqs. (6) and (7).
pffiffiffiffiffiffiffi pffiffiffi pffiffiffi pffiffiffi pffiffiffi
g ¼ CL
LCL f 3 f ¼ a 3 d; b 3 c; c 3
CL b; d 3
a
¼ ðLCL1 ; LCL2 ; LCL3 ; LCL4 Þ ð6Þ
pffiffiffiffiffiffiffi pffiffiffi pffiffiffi pffiffiffi pffiffiffi
g ¼ CL
UCL f þ3 f ¼ a þ 3 a; b þ 3
CL b; c þ 3 c; d þ 3 d
¼ ðUCL1 ; UCL2 ; UCL3 ; UCL4 Þ ð7Þ
The center line and control limits can be given by using a cuts as in Eqs. (8), (9)
and (10).

ga ¼ ðaa ; b; c; da Þ ¼ CLa ; CL2 ; CL3 ; CLa
CL ð8Þ
1 4
pffiffiffiffiffiffiffiffiffi pffiffiffiffiffi pffiffiffi pffiffiffi pffiffiffiffiffi
ga ¼ CL
LCL ga 3 ga ¼ aa 3 da ; b 3 c; c 3
CL b; da 3
aa
ð9Þ
¼ LCLa1 ; LCL2 ; LCL3 ; LCLa4
pffiffiffiffiffiffiffiffiffi pffiffiffiffiffi pffiffiffi pffiffiffi pffiffiffiffiffi
g a ¼ CL
UCL ga þ 3 ga ¼ aa þ 3 aa ; b þ 3
CL b; c þ 3 c; da þ 3 da
ð10Þ
¼ UCLa1 ; UCL2 ; UCL3 ; UCLa4
Results of these equations can be illustrated as in Fig. 1.

The decision about whether the process is in control can be made according to the
percentage area of the sample which remains inside the UCL g and/or LCL g defined as
fuzzy numbers. When the fuzzy sample is completely involved by the fuzzy control
limits, the process is said to be “in-control”. If a fuzzy sample is totally excluded by the
fuzzy control limits, the process is said to be “out of control”. Otherwise, a sample is
partially included by the fuzzy control limits. In this case, if the percentage area which
remains inside the fuzzy control limits (bj) is equal or greater than a predefined
acceptable percentage (b), then the process can be accepted as “rather in-control”;
otherwise it can be stated as “rather out of control” [11].
Fig. 1. Representation of fuzzy control limits [11].
2.3 Classical u Control Chart

If we are related to the number of nonconformities on one product, c-control chart is
used. When the sample size is not constant due to the process constraints, u-control
chart is preferred to monitor and evaluate the process. The classical u-control chart
limits proposed by Shewhart are given the following equations:
sffiffiffiffi

u
UCLu ¼ u þ 3 ð11Þ
nj
CLu ¼ u ð12Þ
sffiffiffiffi

u
LCLu ¼ u 3 ð13Þ
nj
cj
uj ¼ ð14Þ
nj
Pm
j¼1 uj
u ¼ ; j ¼ 1; 2; . . .; m ð15Þ
m
where uj is the number of nonconformities per inspection unit and

u is the average
number of nonconformities per inspection unit, nj is the sample size, cj is total non-
conformities in a sample of nj inspection units, and m is the number of samples.
2.4 Fuzzy e
u Control Chart
In this case, the number of nonconforming is expressed as a triangular fuzzy number
uaj ; ubj ; ucj . The fuzzy averages of nonconforming values are calculated by:
P
ua j
ua ¼ ð16Þ
m
P
ub j
ub ¼ ð17Þ
m
P
uc j
uc ¼ ð18Þ
m
The fuzzy ~u-control chart limits are given as follows:

sffiffiffiffiffi sffiffiffiffiffi sffiffiffiffiffi!
~ u¼ ua ub
uc
U CL ua þ 3 ; ub þ 3 ;
uc þ 3 ð19Þ
nj nj nj
~u ¼ ðua ; ub ; uc Þ

CL ð20Þ
sffiffiffiffiffi sffiffiffiffiffi sffiffiffiffiffi!
~ u¼ uc ub
ua
LCL ua 3 ; ub 3 ;
uc 3 ð21Þ
nj nj nj
3 Interval-Valued Penthagorean Fuzzy Sets
q-rung orthopair fuzzy sets (qROFS) [12], Pythagorean fuzzy sets (PFS) and fermatean
fuzzy sets (FFS) are the recent extensions of intuitionistic fuzzy sets (IFS). q-rung
orthopair fuzzy sets are qth power intuitionistic fuzzy sets whereas PFS is second power
intuitionistic fuzzy sets. Based on qROFS, IFS can be abbreviated as P1FS whereas
Pythagorean fuzzy sets as P2FS and fermatean fuzzy sets as P3FS. Penthagorean fuzzy
sets are abbreviated as P5FS thereafter. The definition of IVP5FSs is given in Eq. (22).
Some arithmetic operations with IVP5FSs are given in Eqs. (23–16). Score and
accuracy functions for IVP5FS are given in Eqs. (27) and (28).
Definition 1. Let X be a nonempty fixed set, an IVP5F set A on X can be expressed as
follows:
nD h i h iE o
~¼
A x; lLA~ ð xÞ; lU
~ ð x Þ ; v L
~ ð x Þ; v U
~ ð x Þ j x 2 X ð22Þ
A A A
h i h i
where lLA~ ð xÞ; lU
~
A
ð x Þ and v L
~
A
ð x Þ; v U
~
A
ð x Þ denote the membership and the non-
membership degrees of x 2 X to A, respectively, which satisfies the following
h i h i
condition: for every x 2 X: lLA~ ð xÞ; lU~ ð x Þ ½ 0; 1 , v L
~ ð x Þ; v U
~ ð x Þ ½0; 1,
A
" A

A
5 5 h i 5 5 15
lA~ ð xÞ þ vA~ ð xÞ 1, and pA~ ð xÞ; pA~ ð xÞ ¼ 1 lA~ ð xÞ vA~ ð xÞ
U U L U U U
; 1
5 5 15
lLA~ ð xÞ vA~ ð xÞ
L
denotes the indeterminacy degree x 2 X.

L U

L U

Definition 2. Let ~a1 ¼ lL1 ; lU
1 ; v1 ; v1 and ~a2 ¼ lL2 ; lU
2 ; v2 ; v2 be two
IVP5F numbers, then the operational laws are defined as follows:

L 5 L 5 L 5 L 5 1=5 U 5 U 5 U 5 U 5 1=5 L L U U
~
a1 ~
a2 ¼ l1 þ l2 l1 l2 ; l1 þ l2 l1 l2 ; v1 v2 ; v1 v2 ;
ð23Þ

L L U U
L 5 L 5 L 5 L 5 1=5 U 5 U 5 U 5 U 5 1=5
~
a1 ~
a2 ¼ l1 l2 ; l1 l2 ; v1 þ v2 v1 v2 ; v1 þ v2 v1 v2 ;
ð24Þ
*" 1=5 # +
5 k U 5 k 1=5 h L k U k i
k~a1 ¼ 1 1 lL1 ; 1 1 l1 ; v1 ; v1 ; k [ 0;
ð25Þ
* " #+
h i
L 5 k 1=5
U 5 k 1=5
L k U k
~ak1 ¼ l1 ; l1 ; 1 1 v1 ; 1 1 v1 ; k [ 0:
ð26Þ

L U

Definition 3. Let ~a ¼ lL~a ; lU~a ; v~a ; v~a be an IVP5F number, then the score
function Sð~aÞ and the accuracy function Pð~aÞ of ~
a are defined as follows:
5 U 5 5 5
1 þ lU~a v~a þ 1 þ lL~a vU~a
Sð~aÞ ¼ ð27Þ
4
U 5 L 5 U 5 U 5
l þ l~a þ v~a v~a
Pð~aÞ ¼ ~a ð28Þ
2
4 Interval-Valued Penthagorean Fuzzy Control Charts

for Defects
4.1 IVP5F c Control Charts
Let the triangular IVP5F data for nonconformities per unit product be given as in
Table 1.
Table 1. Triangular IVP5F data for nonconformities.

Possible values of c IVP5FS
L U
L U
ðcL1 ; cM1 ; cU1 Þ l 1; l 1 ; v 1; v 1

cL cU
cL cU
ðcL2 ; cM2 ; cU2 Þ lc 2 ; lc 2 ; vc 2 ; vc 2

… …
L U
L U
ðcLm ; cMm ; cUm Þ lc m ; lc m ; vc m ; vc m
From Table
1,L theUarithmetic

L mean
of the possible values of c is calculated and the
operation min lc ; lc ; max vc ; vc
i i i Ui
is applied as given in Eq. (29).
Pm Pm Pm
i¼1 cLi i¼1 cMi i¼1 cUi

L U

e
c ¼ ; ; c ; max vc ; vc
; min lLc i ; lUi i i
ð29Þ
m m m
Equation (29) gives the triangular IVP5F center line of the control chart. Then, the
control limits can be given as in Eqs. (30) and (31).
Pm Pm Pm

L U

g ¼ ; ; ; min lLc i ; lU
UCL c ; max vc ; vc
i i i
m m m
Pm Pm Pm ð30Þ
i¼1 cLi cM cU
L U
1=2
3 ; i¼1 i ; i¼1 i ; min lLc i ; lU
c
i
; max v c
i
; v c
i
m m m
Pm Pm Pm

L U

e ¼
L CL ; ; ; min lLc i ; lU i
; max v i
; v i
c c c
m m m
Pm Pm Pm ð31Þ
L U
L U
1=2

3 ; ; ; min lc ; lc ; max vc ; vc
i i i i
m m m

L

L
Finding min lLc i ; lU

c
i
¼ lc min ; lU
c
min
and max vLc i ; vU
c
i
¼ vc max ; vU
c
max
, the control
limits are calculated as follows:
0 P ffiffiffiffiffiffiffiffiffiffiffiffiffiffi P
rP rffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm ffi Pm ffiffiffiffiffiffiffiffiffiffiffiffiffiffi! 1
rP
m m m m
c c c c c c
g ¼@B i¼1 Li
þ 3 i¼1 Li
; i¼1 Mi
þ 3 i¼1 Mi
; i¼1 Ui
þ 3 i¼1 Ui
;C
UCL m m m m m m A
L

lc ; lc ; vc ; vc
min Umin Lmax Umax
ð32Þ
0 P ffiffiffiffiffiffiffiffiffiffiffiffiffiffi P
rP rP ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P ffiffiffiffiffiffiffiffiffiffiffiffiffiffi! 1
rP
m m m m m m
c c c c c c
3 ; i¼1 3 ; i¼1 3 ;C
L L M M U i¼1 Ui
e B i¼1 i i¼1 i i i¼1 i i
L CL ¼ @ m m m m m m A
L

lc min ; lUc
min
; vc ;
Lmax Umax
v c
ð33Þ
After this calculation, the data in Table 1 are needed to be checked whether they are
within the control limits given in Eqs. (32) and (33).
4.2 IVP5F u Control Charts

Let the triangular IVP5F data for nonconformities for variable sample sizes be given as
in Table 2 and s be the inspection unit determined by the expert. Then, the modified
sample sizes become ni =s; i ¼ 1; 2; . . .; m and average defects per modified sample
sizes equal to ~ui ¼ ~ki =ðni =sÞ, i ¼ 1; 2; . . .; m.
Table 2. Triangular IVP5F data for nonconformities per inspection unit.

Number of nonconformities, e ki Sample size, ni e
ui
L1 U1
L1 U1

~k1 ¼ ðkL1 ; kM1 ; kU1 Þ; l ; l ; v ; v n1 ~k1 =ðn1 =sÞ
k k k k

~k2 ¼ ðkL2 ; kM2 ; kU2 Þ; lL2 ; lU2 ; vL2 ; vU2 n2 ~k2 =ðn2 =sÞ
k k k k
… … …

~km ¼ ðkLm ; kMm ; kUm Þ; lLm ; lUm ; vLm ; vUm nm ~km =ðnm =sÞ
k k k k
Average defect number based on the modified sample sizes is calculated using
Eq. (34).
Pm
~ki
e
u ¼ Pm i¼1 ð34Þ
i¼1 ðni =sÞ
e
u can be expressed by a triangular IVP5F number as in Eq. (35).
Pm ~ Pm
~k
Pm
~k
k
L U

e
u ¼ Pm i¼1 Li ; Pmi¼1 Mi ; Pmi¼1 Ui ; min lLu i ; lU
u ; max vu ; vu
i i i
ð35Þ
i¼1 ðni =sÞ i¼1 ðni =sÞ i¼1 ðni =sÞ
Equation (35) gives us the triangular IVP5F center line of the control chart. Then,
the variable control limits can be given as in Eqs. (36) and (37).
Pm ~ Pm
~k
Pm
~k
k
L U

gi¼
UCL Pm i¼1 Li ; Pmi¼1 Mi ; Pmi¼1 Ui ; min lLu i ; lU i
; max v i
; v i

u u u
Pm Pm Pm 1=2
~kL ~kM ~kU
L U

3 Pi¼1
m
i
; Pi¼1
m
i
; Pi¼1
m
i
; min lLu i ; lU i
; max v i
; v i
ni i¼1 ðni =sÞ ni i¼1 ðni =sÞ ni i¼1 ðni =sÞ u u u
ð36Þ
Pm ~ Pm
~k
Pm
~k
k
L U

e i¼ Pm i¼1 Li ; Pmi¼1 Mi ; Pmi¼1 Ui ; min lLu i ; lU
L CL u ; max vu ; vu
i i i

Pm Pm Pm 1=2
~kL ~kM ~kU
L U

3 Pi¼1
m
i
; Pi¼1m
i
; Pi¼1
m
i
; min lLu i ; lU
u ; max vu ; vu
i i i
ni i¼1 ðni =sÞ ni i¼1 ðni =sÞ ni i¼1 ðni =sÞ
ð37Þ

L

L
Finding min lLc i ; lU

c
i
¼ lc min ; lU
c
min
and max vLc i ; vU
c
i
¼ vc max ; vU
c
max
, the control
limits are calculated as follows:
Pm sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm ffi P sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm ffi P
~kL ~kL m ~
k Mi ~kM m ~
kU
g
UCL i ¼ Pm i¼1 i
þ3 Pm i¼1 i
; Pm i¼1
þ3 Pm
i¼1 i
; Pmi¼1 i þ
ð
i¼1 i n =s Þ n i i¼1 i ð n =s Þ ð
i¼1 i n =s Þ n i i¼1 ið n =s Þ i¼1 ðni =sÞ
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pm ffi
~kU
L

3 Pm i¼1 i
; lLu min ; lU min
; vu max ; vU max
ni i¼1 ðni =sÞ u u
ð38Þ
00 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi P sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi P 1
Pm Pm Pm
~kL ~kL m ~
k ~kM m ~
k
B @Pm i¼1 i
3 P i¼1 i
; Pm i¼1 M i
3 P i¼1 i
; Pm i¼1 U i
C
B ni m ni m C
B i¼1 ðni =sÞ i¼1 ðni =sÞ i¼1 ðni =sÞ i¼1 ðni =sÞ i¼1 ðni =sÞ C
B
e i¼B C
L CL s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi 1 C
B Pm C
B ~ L
L
C
@ P i¼1 kUi A A
3 m ; lu ; lu ; vu ; vu
min Umin max Umax
ni i¼1 ðni =sÞ
ð39Þ
After this calculation, the data in Table 3 are needed to be checked whether they are
within the control limits given in Eqs. (38) and (39).
Table 3. Triangular IVP5F data for nonconformities per modified inspection unit.
e
ui IVP5FS

L U
ðuL1 ; uM1 ; uU1 Þ lLu 1 ; lU

u ; vu ; vu
1 1 1

L U
ðuL2 ; uM2 ; uU2 Þ lLu 2 ; lU

u ; vu ; vu
2 2 2
… …

L U
ðuLn ; uMn ; uUn Þ lLu n ; lU

u ; vu ; vu
n n n
5 Application
A firm manufacturing TVs controls their defect numbers in the TVs they produced. The
triangular IVP5F defects numbers found in randomly selected 10 TVs are given in
Table 4.
Table 4. Triangular IVP5F defects numbers.

Sample number Triangular IVP5F defects numbers
1 ((2, 3, 4); ([0.70, 0.90], [0.20, 0.40]))
2 ((1, 2, 3); ([0.70, 0.90], [0.30, 0.40]))
3 ((4, 5, 6); ([0.80, 0.95], [0.10, 0.20]))
4 ((2, 3, 3); ([0.70, 0.95], [0.50, 0.60]))
5 ((1, 1, 1); ([0.85, 0.95], [0.05, 0.10]))
6 ((2, 3, 4); ([0.75, 0.80], [0.15, 0.25]))
7 ((2, 3, 5); ([0.90, 0.95], [0.20, 0.25]))
8 ((5, 6, 7); ([0.75, 0.90], [0.15, 0.50]))
9 ((2, 3, 5); ([0.80, 0.90], [0.30, 0.35]))
10 ((2, 2, 3); ([0.65, 0.95], [0.55, 0.70]))
From Table 4 average triangular IVP5F defects number is calculated as ((2.3, 3.1,
4.1); ([0.75, 0.80], [0.55, 0.70])).
The triangular IVP5F control limits are calculated as follows:
g = ((6.85, 8.38, 10.17); ([0.75, 0.80], [0.55, 0.70]))
UCL
e = ((−2.25, −2.18, −1.97); ([0.75, 0.80], [0.55, 0.70]))
L CL
Since, a negative LCL cannot be accepted, L CL e becomes as follows:
e
L CL = ((0, 0, 0); ([0.75, 0.80], [0.55, 0.70]))
The next step is to check if all sample values are within the control limits. Only the
sample 8 has an intersection with the upper control limit. The intersection level is very
close to zero and sample defect number is within the control limits. This indicates an
under-control process. If the intersection level was high, it would have been compared
with a critical value.
6 Conclusion
Defects control charts are one of the most important tools in statistical process control
which are used to monitor the process location and variation and decide if the process is
under control or not. Since the data in defects control charts and the product quality
attributes involve a certain amount of uncertainty and ambiguity by their relativistic
natures, fuzzy defects control charts provide more flexibility for evaluating if an item is
conforming or nonconforming. In this study IVP5F defects control charts have been
developed as a special case of q-rung orthopair fuzzy sets and an application on a TV

production system is illustrated as an example. For further research other control charts
such as EWMA control charts and CUSUM control charts are suggested to be
developed by the new extensions of intuitionistic fuzzy sets.
References
1. Bradshaw Jr., C.W.: A fuzzy set theoretic interpretation of economic control limits. Eur.
J. Oper. Res. 13(4), 403–408 (1983)
2. Wang, J., Raz, T.: On the construction of control charts using linguistic variables. Int.
J. Prod. Res. 28(3), 477–487 (1990)
3. Taleb, H., Limam, M.: On fuzzy and probabilistic control charts. Int. J. Prod. Res. 40(12),
2849–2863 (2002)
4. Gülbay, M., Kahraman, C., Ruan, D.: a-Cut fuzzy control charts for linguistic data. Int.
J. Intell. Syst. 19(12), 1173–1195 (2004)
5. Gülbay, M., Kahraman, C.: An alternative approach to fuzzy control charts: direct fuzzy
approach. Inf. Sci. 177(6), 1463–1480 (2007)
6. Sentürk, S., Erginel, N., Kaya, I., Kahraman, C.: Design of fuzzy ũ control charts. J. Mult.-
7. Sakthivel, E., Kannan, S.K., Logaraj, M.: Application of fuzzy logic approach in statistical
control charts. Global Stoch. Anal. 4(1), 139–147 (2017)
8. Tekşen, H.E., Anagün, A.S.: Interval type-2 fuzzy c-control charts using ranking methods.
Hacettepe J. Math. Statist. 48(2), 510–520 (2019)
9. Aslam, M., Bantan, R.A.R., Khan, N.: Design of a new attribute control chart under
neutrosophic statistics. Int. J. Fuzzy Syst. 21(2), 433–440 (2019)
10. Ercan-Teksen, H., Anagün, A.S.: Intuitionistic fuzzy C-control charts using fuzzy
comparison methods. In: Kahraman, C., Cebi, S., Cevik, O.S., Oztaysi, B., Tolga, A.,
Sari, I. (eds.) Intelligent and Fuzzy Techniques in Big Data Analytics and Decision Making,
INFUS 2019, Advances in Intelligent Systems and Computing, vol. 1029, pp. 1161–1169.
11. Gülbay, M., Kahraman, C.: Development of fuzzy process control charts and fuzzy unnatural
pattern analyses. Comput. Stat. Data Anal. 51, 434–451 (2006)
12. Yager, R.: Generalized orthopair fuzzy sets. IEEE Trans. Fuzzy Syst. 25, 1222–1230 (2017)
Fuzzy AHP Based Prioritization
and Taxonomy of Information Quality Factors
in Accounting Information Systems
Rozi Mizrahi(&), Berna Tektaş Aygülen, and Kayahan Karakul

Izmir Katip Celebi University, Cigli, Izmir, Turkey
{rozi.mizrahi,berna.tektas,
aygulen.kayahan}@ikcu.edu.tr
Abstract. Enterprise Resource Planning (ERP) systems undertake the task of

producing the data they need on time and reliably. Accounting Information
Systems (AIS) is located in the centre of ERP Architecture and AIS, generates
data on the financial status of businesses. The information produced by AIS is
used for decision making by internal and external stakeholders. In this context,
the quality of the information produced by AIS also affects the quality of the
decisions made by stakeholders. The main objective of this study is to develop
the prioritization-based taxonomy of the AIS’s information quality factors using
the fuzzy Analytical Hierarchy Process (AHP) approach. The information pro-
duction processes of AIS directly affect the quality of information produced. For
this reason, critical factors affecting the information quality were determined by
focusing on the information production processes of AIS. Nineteen information
quality factors were extracted from the literature. Then, Multi-Criteria Decision
Making method fuzzy AHP is used to prioritize and develop the taxonomy of
the identified factors. The implications of fuzzy AHP approach are novel in this
research area as it has been successfully used previously in different other
domains e.g. electrical and electronics, supplier selection, agile software
development and personnel selection.
Keywords: Accounting information systems (AIS) Information quality

factors Fuzzy AHP
1 Introduction
In competitive business environment nowadays, it is crucial for enterprises to establish

their strategies for survival. In this respect, knowledge is acting a pioneering role
constituting the essential source of strategies. This knowledge can be gained through
information management systems in a way to use it as a strategic weapon in com-
petitive market [1].
Enterprise Resource Planning (ERP) systems undertake the task of producing the
data they need on time and reliably. Accounting information systems (AIS) is located
in the centre of ERP and it generates new data for other information systems. Com-
panies benefit from AIS in decision making processes on financial issues. The main
1408 R. Mizrahi et al.
purpose of AIS is to provide high quality accounting information and create a trans-
parent financial reporting environment for users. The quality of accounting information
can be defined as the truthfulness, sensitivity and accuracy of the information in the
financial statements [2]. Quality accounting knowledge is critical for users to make
timely, realistic and accurate decisions.
The aim of this research is to find the relative weights of criteria of the information
based quality factors in AIS. The information quality factors has been identified in the
literature but not prioritized yet and their relative weights has not determined before,
this will be originality of the paper. The prioritization and taxonomy.
In order to increase the quality of the information produced by AIS, critical factors
affecting the information quality must be defined [3]. As a result of the literature review
in the field of quality and information quality management, nineteen critical factors
were identified and these factors were grouped in four main groups. Then, fuzzy Multi
Criteria Decision Making (MCDM) technique, Fuzzy AHP is used to determine the
relative importance levels of these factors. The importance of factors affecting AIS
information quality in decision makers (DMs) has determined.
The objective of this study is to determine the relative weights of quality factors in
AIS by using Fuzzy AHP method. There are studies, which determine the quality
factors of AIS, in the literature. The originality of this paper is that; MCDM techniques
is used to determine the relative weights of AIS quality factors according to DMs
views.
This paper is organized into five sections. The conceptual framework of this paper
is explained in Sect. 2. The methodology of paper which is Fuzzy AHP (Analytical
Hierarchy Process), is displaying in Sect. 3. Then, Sect. 4. provides the results of
Fuzzy AHP. Finally, the conclusions and the future work suggestions of the study are
2 Conceptual Framework
Information; is the data arranged in such a way that it can be understood and used by
people [4]. Information acquired at the end of accounting processes is also defined as
accounting information. Accounting information is provided from AIS.
Accounting information is closely related to the internal (partners, managers,
employees) and external (government, lenders, customers, suppliers, competitors,
potential partners etc.) stakeholders of the business. These stakeholders use their
accounting information as data to shape their business decisions. The quality of the
accounting information used in decision making directly affects the quality of the
decisions made [5]. In this context, it is necessary to identify critical factors that affect
the quality of accounting information.
In the literature, there are many studies conducted to determine quality factors in
total quality management and information quality management. XU(5) conducted a
research in Australia in 2002 in which factors determining the information quality of
AIS are reported. In similar vien, Acar [3] extended the XU(5)’s findings to document
the critical success factors of AIS.
Fuzzy AHP Based Prioritization and Taxonomy of Information Quality Factors 1409
In this study; AIS information quality factors determined by Acar and XU are
divided into four main groups. These are organizational factors, intellectual factors,
technological factors and external factors. Organizational Factors can be classified as
Institutionalization level, physical environment, flexibility of organization, senior
management support, quality culture of organization and internal control system.
Intellectual Factors can be classified as education level, competency of staff, teamwork,
loyalty to job and performance evaluation/rewarding. Technological factors can be
classified infrastructure of knowledge technology, software supporting systems, R&D
activities and improvement, security of knowledge. External factors can be classified as
customer relations, market conditions, supplier relations and legal institutions and
regulations.
3 Methodology
In AHP method complex problems can be divided to simplest sub-modules by

expressing the problem with a hierarchical structure [6]. Linguistic variables are used to
reveal the opinions of Decision Makers (DMs) and the method is based on weighting
the criteria and alternatives then ranking them. It is possible to compare the quantitative
variables with the pairwise comparison scale of 1 to 9. Fuzzy extension of method
enables to cope with the vagueness in the opinions of DMs which is naturally exist in
the decision making problems, by using fuzzy numbers instead of crisp ones.
3.1 Chang’s Fuzzy AHP Algorithm

Extended Fuzzy AHP analysis is developed by Chang [7] as is explained in the fol-
lowing steps. Let X ¼ fx1 ; x2 ; . . .; xn g is a set of criteria and U ¼ fu1 ; u2 ; . . .; un g is a
set of goals, every object is associated with each goal. So there are m Extent value for
each object with the following notation (1).
Mg1; ; Mg2; ; . . .; Mgm; i ¼ 1; 2; . . .; n ð1Þ
Where all Mgj; ðj ¼ 1; 2; . . .; mÞ are TFNs. ith fuzzy Extent value is calculated by the
formula given in the Eq. (6)
Xm hXn Xm i1
Si ¼ Mj
j¼1 g; i¼1
Mj
j¼1 g;
ð2Þ
Pm
Fuzzy Addition operation given in the Eq. (3) is applied to obtain j¼1 Mgj; in Eq. (2)
where (lj ; mj ; uj ) are TFNs.
Xm Xm Xm Xm
Mj ¼
j¼1 g;
l;
j¼1 j
m;
j¼1 j
u
j¼1 j
ð3Þ
Second vector in Eq. (2) can be calculated by the formula in Eq. (4).
Xn Xm Xn Xn Xn
i¼1
Mj ¼
j¼1 g;
l;
i¼1 j
m;
i¼1 j
u
i¼1 j
ð4Þ
The reciprocal of vector in the Eq. (4) can be calculated by using the formula in
Eq. (5).
hXn Xm i1
1 1 1
Mj
j¼1 g;
¼ Pn ; Pn ; Pn ð5Þ
i¼1
i¼1 ui i¼1 mi i¼1 li
For defuzzification, i.e. assigning a real number for a fuzzy number centroid method
will be used. This method is based on the gravity center of TFNs, which the normalized
formula is given in the Eq. (6).
Si Sl þ Sm þ Su
w i ¼ Pn ¼ Pn i ¼ 1; . . .; n ð6Þ
i¼1 Si i¼1 Si
Fuzzy Linguistic Scale that is used to transform the opinions of DMs to TFNs is given
in Table 1.
Table 1. Fuzzy transformation scale [1]

Just equal (1, 1, 1) (1/1, 1/1, 1/1)
(1, 2, 3) (1/3, 1/2, 1/1)
Weakly important (2, 3, 4) (1/4, 1/3, 1/2)
(3, 4, 5) (1/5, 1/4, 1/3)
Strongly more important (4, 5, 6) (1/6, 1/5, 1/4)
(5, 6, 7) (1/7, 1/6, 1/5)
Very strong more important (6, 7, 8) (1/8, 1/7, 1/6)
(7, 8, 9) (1/9, 1/8, 1/7)
Absolutely more important (8, 9, 9) (1/9, 1/9, 1/8)
4 Application
It is aimed to find the weights of the criteria of quality factors in AIS. The quality
factors are determined according to literature review as grouped into 4 main factors as
Organizational Factors, Intellectual factors, Technological Factors an External Factors.
The sub criteria are defined as in the following also the hierarchic structure is given in
the Fig. 1. The hierarchic structure of the problem is given in the Fig. 1.
The pairwise comparison scale is applied to the expert whose profession is on
quality factors in AIS. The pairwise comparison of main criteria and the calculated
fuzzy weights are given in the Table 2.
The pairwise comparison of subcriteria and the calculated fuzzy weights are given
in the Table 3, 4, 5 and 6.
Aim: Quality Factors in

AIS
C1:Organisational Factors C2.Intellectual Factors C3.Technological Factors C4.External Factors
C1.1.Instutionalisation C3.1.Infrastructure of
C2.1.Education Level C4.1.Customer Relations
Level Knowledge Technology
C3.2.Software supporting
C1.2.Physical Environment C2.2.Competency of staff C4.2.Market conditions
systems
C1.3.Flexibility of C3.3.R&D activities and

C2.3.Team work C4.3.Supplier Relations
Organisation improvement
C1.4.Senior Management C3.4.Security of C4.4.Legal Instutions and

C2.4.Loyalty to job
Support Knowledge regulations
C1.5.Quality Culture of C2.5.Performance

Organisation evaluation and rewarding
C1.6.Internal Control
System
Fig. 1. The hierarchy of quality factors in AIS.
Table 2. Pairwise comparison of main criteria and fuzzy weights

C1 C2 C3 C4 Fuzzy weights
C1 (1, 1, 1) (1/4, 1/3, 1/2) (1/7, 1/6, 1/5) (2, 3, 4) (0.06, 0.13, 0.21)
C2 (2, 3, 4) (1, 1, 1) (1/4, 1/3, 1/2) (4, 5, 6) (0.13, 0.28, 0.42)
C3 (5, 6, 7) (2, 3, 4) (1, 1, 1) (7, 8, 9) (0.26, 0.54, 0.77)
C4 (1/4, 1/3, 1/2) (1/6, 1/5, 1/4) (1/9, 1/8, 1/7) (1, 1, 1) (0.03, 0.05, 0.69)
In the last stage the final weights of subcriteria are calculated by production rule of
TFNs by production of the global weights with criteria’s weights. The results are
defuzzificated and normalized. As a result, weights of main criteria and all subcriteria
are given in the Fig. 2 and Fig. 3.
Table 3. Pairwise comparison of organizational factors and fuzzy weights

C1.1 C1.2 C1.3 C1.4 C1.5 C1.6 Fuzzy
weights
C1.1 (1, 1, 1) (2, (1/5, (1, 1, 1) (1/4, (1/5, (0.06,
3, 4) 1/4, 1/3) 1/3, 1/2) 1/4, 1/3) 0.09, 0.13)
C1.2 (1/4, (1, (1/6, (1/5, (1/6, (1/8, (0.02,
1/3, 1/2) 1, 1) 1/5, 1/4) 1/4, 1/3) 1/5, 1/4) 1/7, 1/6) 0.03, 0.05)
C1.3 (3, 4, 5) (4, (1, 1, 1) (1, 1, 1) (1/4, (1/6, (0.12,
5, 6) 1/3, 1/2) 1/5, 1/4) 0.17, 0.26)
C1.4 (1, 1, 1) (3, (1, 1, 1) (1, 1, 1) (1/4, (1/6, (0.08,
4, 5) 1/3, 1/2) 1/5, 1/4) 0.11, 0.16)
C1.5 (2, 3, 4) (4, (2, 3, 4) (2, 3, 4) (1, 1, 1) (1, 1, 1) (0.15,
5, 6) 0.24, 0.37)
C1.6 (3, 4, 5) (6, (4, 5, 6) (4, 5, 6) (1, 1, 1) (1, 1, 1) (0.24,
7, 8) 0.35, 0.51)
Table 4. Pairwise comparison of intellectual factors and fuzzy weights

C2.1 C2.2 C2.3 C2.4 C2.5 Fuzzy
weights
C2.1 (1, 1, (1/6, 1/5, (1/5, 1/4, (1/8, 1/7, (1/7, 1/6, (0.03, 0.04,
1) 1/4) 1/3) 1/6) 1/5) 0.05)
C2.2 (4, 5, (1, 1, 1) (1, 1, 1) (1/4, 1/3, (1/5, 1/4, (0.11, 0.16,
6) 1/2) 1/3) 0.24)
C2.3 (3, 4, (1, 1, 1) (1, 1, 1) (1/3, 1/2, (1/5, 1/4, (0.10, 0.14,
5) 1/1) 1/3) 0.23)
C2.4 (6, 7, (2, 3, 4) (1, 2, 3) (1, 1, 1) (1, 2, 3) (0.19, 0.32,
8) 0.51)
C2.5 (5, 6, (3, 4, 5) (3, 4, 5) (1/3, 1/2, (1, 1, 1) (0.22, 0.33,
7) 1/1) 0.51)
Table 5. Pairwise comparison of technological factors and fuzzy weights

C3.1 C3.2 C3.3 C3.4 Fuzzy weights
C3.1 (1, 1, 1) (1/4, 1/3, 1/2) (1/5, 1/4, 1/3) (1/4, 1/3, 1/2) (0.05, 0.08, 0.13)
C3.2 (2, 3, 4) (1, 1, 1) (2, 3, 4) (1/3, 1/2, 1/1) (0.17, 0.30, 0.54)
C3.3 (3, 4, 5) (1/4, 1/3, 1/2) (1, 1, 1) (1/5, 1/4, 1/3) (0.14, 0.22, 0.37)
C3.4 (2, 3, 4) (1, 2, 3) (3, 4, 5) (1, 1, 1) (0.22, 0.40, 0.70)
Table 6. Pairwise comparison of external factors and fuzzy weights

C4.1 C4.2 C4.3 C4.4 Fuzzy weights
C4.1 (1, 1, 1) (3, 4, 5) (1, 1, 1) (1/7, 1/6, 1/5) (0.14, 0.17, 0.28)
C4.2 (1/5, 1/4, 1/3) (1, 1, 1) (1/4, 1/3, 1/2) (1/7, 1/6, 1/5) (0.04, 0.06, 0.08)
C4.3 (1, 1, 1) (2, 3, 4) (1, 1, 1) (1/6, 1/5, 1/4) (0.11, 0.17, 0.24)
C4.4 (5, 6, 7) (5, 6, 7) (4, 5, 6) (1, 1, 1) (0.41, 0.58, 0.81)
0.50
0.45
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
OrganisaƟonal Intellectual Technological External
Fctors Factors Factors Factors
Fig. 2. Weights of main criteria
0.2
0.15
0.1
0.05
Fig. 3. Weights of information quality factors in AIS
5 Conclusion
The qualitative variables are easily evaluated to form the relative weights of criteria by
AHP, also the fuzzy environment satisfied a more realistic view of DM. Moreover
linguistic statements has been quantified by the transformation Scale of AHP.
Technological factor is found to be the most important criterion. In addition to this

result, it is observed that the relative importance of intellectual and external factors
follow the technological factor. According to DMs, the necessary technology invest-
ments must be made first to provide quality accounting information. The technology
factor includes ERP software, hardware, R&D activities and information security. ERP
software are extremely expensive programs. The allocation of the necessary budget for
the purchase of these technologies has been defined as the most important quality factor.
According to DMs, intellectual factors such as the level of education, competence
of staff, teamwork, loyalty to work and performance evaluation/rewarding are among
the second important criteria in determining the quality of accounting information.
External factors follow the intellectual factors in this respect. Organizational factors
have less impact on the quality of accounting information according to DMs.
According to DMs, security of knowledge is the most critical AIS information
quality subcriteria. The second important quality subcriteria is software support systems.
These two criteria are located under the heading of technology main criteria. According
to the DMs, the legal institutions and their regulations, which are included in the external
factors, are the third important subcriteria in determining the quality of AIS information.
Recently, scandals in the capital markets have caused investors’ loss of confidence.
This has also led to increased discussions on the quality of accounting information and
financial reports. In order to meet the funding needs of the capital markets, it is
imperative that investors’ trust be regained. The way of gaining trust is to increase the
quality of accounting information. In order to increase the quality of the accounting
information produced by AIS, the factors affecting the quality should be focused on by
considering their relative weight. Focusing on these factors will contribute to organi-
zations in improving information quality and increasing their effectiveness.
For further research the other MCDM techniques can be used to evaluate the
relative weights of the quality factors of AIS.
References
1. Demirhan, D.: İşletmelerde stratejik bilgi sistemleri yönetimi ve rekabet üstünlüğü elde
edilmesindeki rolü. Ege Akademik Bakış 2(2), 117–124 (2002)
2. Chen, H., Tang, O., Jiang, Y., Lin, Z.: The role of international financial reporting standards
in accounting quality: evidence from the EU. J. Int. Financ. Manag. Account. 21(3), 220–278
(2010)
3. Sürmeli, F., Erdoğan, M., Erdoğan, N., Banar, K., Önce, S.: Muhasebe Bilgi Sistemi Anadolu
Üniversitesi Yayınları, Eskişehir (1998)
4. Acar, D., Özçelik, H.: Muhasebe bilgi kalitesini etkileyen kritik başarı faktörleri. Muhasebe
ve Finansman 49, 10–23 (2011)
5. Xu, H., Koronios, A., Brown, N.: Managing Data Quality in Accounting Information
Systems. Idea Group Inc. (2002)
6. Kahraman, C.: A brief literature review on fuzzy AHP. Int. J. Anal. Hierarchy Process. 10(2),
293–297 (2018)
7. Chang, D.Y.: Applications of the extent analysis method on fuzzy AHP. Eur. J. Oper. Res. 95
(3), 649–655 (1996)
Analytical Techniques to Compute Cp
and Cpm Capability Indices by R Software
Abbas Parchami(B)
Department of Statistics, Faculty of Mathematics and Computer, Shahid Bahonar

University of Kerman, Kerman, Iran
[email protected]
Abstract. In statistical quality control, as in other statistical problems,

we may be confronted with fuzzy concepts. This paper deals with the
problem of process capability estimation, when the observation and the
specification limits are fuzzy rather than crisp. In other words, this paper
illustrate how a researcher can use “FuzzyNumbers” Package in R soft-
ware to exactly plot the membership function of capability indices based
on fuzzy data and crisp specification limits.
Keywords: Quality control · Process capability index · Fuzzy data ·

Specification limits
1 Introduction
In process improvement efforts, the process capability index (process capabil-
ity ratio) is a statistical measure of process capability: the ability of a process
to produce output within specification limits. For details on the classical pro-
cess capability indices see the original studies in literature review [3,4]. More-
over, see [7–9] to review some researches on process capability indices in fuzzy
environment.
In follow, a shortened review about process capability indices is presented
where data/observation are considered fuzzy rather than crisp. Lee et al. gener-
alized the capability index Cp by extension principles based on fuzzy specifica-
tions and fuzzy data [6]. Under a similar conditions, Lee follows his approach to
generalize capability index Cpk [5]. Then, he fuzzified the proposed fuzzy PCI for
making final decision in the examined manufacturing process. A similar approach
to solve this problem based on extension principle presented by Shu and Wu by
fuzzy data [10]. In their approach, the α-cuts of fuzzy index Cpk was calculated
based on the α-cuts of fuzzy data for 0 ≤ α ≤ 1. Also, they investigated on the
capability of the LCD monitors assembly line using their generalized indices. In
this regard, the capability test on the generalized capability index Cp with fuzzy
data have been investigated in [11] by Tsai and Chen.
1416 A. Parchami
2 Process Capability Indices

Let X is the univariate quality characteristic and also its expected value and
standard deviation are denoted by μ and σ, respectively. Two common capability
indices are
U SL − LSL
Cp = (1)
6σ
and
U SL − LSL
Cpm = (2)
6 σ 2 + (μ − T )2
where T is the target value and LSL and U SL are upper and lower specification
limits respectively [1,2]. Substituting the sample mean and standard deviation
provides a point estimate for the indices (1) and (2).
3 Capability Index Cp Based on Fuzzy Data
The main problem in this paper is plotting the membership functions of capabil-
ity indices Cp and Cpm in R software where the observed sample data x1 , ..., xn
are considered non-precise/fuzzy rather than crisp. Consider the following tri-
angular and trapezoidal fuzzy numbers as the observed random sample
x̃1 = T (0.6, 0.7, 0.85, 1.1), x̃2 = T (0.1, 0.2, 0.2, 0.3), x̃3 = T (0.2, 0.3, 0.4, 0.5),
x̃4 = T (0.4, 0.5, 0.5, 1), x̃5 = T (0.7, 0.9, 0.9, 1.2), x̃6 = T (0.2, 0.6, 0.6, 1).
After loud “FuzzyNumbers” package, we start by creation a list of fuzzy data
in R software.
library(FuzzyNumbers)
data = matrix( c(.6, .7, .85,1.1,
.1, .2, .2, .3 ,
.2, .3, .4, .5 ,
.4, .5, .5, 1 ,
.7, .9, .9, 1.2,
.2, .6, .6, 1 ), ncol=4, byrow=TRUE)
n = dim(data)[1]
FD = c()
for(i in 1:6)
FD = c(FD, as.PiecewiseLinearFuzzyNumber(TrapezoidalFuzzyNumber(
data[i,1],data[i,2],data[i,3],data[i,4]), knot.n=20))
FD[[2]]
for(i in 1:6)
plot( FD[[i]], lwd=2, xlim=c(0,1.25), ylim=c(-0.1,1.1), add = (i != 1),
col=’gray43’, xlab=NA, ylab=NA )
T = 0.45; abline(v=T, col=3, lw=2, lty=5)

LSL = 0.17; abline(v=LSL, col=2, lw=2, lty=3)
USL = 1.0; abline(v=USL, col=2, lw=2, lty=3)
Techniques to Compute Capability Indices 1417
1.0
0.8
0.6
0.4
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0 1.2
Fig. 1. Membership functions of fuzzy data (black line), specification limits (red points)
and target value (green dash).
Fuzzy mean, fuzzy variance and fuzzy standard deviation of fuzzy data
x̃1 , ..., x̃6 are introduced in R by following cods and then the membership func-
tions of these statistics are plotted in Fig. 2 (Fig. 1).
x_bar = (FD[[1]]+FD[[2]]+FD[[3]]+FD[[4]]+FD[[5]]+FD[[6]]) / n
s2 = (1/n) * ( ( FD[[1]] - x_bar )^2 +
( FD[[2]] - x_bar )^2 +
( FD[[3]] - x_bar )^2 +
( FD[[4]] - x_bar )^2 +
( FD[[5]] - x_bar )^2 +
( FD[[6]] - x_bar )^2 )
s = fapply( s2, function(x) sqrt(x) )
plot(x_bar, col=1, lwd=3, lty=1, xlim=c(-0.02,1.03), ylim=c(-0.1,1.1))

plot(s2, col=2, lwd=3, lty=2, add=TRUE)
plot(s, col=3, lwd=3, lty=3, add=TRUE)
Now, one can easily plot the membership function of Cp index in Fig. 3 based
on fuzzy data.
Cp = fapply( s, function(x) (USL-LSL)/(6*x) )
plot(Cp, col=1, lwd=3, lty=1, xlim=c(-0.2,5))
1418 A. Parchami
1.0
0.8
0.6
0.4
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0
Fig. 2. Membership functions of fuzzy mean (line), fuzzy variance (dash) and fuzzy
standard deviation (point) of fuzzy data.
1.0
0.8
0.6
0.4
0.2
0.0
0 1 2 3 4 5
Fig. 3. Membership function of fuzzy capability index Cp based on fuzzy observation.
4 Capability Index Cpm Based on Fuzzy Data

In this section, we are going to plot the membership function of index Cpm
based on fuzzy data. Considering Eqs. (1) and (2), one can use from Cpm = k Cp
relation, where k = 1μ−T 2 , to introduce Cpm index in R for more complicate
1+( σ )
calculations (Fig. 4).
Techniques to Compute Capability Indices 1419
k = fapply( ((x_bar - T)/s)^2 , function(x) (1+x)^-0.5 )

Cpm = k * Cp
Cpm
plot(Cp, col=’gray80’, lwd=2, lty=1, xlim=c(0,2))
plot(Cpm, col=1, lwd=3, lty=2, add=TRUE)
1.0
0.8
0.6
0.4
0.2
0.0
0.0 0.5 1.0 1.5 2.0
Fig. 4. Membership functions of fuzzy capability indices Cp (gray line) and Cpm (black
dash) based on fuzzy observation.
5 Conclusions and Suggestions

Although the classical arithmetic operations can extended by the extension prin-
ciple approach, but the complexity of this principle causes some computational
challenges. To solve these computational challenges, R package “FuzzyNumbers”
has been published on CRAN in 2015. This paper has been tried to briefly explain
how one can introduce capability indices in R software by package “FuzzyNum-
bers” when the observed data are fuzzy rather than crisp.
For future research work, one can try to use the same approach of this paper
to extend other capability indices such as Cpk and Cpmk , where the observations
are fuzzy numbers rather than being real numbers.
References
1. Hsiang, T.C., Taguchi, G.: Tutorial on quality control and assurance-The Taguchi
methods. Joint Meetings of the American Statistical Association, Las Vegas,
Nevada, p. 188 (1985)
1420 A. Parchami
2. Juran, J.M.: Juran’s Quality Control Handbook, 3rd edn. McGraw-Hill, New York
(1974)
3. Kotz, S.: Process Capability Indices. Chapman and Hall, New York (1993)
4. Kotz, S., Johnson, N.: Process capability indices - a review, 1992–2000. J. Qual.
Technol. 34, 2–19 (2002)
5. Lee, H.T.: Cpk index estimation using fuzzy numbers. Eur. J. Oper. Res. 129,
683–688 (2001)
6. Lee, Y.H., Wei, C.C., Chang, C.L.: Fuzzy design of process tolerances to maximize
process capability. Int. J. Adv. Manuf. Technol. 15, 655–659 (1999)
7. Parchami, A., Sadeghpour Gildeh, B.: Trends on process capability indices in fuzzy
environment. In: Kahraman, C., Yanik, S. (eds.) Intelligent Decision Making in
Quality Management, vol. 97, Chap. 5, pp. 127–140. Springer, Switzerland (2016)
8. Parchami, A., Mashinchi, M., Sharayei, A.: An effective approach for measuring
the capability of manufacturing processes. Prod. Plan. Control. 21(3), 250–257
(2010)
9. Parchami, A., Sadeghpour Gildeh, B., Nourbakhsh, M., Mashinchi, M.: A new
generation of process capability indices based on fuzzy measurements. J. Appl.
Stat. 41(5), 1122–1136 (2014)
10. Shu, M.H., Wu, H.C.: Quality-based supplier selection and evaluation using fuzzy
data. Comput. Ind. Eng. 57, 1072–1079 (2009)
11. Tsai, C.C., Chen, C.C.: Making decision to evaluate process capability index Cp
with fuzzy numbers. Int. J. Adv. Manuf. Technol. 30, 334–339 (2006)
Heuristics
Proposal of Genetic Algorithm Approach
for Solving Single Machine Scheduling
Problem Under Learning Effect
Derya Deliktas(B) and Mustafa Urhan
Faculty of Engineering, Department of Industrial Engineering,

Kütahya Dumlupınar University, Kütahya, Turkey
Abstract. This paper addresses a genetic algorithm approach for solv-

ing a single machine scheduling problem considering the learning effect
on sequence-dependent setup times. The objective of the proposed model
is to minimize the makespan. Scheduling problems are referred to as
the NP-hard. Since genetic algorithm parameters influence the result of
the solution, the selection of convenient design value for reaching global
optimum includes a notable impact on the performance of the genetic
algorithm. Therefore, a full experimental design is conducted to get the
optimal parameter combination design in the genetic algorithm. After
determining the suitable parameter combinations, a decision support sys-
tem is created to achieve optimal or near-optimal job-sequence within the
automotive industry.
Keywords: Single machine scheduling problem · Learning effect · Full

experimental design · Decision support system · Genetic algorithm
1 Introduction
Scheduling includes a crucial role within the developing industries. In several
industries, the production system can involve a machine bottleneck that affects
all jobs. Since the management of this bottleneck causes delays in customer
order delivery, the optimal or near-optimal scheduling subject is very important
for the industries [1]. Single machine scheduling problem is a simple version of
scheduling problems and has been mostly examined in the literature [2]. In most
of the classic scheduling problem, the process times are considered as constant.
If a worker processes similar jobs over and over again, a job will have less of
this processing time due to the repeated processing and the accumulation of
experience. This is defined as the learning effect in the literature [3].
Setup time can be considered that has two different types such as sequence-
independent and sequence-dependent. Although the setup time depends only
Supported by The Scientific and Technological Research Council of Turkey (TÜBİTAK)
under the application number 1139B411802016.
1424 D. Deliktas and M. Urhan
on the job to be processed in sequence-independent setup time, the setup time

depends not only on the job to be scheduled but also on the immediately preced-
ing job in sequence-dependent setup time [4]. Many scheduling studies, which
have been intensely addressed in the literature, take into account sequence-
dependent setup times, and these times may be evaluated separately from pro-
cessing times [5–10]. In addition to these, recently, there are many studies with
the learning effect both on the processing times [11–18] and on the past sequence-
dependent setup times [19–23] in the scheduling problems.
In this study, it is analyzed a single machine scheduling problem under a
learning effect on sequence-dependent setup times with minimization makespan.
There are very few papers that examine both sequence-dependent setup times
and learning effect [11]. To the best of our knowledge, such a case study con-
sidered the learning effect on the sequence-dependent setup times has not been
done before. Thus, it is presented a new mixed-integer linear program (MILP)
with learning time on sequence-dependent setup times and a genetic algorithm
approach is developed because it is an NP-hard problem [2].
The organization of this study is given as follows: Sect. 2 provides problem
description while the problem modeling, assumptions, notations, mathematical
model, while Sect. 3 introduces the genetic algorithm approach used to solve the
problem. Section 4 provides the experimental results and analysis. Conclusions
and suggestions for future researchers are presented in Sect. 5.
In this study, we consider a case study in an automotive spare part com-
pany. The mechanical press machine in this industry is a machine with the
most processing and a bottleneck machine. The job sequence on this bottleneck
machine is scheduled. There are a set of 15 independent, non-preemptive jobs,
J = {J1 , J2 , · · · , J15 }, and a single machine that can process these jobs. Every
job is exposed to a setup time before being processed on the machine. The setup
time depends on the previous job processed on the machine. Since the setup
processes are repeated by a worker, setup time is decreased as a function of
the next position with learning effect [24]. The proposed model considers some
assumptions as follows: (i ) The machine can be processed just one job at a time,
(ii ) The machine and jobs are available at time zero, (iii ) The processing time
of each job is known in advance and fixed, (iv ) After a job has started on a
machine, it cannot be interrupted until it is completed, and (v ) All jobs have
the same importance.
The problem is formulated as:
M in Cmax (1)
Cj + M (1 − Xjr ) ≥ pj + s0j rα ∀j, r = 1 (2)
Cj − Ci + M (1 − Yijr ) ≥ pj + sij rα ∀r > 1, i, j, i = j (3)

Single Machine Scheduling Problem Under Learning Effect Based on GA 1425
Notations
i Index for jobs i = 1, 2, . . . , n
r Index for positions r = 1, 2, . . . , n
Parameters
pj Processing time of the job j
sij Sequence-dependent setup time that depends on the job i processed just before the job j
α Learning index; α = log2 LE ≤ 0
M A large positive number
Decision variables
Xjr 1, if the job j is assigned to the r th position; 0, otherwise
Yijr 1, if the job i is assigned to the r th position and precedes the job j ; 0, otherwise
Cjr Completion time of the job j in position r
Cmax Makespan
1 + Yijr ≥ Xi(r−1) + Xjr ∀r > 1, i, j, i = j (4)

n
Xjr ∀r (5)
j=1

n
Xjr ∀j (6)
r=1
s0j = 0 ∀j (7)
Cmax ≥ Cj ∀j (8)
Cj ≥ 0, Xjr ∈ {0, 1} , Yijr ∈ {0, 1} ∀i, j, r (9)

Equation (1) determines the objective function with the minimization of the
makespan. Equation (2) calculates the completion time of the job that is assigned
to the first position while Eq. (3) computes the completion times of the remaining
job in the k th (k > 1) position. Equation (4) ensures that Yijr is zero or one.
Equations (5) and (6) ensure that each job can only be assigned to one position
and only one job can be arranged per position. Equation (7) guarantees that
setup time of the first position is zero. Equation (8) calculates the makespan.
Equation (9) defines positive and binary variables.
As mentioned in Sect. 1, obtaining a solution with solving the above mathe-
matical model is not practical for the case study.
3 Genetic Algorithm Approach

Genetic algorithms (GAs) developed by Holland [25] and popularized by
Goldberg [26] are adaptive methods that can be used to solve complex com-
binatorial optimization problems. This section proposes a GA approach to solve
the problem because the MILP cannot obtain the optimum or near-optimum
solution for the problems over 10 jobs in reasonable computation time [27]. The
details of our GA are introduced in the following sections.
3.1 Chromosome Representation and Initial Population

Chromosome representation is the most important part of GA. Permutation
coding is used as a natural chromosomal representation for the problem. In
this type of representation, n represents the number of jobs, while the series in
chromosomes are represented in the genes at length n, the value of which can be
assigned from 1 to n. In this study, for each chromosome, 15 numbers are defined
randomly and without repetition. The initialization is randomly generated in this
stage of design.
3.2 Genetic Operators

The performance of GAs is affected by the effectiveness of genetic operators.
Due to space limitations, the information related to the used operators is given
briefly.
Selection. The tournament selection method is used to execute the selection

operator. In tournament selection, individuals are chosen randomly from the
population according to tournament size, and the best individual from this group
is chosen as a parent [28].
Crossover. Crossover operators are very important to generate new individuals

for the next generation. In this paper, the partially matched crossover (PMX) [29]
is used to ensure the feasibility of the solutions by considering CrossRate, which
is the crossover probability. Figure 1 illustrates an example.
Fig. 1. An example of the PMX operator.
Mutation. Mutation operators enhance the performance of solutions. Swap

mutation operation is performed. For swapping, two genes are selected randomly
and their positions are exchanged considering MutRate, which is the mutation
probability.
Elitism. In this study, an elitism operator is performed and it is provided to

simply pass the best individual in each generation to the new generation without
any changes.
3.3 Termination Criterion

The GA can terminate when the maximum number of generations has been
performed. In this research, it was used as a fixed number of generations in
order to serve as the termination criterion. Number of generations was selected
in 3000 as stopping condition.
3.4 Parameter Settings

Since the proposed GA presents different parameters, a full-factorial design of
experiments is developed for the problem. In our design of experiments, Table 1
presents values selected for the parameters determined through the calibration
experiments. Four factors have three levels as summarized in Table 1. 81 experi-
ments were required to reach a conclusion. The GA variants were independently
run 31 times, and the average results were tallied. The total number of experi-
ments was 2511 (31 × 34 ). The program was implemented using Microsoft Visual
Studio 2013 software and run on a laptop with 2.1 GHz Core i7 CPU and 8.00 GB
RAM. Effective factors and the magnitudes of the effects were found by using
the analysis of variance (ANOVA) with Minitab 19.0 software.
Levels of each factor were determined considering the main effect plot as
shown in Fig. 2. Given the results of the full-factorial experimental design, the
best parametrization for the GA is considered as P opSize = 50, CrossRate =
0.7, T ourSize = 3, and M utRate = 0.10.
Table 1. Factors and levels for the experimental design.
Factors Levels
1 2 3
Population size 10 30 50
Crossover rate 0.4 0.7 1.0
Mutation rate 0.05 0.10 0.15
Tournament size 3 5 7
Fig. 2. Effect of main design variables on average makespan.

4 Experimental Results
In this study, a decision support system (DSS) was improved to solve the
scheduling problem. The decision support tool that uses genetic algorithm app-
roach assists decision-makers in the existence of job-sequence. According to the
parameter configuration obtained from the full-factorial experimental design, the
optimal or near-optimal fitness function value is obtained as 218.822 min with
15-14-6-5-10-1-9-8-4-12-13-2-11-7-3 (Fig. 3).
Fig. 3. Screenshot of the DSS.
5 Conclusion and Suggestions

This study presents a single machine scheduling problem with sequence-
dependent setup times under the learning effect. A novel mixed-integer linear
program is improved with the minimization of the makespan. The proposed
model is solved using a genetic algorithm approach. A full-factorial experimen-
tal design is used to calibrate the parameters of the GA. The simple DSS is
proposed to schedule the jobs of the bottleneck machines.
There are several topics for future research. Different heuristic or meta-
heuristic methods such as ant colony optimization, genetic algorithms, simu-
lated annealing, tabu search are implemented to this model. This research would
be improved considering multi-objectives such as the total tardiness, the total
weighted flow times, and the total earliness and tardiness, etc. Also, the deteri-
oration effect can be added to the proposed model. Single machine case can be
extended to other machine environments for future researchers.
References
1. Sioud, A., Gravel, M., Gagné, C.: A hybrid genetic algorithm for the single machine
scheduling problem with sequence-dependent setup times. Comput. Oper. Res.
39(10), 2415–2424 (2012)
2. Pinedo, M.: Scheduling: Theory, Algorithms and Systems Development. Prentice
3. Biskup, D.: Single-machine scheduling with learning considerations. Eur. J. Oper.
Res. 115(1), 173–178 (1999)
4. Allahverdi, A., Gupta, J.N.D., Aldowaisan, T.: A review of scheduling research
involving setup considerations. Omega 27(2), 219–239 (1999)
5. Nekoiemehr, N., Moslehi, G.: Minimizing the sum of maximum earliness and maxi-
mum tardiness in the single-machine scheduling problem with sequence-dependent
setup time. J. Oper. Res. Soc. 62(7), 1403–1412 (2011)
6. Wang, X., Tang, L.: A hybrid metaheuristic for the prize-collecting single machine
scheduling problem with sequence-dependent setup times. Comput. Oper. Res.
37(9), 1624–1640 (2010)
7. Zhao, C., Tang, H.: Single machine scheduling with past-sequence-dependent setup
times and deteriorating jobs. Comput. Ind. Eng. 59(4), 663–666 (2010)
8. Ang, A.T.H., Sivakumar, A.I., Qi, C.: Criteria selection and analysis for sin-
gle machine dynamic on-line scheduling with multiple objectives and -sequence-
dependent setups. Comput. Ind. Eng. 56(4), 1223–1231 (2009)
9. Anghinolfi, D., Paolucci, M.: A new discrete particle swarm optimization for
the single-machine total weighted tardiness scheduling problem with sequence-
dependent setup times. Eur. J. Oper. Res. 193(1), 73–85 (2009)
10. Biskup, D., Herrmann, J.: Single-machine scheduling against due dates with past-
sequence-dependent setup times. Eur. J. Oper. Res. 191(2), 587–592 (2008)
11. Mustu, S., Eren, T.: The single machine scheduling problem with sequence-
dependent setup times and a learning effect on processing times. Appl. Soft Com-
put. 71, 291–306 (2018)
12. Mani, V., Chang, P.C., Chen, S.H.: Bi-criteria single machine scheduling problem
with a learning effect: Aneja-Nair method to obtain the set of optimal sequences.
Comput. Math. Appl. 58(1), 39–47 (2009)
13. Biskup, D., Simons, D.: Common due date scheduling with autonomous and
induced learning. Eur. J. Oper. Res. 159(3), 606–616 (2004)
14. Kuo, W.-H., Yang, D.-L.: Minimizing the makespan in a single machine scheduling
problem with a time-based learning effect. Inf. Process. Lett. 97(2), 64–67 (2006)
15. Wang, J.-B.: Single-machine scheduling with learning effect and deteriorating jobs.
Comput. Ind. Eng. 57(4), 1452–1456 (2009)
16. Mosheiov, G., Kyparisis, G.J.: Scheduling problems with a learning effect. Eur. J.
Oper. Res. 132(3), 687–693 (2001)
17. Toksari, M.D., Arık, O.A.: Single machine scheduling problems under position-
dependent fuzzy learning effect with fuzzy processing times. J. Manuf. Syst. 45,
159–179 (2017)
18. Pei, J., Liu, X., Liao, B., Pardalos, P.M., Kong, M.: Single-machine scheduling
with learning effect and resource-dependent processing times in the serial-batching
production. Appl. Math. Model. 58, 245–253 (2018)
19. Kuo, W.-H., Yang, D.-L.: Single machine scheduling with past-sequence-dependent
setup times and learning effects. Inf. Process. Lett. 102(1), 22–26 (2007)
20. Koulamas, C., Kyparisis, G.J.: Single-machine scheduling problems with past-
sequence-dependent setup times. Eur. J. Oper. Res. 187(3), 1045–1049 (2008)
21. Wang, J.-B.: Single-machine scheduling with past-sequence-dependent setup times
and time-dependent learning effect. Comput. Ind. Eng. 55(3), 584–591 (2008)
22. Wen-Chiung, L.: A note on single-machine scheduling with general learning effect
and past-sequence-dependent setup time. Comput. Math. Appl. 62(4), 2095–2100
(2011)
23. Hsu, C.-J., Kuo, W.-H., Yang, D.-L.: Unrelated parallel machine scheduling with
past-sequence-dependent setup time and learning effects. Appl. Math. Model.
35(3), 1492–1496 (2011)
24. Eren, T., Güner, E.: Hazırlık zamanlarının öğrenme etkili olduğu durumda bir
akış tipi çizelgeleme problemi. Gazi Üniversitesi Mühendislik-Mimarlık Fakültesi
Dergisi 22(2), 353–362 (2007)
25. Holland, J.H.: Adaptation in Natural and Artificial Systems. University of Michi-
gan Press, Ann Arbor (1975)
26. Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learn-
ing. Addison Wesley, Reading (1989)
27. Joo, C.M., Kim, B.S.: Genetic algorithms for single machine scheduling with time-
dependent deterioration and rate-modifying activities. Expert. Syst. Appl. 40(8),
3036–3043 (2013)
28. Arumugam, M.S., Rao, M.V.C., Palaniappan, R.: New hybrid genetic operators
for real coded genetic algorithm to compute optimal control of a class of hybrid
systems. Appl. Soft Comput. 6(1), 38–52 (2005)
29. Wang, G., Li, X., Gao, L., Li, P.: A multi-objective whale swarm algorithm
for energy-efficient distributed permutation flow shop scheduling problem with
sequence dependent setup times. IFAC-PapersOnLine 52(13), 235–240 (2019)
Hybrid Krill Herd Algorithm with Particle
Swarm Optimization for Image Enhancement
Elnaz Pashaei1(&), Elham Pashaei2, and Nizamettin Aydin3

1
Department of Software Engineering,
Istanbul Aydin University, Istanbul, Turkey
[email protected]
2
Istanbul Gelisim University, Istanbul, Turkey
3
Yildiz Technical University, Istanbul, Turkey
Abstract. Image enhancement, aimed at improving the image contrast and

information quality, is one of the most critical steps in image processing. Due to
insufficient enhancement and the mean shift problem of conventional image
enhancement techniques, new artificial intelligence-based image enhancement
approaches have become an inevitable need in image processing. This paper
employs the krill herd algorithm (KHA) and particle swarm optimization
(PSO) to suggest a novel hybrid approach, called (PSOKHA) for image
enhancement. The suggested PSOKHA method is used in search of optimum
transfer function parameters to increase the quality of the images. For com-
parative evaluation, the performance of the PSOKHA is compared with six latest
successful enhancement methods: PSO, KHA, screened Poisson equation (SPE),
histogram equalization (HE), brightness preserving dynamic fuzzy HE
(BPDFHE), and adaptive gamma correction weighted distribution (AGCWD).
Experiments results in testing images include a medical image, a satellite image,
and a handwritten image, demonstrate that the suggested strategy can produce
better enhanced images in terms of several measurement criteria such as con-
trast, PSNR, entropy, and structure similarity index (SSIM).
Keywords: Image enhancement Particle swarm optimization (PSO) Krill

herd algorithm (KHA) SNPR
1 Introduction
Image enhancement is a mathematical approach that improves the quality of an image

by contrast stretching, removal of the noise, and magnification of pixels. Image
enhancement can facilitate image interpretation both for a human observer and for
machine recognition. Therefore, it plays a prominent role in many fields such as
medical image analysis, human vision, fault detection, remote sensing, and so on [1].
Techniques for the enhancement of images are divided into two groups: spatial domain
and frequency domain techniques. Approaches are carried out directly on pixel values
in the spatial domain, while the approaches in frequency domains use Wavelet or
Fourier transformations [2]. One of the most useful techniques is to increase the
1432 E. Pashaei et al.
contrast of the images in spatial image enhancement. The contrast stretching approach
improves given image contrast by expanding the range of the pixel intensity values
utilizing a linear transformation (mapping) function [3].
Besides of widely used histogram equalization (HE) based methods for enhancing
contrast [4], recently several evolutionary algorithms have been used to increase the
contrast of images, including cuckoo search algorithm [5], particle swarm optimization
(PSO) [6], genetic algorithm [7], artificial bee colony (ABC) [8], firefly algorithm
(FFA) [9], and so on. Although there are many approaches to enhance the contrast of
the images, most of them cause visual artifacts, resulting in images unnatural
appearance. So image enhancement in different domains is still a challenging task [2].
This study puts forward a novel hybrid approach based on PSO and krill herd
algorithm (PSOKHA) for image contrast enhancement. Due to the simplicity and
excellent performance of the KHA, it has been utilized in various domains. However, at
times the KHA can become trapped in some local optima, leading to its modifications
and hybridizations with other metaheuristics [10]. To improve KHA performance, the
PSO algorithm is utilized in the first stage to enhance the given image’s visual quality.
Then the resulting image fed to KHA for further improvement. The evaluation of
enhanced images reveals that the proposed PSOKHA exhibit better performance than
six latest successful enhancement techniques: PSO [6], KHA, HE [11], screened
Poisson equation (SPE) [12], brightness preserving dynamic fuzzy HE (BPDFHE) [13],
and adaptive gamma correction weighted distribution (AGCWD) [14] based on several
measurement criteria. The main contribution of this paper is to apply KHA to optimize
transfer function parameters for the first time and enhance image visual quality by
combing KHA with the PSO algorithm.
This paper is structured as follows: Details of the transformation and fitness
functions of the enhancement technique are given in Sect. 2. An overview of the KHA,
PSO, and the proposed PSOKHA are elaborated in Sect. 3. In Sect. 4, a detailed
discussion of the experimental results is provided, and Sect. 5 deals with the
conclusion.
2 Problem Formulation
In the spatial domain, a transformation function is employed to boost the image con-
trast by changing the intensity values of the original image to new values [2]. The
definition of transformation function is given as:
Gði; jÞ ¼ T ½F ði; jÞ ð1Þ
where ði; jÞth pixel value in the input image and enhanced image are denoted by F ði; jÞ
and Gði; jÞ, respectively. The aforementioned transformation function T is stated as
kM
T ½F ði; jÞ ¼ ½F ði; jÞ c:lði; jÞ þ lði; jÞ^ a ð2Þ
ðrði; jÞ þ bÞ
Hybrid Krill Herd Algorithm with Particle Swarm Optimization for Image Enhancement 1433
where a, c, b, and k are the unknown constants. lði; jÞ, rði; jÞ and M denote local mean,
local standard deviation, and the global mean of the input image respectively, and are
formulated as follows:
1 Xn1 Xn1
lði; jÞ ¼ : ½F ðx; yÞ ð3Þ
nxn x¼0 y¼0
1 Xn1 Xn1
rði; jÞ ¼ ½ : ½F ðx; yÞ lði; jÞ2 0:5 ð4Þ
nxn x¼0 y¼0
1 Xn1 Xn1
M¼ : ðF ðx; yÞÞ ð5Þ
HV x¼0 y¼0
where H V is the total pixel number in the input image, and n n is the size of the
window used for local information extraction. The enhanced image is produced by
applying the transformation function to image using the optimal parameter values of
a; b; c; and k. In our problem, the a; b; c; and k are the variables that should be tuned.
After obtaining the enhanced image, its quality is assessed using a fitness function. The
fitness function in this study is defined as follows:

log log E IsðeÞ neðIe Þ HðIe Þ
hðIe Þ ¼ ð6Þ
HV
where Ie denotes the obtained image after transformation, IsðeÞ is the resulting image

after performing the Sobel edge detection operator to Ie , and E IsðeÞ is sum of the edge
pixel values in IsðeÞ . The neðIe Þ shows count of edge pixels in the enhanced image, and
HðIe Þ represents enhanced image entropy, calculated using Eq. 7
X255
H ðI e Þ ¼ i¼0
ðhi log2 ðhi ÞÞ ð7Þ
where hi is the probability of ith intensity level in Ie .
3 Proposed PSO-KHA Method
The Krill herd algorithm (KHA) was first introduced by Gandomi and Alavi as a new
nature-inspired search optimization algorithm for solving global optimization function
[15]. The idea of the algorithm is to allocate krill individuals to inspect the search space
to find feasible solutions, based on krill herding behavior. The fitness function in KHA
for each solution is measured by the highest density of the herd and the shortest
distance of each krill’s position from food. The krill individuals’ positions in the
population are updated according to three basic motions, which are (i) movement of
other krill individuals, (ii) location of food, and (iii) random diffusion. The movement
of krill individuals can be formulated as a Lagrangian model as follows:
dXi
¼ Ni þ Fi þ Di ð8Þ
dt
where Ni is related to the motion included by other krill individuals, and Di and Fi
indicate the physical diffusion movement and new foraging motion of the ith krill
individual, respectively. More detailed information about the KHA can be found in [10].
PSO is one of the most commonly used metaheuristic algorithms inspired by the
concept of swarm intelligence in some animals such as fish and birds flocks. The PSO has
been successfully utilized in numerous application domains meanwhile, its modified
versions and its hybrid with other algorithms have been vastly investigated in the liter-
ature. Birds are considered as particles with position and velocity in the algorithm.
Particles keep flying in the search space to find the optimum location, while each particle’s
movement in the search space determined by the optimum position of itself and its
colleagues. A more detailed explanation of the steps of the PSO can be found in [16].
The proposed approach utilizes the combined attributes of PSO and KHA to effi-
ciently produce the contrast-enhanced images (PSOKHA). The proposed PSOKHA is
utilized to determine the best values for the transformation function parameters based
on a fitness function. The proposed PSOKHA involves the following steps:
• A random population is of size N is initialized and a maximum iteration number is
set to a predefined value for both PSO and KHA separately. Each solution indi-
vidual in both algorithms consists of four random values within their ranges for the
parameters of transformation function which are a; b; c; and k.
• In the first step, PSO is utilized to enhance the original image. Each individual
creates a new image using the values of a; b; c; and k parameters. Each image’s
quality is evaluated using a fitness function. The solutions will be updated at each
iteration to find optimal parameter values base on the fitness values.
• After the terminate condition of the PSO is met, an enhanced image with new pixel
intensities is obtained. This enhanced image fed to KHA for more enhancement.
The step2 is repeated for KHA, and the best image with the highest fitness value is
reported.
4 Experimental Results and Discussion
To compare the enhanced quality and optimization capabilities of the suggested

enhancement algorithm, the performance evaluation was carried out using the latest
enhancement algorithms. For a fair assessment of the performance, the maximum
number of iteration and population size was set to fixed numbers 100 and 35 respec-
tively, for the optimization algorithms including PSO, KHA, and suggested PSOKHA.
The boundary limits were set to a 2 ½0; 1:5; b 2 ½0; 0:5; c 2 ½0; 1 and k 2 ½0:5; 1:5
for the optimized parameters. The suggested approach was tested on three satellite,
medical, and handwritten images to determine the robustness of the suggested PSO-
KHA algorithm. Several metrics such as contrast, peak signal-to-noise ratio (PSNR),
entropy, and structure similarity index (SSIM) were considered for quantification of
enhancement efficiency of an algorithm. Tables 1, 2, 3, and 4 summarize the results
achieved by comparing all of the above-mentioned quality measurement values for the
six various enhancement methods that have been evaluated using the given images. In
each case, the best and second-best outcomes are bold and underlined, respectively.
The obtained values for all the above mentioned quality measurements confirm the
superiority of the suggested method compared with other existing approaches.
The PSNR values acquired for all comparative images show that using the suggested
procedure, the processed image quality is enhanced well compared with the original
image. The SSIM, which is a greatly appropriate metric for the efficiency evaluation of
an algorithm, shows that the proposed PSOKHA has great superiority over the other
methods for all images. A high entropy value means that the improved image preserves
all the information in the original image. The PSOKHA outperforms all the latest
algorithms in terms of entropy. Moreover, the contrast metric obtained from the gray-
level co-occurrence matrix (GLCM) indicates PSOKHA’s comparative performance
against other algorithms.
Table 1. Comparison of PSNR for experimental images.

Image PSO KHA HE SPE BPDFHE AGCWD Proposed
Satellite 14.12 16.43 12.796 18.12 24.78 13.81 31.77
Medical 18.83 18.77 16.35 12.22 29.98 14.34 27.19
Handwritten 13.9 20.05 10.53 21.01 27.98 16.08 29.90
Figure 1 demonstrates how our proposed method performs on three test images.
This figure also shows the alteration between the reference image histogram and the
enhanced image. Histograms of all the enhanced images are much flatter than the
original images (referring to a more uniform intensity distribution). The subjective
assessment of the images produced with all six enhancement approaches is shown in
Fig. 2. It indicates that the proposed PSOKHA algorithm not only enhances any minor
detail in the image effectively but also maintains the average intensity of the images.
The results also emphasize the robustness of the suggested approach, which makes it
applicable to images from different domains.
Table 2. Comparison of SSIM for experimental images.

Satellite 0.575 0.598 0.799 0.923 0.906 0.866 0.982
Medical 0.435 0.699 0.826 0.530 0.975 0.893 0.978
Handwritten 0.456 0.874 0.405 0.898 0.894 0.966 0.972
Table 3. Comparison of entropy for experimental images.

Satellite 7.52 7.59 5.59 6.33 6.66 6.80 7.23
Medical 7.59 7.62 5.74 6.56 7.32 7.25 7.63
Handwritten 6.92 6.49 5.32 6.50 5.83 6.00 6.74
Table 4. Comparison of Contrast for experimental images.

Satellite 3.98 2.800 1.240 1.257 1.020 1.069 1.297
Medical 2.63 0.400 0.173 0.473 0.163 0.180 0.184
Handwritten 1.24 0.620 0.680 0.610 0.353 0.366 0.466
Fig. 1. Results of the proposed PSOKHA enhancement technique: (left) original image and its
histogram, (right) the enhanced image and its histogram.
(a) (b) (c) (d)
(e) (f) (j) (h)
Fig. 2. Enhancement results of Images: (a) original, (b) PSO, (c) KHA, (d) HE, (e) SPE,
(f) BPDFHE, (j) AGCWD, and (h) proposed PSOKHA.
5 Conclusion
In this paper, a novel hybrid metaheuristic method has been suggested to enhance gray-
level images using KHA and PSO. Three testing images, including a medical image, a
satellite image, and a handwritten image were utilized to assess the proposed PSOKHA
performance. The results of PSOKHA were compared with six commonly used
enhancement techniques: PSO, KHA, HE, SPE, BPDFHE, and AGCWD. The com-
puted performance metrics revealed that the proposed PSOKHA outperforms other
approaches in all cases. Therefore, it can be concluded that the PSOKHA possesses the
potential for enhancing images, and can make images more visible by adjusting their
contrast. As future work, the usefulness of the proposed method can be investigated in
solving other related domains of image processing such as segmentation and noise
reduction. Also, the proposed approach can be extended by introducing several mod-
ifications into the PSO, KHA, and fitness function parts of the proposed PSOKHA to
improve its quality of solutions.
References
1. Gogna, A., Tayal, A.: Comparative analysis of evolutionary algorithms for image
enhancement. Int. J. Metaheuristics 2, 80 (2012). https://doi.org/10.1504/ijmheur.2012.
048219
2. Dhabal, S., Saha, D.K.: Image enhancement using differential evolution based whale
optimization algorithm. In: Advances in Intelligent Systems and Computing, pp. 619–628.
Springer (2020). https://doi.org/10.1007/978-981-13-7403-6_54
3. Suresh, S., Lal, S.: Modified differential evolution algorithm for contrast and brightness
enhancement of satellite images. Appl. Soft Comput. J. 61, 622–641 (2017). https://doi.org/
10.1016/j.asoc.2017.08.019
4. Nithyananda, C.R., Ramachandra, A.C.: Preethi: review on histogram equalization based
image enhancement techniques. In: International Conference on Electrical, Electronics, and
Optimization Techniques, ICEEOT 2016, pp. 2512–2517. Institute of Electrical and
Electronics Engineers Inc. (2016). https://doi.org/10.1109/ICEEOT.2016.7755145
5. Agrawal, S., Panda, R.: An efficient algorithm for gray level image enhancement using
cuckoo search. In: Lecture Notes in Computer Science (including subseries Lecture Notes in
Artificial Intelligence and Lecture Notes in Bioinformatics), pp. 82–89. Springer, Heidelberg
(2012). https://doi.org/10.1007/978-3-642-35380-2_11
6. Gorai, A., Ghosh, A.: Gray-level image enhancement by particle swarm optimization. In:
Proceedings of the 2009 World Congress on Nature and Biologically Inspired Computing,
NABIC 2009, pp. 72–77 (2009). https://doi.org/10.1109/NABIC.2009.5393603
7. Hashemi, S., Kiani, S., Noroozi, N., Moghaddam, M.E.: An image contrast enhancement
method based on genetic algorithm. Pattern Recognit. Lett. 31, 1816–1824 (2010). https://
doi.org/10.1016/j.patrec.2009.12.006
8. Chen, J., Yu, W., Tian, J., Chen, L., Zhou, Z.: Image contrast enhancement using an artificial
bee colony algorithm. Swarm Evol. Comput. 38, 287–294 (2018). https://doi.org/10.1016/j.
swevo.2017.09.002
9. Dhal, K.G., Iqbal Quraishi, M., Das, S.: Development of firefly algorithm via chaotic
sequence and population diversity to enhance the image contrast. Nat. Comput. 15, 307–318
(2016). https://doi.org/10.1007/s11047-015-9496-3
10. Bolaji, A.L., Al-Betar, M.A., Awadallah, M.A., Khader, A.T., Abualigah, L.M.: A
comprehensive review: Krill Herd algorithm (KH) and its applications (2016). https://doi.
org/10.1016/j.asoc.2016.08.041
11. Kim, H.-J., Lee, J.-M., Lee, J.-A., Oh, S.-G., Kim, W.-Y.: Contrast enhancement using
adaptively modified histogram equalization. Presented at the (2006). https://doi.org/10.1007/
11949534_116
12. Morel, J.-M., Petro, A.-B., Sbert, C.: Screened poisson equation for image contrast
enhancement. Image Process. Line 4, 16–29 (2014). https://doi.org/10.5201/ipol.2014.84
13. Sheet, D., Garud, H., Suveer, A., Mahadevappa, M., Chatterjee, J.: Brightness preserving
dynamic fuzzy histogram equalization. IEEE Trans. Consum. Electron. 56, 2475–2480
(2010). https://doi.org/10.1109/TCE.2010.5681130
14. Huang, S.C., Cheng, F.C., Chiu, Y.S.: Efficient contrast enhancement using adaptive gamma
correction with weighting distribution. IEEE Trans. Image Process. 22, 1032–1041 (2013).
https://doi.org/10.1109/TIP.2012.2226047
15. Gandomi, A.H., Alavi, A.H.: Krill herd: a new bio-inspired optimization algorithm.
Commun. Nonlinear Sci. Numer. Simul. 17, 4831–4845 (2012). https://doi.org/10.1016/j.
cnsns.2012.05.010
16. Yang, X.S., Karamanoglu, M.: Swarm Intelligence and bio-inspired computation: an
overview. In: Swarm Intelligence and Bio-Inspired Computation, pp. 3–23. Elsevier Inc.
(2013). https://doi.org/10.1016/B978-0-12-405163-8.00001-6
Distributed Fuzzy Permutation Flow Shop
Scheduling Problem: A Bee Colony Algorithm
M. Emin Baysal1 , Ahmet Sarucan1 , Kadir Büyüközkan2 ,

and Orhan Engin1(&)
1
Konya Technical University, Konya, Turkey
[email protected]
2
Karadeniz Technical University, Trabzon, Turkey
[email protected]
Abstract. The distributed permutation flow shop scheduling problem is a sub-

class of the permutation flow shop scheduling problem. Distributed scheduling
adopts multi-factory with permutation flow shop scheduling environment. At the
distributed permutation flow shop scheduling, due to the human factors, the pro-
cessing times of the jobs on the machines are not known exactly. Thus, in this
study, the processing time of the jobs on machines are considered as a triangular
fuzzy number. Also, the due dates of the jobs are considered as trapezoidal fuzzy
numbers at this research. To solve the distributed fuzzy permutation flow shop
scheduling problem with multi-objective an artificial bee colony algorithm is
proposed. To the best of our knowledge, this is the first study to solve the dis-
tributed fuzzy permutation flow shop scheduling with an artificial bee colony
algorithm. The proposed artificial bee colony algorithm is first calibrated on the
distributed permutation flow shop scheduling problem. The results showed that the
proposed artificial bee colony algorithm is an efficient solution technique for
solving distributed fuzzy permutation flow shop scheduling problems.
Keywords: Fuzzy distributed permutation flow shop Artificial bee colony

algorithm
1 Introduction
The distributed permutation flow shop scheduling problem is attracted the researchers’
attention in recent years. The distributed permutation flow shop scheduling problem
(DPFSSP) was first proposed by Naderi and Ruiz [1]. Distributed scheduling adopts a
multi-factory environment. A distributed permutation flow shop has a two-step decision
making process. In the first step, the jobs are allocated to suitable factories and at the
second step; the jobs are attended to the machine with a permutation. As a result, the
distributed permutation flow shop is more complicated than the permutation flow shop
scheduling problem. Also, the permutation flow shop scheduling problem is an NP-
hard for more than three machines.
Distributed Fuzzy Permutation Flow Shop Scheduling Problem 1441
The first study about DPFSSP in literature has done in 2010. The literature review
about DPFSSP is given as follows. Naderi and Ruiz [1] assumed n jobs have to be
distributed among the F factories. They proposed six different alternative mixed-integer
linear programming models with minimization of the maximum completion time or
makespan performance criteria among the factories. Also, they proposed two simple
factory assignment rules together with 14 heuristics based on dispatching rules. Xu
et al. [2] developed a hybrid immune algorithm with the criteria to minimize the
maximum completion time to solve DPFSSP. They used permutation-based encoding
to schedules. They hybridized the immune algorithm with four different local search
operators. Li and Chen [3] proposed a hybrid genetic algorithm with minimizing
makespan for solving the DPFSSP. They hybridized the genetic algorithm with local
search and plant allocation rules. Li et al. [4] generated a new type of DPFSSP. They
considered different transport timetables and different loading capacities for each fac-
tory. Also, they used simulated annealing based local search with multi different
neighborhood searching methods for solving this new type of DPFSSP to minimize the
maximum completion time. Viagas et al. [5] proposed eight constructive heuristics
algorithm to solve the DPFSSP by minimizing the total flow time performance criteria.
Ruiz et al. [6] developed an iterated greedy method for DPFSSP with makespan
performance criteria. They used construction and destruction procedures with a local
search at the greedy methods. There is only one study from the literature about the
fuzzy logic-based approach for solving DPFSSP. Wang et al. [7] proposed a fuzzy
logic-based hybrid estimation of distribution algorithm to minimize the makespan of
the DPFSSP under machine breakdown. They adopted a fuzzy logic-based evolution
strategy for the diversity of the population.
At this research, a distributed fuzzy permutation flow shop scheduling problem is
considered. The processing times of the jobs on machines and due dates are given as
fuzzy numbers. To solve this problem, an artificial bee colony algorithm is proposed.
To the best of our knowledge, in the literature, this is the first study that considered a
distributed fuzzy permutation flow shop scheduling problem and also for the first time,
an artificial bee colony algorithm is used to solve this problem.
The paper is organized as follows. Section 2 presents an artificial bee colony
algorithm. Computational results are given in Sect. 3. In Sect. 4, the results and future
research are discussed.
2 Artificial Bee Colony Algorithm
The artificial bee colony algorithm is a swarm intelligence method. It was first
developed by Karaboga [8]. The artificial bee colony algorithm (ABCA) is searched
found the best solution in the solution space based on the foraging behavior of honey
bees. ABCA has been applied to solve many combinatorial optimization problems. In
recent years, the ABCA has used to solve the flow shop scheduling problems from the
literature. These are given as follows. Liu and Liu [9] proposed a hybrid discrete
artificial bee colony algorithm to minimize the makespan in permutation flow shop
scheduling problems. Han et al. [10] proposed a novel discrete artificial bee colony
algorithm to solve the flow shop scheduling problem with blocking. Su-jun et al. [11]
1442 M. Emin Baysal et al.
developed a discrete artificial bee colony algorithm to minimize the maximum com-
pletion time for flow shop scheduling problem with intermediate buffers. Ince et al. [12]
proposed a discrete artificial bee colony algorithm for the permutation flow shop
scheduling problem with sequence-dependent setup times. Sidek et al. [13] used the
artificial bee colony algorithm to solve the permutation flow shop scheduling problem
with makespan criteria. An artificial bee colony algorithm is first used in this study to
solve the distributed fuzzy flow shop scheduling problem.
3 Computational Results
In the permutation flow shop scheduling problem, n-jobs are processed in a set of m-
machines in the same order. The most common performance criterion is to minimize
the makespan value. The permutation flow shop scheduling problem is a well-known
NP-hard problem for more than three machines. In the distributed permutation flow
shop scheduling problem, there are multi-factory exist. In the DPFSSP, at first, n-jobs
have to be distributed among the identical F factories and then the n-jobs are processed
in a set of m-machines in the same order. Also, DPFSSP is an NP-hard problem [1].
Due to the human factors, the processing times of the jobs on the machines and due
dates are not known exactly. Thus, at this study, a distributed fuzzy permutation flow
shop scheduling problem is considered. The processing time of the jobs on machines
and due dates are considered as triangular and trapezoidal fuzzy numbers respectively
[14–16]. The formulation of the distributed fuzzy permutation flow shop scheduling
problem is given as follows.
F Number of factories
n Number of jobs ði ¼ 1; . . .; nÞ
m Number of machines ðj ¼ 1; . . .; mÞ
e
P ij Fuzzy processing times
d~ Fuzzy due date
C~ Fuzzy completion time
The fuzzy processing times of operation Pij and due date d~ are given in Fig. 1.
µ µ
1 1
d1 d2
P1i,j, P2i,j, P3i,j,
(b)
(a)
Fig. 1. Fuzzy processing time (a) and due date (b).

~
dj
1 ~
Cj
Agreement index
Fig. 2. Agreement index.
An agreement index is calculated for the fuzzy due date and completion time of
jobs. The agreement index is illustrated in Fig. 2. The agreement index is calculated as
following [14–17].
areaðd~ \ CÞ
~
Agreement index ¼ ð1Þ
~
areaðCÞ
The agreement index measures the degree to which the due date constraint. The
distributed fuzzy permutation flow shop scheduling problem is solved by two objec-
tives with an artificial bee colony algorithm. These objectives are given as follows.
~ max
Minimize the fuzzy completion time; min Z1 ¼ C ð2Þ
Maximize the agreement index; max Z2 ¼ min ðAgreement indexÞ ð3Þ
The proposed artificial bee colony algorithm is first calibrated on the distributed
per-mutation flow shop scheduling problem from the literature. Naderi and Ruiz [1]’s
30 DPFSSP benchmark instances are solved by the proposed ABCA. The results are
given in Table 1. This instance; 2x10x4-1., is 2 factories, 10 jobs, 4 machines, and the
first type of problem. It can be seen from Table 1, the proposed ABCA is found the
best Cmax (makespan) values for all 30 benchmark instances.
Table 1. Solutions of the test problems.
Best Proposed Best Proposed

Instance Instance
(Cmax) ABCA (Cmax) (Cmax) ABCA (Cmax)
2x10x4-1. 421 421 3x10x5-1. 338 338
2x10x4-2. 415 415 3x10x5-2. 364 364
2x10x4-3. 413 413 3x10x5-3. 417 417
2x10x4-4. 436 436 3x10x5-4. 414 414
2x10x4-5. 401 401 3x10x5-5. 352 352
2x10x5-1. 452 452 4x10x4-1. 346 346
2x10x5-2. 439 439 4x10x4-2. 291 291
2x10x5-3. 518 518 4x10x4-3. 289 289
2x10x5-4 502 502 4x10x4-4. 310 310
2x10x5-5. 528 528 4x10x4-5. 249 249
3x10x4-1. 331 331 4x10x5-1. 327 327
3x10x4-2 313 313 4x10x5-2. 360 360
3x10x4-3. 329 329 4x10x5-3. 417 417
3x10x4-4. 364 364 4x10x5-4. 351 351
3x10x4-5. 368 368 4x10x5-5. 379 379
Some of these benchmark instances are arranged a fuzzy DPFSSP according to

processing times and due dates. This instance; F-2x10x4-1., is fuzzy, 2 factories, 10
jobs, 4 machines, and the first type of problem. It can be seen from Table 2. This fuzzy
DPFSSP is solved by the proposed ABCA. The proposed ABCA procedure is given
below;
Define parameter;
Generate random scout bee solutions;
Evaluate the fitness of the solutions;
Keep the “best scout bee” as the “best solution”;
Begin iteration;
Generate random neighbourhood for each scout bees as much as follower
bees;
Evaluate the fitness of the followers;
Replace the neighbourhood solution which is better than the scout bee, as
new scout bee;
Identify the best scout bee;
Replace the best scout bee which is better than best solution, as new best so-
lution;
Loop until Iteration end;
Bring the best solution as a result.
The best parameters set for proposed ABCA is given below;

Number of iterations: 250
Number of scout bees: 25
Number of follower bees: 25
It can be seen in Table 2, the proposed ABCA is found the minimum fuzzy
completion time. Also proposed ABCA obtains the maximum agreement index.
Table 2. Solutions of the fuzzy test problems.

~ max
Fuzzy instance C Best (Cmax) Agreement index
~1 C
C ~2 ~3
C
F-3x10x5-1. 293 339 377 338 0.900
F-3x10x5-2. 311 367 392 364 0.624
F-3x10x5-3. 366 418 439 417 0.595
F-3x10x5-4. 386 417 438 414 0.830
F-3x10x5-5. 299 353 380 352 0.926
F-4x10x5-1. 295 327 358 327 0.251
F-4x10x5-2. 314 362 384 360 0.818
F-4x10x5-3. 361 417 470 417 0.210
F-4x10x5-4. 310 353 372 351 0.677
F-4x10x5-5. 324 379 402 379 0.704
In distributed permutation flow shop, the parameters such as the processing time and
due date are usually assumed to be known, but in many real applications, these
parameters are dynamic due to human factors. In this study, a fuzzy DPFSSP is
considered. For solving this problem an ABCA is proposed. The fuzzy DPFSSP is
solved by two objectives, fuzzy completion time and agreement index. The benchmark
instances are solved by the proposed ABCA. The results showed that the proposed
ABCA is an effective solving method for this problem.
For future research, the proposed ABCA can be used in a real world application.
References
1. Naderi, B., Ruiz, R.: The distributed permutation flowshop scheduling problem. Comput.
Oper. Res. 37(4), 754–768 (2010)
2. Xu, Y., Wang, L., Wang, S., Liu, M.: An effective hybrid immune algorithm for solving the
distributed permutation flow-shop scheduling problem. Eng. Optim. 46(9), 1269–1283
(2014)
3. Li, Y., Chen, Z.: A hybrid genetic algorithm for distributed permutation flowshop scheduling
problem. In: International Conference on Information Technology and Management
Innovation (ICITMI), pp. 843–847 (2015)
4. Li, Z.Y., Duan, W.Z., Ji, M.C., Yang, Y.X., Wang, S.Y., Liu, B.: The distributed
permutation flowshop scheduling with different transport timetables and loading capacities.
IEEE Congr. Evol. Comput., 2433–2437 (2016)
5. Viagas, V.F., Gonzalez, P.P., Framinan, J.M.: The distributed permutation flow shop to
minimize the total flowtime. Comput. Ind. Eng. 118, 464–477 (2018)
6. Ruiz, R., Pan, Q.K., Naderi, B.: Iterated greedy methods for the distributed permutation
flowshop problem. Omega 83, 213–222 (2019)
7. Wang, K., Huang, Y., Qin, H.: A fuzzy logic-based hybrid estimation of distributed
algorithm for distributed permutation flowshop scheduling problems under machine
breakdown. J. Oper. Res. Soc. 67, 68–82 (2016)
8. Karaboga, D.: An idea based on honey bee swarm for numerical optimization. Technical
Report, Computer Engineering Department, Erciyes University, Turkey (2005)
9. Liu, Y.F., Liu, S.Y.: A hybrid discrete artificial bee colony algorithm for permutation
flowshop scheduling problem. Appl. Soft Comput. 13, 1459–1463 (2013)
10. Han, Y.Y., Gong, D., Sun, X.: A discrete artificial bee colony algorithm incorporating
differential evolution for the flow shop scheduling problem with blocking. Eng. Optim. 47
(7), 927–946 (2015)
11. Su-jun, Z., Xing-sheng, G.: An effective discrete bee colony algorithm for flow shop
scheduling problem with intermediate buffers. J. Cent. South Univ. 22, 3471–3484 (2015)
12. Ince, Y., Karabulut, K., Tasgetiren, M.F., Pan, Q.: A discrete artificial bee colony algorithm
for the permutation flow shop scheduling problem with sequence dependent setup times. In:
IEEE Congress on Evolutionary Computing, pp. 3401–3408 (2016)
13. Sidek, N.A., Bareduan S.A., Nawawi, A.: Performance investigation of artificial bee colony
(ABC) algorithm for permutation flowshop scheduling problem (PFSP). In: International
Conference on Mechanical and Manufacturing Engineering, pp. 1–10 (2018)
14. Kahraman, C., Engin, O., Yılmaz, M.K.: A new artificial immune system algorithm for
multiobjective fuzzy flow shop problems. Int. J. Comput. Intell. Syst. 2(3), 236–247 (2009)
15. Engin, O., Gözen, Ş.: Parallel machine scheduling problems with fuzzy processing time and
fuzzy duedate: an application in an engine valve manufacturing process. J. Mult. Valued
Log. Soft Comput. 15(2), 107–123 (2009)
16. Engin, O., Kahraman C., Yılmaz M.K.: A scatter search method for multiobjective fuzzy
permutation flow shop scheduling problem: a real world application. In: Chakraborty, U.K.
(ed.) Computational Intelligence in Flow Shop and Job Shop Scheduling, vol. 230, pp. 169–
189. Springer, Berlin (2009)
17. Engin, O., Yılmaz, M.K., Baysal, M.E., Sarucan, A.: Solving fuzzy job shop scheduling
problems with availability constraints using a scatter search method. J. Mult. Valued Log.
Soft Comput. 21, 317–334 (2013)
Fuzzy Metaheuristics: A State-of-the-Art
Review
Nurşah Alkan(&) and Cengiz Kahraman

Macka, Besiktas, 34367 Istanbul, Turkey
{nalkan,kahramanc}@itu.edu.tr
Abstract. A metaheuristic is a high-level problem-independent algorithmic

framework that provides a set of guidelines or strategies for developing heuristic
optimization algorithms, reaching the optimum solution in the solution space
more quickly by using efficient searches in a high-level work environment.
There are various metaheuristics algorithms in the literature and the use of these
algorithms for many problems in different areas is increasing rapidly. However,
because the problems addressed are complex and uncertain, more effective and
reliable results are needed for practitioners. In this context, fuzzy sets that better
express uncertainties and reduce complexity can be successfully used with
metaheuristic algorithms to achieve more concrete and realistic results. How-
ever, there isn’t a guiding source in the literature for those who want to research
this topic. Therefore, this study aims to guide researchers on fuzzy-based
metaheuristic algorithm applications by presenting literature research. For this
aim, a large number of papers implementing fuzzy-based metaheuristic algo-
rithms have been summarized with graphical figures by analyzing with respect
to some characteristics such as subject area, published journal, publication year,
and source country.
Keywords: Metaheuristics Fuzzy logic Genetic algorithms Literature

review
1 Introduction
Combinatorial optimization researched in areas such as artificial intelligence and

operations research is applied for problems arising in several relevant contexts such as
economics, engineering, science. Various studies have been conducted to help solve
these problems effectively to produce practical and theoretical information [1]. Meta-
heuristic algorithms that find the optimal solution of hard optimization problems in a
reasonable time represent a core research field of combinatorial optimization [2]. In
particular, it is very advantageous to use metaheuristic algorithms in problems with
wide solution space which makes it impossible to find every possible solution.
Metaheuristic algorithms help in producing of different solutions in complex opti-
mization problems with the ability to achieve global results without being fitted to local
optimum points. Metaheuristics begin the search process by obtaining the first solution
and then improve the search directed under certain principles [3]. There are a lot of
metaheuristic algorithms in the literature, but better metaheuristic approaches have

been needed due to the increase in the complexity of large-scale computing problems in
recent years [4]. Besides, uncertainty emerging with increased complexity has revealed
the need to generate more effective and more reliable results. In this context, the
utilization with fuzzy approach of metaheuristic algorithms will enable generated more
concrete and more realistic results by providing to better express uncertainties and
reduce complexity. The combination of fuzzy approaches with metaheuristic algo-
rithms has recently taken attention by researchers and the number of studies addressed
these two topics together have started to increase day by day. This has revealed the
need to give researchers an idea of how these two topics can be used together.
Therefore, this study aims to present a state of the art of literature review for
researchers. This article gives an idea related to the adoption of these methods for what
kinds of purposes and with what factors. The recent research trends of fuzzy meta-
heuristics algorithms developed are also surveyed to provide a roadmap to researchers
studying in this field. The originality of this study comes from presenting the a-state-of-
art of the fuzzy studies carried out on the recently developed metaheuristics.
The rest of this paper is organized as follows: the general structure of the meta-
heuristic algorithms and their new developments are explained in Sect. 2. Section 3
includes an analysis of the literature for researchers who want to use together with
fuzzy systems the new metaheuristic algorithms. Finally, the obtained results and future
research suggestions are explained in Sect. 4.
2 Metaheuristic Algorithms
Metaheuristics are widely accepted as an efficient approach for many hard optimization
problems and have become more popular due to their less complexity and increased
efficiency compared to other traditional approaches. They are repetitive procedures that
intelligently guide a subordinate heuristic to properly explore and exploit the search
space by combining different concepts. There are many classifications of meta-heuristic
algorithms in the literature. As shown in Fig. 1, metaheuristic algorithms can be mainly
classified into two main categories: Single solution-based and population-based [4].
Single solution-based algorithms only use a single solution to reach some satisfactory
solution. The most popular single-based metaheuristic algorithms are simulated
annealing, tabu search, greedy randomized adaptive search procedure. Population-
based algorithms starting with a set of multiple solutions tend to perform more search.
The most popular algorithms are genetic algorithms, particle swarm optimization, ant
colony optimization, artificial bee colony algorithm. There are various metaheuristic
algorithms developed in recently, besides the algorithms placed in literature. Some of
the recently developed algorithms have been listed in Table 1. Grey Wolf Optimization
(GWO) algorithm is a new population-based metaheuristic algorithm that imitates the
leadership hierarchy and hunting mechanism of grey wolves in nature [5]. Ant Lion
Optimization (ALO) algorithm is a novel nature-inspired algorithm that mimics the
hunting mechanism of antlions in nature modeling the interaction between nations and
ants [6]. Moth-Flame Optimization (MFO) is a nature-inspired swarm intelligence-
based metaheuristic algorithm which imitates the movement process of moths by
Fuzzy Metaheuristics: A State-of-the-Art Review 1449
utilizing the phenomenon of Transverse Orientation [7]. Whale Optimization Algo-

rithm (WOA) is a population-based algorithm that imagery the hunting behavior of
whales in recognizing prey location and encircling them [8]. Jaya Algorithm (JA) is
based on the concept that the solution created for a particular problem is close to the
best solution and far from the worst solution [9]. A novel population-based opti-
mization algorithm, Sine Cosine Algorithm (SCA) starts the optimization process by
producing various random solutions and then utilizes a mathematical model based on
sine and cosine functions to determine the best solution [10].
Fig. 1. Classification of metaheuristic algorithms
Bird Swarm Algorithm (BSA) is based on the swarm intelligence extracted from the
social behaviors and interactions in bird swarms. The algorithm introduces population
regeneration strategies that mimic the foraging behavior of birds and the behavior of
fleeing from predators through social interactions [11]. Electromagnetic Field Opti-
mization (EFO) algorithm based on the electromagnetic field principle, the electro-
magnetic particle avoids the worst solution and moves towards the best solution due to
the attraction and repulsion forces in the electromagnetic field, and then all the elec-
tromagnetic particles gather around the optimum solution [12]. A bio-inspired opti-
mization algorithm, Salp Swarm Algorithm (SSA) is based on the swarming behavior of
salps when navigating and foraging in oceans [13]. Grasshopper Optimization Algo-
rithm (GOA) is used to solve optimization problems by mathematically modeling and
mimicking grasshopper herds’ behavior in nature [14]. Drone Squadron Optimization
(DSO) inspired by the control of drone squadron is a new meta-heuristics method which
featured with the self-adaptive ability [15].
Table 1. Metaheuristic algorithms recently developed

Algorithms Year of proposal
Grey Wolf Optimization [5] 2014
Ant Lion Optimization [6] 2015
Moth-Flame Optimization [7] 2015
Whale Optimization Algorithm [8] 2016
Jaya Algorithm [9] 2016
Sine Cosine Algorithm [10] 2016
Bird Swarm Algorithm [11] 2016
Electromagnetic Field Optimization [12] 2016
Salp Swarm Algorithm [13] 2017
Grasshopper Optimization Algorithm [14] 2017
Drone Squadron Optimization [15] 2018
Although there are a lot of studies on metaheuristic algorithms developed recently

in the literature, it has been necessary to handle metaheuristic algorithms with fuzzy
approaches to produce more effective and reliable results due to the increase of com-
plexity and uncertainty in the problems. Dong et al. [16] have proposed an improved
DSO-based approach to optimize interval type-2 fuzzy logic controller parameters and
they have performed a comparative analysis to verify the DSO. Simulation results have
shown that DSO performs better than other popular algorithms. Dhyani et al. [17] have
proposed an evolving Fuzzy-PID control design for the trajectory tracking problem of a
serial manipulator in the study. The scaling factors of the fuzzy controller have been
evolved by using recently proposed stochastic nature-inspired optimizers which are
MFO, novel bat algorithm, and SCA. Sadeghi-Moghaddam et al. [18] have used the
method of converting continuous numbers into a discrete form for the transportation
problem. they have tackled as fuzzy both fixed costs and variable costs and have been
used to evaluate the performance of algorithms based on genetic algorithm, simulated
annealing, particle swarm optimization, and WOA by considering different problem
sizes. Singh et al. [19] have evolved fuzzy logic system using artificial bee colony
optimization, genetic algorithm, firefly algorithm, particle swarm optimization, and
WOA. The simulations results have shown that whale optimization algorithm requires
less time and fuzzy system predictions are more precise than others. Tak [20] has
presented a new forecasting method that employs intuitionistic fuzzy c-means clus-
tering and a GWO in recurrent type-1 fuzzy functions in the study.
3 Analysis of Reviews
In this section, the publications realized to evaluate optimization problems by using the
metaheuristic algorithms with fuzzy approaches are analyzed according to different
features such as publication year, journal name, country, and metaheuristic algorithm.
The results of the analysis have been presented through tables and figures. In the first step,
the Scopus database has been analyzed to present a comprehensive review of meta-
heuristic algorithms used to solve optimization problems. The analysis of papers pub-
lished in international journals indexed in this database has been realized by checking one
by one. In the context, the analysis processes have been realized based on following
keywords: “metaheuristic + fuzzy, Grey Wolf Optimization + fuzzy, Ant Lion Opti-
mization + fuzzy, Moth-Flame Optimization + fuzzy, Whale Optimization Algo-
rithm + fuzzy, Jaya Algorithm + fuzzy, Sine Cosine Algorithm + fuzzy, Bird Swarm
Algorithm + fuzzy, Electromagnetic Field Optimization + fuzzy, Salp Swarm Algo-
rithm + fuzzy, Grasshopper Optimization Algorithm + fuzzy, Drone Squadron Opti-
mization + fuzzy”. Since optimization and optimizer can be used mutually, keywords
have also been accepted using the optimizer. 388 papers have been obtained according to
these keywords used in the analysis. However, after checking the content of the papers,
289 papers have been found in total by removing irrelevant papers.
Firstly, the literature has been analyzed in terms of the publication year. The
analysis results have been presented in Fig. 2 by considering all fuzzy metaheuristic
algorithms used to solve optimization problems. The number of studies utilizing fuzzy
metaheuristic algorithms increases year by year as observed in Fig. 2. The number of
studies, which was 21 in 2009, has reached 118 in 2019 according to analyze results.
Additionally, the percentage of studies using fuzzy approach in terms of publication
year has been presented by considering novel metaheuristic algorithms given in
Table 1 in Fig. 3.
Fig. 2. Number of publications using fuzzy metaheuristic algorithms according to years
The results show that the number of studies utilizing fuzzy novel metaheuristic
algorithms increases year by year and the number of studies on fuzzy GWO algorithm
is higher than others each year. GWO algorithm based on fuzzy approach is the study
area that attracts more attention by researchers as observed in Fig. 3. Furthermore, the
overall percentage distribution of algorithms given in Table 1 has been presented in
Fig. 4. The algorithms with the least study in total have been determined as BSA, DSO,
EFO.
The number of papers according to countries has been presented in Table 2. India is
the country that has the largest number of papers by using new metaheuristic algo-
rithms in the fuzzy environment. China, Iran, Egypt, Mexico, Australia, Turkey, and
Saudi Arab are countries following India, respectively. The percentage distribution of
publications of these countries has been shown in Fig. 5.
Fig. 3. Number of publications using fuzzy novel metaheuristic algorithms according to years
Fig. 4. Overall percentage distribution of publications related to novel metaheuristic algorithms
Table 2. Number of publications according to countries

Country Numbers of publication
India 115
China 40
Iran 39
Egypt 26
Mexico 14
Australia 11
Turkey 10
Saudi Arabia 9
Fig. 5. Percentage distribution of publications according to countries
After that, the number of publications applying novel metaheuristic algorithms

based on fuzzy approach according to the journals have been presented in Table 3.
Advances in Intelligent Systems and Computing journal has the most publications made
on fuzzy novel metaheuristic algorithms. Additionally, the percentage distribution of
publications according to journals have also been given in Fig. 6.
Table 3. Number of publications based on journals

Journal Numbers of
publication
Advances in Intelligent Systems and Computing 15
Soft Computing 10
Applied Soft Computing Journal 8
IEEE Access 8
Studies in Computational Intelligence 5
Lecture Notes in Computer Science Including Subseries Lecture Notes 4
in Artificial Intelligence and Lecture Notes in Bioinformatics
Evolutionary Intelligence 4
Journal of Advanced Research in Dynamical and Control Systems 4
Lecture Notes in Electrical Engineering 4
Neural Computing and Applications 4
Fig. 6. Percentage distribution of publications according to journals

4 Conclusion
Although there are a lot of metaheuristic algorithms in the literature along with meta-
heuristic algorithms occurring recently, better metaheuristic approaches have been nee-
ded due to the increase in the complexity of large-scale problems. To handle the
uncertainty emerging with increased complexity, the utilization with metaheuristic
algorithms of fuzzy systems will generate more effective and more reliable results. This
work has been revealed the studies addressing in the literature the use of fuzzy systems
together with metaheuristic algorithms developed in recently. In this context, these
studies have been analyzed by grouping them according to publication year, published
journal, source country, type of novel metaheuristic algorithms. GWO has been found as
the most studied algorithm among the novel metaheuristic algorithms. EFO and DSO
algorithms have been rarely used in fuzzy metaheuristic publications. This situation
presents an opportunity for researchers to study in this area. For further research, novel
fuzzy metaheuristic algorithms are proposed for the solution of complex optimization
problems.
References
1. Ramos-Figueroa, O., Quiroz-Castellanos, M., Mezura-Montes, E.: Metaheuristics to solve
grouping problems: a review and a case study. Swarm Evol. Comput. 53 (2020)
2. Elshaer, R., Awad, H.: A taxonomic review of metaheuristic algorithms for solving the
vehicle. Comput. Ind. Eng. 140, 106242 (2020)
3. Onar, S., Öztaysi, B., Kahraman, C., Yanık, S., Senvar, Ö.: A literature survey on
metaheuristics in production systems. In: Operations Research/Computer Science Interfaces
Series, pp. 1–24 (2016)
4. Kumar, A., Bawa, S.: A comparative review of meta-heuristic approaches to optimize the
SLA violation costs for dynamic execution of cloud services. Soft. Comput. 24(6), 3909–
3922 (2020)
5. Mirjalili, S., Mirjalili, S., Lewis, A.: Grey wolf optimizer. Adv. Eng. Softw. 69, 46–61
(2014)
6. Mirjalili, S.: The ant lion optimizer. Adv. Eng. Softw. 83, 80–98 (2015)
7. Mirjalili, S.: Moth-flame optimization algorithm: A novel nature-inspired heuristic paradigm.
Knowl. Based Syst. 89, 228–249 (2015)
8. Mirjalili, S., Lewis, A.: The whale optimization algorithm. Adv. Eng. Softw. 95, 51–67
(2016)
9. Venkata Rao, R.: Jaya: a simple and new optimization algorithm for solving constrained and
unconstrained optimization problems. Int. J. Ind. Eng. Comput. 7, 19–34 (2016)
10. Mirjalili, S.: SCA: a sine cosine algorithm for solving optimization problems. Knowl. Based
Syst. 96, 120–133 (2016)
11. Meng, X., Gao, X., Lu, L., Liu, Y., Zhang, H.: A new bio-inspired optimisation algorithm:
Bird Swarm Algorithm. J. Exp. Theor. Artif. Intell. 28(4), 673–687 (2016)
12. Abedinpourshotorban, H., Mariyam Shamsuddin, S., Beheshti, Z., Jawawi, D.: Electromag-
netic field optimization: a physics-inspired metaheuristic optimization algorithm. Swarm
Evol. Comput. 26, 8–22 (2016)
13. Mirjalili, S., Gandomi, A., Mirjalili, S., Saremi, S., Faris, H., Mirjalili, S.: Salp Swarm
Algorithm: a bio-inspired optimizer for engineering design problems. Adv. Eng. Softw. 114,
163–191 (2017)
14. Saremi, S., Mirjalili, S., Lewis, A.: Grasshopper optimisation algorithm: theory and
application. Adv. Eng. Softw. 105, 30–47 (2017)
15. de Melo, V., Banzhaf, W.: Drone squadron optimization: a novel self-adaptive algorithm for
global numerical optimization. Neural Comput. Appl. 30(10), 3117–3144 (2018)
16. Dong, H., Gao, L., Shen, P., Li, X., Lu, Y., Dai, W.: An interval type-2 fuzzy logic controller
design method for hydraulic actuators of a human-like robot by using improved drone
squadron optimization. Int. J. Adv. 16(6), 1–16 (2019)
17. Dhyani, A., Panda, M., Jha, B.: Design of an evolving Fuzzy-PID controller for optimal
trajectory control of a 7-DOF redundant manipulator with prioritized sub-tasks. Expert Syst.
Appl. (in press)
18. Sadeghi-Moghaddam, S., Hajiaghaei-Keshteli, M., Mahmoodjanloo, M.: New approaches in
metaheuristics to solve the fixed charge transportation problem in a fuzzy environment.
Neural Comput. Appl. 31, 477–497 (2019)
19. Singh, S., Singh, S., Banga, V.K.: Design of fuzzy logic system framework using
evolutionary techniques. Soft. Comput. 24(6), 4455–4468 (2019)
20. Tak, N.: Type-1 recurrent intuitionistic fuzzy functions for forecasting. Expert. Syst. Appl.
140 (2020)
Fuzzy and Evolutionary Algorithms
for Transport Logistics Under Uncertainty
Yuriy Kondratenko , Galyna Kondratenko , Ievgen Sidenko ,

and Mykyta Taranov(&)
Intelligent Information Systems Department, Petro Mohyla Black Sea National

University, 68th Desantnykiv Str., 10, Mykolaiv 54003, Ukraine
{yuriy.kondratenko,halyna.kondratenko,ievgen.sidenko,
mykyta.taranov}@chmnu.edu.ua
Abstract. This paper deals with the analysis of fuzzy and evolutionary
approaches for efficiently solving vehicle routing problems (VRP) with con-
straints on vehicle’s capacity (CVRP) and time-windows (VRPTW). Authors
focused their research on CVRP for marine bunkering tankers, in particular, on
the planning of tanker’s routes under uncertainty of fuel demands at nodes.
Triangular fuzzy numbers (TFNs) are proposed for modeling uncertain
demands. In this case, the maximum possible number of customers is calculated,
which can be served based on the subtraction operation with TFNs. In the paper,
the authors also analyzed the planning of transport routes with time-windows.
Currently, there are several methods and algorithms for planning of transport
routes with time-windows, in particular: saving and sweeping algorithms, ant
colony optimization (ACO) algorithm, artificial bee colony (ABC) algorithm,
etc. In this paper, the authors discussed the features of using the ACO algorithm
and the ABC algorithm to solve the vehicle routing problems with time-
windows and the influence of their application on the results. The modeling
results confirm the efficiency of the proposed fuzzy and evolutionary algorithms.
Keywords: Fuzzy model Triangular fuzzy number Vehicle routing

problem Planning Transport route Time-window Ant colony optimization
algorithm Artificial bee colony algorithm
1 Introduction
Cargo transportation (from food to industrial) is an integral part of the life of any
person. The need to solve transport problems (TP), with minimizing the cost of
transportation, is determined by the economic effect, as this allows increasing the profit
of the logistics company [1, 2]. The tasks of transport logistics include choosing a
vehicle for cargo delivery, planning and optimizing transport routes, including
uncertainty, choosing and optimizing a fleet of vehicles, creating logistics chains for
multimodal transportation, etc. One of the sub-tasks of planning and optimizing
transport routes are vehicle routing problems (VRP) with constraints on vehicle’s
capacity and time-windows [3, 4]. VRP are combinatorial optimization problems in
which for a fleet of vehicles located in one or more depots, a set of routes to several
Fuzzy and Evolutionary Algorithms for Transport Logistics Under Uncertainty 1457
consumers should be defined. The relevance of this paper is associated with the
practical significance of the VRP with considerable complexity. VRP is a well-known
integer-programming task belonging to the class of NP-hard problems, which means
that the computational complexity of the task exponentially depends on the size of the
input data [2, 4, 5]. The authors consider the use of a fuzzy technique for solving the
capacitated VRP (CVRP). In this case, authors focused their research on CVRP for
marine bunkering tankers, in particular, on the planning of tanker’s routes in conditions
of uncertain fuel demands at nodes. In the paper, the authors also analyzed the planning
of transport routes in VRP with time-windows (VRPTW). In this case, the authors
consider the use of evolutionary algorithms for solving the VRPTW.
2 Related Works and Problem Statement
It is known that classical methods and approaches to solving VRP, as a rule, are single-
threaded and cannot parallelize. Accordingly, the execution time grows exponentially
from the dimension of the problem. Exact methods for solving TP, allow to find a
solution only for problems with a small number of customers [2, 3]. For solving
problems of large dimensions, exact methods are not effective due to their large time
costs. However, right now, in conditions of the globalization of the economy, effective
approaches to solving problems of large dimensions are required. Such approaches
include fuzzy logic, fuzzy data models, genetic algorithms, evolutionary strategies and
methods, data mining, neural networks [6–8], etc.
In the paper [9] to optimize cargo transportation, the authors proposed an approach
of linear graphs. This approach is applied using two components. The first component
is fuzzy methods and technologies for describing input variables; the second compo-
nent is genetic algorithms for optimizing transport routes under uncertainty. Unlike
exact methods of mathematical programming, evolutionary methods allow finding
solutions close to optimal in an acceptable time.
The authors demonstrate the use of a soft computing approach for solving truck and
trailer routing problem [10]. The application of the current approach is that the
decision-maker allows the constraints to be “the maximum possible”. One of the
important tasks is the vehicle routing problem with fuzzy demands (VRPFD) using the
theory of fuzzy sets [11–14]. This article [15] proposes another approach, which
includes the structure of fuzzy sets, fuzzy optimization, modeling inaccuracy and
flexibility, and analyzes practical approaches to solving VRP under uncertainty. The
authors [16] give an example of the use of fuzzy logic methods for modeling logistics
flows, in particular, the composition of fuzzy relationships for the subsequent
automation of decision-making processes. Customer orders (demands) can be presented
in the form of random fuzzy data, and the travel time between customers is calculated
taking into account the probability distribution [13, 17]. The fuzzy optimization models
considered by the authors [18] allow solving the problems of transport route planning
using metaheuristics in a compressed form.
Recently, various evolutionary optimization methods and algorithms (particle
swarm optimization (PSO), ant colony optimization (ACO), artificial bee colony
(ABC)) are often used to solve the VRPFD and VRPTW [19–21]. The article [22]
1458 Y. Kondratenko et al.
presents a modification of the ACO algorithm using fuzzy logic and the formation of
fuzzy production rules to calculate the possibility of the transition of an ant from one
node to another. This allows taking into account the uncertainty of the transportation
cost at different sections of the route. In the paper [23] fuzzy VRPTW with an uncertain
travel time is considered. The authors propose a combination of a fuzzy algorithm and
an evolutionary algorithm to solve the VRPTW while maintaining the constraints.
The purpose of the paper is the analysis of fuzzy models and evolutionary
approaches for efficiently solving transportation and vehicle routing problems with
constraints on vehicle’s capacity and time-windows.
3 CVRPFD for Marine Bunkering Tankers
Consider the task of planning routes with fuzzy demands (CVRPFD) using an example
of marine bunkering tankers, in particular, planning of tanker’s routes in conditions of
uncertain fuel demands at nodes (ports).
At the initial stage, the tanker has a certain amount (a crisp number) of cargo (in
this case it’s fuel) that corresponds to the cargo capacity of the tanker Q0. Fuel demand
Ai at the i-th port is fuzzy and represented as a triangular fuzzy number (TFN). To
determine the amount of remaining fuel on the tanker after servicing the 1st port, the
TFN A1 (fuel demand) must be subtracted from a crisp number Q0 (total amount of fuel
on the tanker) [7, 12–14]. After that, the remaining fuel Q1 on the tanker also becomes
TFN. For example, the result of servicing the third port using subtracting
(Q3 Q2 A3) the two TFNs (Q2 , A3) is shown in Fig. 1.
Fig. 1. The result of servicing the third port with fuel demand as TFN A3.
In this case, TFN Q2 = ð220; 280; 280Þ is a model of the remaining fuel on the
tanker after servicing the second port, TFN A3 = ð160; 180; 200Þ is a model of the next
fuel demand at the third port of type “about 180” or “between 160 and 200”, TFN
Q3 = ð20; 100; 120Þ is a model of the remaining fuel on the tanker after servicing the
third port. The tanker can service the ports as long as the remaining fuel on the tanker
exceeds the port demand. When the appropriate condition is not fulfilled, a conflict
situation occurs and the service of ports by this tanker is terminated. The last demand
(where the conflict situation occurred) is not counted in the total number of completed
demands.
The process of servicing ports (for example, four) with fuzzy fuel demands is
presented in Fig. 2.
Fig. 2. The process of servicing ports with fuzzy fuel demands.
A conflict situation occurred during the servicing the fourth port, because the fuel
demand A4 = ð60; 80; 110Þ exceeded the remaining fuel Q3 = ð90; 150; 230Þ.
The result of the developed software can also be seen in another example with 6
ports (Fig. 3)
Fig. 3. Developed software for servicing ports with fuzzy fuel demands.
4 Evolutionary Algorithms for Solving the VRPTW
Consider the VRPTW task using the ACO and the ABC algorithms. In this case, time-
windows are presented as fuzzy intervals.
Currently, there are several known methods and approaches to solving the problem,
in particular, (a) exact and heuristic methods that include the method of branches and
boundaries, the method of branches with cutting, the method of complete search,
saving and sweeping algorithms, (b) meta-heuristic methods that include ACO algo-
rithm, ABC algorithm, wolf pack algorithm, PSO algorithm, genetic algorithm, tabu
search algorithm, and others [24, 25]. Using exact and heuristic methods takes a lot of
time and resources to find the optimal solution for a large number of nodes. Therefore,
it is advisable to use meta-heuristic methods, since they provide a sufficiently high
accuracy and require less time and resources when properly selecting and adjusting the
internal parameters [2, 5, 25].
To solve this VRPTW, an application that implements the ACO and the ABC
algorithm was developed. For the effective work of the application, the internal
parameters of the respective algorithms must be set.
An eleven-node task (for example) was selected to test the effect of internal
parameters on the route planning result. To test the effect of internal parameters on the
result of route planning, a task with eleven nodes was selected, which includes one
“zero” node-warehouse and ten customer nodes, to which it is necessary to deliver the
ordered quantity of cargo in a specific period of time (according to time-windows), for

example, [1, 3], which means “from 1 to 3 pm”.
Analyzed the influence of all the parameters of the ACO algorithm on the result of
planning, the authors determined the optimal values: greed a = 0.3, herd b = 0.75, the
rate of weathering of the pheromone is 0.001, the number of ant-agents is 3, the number
of iterations of the algorithm is 500, and the maximum permissible error of the result is
0.0001.
To further compare the ACO and the ABC algorithms, the same test sample should
be selected to test the ABC algorithm. Analyzed the influence of all parameters of the
ABC algorithm on the result of planning, the authors determined the optimal values:
the ratio of different types of bees is 2:5:3, maximum number of bees is 200, the
number of iterations of the algorithm is 500, the maximum number of visits to the
neighborhood of the best decisions by one bee is 100, the maximum permissible error
of the result is 0.01, and the probability that the bee will make a mistake is 0.01.
Next, these algorithms are compared according to the main criteria (Fig. 4). Figure 4,
a shows the deviations of the route length from the best solution found by both algorithms.
ACO algorithm can be considered a little better than the ABC method since in most
solutions it gives a smaller deviation than the ABC algorithm. The algorithms are com-
pared in terms of the total time spent on route maintenance (Fig. 4b). It clearly shows that
the ABC algorithm provides solutions that are much closer in time to the best solution
found in both algorithms.
Fig. 4. Comparison of ACO and ABC algorithms by: a) deviations from the best found length;
b) deviations from the best found time; c) the effectiveness of the use of time; d) transport loading
efficiency; e) solution search time
The ACO algorithm gives a higher efficiency of the use of time (Fig. 4c). The ACO
algorithm in most cases gives the solutions with a slightly higher efficiency of transport
loading, but since the values are almost equal, it is accepted that they are approximately
the same (Fig. 4d). The ACO algorithm is faster to find decisions than the ABC
algorithm (Fig. 4e).
Therefore, the results of the comparison (Fig. 4) show that the ACO algorithm
produces better results than the ABC algorithm, and therefore is more efficient for the
planning of transport routes in VRP with time-windows.
5 Conclusions
The authors analyzed of fuzzy and evolutionary approaches for efficiently solving
CVRP and VRPTW tasks. Authors considered the CVRPFD using an example of
marine bunkering tankers, in particular, planning of tanker’s routes in conditions of
uncertain fuel demands at nodes (ports). TFNs are proposed for modeling uncertain
demands. In this case, the maximum possible number of customers is calculated, which
can be served based on the subtraction operation with TFNs. The authors developed
software to support the process of servicing ports with fuzzy fuel demands.
In this paper, the authors discussed the features of using the ACO algorithm and the
ABC algorithm to solve the vehicle routing problems with time-windows and the
influence of their parameters on the results. The modeling results confirm the efficiency
of the proposed fuzzy and evolutionary algorithms.
References
1. Gath, M.: Optimizing Transport Logistics Processes with Multiagent Planning and Control.
Springer, Wiesbaden (2016). https://doi.org/10.1007/978-3-658-14003-8
2. Kappauf, J., Lauterbach, B., Koch, M.: Transport logistics. In: Logistic Core Operations with
SAP, pp. 11–98. Springer, Heidelberg (2012)
3. Golden, B., Raghavan, S., Wasil, E.: The Vehicle Routing Problem: Latest Advances and
New Challenges. Springer, Boston (2008). https://doi.org/10.1007/978-0-387-77778-8
4. Roch, C., Langer, S.: The capacitated vehicle routing problem. Digitale Welt 3, 30–33
(2019). https://doi.org/10.1007/s42354-019-0165-z
5. Yang, C., Guo, Z., Liu, L.: Comparison study on algorithms for vehicle routing problem
with time windows. In: 21st International Conference on Industrial Engineering and
Engineering Management, Paris, pp. 257–260. Atlantis Press (2015)
6. Solesvik, M., Kondratenko, Y., Kondratenko, G., Sidenko, I., Kharchenko, V., Boyarchuk,
A.: Fuzzy decision support systems in marine practice. In: IEEE International Conference on
Fuzzy Systems (FUZZ-IEEE), Naples, Italy, pp. 1–6 (2017)
7. Kondratenko, Y.P., Kondratenko, N.Y.: Synthesis of analytic models for subtraction of fuzzy
numbers with various membership function’s shapes. In: Gil-Lafuente, A., Merigó, J., Dass,
B., Verma, R. (eds.) Applied Mathematics and Computational Intelligence. FIM 2015.
Advances in Intelligent Systems and Computing, vol. 730, pp. 87–100. Springer, Cham
(2018). https://doi.org/10.1007/978-3-319-75792-6_8
8. Kondratenko, Y.P., Klymenko, L.P., Sidenko, I.V.: Comparative analysis of evaluation
algorithms for decision-making in transport logistics. In: Jamshidi, M., Kreinovich, V.,
Kacprzyk, J. (eds.) Advance Trends in Soft Computing. Studies in Fuzziness and Soft
9. Kobersy, I., Shkurkin, D.: Application of genetic algorithms for optimization of transport
problems. News of the SFU. Technical Science, pp. 172–176 (2012). (in Russian)
10. Torres, I., Rosete, A., Cruz, C., Verdegay, J.: Fuzzy constraints in the truck and trailer
routing problem. In: 4th International Workshop on Knowledge Discovery, Knowledge
Management and Decision Support, pp. 71–78. Atlantis Press (2013)
11. Erbao, C., Mingyong, L.: A hybrid differential evolution algorithm to vehicle routing
problem with fuzzy demands. J. Comput. Appl. Math. 231(1), 302–310 (2009). https://doi.
org/10.1016/j.cam.2009.02.015
12. Teodorovic, D., Pavkovic, G.: The fuzzy set theory approach to the vehicle routing problem
when demand at nodes is uncertain. Fuzzy Sets Syst. 82, 307–317 (1996)
13. Werners, B., Kondratenko, Y.: Alternative fuzzy approaches for efficiently solving the
capacitated vehicle routing problem in conditions of uncertain demands. In: Berger-Vachon,
C., Gil Lafuente, A., Kacprzyk, J., Kondratenko, Y., Merigó, J., Morabito, C. (eds.)
Complex Systems: Solutions and Challenges in Economics, Management and Engineering.
Studies in Systems, Decision and Control, vol. 125, pp. 521–543. Springer, Cham (2018).
14. Kondratenko, G.V., Kondratenko, Y.P., Romanov, D.O.: Fuzzy models for capacitive
vehicle routing problem in uncertainty. In: 17th International DAAAM Symposium
“Intelligent Manufacturing and Automation: Focus on Mechatronics & Robotics”, Vienna,
Austria, pp. 205–206 (2006)
15. El-Sherbeny, N.: Imprecision and flexible constraints in fuzzy vehicle routing problem. Am.
J. Math. Manag. Sci. 31(1–2), 55–71 (2011)
16. Dli, M., Gimarov, V., Glishko, S., Ivanova, I.: Fuzzy-production ant algorithm for
optimizing an enterprise transport network. Transp. Bus. Russ. 5, 135–136 (2013). (in
Russian)
17. He, Y., Xu, J.: A class of random fuzzy programming model and its application to vehicle
routing problem. World J. Model. Simul. 1(1), 3–11 (2005)
18. Brito, J., Moreno, J., Verdegay, J.: Transport route planning models based on fuzzy
approach. Iran. J. Fuzzy Syst. 9(1), 141–158 (2012)
19. Peng, Y., Qian, Y.: A particle swarm optimization to vehicle routing problem with fuzzy
demands. J. Converg. Inf. Technol. 5(6), 112–119 (2010)
20. Cheng, R., Gen, M.: Vehicle routing problem with fuzzy due-time using genetic algorithm.
Jpn. J. Fuzzy Theory Syst. 7, 1050–1061 (1995)
21. Gupta, R., Singh, B., Pandey, D.: Fuzzy vehicle routing problem with uncertainty in service
time. Int. J. Contemp. Math. Sci. 5(11), 497–507 (2010)
22. Nikishov, S.: Application of fuzzy composition for logistics flows modeling. Quest. Innov.
Econ. 7(3), 247–256 (2017). (in Russian)
23. Sandhya, B., Katiyar, V.: Integrating fuzzy and ant colony system for fuzzy vehicle routing
problem with time windows. Int. J. Comput. Sci. & Appl. (IJCSA) 4(5), 73–85 (2014).
https://doi.org/10.5121/ijcsa.2014.4506
24. Bräysy, O., Dullaert, W., Gendreau, M.: Evolutionary algorithms for the vehicle routing
problem with time windows. J. Heuristics 10, 587–611 (2004)
25. Ombuki, B., Ross, B., Hanshar, F.: Multi-objective genetic algorithms for vehicle routing
problem with time windows. J. Appl. Intell. 24, 17–30 (2006)
Using GA for Locating Post-disaster Field
Hospitals: The Case of Istanbul
Yesim Komurcu1 and Seda Yanik2(&)

1
Halic University, Beyoglu, 34421 Istanbul, Turkey
[email protected]
2
Istanbul Technical University, Macka, Besiktas, 34367 Istanbul, Turkey
[email protected]
Abstract. Most of the post-disaster losses occur due to the lack of humanitarian
relief planning or inadequate implementations. Locating temporary field hos-
pitals (FHs) and assignment of the injured people to the hospitals are key issues
in natural disaster management. The capacities of existing emergency units of
hospitals (EH) are expected to be not enough for the injured people in case of a
serious earthquake in Istanbul. Thus, the FHs which will serve as additional
capacity are needed. The objectives of this study are to minimize the total travel
cost and the total cost of establishing FHs in order to respond to all of the injured
as quickly as possible. To this aim, a heuristic solution methodology, genetic
algorithm (GA) is developed in MATLAB. Different crossover and replacement
strategies have been tested in order to improve the performance of the algorithm.
In the experimental study, the performances of different GAs strategies are
compared using experimental datasets where optimal results are known.
According to the result of the comparison, the best GA is selected and applied
the for the real Istanbul dataset.
Keywords: Humanitarian relief Location-allocation Genetic algorithm
1 Introduction
1.1 Natural Disaster

Natural disasters affect the lives of human beings all around the world and they cause
big losses. Most of the post-disaster losses occur due to the lack of humanitarian relief
planning or inadequate implementations. A timely and effective response to disaster
requires the interaction and coordination of many parties. After a natural disaster such
as earthquakes, fire, etc. the medical intervention for the injured people should be given
as quick as possible. The capacities of existing emergency units of the EHs are
expected to be not enough for the injured people in case of a serious earthquake in
Istanbul. In order to treat the injured people quickly, emergency units are needed to be
established close to the locations of the injured population [1].
Salman and Gul [2] utilize multi period mixed integer programming to determine
the location and size of emergency FH to be established after a disaster to cope with
demand surge. Salman and Yucel [3] apply tabu search algorithm for Istanbul earth-
quake with link failure cases. Zahiri et al. [4] develop a robust possibilistic
Using GA for Locating Post-disaster Field Hospitals 1465
programming model for multi-period location-allocation problem in an organ transplant

supply chain. Sha and Huang [5] propose a multi-period location-allocation model and
use heuristic algorithm based on Lagrangian relaxation for emergency blood supply
schedule. Saeidian et al. [6] compare the performance of GA and Bees Algorithm for
finding the optimum location of relief centers and allocating the parcels to them. Zhang
et al. [7] uses a GA-based multi-objective optimization approach to find the tradeoffs
between the objectives in locating health-care facilities.
This study deals with the post-disaster relief planning for the expected earthquake
in Istanbul which is located on a major fault line and hosts more than 17 million
populations. The FHs which will be quickly constructed after the disaster aim to serve
as additional capacity. The problem we study is finding the optimum locations of the
FHs considering the EHs and the assignment of the injured people to the hospitals.
A mixed integer model is formulated for the problem. The objectives of the model are
to minimize the total travel cost and the total cost of establishing FHs in order to
respond to all of the injured as quickly as possible. Since, finding the optimal result for
the real-sized problems is not possible with the mixed integer programming, a heuristic
solution methodology, genetic algorithm (GA) is developed in MATLAB. Different
crossover and replacement strategies have been tested in order to improve the per-
formance of the algorithm. In the experimental study, the performances of different
GAs strategies are compared using experimental datasets where optimal results are
known. According to the result of the comparison, the best GA is selected and applied
for real Istanbul dataset.
The study is organized as follows: In Sect. 2 model formulation and the solution
methodology are presented. In Sect. 3, a case study is described and in Sect. 4 the
computational results are presented. In Sect. 5, conclusions are summarized.
2 Model Formulation and Solution Methodology
In this section, we first present the model formulation for the problem of locating FHs
after an earthquake and the allocation of the injured people to the hospitals. Then, the
GA design which we use as a heuristic solution methodology is described.
2.1 Location – Allocation Model Formulation

The sets, indices, parameters and the decision variables of the mathematical model are
given and then the model formulation is presented in what follows:
Sets and Indices

H, h: Set and index of hospital locations, h 2 H
I, I: Set and index of EH locations, I H
J, j: Set and index of temporary FH locations, J H
P, p: Set and index of patient/injured locations (PL), p 2 P
1466 Y. Komurcu and S. Yanik
Parameters
s: Setup cost of one FH
n: Total number of FH that can be set up
cj : The capacity of the FH at location j
pi : The capacity of EH at location i
dph : Travel distance from location p to location h
r: Travel distance limit
bph : 1, if dph r; 0, otherwise.
yp : Number of injured people in location p.
Decision Variables
wph : Number of assigned injured people from location p to location h

xj : 1; if open FH in location j
0; otherwise:
Using the above notation, the mixed integer model for post-disaster FH location
allocation model is formulized as follows:
X X X
Minimize z ¼ j2J
s : xj þ p2P
d : wph
h2H ph
ð1Þ
subject to
X
p2P
wpi pi 8i 2 I ð2Þ
X
p2P
wpj xj cj 8j 2 J ð3Þ
X
j2J
xj n ð4Þ
X
h2H
wph bph ¼ yp 8p 2 P ð5Þ
wph 0 and integer ð6Þ
xj 2 f0; 1g ð7Þ
The objective function minimizes the total travel cost of the transported injured
people and the total setup cost of opened FHs, in Eq. (1). In Eq. (2), the number of
injured people assigned to the EHs does not exceed the capacity, in Eq. (3), the number
of injured people assigned to the temporary FHs does not exceed the capacity.
Equation (4) limits the total capacity of the opened FH in all locations by in Eq. (5) is a
constraint to serve all injured people. Equation (6) and Eq. (7) are sign constraints.
2.2 Solution Methodology

The objectives of the problem are to minimize the total travel cost and the total cost of
establishing FHs in order to respond to all of the injured as quickly as possible. To this
aim, a heuristic solution methodology, GA is developed in MATLAB.
The solutions are encoded using a chromosome consisting of a number of genes in
parallel to the number of injured population locations. And the genes are encoded using
a real number which shows the number of the FH to which the people at the specific
injured population location is assigned. In Table 1, the design of the GA is
summarized.
Table 1. Genetic algorithm parameters.

Genetic algorithm Information of parameters
parameters
Encoding Real number encoding
Population size 100
Selection operator Roulette wheel selection
Replacement Worst 50% individual of population and worst 1 individual of
population.
Crossover Two-point crossover and union crossover.
Mutation 1% of population size
Iteration number 100
Stop criterion 3h
2.3 Experimental Study

The experimental study is employed in order to compare four different variants of the
GA design with the optimal result of a dataset. The four variants of the GA are
generated by the combination of the two crossover approaches, namely two-point
crossover (TPC) or union crossover (UC) approaches and the two replacement
strategies, namely replacement of worst 50% individual of population or replacement
of worst 1 individual of population from generation to generation. The size and defi-
nition of the dataset used in this experimental study is given in Table 2.
The four different GAs are run for several times and the averages of the obtained
results are summarized in Table 3. The results show that TPC does not give the optimal
result. Thus, we eliminate the use of TPC and continue to use the UC by which the
optimal solution is obtained. Only using the UC, the replacement strategies, namely
worst 50% replacement and worst 1 replacement are compared. Both of these
replacement strategies give the optimal result but the replacement of 1 population finds
the optimal solution faster as shown in Fig. 1. In the experimental study, GA’s stopping
criterion is set as 1 h. Dotted lines in Fig. 1 show the average computation time of the
10 runs. Using the UC & worst 1 replacement strategies, average run time is 34 min.
For UC & worst %50 replacement, average run time is 60 min.
Table 2. Size and definition of the experimental dataset.

Data of experimental study Number of data
Number of EH 5
Number of potential temporary FH 17
Number of patient location 11
Setup cost of temporary FH 3,000 Turkish lira (TL)
Capacities of EH 350–650 person/hospital
Capacities of temporary FH 300 person/hospital
Table 3. Comparison of the results obtained by the four different GA designs.

Run TPC & worst 50% TPC & worst 1 UC & worst 50% UC & worst 1
number replacement replacement replacement replacement
1 8042 8956 7888 7888
2 8744 8282 7888 7888
3 8304 8274 7888 7888
4 8580 8318 7888 7888
5 9324 7892 7888 7888
6 8138 8762 7888 7888
7 8126 8226 7888 7888
8 8562 8084 7888 7888
9 8242 8474 7888 7888
10 8050 8270 7888 7888
Fig. 1. UC comparison of computation time (minute) according to run number.

3 Case Study
The case study is conducted for a serious earthquake scenario in Istanbul. Istanbul is
one of the largest cities in the world with a population 17 million people. There exist
pre-disaster preparations for an earthquake in Istanbul. Also anticipations related to
post-disaster situation are generated and post-disaster planning is made. In our case
study, we aim to find the optimal FH locations and assignment of PLs to the hospitals
in two districts of Istanbul, namely Bahcelievler and Kucukcekmece. These two dis-
tricts are expected to be severely affected by an earthquake. Using Japan International
Cooperation Agency (JICA) reports of Istanbul earthquake scenarios, we collected
required datasets for our case study as follows:
Number of patients: According to the JICA Report, 2% of Bahcelievler population
and 1% of Kucukcekmece population will be seriously injured after an earthquake.
Using the population data of the districts, we calculate the number of injured people in
each district. Then we distribute the injuries to each subdistrict of the districts with
respect to the populations of the subdistricts. When the number of injuries in each
subdistrict is less than 200 people, we represent the injuries at one point which is the
center of the subdistrict. If the number of injuries in a subdistrict is greater than 200, we
assign more than one point to that subdistrict where each point represents the location
of maximum 200 injured people.
EHs and capacities: In case of disasters, the multi-floor structure of the hospitals
can’t be utilized. Only the emergency services are in use and the capacity of the
emergency units in EHs depends on the bed number of hospitals. Using the number of
beds of EHs, we calculate the capacities of the emergency units of the EHs.
Potential FH and capacities: The mobile FHs can be established on football fields,
school gardens and parks. In these areas it is assumed that mobile fields hospitals can
be constructed with an area of 520 m2 and serve 208 patients.
FH setup cost: Information on the costs of tent, motor and isolation costs are
collected and the cost for a 520 m2 FH is calculated to be approximately 650 K TL.
Distance between hospital and PL: Euclidean distance is calculated using the
latitude and longitude information of the hospitals and patient locations.
A summary of the dataset related to the Istanbul case study is given in Table 4.
Table 4. Dataset for the Istanbul case study.

Real Istanbul data Number of data
Number of EH 26
Number of potential temporary FH 243
Number of patient location 118
Setup cost of temporary FH 650,000 TL
Capacities of EH 160–300 person/hospital (changeable)
Capacities of temporary FH 208 person/hospital
Location of potential temporary FH Schoolyard, football field, parking area
4 Computational Results
In the computational study, we conduct two sets of runs changing the criteria of
assignment of patients to the hospitals as follows: (i) without a travel distance limit
(wodGA), (ii) with a travel distance limit (wdGA). Then we compare the results of both
in terms of travel costs of the patient-hospital assignment and setup costs of the FHs.
When GA is run without the distance limit (wodGA), the best solution with respect
to total cost is found in run number 2. According to this best solution, EHs serve 32
PLs. 86 new FHs are opened in order to serve the rest of the PLs. 94.7% of the total
cost is the setup cost and 5.3% is the travel cost. When GA is run with the distance
limit (wdGA), the best solution assigns 29 PLs to the EHs. For the rest of the PLs, 89
new FHs are opened. 95.4% of the total cost is the setup cost and 4.6% is the travel
cost. The results of 5 runs with and without the travel limit are shown in Table 5.
Table 5. GA results without (i) and with (ii) travel distance limit (currency: Turkish lira).
Run Total cost Setup cost Travel cost Total cost Setup cost Travel cost
number (wodGA) (wodGA) (wodGA) (wdGA) (wdGA) (wdGA)
1 59,145,545 55,250,000 3,895,545 61,143,416 57,850,000 3,293,416
2 59,023,863 55,900,000 3,123,863 60,638,902 57,850,000 2,788,902
3 59,164,993 55,250,000 3,914,993 61,262,931 57,850,000 3,412,931
4 59,162,232 55,250,000 3,912,232 61,252,393 57,850,000 3,402,393
5 59,153,377 55,250,000 3,903,377 61,264,053 57,850,000 3,414,053
Average 59,130,002 55,380,000 3,750,002 61,112,339 57,850,000 3,262,339
The solution without the distance limit has an average of unutilized capacity of 30
persons/hospital. 85.9% of the capacity of EHs and newly opened FHs are used. When
distance limit is incorporated into the model, the average unutilized capacity increases
to 33 persons/hospital. 84.3% of the capacity of EHs and opened FHs are used.
Without the distance limit, only 61% of the patients are assigned to hospitals that
are in 5 km radius from the PL as seen in Fig. 2. Thus we include the travel distance in
the model in order to ensure that all patients are assigned to hospitals that are in 5 km
radius. In Fig. 3, number of PLs and their distances to the assigned hospitals are given
both in case of solving the model with and without the distance limit.
Fig. 2. Distance between assigned PL and hospital (m), respectively, (a) without the distance
limit (b) with the distance limit.
Fig. 3. Number of PLs assigned to hospitals on the basis of distance, (a) without the distance
limit (b) with the distance limit.
5 Conclusions
Earthquake is a natural disaster of which the intensity and timing can’t be predicted.
However, loss of life and property can be minimized by effective disaster management
after the earthquake. Istanbul is a very crowded city on the fault lines. Since the
capacity of permanent hospitals is expected to be not enough for the injured people in
case of a serious earthquake in Istanbul, FHs should be established in open areas.
Identifying the optimal locations of these FHs and assignment of the injured to hos-
pitals are important to decrease the response time.
In the study, the performance of four GA designs with different types of crossover
and replacement strategies are compared using an experimental dataset. Then, the best
GA is applied to real Istanbul data. Because rapid response to injured is more valuable
than cost of setup, distance limit is added to the model and two models with and
without distance limit are evaluated. It is found that a distance limit of 5 km. between
the hospitals and patients can be ensured with a 2.75% increase in the total cost.
In the future research, factors affecting the travel (such as damage to the road,
demolition of the building to the road) can be taken into account. Moreover, the supply
transportation to the hospitals from specified warehouses can also be included in the
location model.
References
1. Boonmee, C., Arimura, M., Asada, T.: Facility location optimization model for emergency
humanitarian logistics. Int. J. Disaster Risk Reduct. 24, 485–498 (2017)
2. Salman, F.S., Gul, S.: Deployment of field hospitals in mass casualty incidents. Comput. Ind.
Eng. 74, 37–51 (2014)
3. Salman, F.S., Yucel, E.: Emergency facility location under random network damage: insights
from the Istanbul case. Comput. Oper. Res. 62, 266–281 (2015)
4. Zahiri, B., Moghaddam, R.T., Pishvaee, M.S.: A robust possibilistic programming approach
to multi-period location-allocation of organ transplant centers under uncertainty. Comput. Ind.
Eng. 74, 139–148 (2014)
5. Sha, Y., Huang, J.: The multi-period location-allocation problem of engineering emergency
blood supply systems. Syst. Eng. Procedia 5, 21–28 (2012)
6. Saeidian, B., Mesgari, M.S., Ghodousi, M.: Evaluation and comparison of Genetic Algorithm
and Bees Algorithm for location–allocation of earthquake relief centers. Int. J. Disaster Risk
Reduct. 15, 94–107 (2016)
7. Zhang, W., Cao, K., Liu, S., Huang, B.: A multi-objective optimization approach for health-
care facility location-allocation problems in highly developed cities such as Hong Kong.
Comput. Environ. Urban Syst. 59, 220–230 (2016)
Optimization
Optimization of Spatial-Time Planning
Resource Allocation Under Uncertainty
Olesiya Kosenko , Alexander Bozhenyuk(&) ,

Stanislav Belyakov , and Margarita Knyazeva
Southern Federal University, Nekrasovsky str. 44, 347922 Taganrog, Russia

Abstract. This paper is devoted to the problem of project planning and optimal
resource allocation under fuzzy estimated parameters. The effective functioning
of the enterprise directly depends on the early supply and optimal resource
delivery of various types. At the same time, one needs to take into consideration
scheduling flexibilities of activity planning, inaccurate data, uncertain resource
levelling and the resource availability. The paper takes into consideration the
dynamical temporal aspect of project planning problem that introduces signifi-
cant uncertainty when planning the activities for the enterprise and obtaining the
optimal solution for resource allocation problem.
Keywords: Project planning Resource allocation Multi-indexing

Efficiency Uncertainty Fuzzy optimization Dynamic problem
1 Introduction
A feasible assessment of the time and cost parameters for the production system, carried
out as a part of the network project elements interaction of the distribution production
system, is relevant. One of the main problems of network interaction and resource
allocation is to provide a certain consumer with the necessary resource of the right
quantity with minimal cost [1]. In complex systems, distributions and allocations made
in one period can affect distributions in subsequent periods. In this case, the dynamic
components of the system must be taken into account when solving the optimization
problem [2], as well as the time to find optimal allocation solution is important as well.
The purpose of this article is to find the optimal solution to the distribution problem,
which would take into account as many parameters as possible, affecting the final result
of solving the problem. Moreover, the parameters of the resource allocation problem
are presented in the form of fuzzy intervals.
This paper is organized as follows: Sect. 2 considers potential method application
for multi-index distribution problems. Section 3 formalizes multi-index problem based
on fuzzy interval parameters estimation. In Sect. 4 the method for optimal multi-index
problem decision based on fuzzy interval parameters estimation is introduced. Section 5
introduces computational experiments for the developed method. And the last section
considers conclusion and future research suggestions.
1476 O. Kosenko et al.
2 Analysis of Methods for Solving Distribution Problems
Linear programming methods for mathematical optimization are considered as an

effective tool for solving resource allocation problems, transportation problems ant
others, especially those found in industrial production systems [3]. The distribution and
allocation problems are the special cases of the linear programming problem, the main
idea of which is to optimize the allocation of resources with minimal cost. The solution
of the problems is provided by one of the linear programming methods, named the
Hungarian Algorithm for Assignment Problem, Index Method or Potential Method.
In a two-index transportation problem (the first index determines the producers or
suppliers of a homogeneous resource, the second index determines the consumers of
this resource), the solution of equations can be obtained as follows:
vj ui cij ; i ¼ 1; I; j ¼ 1; J;
vj ui ¼ cij ; 8xij [ 0;
where vj and ui are the capacities of the corresponding producers and consumers
centers [4].
The classical distribution and allocation problem does not fully describe the spe-
cifics of needs related to the movement of resources, since it is determined only by the
resources availability of producers and the demand level of consumers.
In practice, the problem often arises of transportation (moving) heterogeneous
resources using transmitting elements (vehicles) of various types through storage and
redistribution centers, that leads to an increase in the number of indices of the problem
being solved.
The general formulation of the multi-index resource allocation problem is defined
as follows [5, 6]: the idea is to find a vector X ¼ ðx1 ; x2 ; . . .xk Þ, minimizing the
objective function
n1 X
X n2 X
nk
F ð xÞ ¼ ... ci1 i2... ik xi1 i2... ik ;
i1 ¼1 i2 ¼1 ik ¼1
with respect to the constraints:
X
n1
xi1 i2... ik ¼ bi2 i3... ik ; iq 2 Iq ; iq ¼ ð2; 3; . . .; k Þ;
i1 ¼1
X
n2
xi1 i2... ik ¼ bi1 i3... ik ; iq 2 Iq ; iq ¼ ð1; 3; . . .; k Þ;
i2 ¼1
X
nk
xi1 i2... ik ¼ bi1 i2... ik1 ; iq 2 Iq ; iq ¼ ð1; 2; . . .; k 1Þ;
ik ¼1
Optimization of Spatial-Time Planning Resource Allocation Under Uncertainty 1477
and non-negative transportation volumes of cargo:
xi1 i2... ik 0; i1 2 I1 ; i2 2 I2 ; ik 2 Ik ;
where ci1 i2... ik is cost matrix for the resource unit allocation xi1 i2... ik with respect to
ði1 ; i2 ; . . .; ik Þ; and bi1 i2... ik1 is the total amount of resources that are transported with
respect to transportation volumes ði1 ; i2 ; . . .; ik Þ.
The Potential Method that was developed for solving distribution problems is used
for two-dimensional and sometimes three-dimensional problems with small inputs.
Also, when applying the Potential Method, it is necessary to consider convergence
analysis [7]. In paper [8], the nature of convergence is discussed and several methods in
distribution problems are proposed, but all of them are applicable for two-index
problems of small input. The recommended methods for searching the best solution of
a multi-index distribution problem, according to studies [9, 10], are approximate
methods.
3 Multi-index Distribution Problem Formulation with Fuzzy

Parameters
To solve the distribution problem, it is proposed to formalize the initial parameters of

the problem as intervals.
Definition. The fuzzy interval is specified on the set Q by four parameters as follows [11]:
~ ¼ ða; b; A; BÞ;
Q
where a – is the left value of uncertainty, and b is the right value of uncertainty in the
form of the interval, A – is a lower modal value and B – is the upper modal value of
fuzzy interval [10]. Values A and B is the core of a fuzzy interval.
Let H ¼ f1; 2; . . .; mg be the set of real numbers. To each number h 2 H we
associate the parameter ih, which is determined by the index using one of the set values
f1; 2; . . .; nh g ¼ Ih . The set of the index-values (i1 ; i2 ; . . .; im ) can be set as m-indexed
element and let it be denoted as be Z [11] so that Z ¼ ði1 ; i2 ; . . .; im Þ, where i1 2
f1; 2; . . .; n1 g ¼ I1 ; i2 2 f1; 2; . . .; n2 g ¼ I2 ; . . .im 2 f1; 2; . . .; nm g ¼ Im : Q
The number of different sets of indexes values (i1 ; i2 ; . . .; im ) is N ¼ m h¼1 nh .
These sets form the set E of m-indexes Z, which can be defines as direct product of sets
I1 ; I2 ; . . .; Im as follows: E ¼ fZ1 ; Z2 ; . . .; ZN g ¼ I1 I2 . . . Im .
To each m-index element Z 2 E let’s put into correspondence a fuzzy number
~xi1 ;i2 ;...;im ¼ ~xz . The set of all possible values of indexes i1 ; i2 ; . . .; im is called m-index
matrix and can be denoted as f~xi1 ;i2 ;...;im g ¼ f~xz g.
Thus, a multi-index (m-index) transportation and distribution problem with fuzzy
parameters can be formulated as follows:
X
n1 X
n2 X
nm
~ ðX Þ ¼
F ... ~ci1 i2 ...im ~xi1 i2 ...im ! min;
i1 ¼1 i2 ¼1 im ¼1
Pn1 Pn 2
subject to i1 ~xi1 i2 ...im ¼ ~bi1 i2 ...im ; i1 2 I1 ; i2 ~xi1 i2 ...im ¼ ~
bi1 i3 ...im ; i2 2 I2 ; . . .;
Pnm
~
x ¼ ~
b ; i 2 I :
im i1 i2 ...im i1 i2 ...im m m
The transportation volumes are non-negative: ~xi1 i2 ...im 0; i1 2 I1 ; i2 2 I2 ; im 2 Im ,
where ~ci1 i2 ...im - is the set of fuzzy transportation costs for the resource unit ~xi1 i2 ...im ; and
~bi i ...i - is the set of fuzzy resource volumes, that are transported subject to
1 2 m
(i1 ; i2 ; . . .im ).

Let’s introduce m-indexed matrices of fuzzy numbers f~cZ g; ~ bZ and a u subsets
Zj, j 2 f1; 2; . . .; ug. The multi-index distribution problem is based on set X ~ ¼ f~xZ g
search to minimize linear function:
X
~ ðX Þ ¼
F ~cZ ~xZ ! min; ð1Þ
E
subject to
X X X
E1
~xZ ¼ ~bZ1 ; Z1 2 E1 ; E2
~xZ ¼ ~bZ2 ; Z2 2 E2 ; . . .; Eu
~xZ ¼ ~
bZu ; Zu 2 Eu ; ð2Þ
~xZ 0; Z 2 E: ð3Þ
The multi-index distribution problem stated with Eqs. (1)–(3) consists of N ¼

Q
m
nh variables. The Eq. (2) consists of u groups linear equations, while j-group
h¼1
consists of Uj equations. The total number of equations in the system is:
X
u
U¼ Uj :
j¼1
The m-index distribution problem Rs can be presented in vector-matrix form. To

each variable ~xZ , Z 2 E let’s put into correspondence U-dimension vector
~ Z ¼ ð~
P ~ UZN Þ, each component of which w
w1Z1 ; . . .; w ~ UZN , k 2 f1; 2; . . .; U g is equal to
zero, if the variable ~xZ is not included into k-th equation, otherwise it will be equal to
the coefficient ~aZ . Thus we can obtain the following vector:
~ Z is composed of u -groups of elements, which are coefficients of xF in the

Vector P
correspondent group of equations. However, i-th group consists of only one non-zero
~ UN :
~ Z , Z 2 E into matrix form P
element ~aZj . Let’s convert vector representation P
2 3
w~ 1Z1 ~ 1Z2
w ... ~ 1ZN
w
6w~ ~ 2Z2
w ... ~ 2ZN 7
w
~ ¼ 6 2Z1
P 7 ¼ ðw
~ kZh Þ; k 2 f1; 2; . . .; U g; Zh 2 E:
4 ... ... ... ... 5
~ UZ1
w ~ UZ2
w ... ~ UZN
w
We introduce the vector of constraints:
Interpreting m-indexed matrices f~cz g and f~xZ g as the vectors C ~¼

~ ð~x1 ; ~x2 ; . . .; ~xN Þ the multi-index linear programming problem can be
ð~c1 ; ~c2 ; . . .; ~cN Þ; X
stated as follows:
~ XÞ
Fð ~ X;
~ ¼C ~
subject to
~X
P ~ ¼ B;
~ X~ 0
Matrix P~ consists of u-groups of lines, and j-th group of conditions (8) and each of
it contains one unit-number in each column. The sum of all elements in the equation for
P P
the j-th group of conditions leads to the equation: Z2E ~xZ ¼ Zj 2Ej ~ bZj j .
System of Eqs. (2) contains at least т – 1 linear-dependent equations from different
groups. The number of dependent equations is q u 1 due to the matrix structure P. ~
4 The Method of Indexed-Elements for Multi-index

Distribution Problem with Fuzzy Parameters
Let’s consider the proposed method of indexed-elements, which takes into consider-
ation the uncertainty of the initial data and let’s consider a five-index distribution
problem. The method of index elements is based on the combination of the zero-
transformation method for solving three-index transportation problems and the fuzzy-
interval approach when estimating problem parameters [10].
To determine the parameters of the five-index resource allocation problem, we
present the objective function as follows:
XI XJ XK XL XR
~ ðX Þ ¼
F ~c ~x ! min; ð4Þ
i¼1 j¼1 k¼1 l¼1 r¼1 ijklr ijklr
where ~cijklr , i 2 I; j 2 J; k 2 K; l 2 L; r 2 R is a fuzzy set given on the set of real

numbers Hc and it defines the fuzzy transportation cost of one unit of r-th resource
from i-th production center to the j-th consumption center through the k-th distribution
center with the help of l-type transmission element;
~xijklr , i 2 I; j 2 J; k 2 K; l 2 L; r 2 R is a fuzzy set given on the set of real numbers

Hx that defines fuzzy amount of r-th resource, that is transported from i-th i-th pro-
duction center to the j-th consumption center through the k-th distribution center with
the help of l-type transmission element.
Fuzzy objective function (4) is limited on the set of feasible schedules, because of
~xijklr ~ 0 and ~cijklr ~ 0.
To determine the minimum cost of resource transportation, we developed the
method for solving the multi-index resource allocation problem, which consists of the
following steps:
1. Determination of minimal elements a~i ; b ~ ; ~c ; l ~
Step j k ~ l ; /r in appropriate sections.
Step 2. Construction of a matrix containing zero index elements ~cijklr ¼ 0.
Step 3. Selection of leading index element ~x0ijklr :
Step 4. Adjustment of values of constraints for leading index element ~x0ijklr :
Step 5. Exception of section, corresponding to maximal index.
Step 6. Determination of objective function components.
This method was developed for a five-index problem, taking into account the fuzzy
interval problem parameters, which determines the minimum cost of distribution of r-th
resource, that is transported from i-th production center to the j-th consumption center
through the k-th distribution center with the help of l-type transmission element. The
problem of linear programming will be fuzzy-solvable if there is a fuzzy optimal
schedule for the problem.
5 Example and the Case-Study for the Method of Indexed-

Elements
To study and compare the results of various methods for solving multi-index distri-
A five-index resource allocation problem
bution problems, let’s consider the example.
is given, with the following sets: A ¼ aj ¼ f7; 7; 16g – sets of constraints for the

first parameter; B ¼ bj ¼ f1; 12; 9; 8g – sets of constraints for the second
parameter; D ¼ fdk g ¼ f34g – sets of constraints for the third parameter;
V ¼ fvr g ¼ f3; 5; 22g – sets of constraints for the forth parameter;
Q ¼ fql g ¼ f6; 4g – sets of constraints for the fifth parameter.
Let’s solve this five-index problem of resource allocation using the fuzzy-estimated
potential method, the zero-transformation method and the developed fuzzy index-
element method [10].
Fuzzy formulated five-index potential method for solving distribution problem
can be stated as follows:
XI X
J X
K X
R X
L
~ X
F ~ ¼ ~cijkrl~xijkrl :
i¼1 j¼1 k¼1 r¼1 l¼1
Fuzzy schedule X e ¼ f~x g is a set of fuzzy variables of the problem. To make it

ijkrl
possible to compare the obtained solutions of the problem, we set the left fuzzy
coefficient, the right fuzzy coefficient, the lower modal value and the upper modal value
of the fuzzy interval, which determines the value of the initial parameter of the dis-
tribution problem as a fuzzy. So the value of the center of gravity of the constructed
trapezoid according to the fuzzy interval data, along the abscissa, was equal to the
corresponding value of the problem in a crisp non-fuzzy formulation. The result of
solving this problem with fuzzy parameters is to determine the optimal fuzzy resource
allocation
schedule or plan and the optimal value of the objective function is
~ e
F X ¼ ð31; 98; 132; 38Þ.
Five-Index Zero-Transformation Method for Solving Distribution Problem

In the example solved by the potential method, the initial values of the objective
function were used without any transformations. In the null transform method, the
coefficients cijkrl for each iteration can be counted as follows:

c1ijkrl ¼ cijkrl ai þ bj þ ck þ /r þ ll ;
and coefficients ai ; bj ; ck ; ur ; ll are the calculated parameters in the corresponding

section of the cost matrix, which are determined as follows:

ai ¼ min cijkrl ; i 2 f1; 2; . . .; I g;
jkrl

bj ¼ min cijkrl ai ; j 2 f1; 2; . . .; J g;
ikrl

ck ¼ min cijkrl ai bj ; k 2 f1; 2; . . .; K g;
ijrl

/r ¼ min cijkrl ai bj ck ; r 2 f1; 2; . . .; Rg;
ijkl

ll ¼ min cijkrl ai bj ck /r ; l 2 f1; 2; . . .; Lg:
ijkr
n o
After eight iterations optimal solution was found: X ¼ xijkrl = fx11111 ¼ 1;
x13111 ¼ 1; x12131 ¼ 5; x22131 ¼ 7; x33112 ¼ 1; x33122 ¼ 5; x33132 ¼ 2; x34132 ¼ 8g.
The optimal value of the objective function is F(X*) = 114.
Five-Index Fuzzy Indexed-Elements Method for Solving Distribution Problem
According to the proposed algorithm for solving the multi-index resource allocation
problem under the uncertainty of the initial parameters, the leading index elements will be
calculated
n ando a fuzzy plan for solving the resource allocation problem isgenerated as
e
X ¼ ~x
, with the optimal value of the objective function F X ~ e ¼ ð32; 100;
ijkrl
136; 35Þ. The solution of the distribution problem developed by the indexed-elements
method for fuzzy-interval parameters fully corresponds to the results of solutions in a
crisp non-fuzzy formulation, which is consistent with the basic principle of fuzzy-interval
mathematics - the coverage of all possible solutions.
6 Conclusion
To solve the problem of planning and allocation optimization of complex diversified

sets of activities that require the participation of a large number of performers and the
costs of limited resources, the application of classical methods is difficult. It is also
necessary to take into account the uncertainty of the initial data of the problem. The
proposed method of indexed-elements allows considering all the limitations of a
multidimensional distribution problem, formalized in a fuzzy multi-index form and
determine the optimal solution for the operational planning step of production and
distribution processes. Future research for the estimation of practical usability is sup-
posed to be held based on computational efficiency estimation for the large-scale
distribution problems, accuracy of decision and the number of iterations while solving
multi-index distribution problems.
Research according to the research projects N20-01-00197 and N19-07-00074.
References
1. Bucciarelli, E., Chen, S., Corchado, J. (eds.): Decision Economics: Complexity of Decisions
and Decisions for Complexity. Springer, Cham (2020)
2. Sung, W-P., Kao, J.C.M., Chen R. (eds.): Frontiers of Energy and Environmental
Engineering. CRC Press, Taylor & Francis Group (2013)
3. Sholarin, E.A., Awange, J.L.: Environmental Project Management: Principles, Methodology,
and Processes. Springer, Cham (2016)
4. Fletcher, R.: Practical Methods of Optimization. Wiley, Wiley (2000)
5. Kosenko, O., Sinyavskaya, E., Shestova, E., Kosenko, E. Chemes, O.: Method for solution
of the multi-index transportation problems with fuzzy parameters. In: XIX IEEE Interna-
tional Conference on Soft Computing and Measurements (SCM), pp. 179–182 (2016)
6. Kosenko, O., Shestova, E., Sinyavskaya, E., Kosenko, E., Nomerchuk, A., Bozhenyuk, A.:
Development of information support for the rational placement of intermediate distribution
centers of fuel and energy resources under conditions of partial uncertainty. In: XX IEEE
International Conference on Soft Computing and Measurements (SCM), p. 224 (2017)
7. Laplante, P.A.: Software Engineering for Image Processing Systems. CRC Press, Boca
Raton (2005)
8. Denardo, E.V.: Linear Programming and Generalizations: A Problem-based Introduction
with Spreadsheets. Springer, Boston (2011)
9. Mutanov, G.: Mathematical Methods and Models in Economic Planning, Management and
Budgeting. Springer, New York (2014)
10. Kosenko, O.V.: Improving the efficiency of methods for solving multi-objectives resource
allocation. In: Proceedings of the SFU. Engineering Sciences, vol. 4, no. 177, pp. 73–81
(2016)
11. Dubois, D., Prade, H.: Fuzzy Sets and Systems. Academic Press, New York (1980)
Solving Fuzzy Multi-objective Linear
Programming Problems Using
Multi-player Zero-Sum Game
Gizem Temelcan1(B) , Inci Albayrak2 , Hale Kocken2 ,

and Mustafa Sivri2
1
Computer Technologies Programme, Istanbul Aydin University,
Kucukcekmece, Istanbul, Turkey
[email protected]
2
Department of Mathematical Engineering, Yildiz Technical University,
Esenler, Istanbul, Turkey
{ibayrak,hgonce,msivri}@yildiz.edu.tr
Abstract. The difficulty in solving either multi-objective linear pro-

gramming (MOLP) problems or fuzzy multi-objective linear program-
ming (FMOLP) problems is the trade-off among objectives. To deal with
this difficulty, we proposed a new algorithm for solving FMOLP prob-
lems by using zero-sum game. First, FMOLP problem given is converted
to a crisp MOLP problem by using ranking function, and then a payoff
matrix is constructed to find the weights of each objective function of
the MOLP problem. After that, each weight is multiplied with the cor-
responding objective function, a single-objective LP problem is obtained
and thus, FMOLP problem is solved. The proposed algorithm is illus-
trated by numerical examples for the FMLOP problems having fuzzy
constraints or not.
Keywords: Fuzzy multi-objective linear programming problem ·

Ranking function · Zero-sum game
1 Introduction
Many real life problems such as economics, supply chain management, industrial
neural network design, automotive, medical treatments, scheduling, control etc.
are defined by multiple conflicting objectives. These multi-objective problems
are a type of linear or nonlinear programming problems which have more than
one objective function and subject to constraints. The difficulty in solving these
multi-objective problems is the trade-off among objectives. Since the real life
problems cannot be modeled determinedly, no one exactly knows the precise
values of parameters. Thus, it can be convenient to consider these parameters as
fuzzy, and also be more realistic than conventional one. Therefore, a weighted
approach is developed which assigns nonnegative weights to the objective func-
tions according to their importance, and a single objective function is obtained
1484 G. Temelcan et al.
by multiplying these weights with the corresponding objective functions. This

process, which is a conversion from a multi-objective to a single-objective, is
executed by using game theory which determines the weights. Game theory is a
study of strategic interactions among players, and in the multi-objective prob-
lems these players become objective functions.
Many approaches are proposed to solve MOLP problems in the literature.
Belenson and Kapur [21] developed a technique for solving MOLP problems by
applying a LP approach for the solution of two player zero-sum games with mixed
strategies. Tanaka and Asai [7] formulated fuzzy LP (FLP) problem with fuzzy
numbers, and solved the nonlinear programming problem obtained by using
max–min operator. Rao and Freiheit [15] developed an algorithm for solving
multi-objective problems. The algorithm is a modification to the game theory
in which two optimization steps are combined. Lai and Hwang [8] considered
a numeric investment problem which is a MOLP model to solve a LP problem
with imprecise objective and/or constraint coefficients that are taken as triangu-
lar fuzzy numbers. Athan and Papalambros [16] presented generalized weighted
criteria methods which keep the advantages of linear method without suffer-
ing from limitation, that is, in a nonconvex attainable region, forming a linearly
weighted objective function from a multi-objective optimization problem can fail
to capture Pareto optimal points. Das et al. [3] proposed a solution procedure
for the multi-objective transportation problems in which cost coefficients of the
objective functions and source and destination parameters are expressed as inter-
val values. Cadenas and Verdegay [22] considered a MOLP problem in which cost
coefficients are taken as fuzzy numbers, and the problem proposed in their study
assumes that the objective functions can be defined by different decision mak-
ers. Wu et al. [6] presented a new approximate algorithm for solving FMOLP
problems involving fuzzy parameters in any form of membership functions in
both objective functions and constraints. Arikan and Gungor [9] introduced a
two-phase procedure to solve FMOLP problems. In the first phase, a group of
vector optimization models is constructed by utilizing fuzzy parametric program-
ming while in the second phase each model is solved by FLP. Annamdas and
Rao [17] proposed an algorithm to solve multi-objective optimization problems
involving continuous, discrete and/or mixed design variables. In the algorithm,
a modified game theory approach, coupled with modified particle swarm opti-
mization, is used. Marler and Arora [18] provided insight into how the weighted
sum method works, and researched the basic importance of weights with respect
to preferences, Pareto optimal set, and objective function values. Suprajitno [4]
proposed MOLP problems having interval coefficients and variables and solved
the problems by using modified simplex method. Cheng et al. [2] presented a
method for solving FMOLP problems. They transformed fuzzy constraints and
objectives into crisp by using deviation degree measure, and aggregated a multi-
objective problem into a single-objective by using weighted max-min approach.
Porchelvi and Vasanthi [11] considered a FMOLP problem in which all variables
of the objective function and constraints are trapezoidal fuzzy numbers, applied
simplex method and obtained the optimal solution by using ranking functions.
Pandit [14] proposed a method to solve fully FMOLP problems involving fuzzy
parameters in triangular type. The problem is converted into a MOLP problem
Solving FMOLP Problems Using Multi-player Zero-Sum Game 1485
which is solved by using Pareto optimality technique. Pandian and Jayalak-

shmi [19] presented a moving optimal method to obtain efficient solutions to
the MOLP problems. The method is helpful for decision makers in order to
choose an efficient solution according to their level of satisfaction on the objec-
tive functions. Luhandjula and Rangoaga [1] proposed an approach to solve fuzzy
multi-objective programming problems having fuzzy-valued objective functions.
Their approach approximates fuzzy numbers by their respective nearest interval
approximations. Matejaš and Perić [20] presented an iterative method based on
game theory for solving MOLP problems with an arbitrary number of decision
makers. Porchelvi and Rukmani [10] focused on solving intuitionistic FMOLP
problem and intuitionistic multi-objective fuzzy linear fractional programming
problem by using weighting factor, and proved that the solutions obtained are
independent of weights. Ezzati et al. [13] proposed a new algorithm based on a
new lexicographic ordering on triangular fuzzy numbers for solving fully FLP
problem by transforming it to a MOLP problem. Yang et al. [12] considered fully
FMOLP problems where all of the parameters and variables are triangular fuzzy
numbers, adopted lexicographic order relation used in [13] for ranking triangu-
lar fuzzy numbers, and solved the problem by using min operator. Charkhgard
et al. [5] presented a LP-based algorithm for MOLP problems having linear
multi-objective function and affine constraints.
In this paper, we developed an algorithm for solving FMOLP problems. Two
types of problems are considered; one having fuzzy cost coefficients, crisp con-
straints and variables, and the other containing fuzzy cost coefficients and con-
straints, and crisp variables. In the algorithm, the FMOLP problem is converted
into a crisp MOLP problem by using ranking function. The MOLP problem
is separated into sub-LP problems for each objective function, and individual
optimal solutions are obtained. A payoff matrix is constructed by considering
the individual optimal solutions as players and individual optimal values as
strategies. The payoff matrix is solved by using zero-sum game for obtaining
the weights to convert the MOLP to a single-objective LP problem. The single-
objective LP problem is solved, and the optimal solution is substituted into the
objective functions of original FMOLP problem to determine the fuzzy optimal
solutions.
This paper is organized as follows: In Sect. 2, some basic definitions are given.
In Sect. 3, the proposed algorithm is presented. To show the application of the
proposed algorithm, some numerical examples are solved in Sect. 4, and conclu-
sion is presented in Sect. 5.
2 Preliminaries
Definition 1. ã = (a1 , a2 , a3 ) is a triangular fuzzy number if its membership
function is ⎧ r−a1
⎨ a2 −a1 , a1 ≤ r ≤ a2
µã (r) = ar−a 3
, a2 ≤ r ≤ a3 (1)
⎩ 3 −a2
0, otherwise.
Definition 2. A FMOLP problem is defined as:
M in (z̃1 (x), z̃2 (x), ..., z̃k (x)) (2a)
s.t.
n

ãij xj ≤ b̃i , i = 1, 2, ..., m (2b)
j=1
xj ≥ 0, j = 1, 2, ..., n (2c)
n

where z̃k (x) = c̃kj xj , (k = 1, 2, ..., k); c̃kj , ãij , b̃i are triangular fuzzy num-
j=1
bers, and xj , (j = 1, 2, ...n) are crisp variables.
Definition 3. A ranking function R is a mapping from the set of fuzzy numbers

to the set of real numbers, and in which a natural order exists. Let ã = (a1 , a2 , a3 )
a1 + 2a2 + a3
be a triangular fuzzy number. The ranking function of ã is R(ã) = .
4
Definition 4. For a MOLP problem, the point x∗ ∈ X is a Pareto optimal
solution iff there is no x0 ∈ X such that zp (x∗ ) ≤ zp (x0 ) for all p, and zq (x∗ ) <
zq (x0 ) for at least one q.
Definition 5. A feasible solution x∗ of FMOLP problem (2) is said to be a

Pareto optimal solution iff there is no another feasible solution x0 such that
R(z̃p (x∗ )) ≤ R(z̃p (x0 )) for all p, and R(z̃q (x∗ )) < R(z̃q (x0 )) for at least one q.
3 Proposed Algorithm
The proposed algorithm is given as follows:
Step 1. Load FMOLP problem given in (2) having fuzzy constraints or not.
Step 2. Convert the FMOLP problem (2) to a crisp MOLP problem by using
ranking function given in Definition 3.
Step 3. Separate the crisp MOLP problem into sub-LP problems for each objec-
tive function and find individual optimal solutions for each sub-LP problem.
Step 4. Construct a payoff matrix considering the individual optimal solutions
as players and individual optimal values as strategies.
Step 5. Solve the k-player zero-sum game obtained by the payoff matrix, and
determine weights for each objective function to obtain a single-objective LP
problem.
Step 6. Solve the single-objective LP problem and find an optimal solution.
Step 7. Substitute the optimal solution into the objective functions of the
FMOLP problem (2), respectively, and determine the fuzzy optimal solutions.
4 Numerical Examples
4.1 Solving FMOLP Problems Having Only Fuzzy Cost Coefficients
Example 1. Consider the following MOLP problem [1]:
M in (0.5, 1, 1.5)x1 + (1.8, 2, 3)x2 (3a)
M in (0, 2, 3)x1 + (0, 1, 3)x2 (3b)

s.t.
x1 + x2 ≥ 6 (3c)
2x1 + x2 ≥ 9. (3d)
Step 1–2. Use ranking function for each objective function and obtain a crisp
MOLP problem:
M in z1 = x1 + 2.2x2 (4a)
M in z2 = 1.75x1 + 1.25x2 (4b)
s.t.
x1 + x2 ≥ 6 (4c)
2x1 + x2 ≥ 9. (4d)
Step 3. Solve each sub-LP problem and find individual optimal solutions and
optimal values, respectively. These solutions are given in Table 1.
Table 1. Individual optimal solutions and corresponding optimal values.
Objective function Objective function value Optimal solution (x1 , x2 )

z1 6 (6, 0)
z2 9 (3, 3)
Step 4–5. Solve two-player zero-sum game to obtain the weights for converting
MOLP problem to the single-objective LP problem. The payoff matrix is given
in Table 2.
Table 2. The payoff matrix and weights of objective functions of Example 1.
Optimal solution (x1 , x2 ) z1 z2

(6, 0) 6 10.5
(3, 3) 9.6 9
Weights 0.29 0.71
Step 6–7. Solve the following single-objective LP problem constructed by using

the weights obtained from the payoff matrix presented in Table 2.
M in z = 0.29(x1 + 2.2x2 ) + 0.71(1.75x1 + 1.25x2 ) (5a)
s.t.
x1 + x2 ≥ 6 (5b)
2x1 + x2 ≥ 9 (5c)
The optimal solution of the LP problem (5) is (x∗1 , x∗2 )
= (3, 3) and the optimal
value is z ∗ = 9.18. The fuzzy optimal solutions of FMOLP problem (3) are
z̃1 = (6.9, 9, 13.5) and z̃2 = (0, 9, 18), respectively.
4.2 Solving FMOLP Problem Having Fuzzy Objective Functions

and Constraints
Example 2. Consider the following MOLP problem [2]:
M in z̃1 = (5, 7, 9)x1 + (4, 5, 6)x2 + (1, 2, 3)x3 (6a)
M in z̃2 = (4, 6, 7)x1 + (2, 4, 7)x2 + (4, 5, 6)x3 (6b)
M in z̃3 = (2, 4, 6)x1 + (3, 4, 6)x2 + (2, 3, 5)x3 (6c)
s.t.
(2, 5, 7)x1 + (2, 3, 4)x2 + (1, 2, 3)x3 =
˜ (8, 16, 24) (6d)
(1, 2, 3)x1 + (1, 2, 3)x2 + (1, 3, 4)x3 ≤
˜ (7, 17, 22) (6e)
(2, 3, 4)x1 + (1, 2, 4)x2 + (2, 3, 4)x3 ≥
˜ (12, 18, 25). (6f)
Step 1–2. Use ranking function in the objective functions and constraints and
obtain the following crisp MOLP problem.
M in z̃1 = 7x1 + 5x2 + 2x3 (7a)
M in z̃2 = 5.75x1 + 4.25x2 + 5x3 (7b)
M in z̃3 = 4x1 + 4.25x2 + 3.25x3 (7c)
s.t.
4.75x1 + 3x2 + 2x3 = 16 (7d)
2x1 + 2x2 + 2.75x3 ≤ 15.75 (7e)
3x1 + 2.25x2 + 3x3 ≥ 18.25 (7f)
Step 3. Solve each sub-LP problem and find individual optimal solutions and
optimal values, respectively. These solutions are given in the Table 3.
Step 4–5. Solve three-player zero-sum game to obtain the weights for convert-
ing MOLP problem to the single-objective LP problem. The payoff matrix is
presented in Table 4.
Step 6–7. Convert the MOLP problem to the single-objective LP problem by using
the weights given in Table 4 and obtain the optimal solution as (x∗1 , x∗2 , x∗3 ) =
(1.2896, 0.1913, 4.6503). Fuzzy optimal values z̃1∗ = (11.8635, 19.2843, 26.7051),
z̃2∗ = (24.1422, 31.7543, 38.2681) and z̃3∗ = (12.4537, 19.8745, 32.1369) are found,
respectively, by substituting the optimal solution into objective functions of (6).
Table 3. Individual optimal solutions and corresponding optimal values.
Objective function Objective function value Optimal solution (x1 , x2 , x3 )

z1 19.2841 (1.2896, 0.1913, 4.6503)
z2 31.5517 (1.3793, 0, 4.7241)
z3 21.0369 (1.2896, 0.1913, 4.6503)
Table 4. Payoff matrix of three player zero sum game and weights for Example 2.
Optimal solutions (x1 , x2 , x3 ) z1 z2 z3

(1.2896, 0.1913, 4.6503) 19.2841 31.4797 21.0369
(1.3793, 0, 4.7241) 18.2033 31.5517 19.408
(1.2896, 0.1913, 4.6503) 18.2033 31.4797 21.0369
Weights 1 0 0
5 Conclusion
In this paper, an algorithm based on game theory for solving FMOLP problems
is presented. In the algorithm, fuzzy cost coefficients of the objective functions
and fuzzy coefficients of the constraints, if there exists, are defuzzified by using
ranking function, and a crisp MOLP problem is obtained. The multi-objective
problem is separated into sub-LP problems, and individual optimal solutions are
found. A payoff matrix is constructed by using the individual optimal solutions,
and a multi-player zero-sum game is solved to obtain the weights corresponding
to each objective function. The weights, which are the degrees of participation of
each objective function, are determined to construct a single-objective function.
Thus, the single-objective LP problem is solved subject to composed constraints
and the optimal solution is found. By substituting the optimal solution into each
objective of FMOLP problem, fuzzy optimal values are determined. Generating
multiple solutions can be offered as an advantage of the proposed algorithm. As
a future direction, the solutions of FMOLP problems are aimed to be compared
by using various ranking functions from the literature.
References
1. Luhandjula, M.K., Rangoaga, M.J.: An approach for solving a fuzzy multi objective
programming problem. Eur. J. Oper. Res. 232(2), 249–255 (2014)
2. Cheng, H., Huang, W., Zhou, Q., Cai, J.: Solving fuzzy multi objective linear
programming problems using deviation degree measures and weighted max-min
method. Appl. Math. Model. 37(10–11), 6855–6869 (2013)
3. Das, S.K., Goswami, A., Alam, S.S.: Multi objective transportation problem with
interval cost, source and destination parameters. Eur. J. Oper. Res. 117(1), 100–
112 (1999)
4. Suprajitno, H.: Solving multi objective linear programming problem using interval
arithmetic. Appl. Math. Sci. 6(80), 3959–3968 (2012)
5. Charkhgard, H., Savelsbergh, M., Talebian, M.: A linear programming based algo-
rithm to solve a class of optimization problems with a multi-linear objective func-
tion and affine constraints. Comput. Oper. Res. 89, 17–30 (2018)
6. Wu, F., Lu, J., Zhang, G.: A new approximate algorithm for solving multiple objec-
tive linear programming problems with fuzzy parameters. Appl. Math. Comput.
174(1), 524–544 (2006)
7. Tanaka, H., Asai, K.: Fuzzy linear programming problems with fuzzy numbers.
Fuzzy Sets Syst. 13(1), 1–10 (1984)
8. Lai, Y.J., Hwang, C.L.: A new approach to some possibilistic linear programming
problems. Fuzzy Sets Syst. 49(2), 121–133 (1992)
9. Arikan, F., Gungor, Z.: A two-phase approach for multi-objective programming
problems with fuzzy coefficients. Inf. Sci. 177(23), 5191–5202 (2007)
10. Porchelvi, R.S., Rukmani, S.: Solution procedures for multi-objective intuitionistic
fuzzy linear and fractional linear programming problems using weighting factor.
Int. J. Fuzzy Math. Arch. 9(2), 179–187 (2015)
11. Porchelvi, R.S., Vasanthi, L.: Solving a multi objective fuzzy variable linear pro-
gramming problem using ranking functions. Int. J. Sci. Res. 2(1), 733–735 (2013)
12. Yang, X.P., Cao, B.Y., Lin, H.T.: Multi-objective fully fuzzy linear programming
problems with triangular fuzzy numbers. In: 2014 11th International Conference
on Fuzzy Systems & Knowledge Discovery, China, pp. 171–177. IEEE (2014)
13. Ezzati, R., Khorram, E., Enayati, R.: A new algorithm to solve fully fuzzy linear
programming problems using the MOLP problem. Appl. Math. Model. 39(12),
3183–3193 (2015)
14. Pandit, P.: Multi-objective linear programming problems involving fuzzy parame-
ters. Int. J. Soft Comput. Eng. 3(2), 99–101 (2013)
15. Rao, S.S., Freiheit, T.I.: A modified game theory approach to multiobjective opti-
mization. J. Mech. Des. 113(3), 286–291 (1991)
16. Athan, T.W., Papalambros, P.Y.: A note on weighted criteria methods for compro-
mise solutions in multi-objective optimization. Eng. Optim. 27(2), 155–176 (1996)
17. Annamdas, K.K., Rao, S.S.: Multi-objective optimization of engineering systems
using game theory and particle swarm optimization. Eng. Optim. 41(8), 737–752
(2009)
18. Marler, R.T., Arora, J.S.: The weighted sum method for multi-objective optimiza-
tion: new insights. Struct. Multidiscip. Optim. 41(6), 853–862 (2010)
19. Pandian, P., Jayalakshmi, M.: Determining efficient solutions to multiple objective
linear programming problems. Appl. Math. Sci. 7(26), 1275–1282 (2013)
20. Matejaš, J., Perić, T.: A new iterative method for solving multiobjective linear
programming problem. Appl. Math. Comput. 243, 746–754 (2014)
21. Belenson, S.M., Kapur, K.C.: An Algorithm for solving multicriterion linear pro-
gramming problems with examples. J. Oper. Res. Soc. 24(1), 65–77 (1973)
22. Cadenas, J.M., Verdegay, J.L.: Using ranking functions in multiobjective fuzzy
linear programming. Fuzzy Sets Syst. 111(1), 47–53 (2000)
A Bioinspired Algorithm for Improving
the Effectiveness of Knowledge Processing
Ilona Kursitys(&), Yury Kravchenko, Elmar Kuliev,

and Alexandr Natskevich
Southern Federal University, Rostov-on-Don, Russia

[email protected]
Abstract. The paper deals with an approach to improve the effectiveness of

knowledge processing in terms of large dimensions. The authors suggest the
model of classification of the information resources to be used as a prepro-
cessing stage for their further integration. The amount of information produced,
transferred and processed by people and various technical devices rapidly grow
every year, so the problem of improving the knowledge processing efficiency is
very important these days. Ontological structures as used in this work to rep-
resent knowledge of the information systems, because they allow us to consider
the semantics of the processed knowledge. The authors propose performing the
classification using two components of semantic similarity between the objects
of the ontologies. They are equivalent and horizontal semantic similarity com-
ponents. To solve the classification task, we use bioinspired algorithms since
they are proven to be effective in terms of solving the optimization problems of
large dimensions. One of the suggested bioinspired algorithms developed for
solving the mentioned tasks is the bacteria optimization algorithm. The paper
describes the algorithm and provides the results of its work. The experiments
show that the bacteria algorithm gives effective results with polynomial time
complexity.
Keywords: Classification Information resources Bacteria algorithm

Ontology Semantic similarity Bagging
1 Introduction
In conditions of mass distribution of the information technologies in every sphere of

our life, the problem of development of new methods and algorithms for information
and knowledge processing stays relevant today. The data science specialists estimate
the amount of information in the global net as approximate to 44 Zb (ZetaBytes). This
fact implies that providing the access for large amount of different distributed
knowledge is one of the most important issues today [1].
Knowledge is considered as structured and systematized information intended for
solving tasks and performing functions of the information systems [2]. The paper deals
with declarative knowledge, and the authors focus on its content aspect. Such an
approach requires to apply the principles of the “Semantic Web” concept [3], which is
based on the ideas of understanding the semantics of the processed information.
1492 I. Kursitys et al.
In terms of the paper, the sources of the knowledge in the information system are
represented by its information resources.
The paper aims at improving the effectiveness of knowledge processing by ana-
lyzing such subtasks as integration, classification, and estimation of semantic similarity
between the elements of knowledge. The originality of the paper is provided by new
models and architectures, and a modified bioinspired algorithm that can provide better
solutions than the analogs.
The rest of the paper is organized in the following way. The second section is
reviews the literature on the stated problems and analyzes the benefits and shortcom-
ings of the existing solutions. The third section presents the knowledge classification
problem statement and models and architecture of algorithm composition (bagging) for
its solving. In the fourth section, the authors demonstrate the developed modified
bioinspired algorithm to be used in the proposed architecture of bagging. The fifth
section deals with the experimental research conducted to estimate the effectiveness of
the modified algorithm. In the conclusion, the authors summarize the article and present
the ideas for further research.
2 Literature Review
The most popular model for processing the information of large scopes is the Apache
Hadoop project which is a shareware set of utilities, libraries, and software for the
development and executing the distributed programs. The idea includes storing and
processing the data with a lot of computational nodes united with the distributed file
system Hadoop Distributed File System (HDFS) [4].
The aids of parallel processing of information and knowledge are developed in the
MapReduce interface. The application is divided into several elementary tasks, then
calculated on separate cluster nodes, and finally united forming the results [5]. How-
ever, the results of each step are written on the disk which makes the writing and
reading time exceed the calculation time [6].
The new methods of processing large amount of data and knowledge can be
referred as a common technology called NoSQL (Not Only SQL). NoSQL are char-
acterized by working with non-structured or weakly structured information, lack of the
mechanisms of providing the integrity of the data, and implementation as distributed
systems with the ability to be horizontally scaled. The benefits include the possibility to
process large amount of non-structured information but in non-defined time [7].
The program solutions for improving the quality of knowledge processing include
different models and algorithms for solving tasks of classification, clustering, inte-
gration, etc. Such methods include Byes classifiers [8], neural networks methods [9],
nearest neighbor methods [10], support vector machine methods [11], and so on. The
most promising methods of solving the mentioned tasks include algorithms of random
search or heuristic algorithms, which involve the methods of evolutionary, genetic, and
bioinspired search.
The next section presents the problem statement, models, and architecture for
solving the mentioned tasks.
A Bioinspired Algorithm for Improving the Effectiveness of Knowledge Processing 1493
3 A Model and an Architecture for Improving

the Effectiveness of Knowledge Processing
The source of knowledge in the information system is represented as information

resources. Thus, the problem is reduced to increasing the effectiveness of the infor-
mation resources processing. Integration is considered as the main task in the process
of information resources processing. The authors propose solving the problem by
introducing the stage their preliminary classification [12].
Let I ¼ fi1 ; i2 ; . . .; in g be a set of information resources in the information system,
X ¼ fx1 ; x2 ; . . .; xm g be a set of classes. The task is to divide the set of information
resources into the classes of semantically similar ones on the basis of several com-
ponents of semantic similarity F ¼ ff 1 ; f 2 ; . . .; f k g.
In terms of the Semantic Web concept, it is reasonable to use ontological models to
represent the information resources [13–15]. The authors propose a modified onto-
logical model which can increase the accuracy of classification.
Let us denote the ontology as O ¼ \C; E; A; M; R [ , where C is the set of
concepts; E is the set of entities; A is the set of attributes; M is the set of markers which
include the ontology to a certain class and express the degree of inclusion; R is the set
of relations between concepts, entities, and attributes.
The set of markers M consists of the following elements
M ¼ fm1 ; m2 g; ð1Þ
where
8
>
> 0; if the ontology O is not considered yet;
<
1; if the O is included in class 1 ðequivalent ontologiesÞ;
m1 ¼ ð2Þ
>
> 2; if the O is included in class 2 ðsimilar Þ;
:
3; if the O is not included in any class ðdifferentÞ:
m2 ¼ ½0; 1: ð3Þ
Considering the ontological model presented above, we can reduce the stated
problem to classification of the ontologies. Let us describe the proposed classes [12]:
1. Equivalent ontologies, which elements provide the maximum intersection of
P P
attributes ni¼1 m k¼1 Ai \ Ak ¼ max.
1 2
2. Similar ontologies, which elements have partial intersection of the attribute sets
and similar values of horizontal component of semantic similarity.
3. Different ontologies, which elements have no semantic similarity at all.
In terms of the stated problem of ontologies classification, the paper proposes
maximization of two fitness functions:
• the total intersection of the concepts attributes:

Xn Xm 1
f1 A1 ; A2 ¼ i¼1
A \ A2j ! max
j¼1 i
ð4Þ
• the number of the matching concepts in the hierarchy of the ontologies:

f2 C 1 ; C 2 ! max ð5Þ
To increase the accuracy, the authors propose using the composition of the algo-
rithms – bagging – which allows us to obtain the solutions by several algorithms
simultaneously. The architecture is presented in Fig. 1.
Select the attributes
Select the number of algorithms n
Select the Select the Select the Select the

algorithm A1 algorithm A2 algorithm Ai algorithm An
A11 A21 Ai1 An1
Р11 Р21 Рi1 Рn1
Voting
Level of the equivalent

Aggregated solution component of semantic
Р1 similarity
Select the number of algorithms m
Select the Select the Select the Select the

algorithm A1 algorithm A2 algorithm Aj algorithm Am
A12 A22 Aj2 Am 2
Р12 Р22 Рj2 Рm 2
Level of the horizontal

Voting component of semantic
similarity
Final solution Р2
Fig. 1. Architecture of bagging the algorithms for solving the classification task.
The authors propose using bioinspired algorithms in the architecture because they
are effective for solving the tasks of high dimensions in the appropriate time.
4 Bacterial Algorithm for Solving the Classification Task
According to the two-level architecture of bagging, the authors propose the following
steps of algorithm work:
1. Solving the task according to the equivalent measure of similarity. The solutions are
encoded by their attributes sets. Then we select the initial set of attributes A1 and
other sets of attributes of the ontology O2. The initial population is created, and the
algorithm work starts. The most similar concepts are input in the list.
2. Solving the task according to the horizontal measure of similarity. The solutions are
represented as a two-dimensional array with the numbers of the equivalent concepts
of two ontologies considered as equivalent according to the firs similarity com-
ponent. The algorithm finds such sequence of the concepts which number can form
a hierarchy in terms of both ontologies.
The bacteria algorithm includes the following steps: chemotaxis, swarming,
reproduction, reduction and distribution. Chemotaxis in an adaptive process of a
bacteria moving to the distributed nutrition in the environment. The swarming process
is the intercellular exchange. The reproduction process implements the natural selec-
tion: only the fittest species can survive and share their genetic material with the next
generation. The worst species are deleted from the population. During the reduction
and distribution process some part of bacteria is randomly selected and deleted to
appear then in random search space positions for diversity [16, 17] (Fig. 2).
Begin
Enter the parameters of

the algorithm
Create the initial
population
Count the reduction and
distribution m=m+1
No
m<Ned End
Yes
Count the reproduction,
k=k+1
Reduction and No
disribution
k<Nre
Да
Count the chemotaxis,
j=j+1
No
j<Nch Reproduction
Yes
Bacteria count i=i+1
Yes
(i=S)
No
Calculate the fitness
function F(i,j,k,m)
Tumbling
Recalculate the fitness

function F(i,j+1,k,m)
Yes Runs count s=0
No
s<Ns
s=s+1
No
s=Ns F(i,j+1,k,m)<Flast
Yes
Flast=F(i,j+1,k,m)
Run
Fig. 2. Scheme of the developed bacteria algorithm.

5 Experimental Research
To evaluate the effectiveness of the proposed architecture and bioinspired algorithm,

the authors carried out a set of computational experiments using different numbers of
ontologies, concepts, and attributes. The developed bacteria algorithm was compared
with greedy algorithm and firefly algorithm developed in [12]. The Table 1 shows the
results of the experiments.
To calculate the accuracy of the developed algorithm, the authors use the criteria of
average value of semantic similarity between the ontology concepts classified as
equivalent or similar. Figure 3 presents the diagram of the algorithms quality.
As shown in Table 1 and Fig. 3, the bacteria algorithm is 7% better than the firefly
algorithm and 80% better than the greedy algorithm.
Table 1. Results of the algorithms work.

Number of Number of Greedy Firefly Bacteria
concepts relations algorithm algorithm algorithm
1000 300 0,00005 0,0007 0,00083
2000 500 0,000008 0,00009 0,00017
3000 700 0,0000004 0,0000026 0,0000074
4000 900 0,00000002 0,00000025 0,00000032
5000 1000 0,000000004 0,000000053 0,000000059
6000 1500 0,0000000005 0,000000048 0,00000007
7000 2000 0,00000000007 0,00000000958 0,00000002
Fig. 3. Diagram comparing the algorithm’s quality.

6 Conclusion
The paper aimed to develop the method to improve the effectiveness of knowledge
processing in terms of large dimensions. The main challenge in this field is huge
amount of information and knowledge generated, processed and stored in the dis-
tributed information systems.
The authors suggest an approach which includes classification of the information
resources as a pre-processing stage of their further processing. The model of classifi-
cation and the problem statement is presented in this paper. Ontological structures were
used in this work to represent knowledge of the information systems, because they
allow us to consider the semantics of the processed knowledge. The modified onto-
logical model developed for the classification task is presented in this paper. The
authors propose using two components of semantic similarity between the objects of
the ontologies as the classification criterion.
The authors developed a bioinspired algorithm based on the bacteria behavior. The
algorithm is modified by changing the angle of bacteria tumbling. The paper describes
the algorithm and provides the results of its work. The experiments show that the
bacteria algorithm gives effective results with polynomial time complexity.
Further research will be related to development of new bioinspired algorithms for
solving the classification tasks such as artificial immune system and cuckoo search
algorithms. Also, the authors plan to develop the models and methods for knowledge
clustering.
Acknowledgements. The reported study was funded by RFBR according to the research project
№ 18-29-22019.
References
1. Wassén, O.: Big Data facts – How much data is out there? https://www.nodegraph.se/big-
data-facts/. Accessed 20 Nov 2019
2. Kravchenko, Y.A., Kureichik, V.V.: Knowledge management based on multi-agent
simulation in informational systems. In: 8th IEEE International Conference on Application
of Information and Communication Technologies. AICT 2014, pp. 264–267 (2014)
3. Semantic Web. https://www.w3.org/standards/semanticweb/. Accessed 25 Sept 2019
4. The Apache Software Foundation Project. http://www.apache.org/foundation. Accessed 22
May 2018
5. Apache Hadoop project website. http://hadoop.apache.org. Accessed 22 May 2018
6. Dean, J., Ghemawat, S.: MapReduce: Simplified data processing on large clusters. Commun.
ACM 51, 137–150 (2004)
7. Sadalage, P., Fowler, M.: NoSQL Distilled: A Brief Guide to the Emerging World of
Polyglot Persistence. Addison-Wesley Professional, Boston (2012)
8. Tipping, M.E.: Sparse Bayesian learning and the relevance vector machine. J. Mach. Learn.
Res. 1, 211–244 (2001)
9. Bishop, C.: Pattern Recognition and Machine Learning. Springer, New York (2010)
10. Cunningham, P., Delany, S.J.: k-Nearest neighbour classifiers. Mult. Classif. Syst 34(11),
1–17 (2017)
11. Wu, Q., Zhou, D.X.: SVM soft margin classifiers: linear programming versus quadratic
programming. Neural Comput. 17(5), 1160–1187 (2005)
12. Kravchenko, Yu.A, Kursitys, I., Markov, V.V.: Bioinspired algorithm for acquiring new
knowledge based on information resource classification. In: International Russian Automa-
tion Conference (RusAutoCon), pp. 1–5 (2019)
13. Maedche, A., Zacharias, V.: Clustering ontology-based metadata in the semantic web. In:
Proceedings PKDD-2002 LNAI 2431, pp. 348–360 (2002)
14. Kureychik, V., Semenova, A.: Combined method for integration of heterogeneous ontology
models for big data processing and analysis. Adv. Intell. Syst. Comput. 573, 302–311 (2017)
15. Kravchenko, Y., Kursitys, I., Bova, V.: The development of genetic algorithm for semantic
similarity estimation in terms of knowledge management problems. Adv. Intell. Syst.
Comput. 573, 84–93 (2017)
16. Das, S., Biswas, A., Dasgupta, S., Abraham, A.: Bacterial foraging optimization algorithm:
theoretical foundations, analysis, and applications. In: Abraham, A., Hassanien, A.-E.,
Siarry, P., Engelbrecht, A. (eds.) Foundations of Computational Intelligence, vol. 3, pp. 23–
55 (2009)
17. Amghar, Y.T., Fizazi, H.: A hybrid bacterial foraging optimization algorithm and a radial
basic function network for image classification. J. Inf. Process. Syst. 13(2), 215–235 (2017)
Spatial Coherence Radius of Plane Wave
Propagating in Anisotropic Maritime
Atmospheric Turbulence
Bing Guan1, Fei Chen2, and Jaeho Choi2(&)

1
School of Computer Science and Engineering,
Huizhou University, Huizhou, China
2
Department of ECE, CAIIT, Jeonbuk National University, Jeonju, Korea
[email protected]
Abstract. Maritime atmospheric turbulence has been theoretical investigated,

since it can affect optical waves more obviously than it propagate in terrestrial
turbulence. In this study, we obtain the new analytical expression for the wave
structure function of plane wave propagating in the anisotropic maritime
atmospheric turbulence. Based on this expression, the formula of the spatial
coherence radius of plane wave is derived. Meanwhile, the turbulence strength,
the anisotropic parameter, the power law exponent, and the turbulence scales are
considered in the numerical results.
Keywords: Spatial coherence radius Wave structure function Optical

wave Maritime atmospheric turbulence
1 Introduction
Scientists and researchers use classic Kolmogorov and non-Kolmogorov power spec-
trum for doing experimental and theoretical research about laser beam propagation
through terrestrial and high-altitude environments for a long time [1–3]. Lately, we
have found that more research attention is drawn to the theoretical survey of the
property of laser beam propagation through maritime atmosphere turbulence. A new
atmospheric spectral model is developed to analyze irradiance fluctuation of optical
wave propagation through the marine environment [4]. Later, Toselli et al. use this
maritime power spectral model for evaluating long term beam spread and beam wander
of Gaussian beam wave [5]; Cui developed and analyzed temporal power spectra of
irradiance scintillation for infrared optical waves [6]. Cheng et al. calculated the
scintillation and aperture averaging for Gaussian beams in non-Kolmogorov or gen-
eralized maritime atmospheric turbulence [7]. Furthermore, long-exposure modulation
transfer function and temporal frequency spread are discussed when the optical waves
propagate through anisotropic non-Kolmogorov atmospheric turbulence [8, 9].
In this work, wave structure function (WSF) of the plane wave that propagates in
the anisotropic maritime turbulence is investigated. Then, the new expression of the
spatial coherence radius (SCR) of the plane wave propagating in maritime atmosphere
environment is derived. Then, we analyze the influence of the power law exponent, the
1500 B. Guan et al.
anisotropic factor, the turbulence strength and the inner scale of the turbulence on the
WSF and SCR.
2 Anisotropic Non-kolmogorov Power Spectrum
An analytical model for the refractive index power spectrum which is referred to as the
modified atmospheric spectrum is defined as follows [1]:
j2 2 j j 7
)[j þ j20 ] 6 ½1 þ a1 þ a2 ð Þ6
11
un ðj; aÞ ¼ 0:033Cn2 exp( ð1Þ
jl
2 jl jl
where j is the scalar spatial frequency in rad/m, Cn2 is the refractive index structure
parameter in m3 , a1 = 1.802 and a2 = −0:254, jl ¼ 3:3
2
l0 , and j0 ¼ L0 , where l0 and L0
2p
are inner scale and outer scale parameters, respectively. In [4] later, the coefficients
have been changed to a1 = −0.061 and a2 = 2.836 as an effort to make it a maritime
atmospheric spectrum for a weak optical turbulence with small inner scale values.
The spectrum model in Eq. (1) is extended to the one for the non-Kolmogorov
turbulence and it is defined as follows [7]:
~ 2 expðj Þ½j2 þ j2 2 ½1 þ a1 j þ a2 ð j Þ32

2 a a
un ðj; aÞ ¼ AðaÞC n ð2Þ
jH
2 0
jH jH
where jH ¼ CðaÞ=l0 , CðaÞ is defined as follows:
3a 3a 4a 4a 12 3a 4 a a51

CðaÞ ¼ fpAðaÞ½Cð Þð Þ þ a1 Cð Þð Þ þ a2 Cð Þð Þg
2 3 2 3 4 2
ð3Þ
In our study, anisotropic parameter f is introduced, so that the resulting power

spectrum for the anisotropic maritime atmospheric turbulence can be defined as
follows [9]:
~ 2 f2a expðj Þ½j2 þ j j j

2 a a
un ðj; a; fÞ ¼ AðaÞC n ~20 2 ½1 þ a1 þ a2 ð Þ32 ð4Þ
~H
j 2 ~H
j ~H
j
~H ¼ jH =f and j
where j ~0 ¼ j0 =f.
3 WSF and SCR of Plane Waves
The WSF for the isotropic non-Kolmogorov turbulence can be expressed as

follows [1]:
Spatial Coherence Radius of Plane Wave 1501
1
Z
Dpl ðq; aÞ ¼ 8p2 k2 L jun ðj; aÞ½1 J0 ðjqÞdj: ð5Þ
0
where q the scalar separation distance between two is points in 2-D plane; k ¼ 2p=k is
the optical wave number; L is the path length; J0 ðÞ is the zero order Bessel function of
the first kind.
In this work, Eq. (4) is used as the expression for the anisotropic non-Kolmogorov
power spectrum in our derivation of new wave structure function expression for plane
wave. Then, Eq. (5) can be rewritten as follows:
1
Z
Dpl ðq; a; fÞ ¼ 8p2 k2 L jun ðj; a; fÞ½1 J0 ðjqÞdj: ð6Þ
0
By substituting Eq. (4) into Eq. (6), and expanding J0 ðÞ as a series representation,
one can obtain the new WSF expression for the plane waves as follows:
a
~ 2 f2a fq2 Cð2 2Þ~
j0 4a
Dpl ðq; a; fÞ ¼ 4p2 k2 LAðaÞC n 4 CðaÞ
~2a
þj H Cð1 a=2Þ½1 J1 ðaÞ
2
q
2 Cða2 2:5ÞCð2:5Þb1 j
~4a
þ 4 Cða2Þ
0
~2a
þ a1 j H Cð1:5 a=2Þ½1 J2 ðaÞ ð7Þ
q ~
2 Cð0:75a 3:5ÞCð3:5 0:25aÞb j 4a
þ 4 Cða2Þ
2 0
~2a
þ a2 j H Cð2:5 0:75aÞ½1 J3 ðaÞg
~2H =4Þ
J1 ðaÞ ¼ 1 F1 ð1 a=2; 1; q2 j
~2H =4Þ
J2 ðaÞ ¼ 1 F1 ð1:5 a=2; 1; q2 j ð8Þ
J3 ðaÞ ¼ 1 F1 ð2:5 ~2H =4Þ
0:75a; 1; q2 j
Then, the new expression for SCR of the plane wave is derived. The SCR
derivations begin with the new WSF derived in the above paragraph. That WSF needs
to be approximated and simplified to be used effectively in the numerical computations
when one performs computer simulations. Consider the WSF in Eq. (7). That equation
involves the confluent hypergeometric functions 1F1() [10]. By using the approximate
format of the confluent hypergeometric function, then the WSF can be approximately
expressed as follows:
~ n2 f2a fq Cða=2 2Þ~

j4a
2
Dpl ðq; a; fÞ ¼ 4p2 k 2 LAðaÞC 0
~ 2a
þj H Cð1 a=2Þ
4 Cða=2Þ
22a qa2 j
~ a2
H q2 Cða=2 2:5ÞCð2:5Þb1 j ~4a 23a qa3 j~a3
H
½1 þ 0
~ 2a
þ a1 j H Cð1:5 a=2Þ½1
Cða=2Þ 4 Cða=2Þ Cða=2 1=2Þ
q2 Cð0:75a 3:5ÞCð3:5 0:25aÞb2 j ~4a ~1:5a5
251:5a q1:5a5 jH
þ 0
~ 2a
þ a2 j H Cð2:5 0:75aÞ½1 g
4 Cða=2Þ Cð3a=4 3=2Þ
ð9Þ
From the WSF of the optical waves, the SCR q0 is defined by the 1/e point of the
complex degree of coherence [1] and Dðq0 ; LÞ ¼ 2. Based on the approximation
1502 B. Guan et al.
expression of the WSF for the plane waves defined in Eq. (9), the new SCR expression
of plane wave for the case of L0 ¼ 1 is derived and it is defined as follows:
2a2 Cða2Þ 2a

q0 qpl f ½~
j Cð1 a=2Þ þ a1 Cð1:5 a=2Þ
Cð1 a2Þ H
ð10Þ
2 1
þ a2 Cð2:5 0:75aÞ ga2 :
2 2 ~
4p k LAðaÞCn f
2 2a
4 Numerical Analysis and Results
There are two sets of new expressions derived and they are set for evaluation with
respect to various characterizing parameters. The WSF is defined in Eq. (7) and the
SCR is defined in Eq. (10). We have made some general assumptions in the numerical
simulations: the optical waves propagate with the generalized structure parameter Cn2 ¼
1:4 1014 m3a ; the scalar separation distance is q ¼ 3 cm the wavelength k ¼
1:65 106 m; the inner scale of the eddy size is 1 mm, the outer scale of the eddy size
is 10 m; the optical path lengths vary from 100 m to 8 km; and the power law exponent
a varies from 3 to 4; and the case of l0 \q \L0 is used for the SCR simulations.
The first sets of simulations are performed using the new expression of wave
structure function defined in Eq. (7). The focus of the evaluation is to analyze the
behaviors of the WSF in terms of various characterization parameters. Those include
the power law exponent a, the turbulence strength r2R , and the anisotropic factor f.
Figure 1 shows the behavior of the WSF with respect to the increasing power law
exponent a, when the anisotropic factor f ¼ 1, which actually makes the turbulence
isotropic. The WSF increases when a varies from 3 to 3.3, and then decrease gently
afterwards. One can see the smooth bumps that the WSF gets to their maximum when
a 3:3. On the other hand, Fig. 2 shows the behaviors of the WSF withrespect to the
turbulence strength r2R . The WSFs monotonically increase as the turbulence strength
increases. And the anisotropic parameter influences the value of WSF obviously
decreases when the anisotropy increases. The second set of simulations is performed
using the new expressions of spatial coherence radiuses defined in Eq. (10). Similar to
Fig. 1. Wave structure function of plane wave for increasing power exponent.
Fig. 2. Wave structure function of plane wave for increasing turbulence strength.
the simulations using the WSFs, the focus of the evaluation is also to analyze the
behaviors of SCRs in terms of various characterization parameters. Those include the
turbulence strength r2R , the anisotropic factor f, and the inner scale of the atmosphere
turbulence. Figure 3 shows the behavior of the SCR with respect to the increasing
turbulence strength r2R and the increasing anisotropic factor f. On the contrary to the
WSFs cases, the SCRs decrease as the turbulence strength gets stronger. Also, it is
important to note that the SCRs are also significantly influenced by the anisotropic
factor, and the curve become smooth when the anisotropy inceases. Figure 4 shows the
behavior of the SCR with respect to the inner scale of the eddy size l0. In these
simulations, we have set that the anisotropic factor f ¼ 1. The simulation results show
that the SRCs for the plane waves are also influenced by the inner scale of the eddy
size; the curves increase monotonically as l0 increases. Also, we can observe that the
anisotropic parameter influences the SCRs similar with Fig. 3.
Fig. 3. Spatial coherence radius of plane wave for various turbulence strength and anisotropy.
1504 B. Guan et al.
Fig. 4. Spatial coherence radius of plane wave for various turbulence scale and anisotropy.
5 Conclusion
In this work, a new expression has been presented for the wave structure function and
also for the spatial coherence radius of the free space plane wave propagating in a
horizontal path, which is disturbed by an anisotropic maritime turbulence. Those newly
derived analytic expressions of WSF and SCR are evaluated and their behaviors are
observed by varying four major charactering parameters, which are the power law
exponent a, the turbulence strength r2R , the anisotropic factor f, and the inner scale of
the eddies l0, respectively. Those four parameters individually or in their combinations
have extensive impacts on the magnitudes of the WSF and SCR. In conjunction to
current work presented in this work, we are also looking at the temporal frequency
spreads of optical waves propagating in a free space with a moderate to strong tur-
bulence. The propagation paths can be slant and horizontal. Moreover, the study on the
effects of short-exposure MTFs on an imaging system is also underway.
Acknowledgment. This work has been supported partly by BK21 and Jeonbuk National
University of Korea.
References
1. Andrews, L., Phillips, R.: Laser Beam Propagation through Random Media, 2nd edn. SPIE,
Bellingham (2005)
2. Toselli, I., Agrawal, B., Restaino, S.: Light propagation through anisotropic turbulence.
J. Opt. Soc. Am. A 28(3), 483–488 (2011)
3. Kotiang, S., Choi, J.: Wave structure function and long-exposure MTF for laser beam
propagation through non-Kolmogorov turbulence. Opt. Laser Technol. 74, 87–92 (2015)
4. Grayshan, K., Vetelino, F., Young, C.: A marine atmospheric spectrum for laser
propagation. Waves Random Complex Media 18(1), 173–184 (2018)
5. Toselli, I., Agrawal, B., Restaino, S.: Gaussian beam propagation in maritime atmospheric
turbulence: long term beam spread and beam wander analysis. Free-Space Laser
Communications X. International Society for Optics and Photonics 2010, vol. 7814,
p. 78140R (2010)
6. Cui, L.: Temporal power spectra of irradiance scintillation for infrared optical waves’
propagation through marine atmospheric turbulence. JOSA A31(9), 2030–2037 (2014)
7. Cheng, M., Guo, L., Zhang, Y.: Scintillation and aperture averaging for Gaussian beams
through non-Kolmogorov maritime atmospheric turbulence channels. Opt. Express 23(25),
32606–32621 (2015)
8. Cui, L., Fei, W., Xue, B.: Theoretical investigations of infrared optical wave modulation
transfer function models in anisotropic marine turbulence. JOSA A35(9), 1496–1503 (2018)
9. Guan, B., Choi, J.: Temporal frequency spread of optical waves propagating in anisotropic
maritime atmospheric turbulence. Appl. Opt. 58(11), 2913–2919 (2019)
10. Andrews, L.: Special Functions of Mathematics for Engineers, 2nd edn. McGraw-Hill, New
York (1997)
A Comparative Performance Analysis
of Consensus Models Based
on a Minimum Cost Metric
Á. Labella(B) , R. M. Rodrı́guez , and L. Martı́nez
University of Jaén, Campus Las Lagunillas s/n, 23071 Jaén, Spain

[email protected]
Abstract. Consensus reaching processes (CRPs) are key in the reso-

lution of many group decision making problems, since they guarantee
a solution in which most of decision makers agree. For this reason, a
great number of consensus models have been proposed in the specialized
literature, being it difficult to make a proper comparison among all of
them and determine which model best fits a given problem. Recently, a
new cost metric based on comprehensive minimum cost consensus models
has been proposed, which compares the solution obtained by a consensus
model with the agreed one with minimum cost. Therefore, this contribu-
tion aims to carry out a reliable comparative analysis of the performance
of several consensus models by using the cost metric in order to identify
their strong points and drawbacks.
Keywords: Consensus reaching process · Minimum cost consensus ·

Cost metric
1 Introduction
Nowadays, it is pretty common to find decision making problems in which several
decision makers with different points of view provide their opinions with the aim
of obtaining a common solution, this kind of problems are called, group decision
making (GDM) problems. Traditionally, a GDM problem is solved by a selec-
tion process [2] ignoring the agreement among decision makers. Consequently,
decision makers could feel that their preferences were not considered to obtain
the best alternative as solution of the problem [1]. To overcome this situation,
a consensus reaching process (CRP) is included in the resolution scheme of a
GDM problem. A CRP is an iterative process in which decision makers discuss
and modify their initial preferences to achieve an agreement and an acceptable
solution for all of them. There are different consensus models in the literature
[3,4,8] and they can be classified according to the taxonomy introduced in [8]:
This work is partially supported by the Spanish Ministry of Economy and Compet-
itiveness through the Spanish National PGC2018-099402-B-I00 and the Postdoctoral
fellow Ramón y Cajal (RYC-2017-21978).
A Comparative Analysis of Consensus Models Based on a MCC Metric 1507
– Consensus with feedback and without feedback

– Consensus measures based on distances to the collective opinions and based
on distances among decision makers.
Usually, authors propose new consensus models assuring that their models
provide better results in comparison with others, but they do not use any metric
to support that statement. The first attempt to analyse and compare the perfor-
mance of different CRPs was introduced in AFRYCA [5,8]. AFRYCA is a tool
that implements the most widely used CRPs and considers several measures
to evaluate CRPs such as, number of rounds necessary to reach the consen-
sus, the number of changes carried out across the CRP and several consistency
measures. Nevertheless, these criteria are quite simple and cannot objectively
measure the performance of different consensus models. To cope this challenge,
recently Labella et al. [4] have proposed a metric based on the cost of modifying
decision makers’ preferences. It uses a comprehensive minimum cost model to
compute the optimal consensus solution, the one that involves the least possible
changes in decision makers’ preferences, and compare it with the one obtained by
the analyzed consensus model. Therefore, the aim of this contribution is to ana-
lyze the performance of several consensus models by means of the cost metric and
point out their advantages and disadvantages. To do so, we will use AFRYCA
that implements some well-known CRPs and the cost metric mentioned.
The structure of this contribution is as follows: Sect. 2 revises some basic
concepts about GDM, CRP and the cost metric. Section 3 shows the comparative
analysis among the consensus models by using AFRYCA. Finally, Sect. 4 points
out the conclusions and future works.
2 Preliminaries
This section revises some preliminary concepts about GDM, CRP and a cost
metric used to evaluate the performance of the consensus models.
2.1 Group Decision Making and Consensus Reaching Process

Usually, there are several decision makers with different points of view and
knowledge involved in decision making problems with the aim of obtaining a
common solution, this leads to GDM [6]. In a GDM problem a set of decision
makers E = {e1 , . . . , em } provide their preferences over a set of alternatives
X = {x1 , . . . , xn } by means of a fuzzy preference relation (FPR), because it is
simple and easy to build.
Definition 1. [7] A FPR P k , associated to a decision maker ek on a set of
alternatives X, is a fuzzy set on X× X, characterized by the membership function
μP k : X × X → [0, 1]. ⎛ k ⎞
p11 . . . pk1n
⎜ .. ⎟ ,
P k = ⎝ ... . . . .⎠
pkn1 . . . pknn
1508 Á. Labella et al.
where each assessment pkij represents the degree of preference of the alternative
xi over xj according to decision ek . The FPR is usually assumed to be reciprocal,
i.e., pkij + pkji = 1, ∀i, j = 1, 2, . . . , n, k = 1, 2, . . . , m.
Traditionally, a GDM problem is solved by a selection process which is

divided into two phases: (i) aggregation in which decision makers’ preferences
are fused to get a collective opinion, and (ii) exploitation in which the best alter-
native is selected according to the collective opinion. Nevertheless, this process
does not take into account the agreement among decision makers which implies
that some of them could feel that their preferences were not considered. In real
world problems, disagreement among decision makers is inevitable, for this rea-
son, it is important to remove the disagreement among decision makers and
obtain solutions more appreciated and accepted by the group. To do so, a CRP
is included in the solving scheme of GDM problems. A CRP is a dynamic and
iterative process in which decision makers discuss and modify their preferences
to reach a collective opinion which is accepted by the whole group. A general
scheme of a CRP is explained below:
– Consensus measurement: the consensus level is calculated by consensus mea-

sures based on aggregation operators and distance measures.
– Consensus control: the consensus level obtained is compared with the consen-
sus threshold μ ∈ [0, 1] set a priori. If the consensus level is greater than the
threshold, a selection process is applied, otherwise, another round is necessary.
– Consensus progress: the decision makers who are far away from the collec-
tive opinion are identified and some advises are generated to change their
preferences and increase the consensus level in the next round.
The computation of the consensus level is very important. According to Palo-

mares et al. in [8] the consensus measures can be classified in two types:
– Consensus measure based on the distance of each decision maker to the col-
lective opinion.
– Consensus measure based on the distances among decision makers.
2.2 Metric Based on Minimum Cost
There are many consensus models [3,4,8], however, when a new consensus model
is proposed, authors affirm that it is better than previous ones, but there is not
any measure that analyzes its performance in comparison with others to know
which one is better. A first attempt was introduced in AFRYCA [5,8], which
considers several criteria to evaluate the performance such as, the number of
rounds necessary to achieve the consensus, the number of changes carried out
across the consensus process, and some consistency measures. Nevertheless, these
measures are quite simple and cannot compare consensus models in a proper
way. Recently, Labella et al. [4] have introduced a metric based on the cost of
changing decision makers’ preferences to evaluate CRPs. This metric uses an
optimal solution obtained of a comprehensive minimum cost consensus model

defined as well.
Two MCC models were introduced depending on the consensus measure used.
Let P k be the FPR provided by a decision maker, and it is adjusted to
k
P = (pkij )n×n , k = 1, . . . , m, to reach a solution accepted by the group, and
P = (pij )n×n is the adjusted FPR of the collective FPR.
– Consensus measure based on the distance between decision makers’ prefer-
ences and the collective opinion.
(M − 1)

m n−1 n
min ck |pkij − pkij |
k=1
⎧ i=1 j=i+1
⎪ m
⎪
⎪ pij = wk pkij
⎪
⎨ k=1
k
s.t. |pij − pij | ≤ ε, k = 1, . . . , m, i = 1, . . . , n − 1, j = i + 1, . . . , n
⎪
⎪
m n−1 n
⎪
⎪ 2
wk |pkij − pij | ≤ γ.
⎩ n(n−1)
k=1 i=1 j=i+1
where (c1 , . . . , cm ) are the cost of moving each decision maker’ opinion 1 unit,
ε ∈ [0, 1] is the maximum acceptable distance of each decision maker to the
collective opinion, γ = 1 − μ, being m μ the consensus threshold and wi ∈ [0, 1]
is the decision maker’ weight, i=0 = 1.
– Consensus measure based on the distance among decision makers.
(M − 2)

m n−1 n
min ck |pkij − pkij |
k=1 i=1 j=i+1
⎧
⎪ m
⎪
⎪ pij = wk pkij
⎪
⎨ k=1
k
s.t. |pij − pij | ≤ ε, k = 1, . . . , m, i = 1, . . . , n − 1, j = i + 1, . . . , n
⎪
⎪
⎪ 2 n−1
⎪
n m−1 m
wk +wl k l
⎩ n(n−1) m−1 |pij − pij | ≤ γ.
i=1 j=i+1 k=1 l=k+1
The metric computes the difference in cost between the MCC model (M−1
or M−2) and the solution obtained with the consensus models to evaluate.
Let P = (P 1 , . . . , P m ) be the initial decision makers’ preferences, and P =
1 m
(P , . . . , P ) the optimal adjusted FPRs of the MCC model (M−1) or (M−2),
k
where P k and P are the initial and adjusted FPRs of the decision maker ek ,
k
k = 1, 2, . . . , m, respectively. The distance between P k and P is calculated as
k
2
n−1
n
k
d P k, P = pij − pkij , k = 1, . . . , m. (1)
n(n − 1) i=1 j=i+1
and the relative distance between P and P , is defined as follows:

m k

D P, P = d P k, P . (2)
k=1
In similar manner, let P = (P1 , P2 , . . . , Pm ) be the agreed solution obtained
in a consensus model, the distance between P k and Pk is calculated as
2
n−1
n
k
d P k , Pk = pij − pkij , k = 1, . . . , m. (3)
n(n − 1) i=1 j=i+1
and
m
D P, P = d P k , Pk . (4)
k=1
Finally, the cost metric is defined as follows:

⎧
⎪
D P,P
⎨ 1 − ( ) , if D P, P ≤ D P, P
D (P,P )
φ P, P = D(P,P ) (5)
⎪
⎩ − 1, if D P, > D P, P .
P
)
D (P,P

φ P, P ∈ [−1, 1], when φ P, P < 0 means there is excessive changes

in decision makers’ preferences, on the contrary φ P, P > 0 means decision
makers’ preferences can be closer each other. If P = P , the CRP solution is the
best.
3 Comparative Analysis
This section shows a comparative analysis among several CRPs by using the cost
metric proposed by Labella et al. [4] together other metrics. Three CRPs have
been selected, Palomares et al. [9], Quesada et al. [10] and Rodrı́guez et al. [11].
Such a selection is based on two reasons:
– The similarity between them in terms of consensus computation, since all of
them use the consensus measure based on the distance among decision makers
that facilitates to carry out a fair comparative analysis
– The consensus models present strengths and drawbacks in different aspects
in their performances, which implies a deep analysis that cannot be carried
out with simple metrics.
3.1 Description of the Problem and Parameters

The previous CRPs are used to solve a GDM problem composed by 30 decision
makers, E = {e1 , . . . , e30 }, who provide their preferences by means of FPRs,
over 3 alternatives X = {x1 , x2 , x3 }. The consensus threshold is predefined as
μ = 0.85 and the maximum numbers of rounds to reach consensus is h = 15.
The parameter ε related to the comprehensive MCC models is ε = 0.1. Finally,
each consensus model uses different parameters to simulate the CRP. For sake
of space, the definition of the parameters is not introduced here, but their values
are represented in Table 1 (see [9,10] and [11] for further details).
Table 1. CRPs’ parameters value.
Palomares et al. [9] Quesada et al. [10] Rodrı́guez et al. [11]

ε∗ = 0.05 ε∗ = 0.05 δ = 0.7
α = 0.2 α = 0.2 a=3
β = 0.6 β = 0.6 b = 10
increment = 0.1 increment = 0.1 β ∗ = 1.8
hstart = 4
η = 0.5
g = 0.5
Remark 1. Note that we keep the notation of the CRPs’ parameters used by the
authors. For this reason, ε∗ and β ∗ have been represented with the symbol ∗ in
Table 1 in order to avoid confusion with other parameters.
Remark 2. For the CRPs simulations, we have considered that all the experts
accept always the suggestions provided by the consensus models to guarantee a
fair comparison among them.
3.2 Comparison
A comparative analysis about the CRPs performance is carried out in this

section. Firstly, the GDM problem is solved with the different consensus models
by using AFRYCA. The results obtained are represented in Table 2.
Table 2. CRPs’ parameters value.
Palomares et al. [9] Quesada et al. [10] Rodrı́guez et al. [11]

Number of rounds 6 9 12
Number of changes 143 198 345
Cost metric 0.41 0.46 0.27
Consensus level 0.87 0.86 0.86
Ranking x3 x1 x2 x3 x1 x2 x3 x1 x2
Table 2 provides relevant information about the CRPs. Regarding the rank-
ing of the alternatives, all of them choose alternative x3 as the best solution of the
problem. However, Palomares et al.’s model needs just 6 rounds to achieve the pre-
defined consensus, being the fastest. The consensus level achieved for each model
is quite similar, although the Palomares et al.’s model reaches such a level with
a less number of changes in the decision makers’ preferences (143). According to
the previous issues, it would seem logical to think that Palomares et al. is the best
analyzed consensus model because it reaches a higher level of consensus with less
changes. On the contrary, Rodrı́guez et al.’s model might seem the worst, since it
needs more rounds and changes than any other model. Nevertheless, the cost met-
ric shows that Rodrı́guez et al.’s model is much closer to the optimal solution than
the other consensus models. This issue is graphically represented in Fig. 1, which
shows the evolution of the decision makers’ preferences for each CRP and the visu-
alization of the solution with minimum cost.
Fig. 1. Decision makers’ preferences evolution.
Figure 1 shows the optimal cost solution for the problem on the top of the
figure, noted as MCC solution, which is the agreed solution provided by model
(M−2). If we compare such a solution with the one provided by the Palomares et
al.’s consensus model, the decision makers’ preferences are dispersed. This means
that it achieves the desired consensus because it compensates the decision makers
whose opinions are furthest away with others whose opinions are quite similar.
Obviously, it implies that the consensus model needs less changes to achieve the
consensus, but the solution is not homogeneous. On the other hand, the models
of Quesada et al. and Rodrı́guez et al. present more homogeneous solutions in
which decision makers’ preferences are closer each other, but Rodrı́guez et al.
model presents clearly the closest one to the optimal solution.
Therefore, although at first sight it might seem that the Palomares et al.’s
model is the best one by number of rounds and changes required, it is also
necessary to take into account other metrics, such as the cost metric, that allows
to measure other characteristics of the performance of consensus models and
carried out a more reliable and deeper analysis.
4 Conclusions
The increasing key role of CRPs in GDM problems has resulted in many con-
sensus approaches with their own peculiarities to support such processes. For
this reason, it is difficult to make proper comparisons among consensus models
to choose the best suitable for a given problem. Several authors use simple met-
rics to justify that their models are better than other but, often they are not
enough to guarantee such a statement in an objective way. Recently, a new cost
metric based on comprehensive MCC models has been proposed, which allows
to measure properly the performance of consensus models.
This contribution has carried out a comparative performance analysis of sev-
eral consensus models by using the cost metric in order to show the importance of
using proper metrics to compare consensus models and identify their advantages
and drawbacks. To do so, AFRYCA that implements some consensus models, has
been used to solve the GDM problem. As future research, we intend to optimize
the MCC models in order to evaluate the performance of CRPs on large-scale
group decision making problems with hundreds of experts.
References
1. Butler, C.T.L., Rothstein, A.: On Conflict and Consensus: A Handbook on Formal
Consensus Decision Making. Food Not Bombs Publishing, Takoma Park (2006)
2. Herrera, F., Herrera-Viedma, E., Verdegay, J.: A sequential selection process in
group decision making with linguistic assessments. Inf. Sci. 85(4), 223–239 (1995)
3. Herrera-Viedma, E., Cabrerizo, F., Kacprzyk, J., Pedrycz, W.: A review of soft
consensus models in a fuzzy environment. Inf. Fusion 17, 4–13 (2014)
4. Labella, Á., Hongbin, L., Rodrı́guez, R., Martı́nez, L.: A cost consensus metric for
consensus reaching processes based on a comprehensive minimum cost model. Eur.
J. Oper. Res. 281, 316–331 (2020)
5. Labella, Á., Estrella, F.J., Martı́nez, L.: Afryca 2.0: an improved analysis frame-
work for consensus reaching processes. Prog. Artif. Intell. 1–14 (2017)
6. Lu, J., Zhang, G., Ruan, D., Wu, F.: Multi-Objective Group Decision Making.
Imperial College Press, London (2006)
7. Orlovsky, S.: Decision-making with a fuzzy preference relation. Fuzzy Sets Syst.
1(3), 155–167 (1978)
8. Palomares, I., Estrella, F., Martı́nez, L., Herrera, F.: Consensus under a fuzzy
context: taxonomy, analysis framework AFRYCA and experimental case of study.
Inf. Fusion 20, 252–271 (2014)
9. Palomares, I., Martı́nez, L., Herrera, F.: A consensus model to detect and manage
noncooperative behaviors in large-scale group decision making. IEEE Trans. Fuzzy
Syst. 22(3), 516–530 (2014)
10. Quesada, F., Palomares, I., Martı́nez, L.: Managing experts behavior in large-
scale consensus reaching processes with uninorm aggregation operators. Appl. Soft
Comput. 35, 873–887 (2015)
11. Rodrı́guez, R.M., Labella, Á., De Tré, G., Martı́nez, L.: A large scale consensus
reaching process managing group hesitation. Knowl.-Based Syst. 159, 86–97 (2018)
Pedestrian Detection Based on Improved Mask
R-CNN Algorithm
Wenjun Yu1, Sumi Kim2, Fei Chen1, and Jaeho Choi1(&)

1
Department of ECE, Jeonbuk National University, CAIIT,
Jeonju, Republic of Korea
[email protected]
2
Seoyeong University, Gwangju, Republic of Korea
Abstract. As the modern society evolves around digital media, object recog-
nition becomes one of the important areas for computer vision. Pedestrian
detection particularly draws much attention because it is closely related to
everyday life. Recently, pedestrian detection has achieved great success in
intelligent monitoring, intelligent driving, and environmental protection.
Although, there are several pedestrian detection algorithms based on deep
learning, the pedestrian detection is still a huge challenge. Background occlu-
sion, pedestrians’ various changing postures and objects’ occlusion give sig-
nificant impact on the recognition results; it still brings up much attention. In this
paper, to reduce the influence of external factors, we propose a new method
based on Mask R-CNN algorithm. The proposed system was evaluated on the
Daimler pedestrian dataset for training and on the public Caltech and INRIA
pedestrian datasets for testing. The experimental results have showed that the
proposed algorithm achieves better detection accuracy than the conventional
ones.
Keywords: Pedestrian detection Mask R-CNN Feature concatenation

Hard negative mining
1 Introduction
In recent years, pedestrian detection has developed rapidly and been applied more and
more widely. Many applications rely on pedestrian detection, such as smart vehicles,
object monitoring, robotics, and video surveillance. Because of its important role,
pedestrian detection has attracted full attention in the computer vision field. But the
diversity of pedestrian body pose, object occlusions, clothing, lighting change and
complicated backgrounds in the video sequence or image, those of them can have a
huge impact on recognition, so the pedestrian detection still has a challenging in
computer vision. Traditional methods have achieved good results, a high accuracy.
Usually they use single or multiple underlying features for feature extraction and then
use sliding window and linear support vector machine (SVM) [1] to classification. We
can see the advantages of this approach, but the results from using some of the new
pedestrian datasets are not ideal and still needs to be improved.
1516 W. Yu et al.
Recently, the deep convolutional neural network (CNN) is more and more popular
in computer vision. Object detections use deep learning methods always have the better
results than others. Those are Fast-R-CNN [2] and Faster-R-CNN [3]; they achieve the
state-of-the-art performance in object detection. In fact, these methods have two main
parts: The first one is object region proposal. It is used to scan the entire image to
generate a set of candidate image regions or bounding boxes at different locations and
scales that could possibly contain the target objects; the second one is image classi-
fication. This determines whether the proposed areas are truly objects or not.
Due to the more in-depth exploration of deep learning for pedestrian detection,
Mask-R-CNN [4] was created. Mask R-CNN is an advanced version of Faster R-CNN.
For all of proposal box in Faster R-CNN, the semantic segmentation part uses the FCN
and the segmentation task is performed simultaneously with the positioning and
classification task. In Faster R-CNN, the RoI Pooling is replaced by ROI Align,
because the RoI Pooling is not pixel-to-pixel alignment, it may not have much effect on
bounding box, but it does have a significant influence on the accuracy of the mask. The
accuracy of the mask increased significantly from 10% to 50% with RoI Align.
In this paper, we propose a new pedestrian detection based on Mask R-CNN [4].
The proposed model is trained on public Caltech and INRIA pedestrian datasets. The
effectiveness of this method is proved by experiments, and the results show that our
method has a significant effect on pedestrian detection. In the rest of this paper, Sect. 2
briefly introduces the related work of pedestrian detection and the latest progress of
deep learning methods. Section 3 introduce the proposed pedestrian detection method.
Section 4 gives the experimental process and results of the model. Finally, the whole
paper is summarized in Sect. 5.
2 Related Work
In fact, Mask R-CNN, developed from Faster R-CNN (2016), added a Mask Prediction
Branch on Faster R-CNN, improved ROI Pooling and newly presented ROI Align.
Faster R-CNN followed from Fast R-CNN (2015), and Fast R-CNN from R-CNN
(2014). In 2014, deep learning is in its third year. In R-CNN proposed by Ross
Girshick in CVPR in 2014, convolutional neural network was used for target detection.
The contribution of R-CNN has two main aspects: 1) the convolutional neural network
is used to extract features; 2) the bounding box regression is used to modify the target
bounding box. But it also has the problems: 1) time-consuming selective search, which
costs 2 s for an image and 2) time-consuming serial CNN forward propagation. For
each RoI, an AlexNet lift feature is required, which costs about 47 s for all RoI lift
features. 3) the three modules are trained separately, and during the training, the storage
space is greatly consumed.
As to circumvent this problem, Ross proposed Fast R-CNN and improved it in
2015. The contribution of Fast R-CNN can be divided into two aspects: 1) instead of
the serial feature extraction method of R-CNN, a neural network is directly used to
extract features from the full picture. 2) in addition to selective search, other parts can
be trained together. However, Fast R-CNN also has disadvantages, which is reflected in
the time consuming selective search still exists.
Pedestrian Detection Based on Improved Mask R-CNN Algorithm 1517
Faster R-CNN, published in 2016, made the following innovations: Firstly, the
shared convolutional layer is used to extract features for the full picture, and then the
feature maps obtained are sent to RPN, which generates the box to be detected (the
location of the specified RoI) and makes the first correction to the bounding box of RoI.
Then, there comes the architecture of Fast R-CNN. RoI Pooling Layer selects the
features corresponding to each RoI on the feature map according to the output of RPN
and sets the dimension as a fixed value. Finally, the boxes are classified using the FC
Layer and a second fix to the target bounding box is made. In particular, the Faster
R-CNN truly implemented end-to-end training. In this work, we propose a better
algorithm for pedestrian detection contributed to the Mask-R-CNN.
3 Proposed Method
3.1 Pedestrian Detection Method

The proposed pedestrian detection model is based on the deep learning architecture of
Mask R-CNN, which is one of the superb deep learning methods proposed for object
detection. The Mask R-CNN is first shown by Kaming in 2017 and widely used for
target detection. It became a popular model that realizes a stage-to-stage detection
process of target detection using series of R-CNNs.
In Fig. 1, the basic Faster R-CNN is shown in black, and the its modification in red.
In Mask R-CNN, ROI Align is replaced with ROI Pooling layer and a parallel FCN
layer (mask) is added. So the Mask R-CNN has several features: 1) add branch network
on the basis of border recognition for semantic Mask recognition; 2) the training is
simple, compared with the faster, it only adds a small overhead and can run up to 5 fps;
3) it can be also used for attitude estimation; 4) without the help of trick, the effect of
each task is better than all the single-model entries. Figure 2 and 3 include illustrations
for Faster R-CNN and Mask R-CNN networks.
Fig. 1. The gray part is the original faster R-CNN, and the red part is the modification on the
faster network.
1518 W. Yu et al.
Fig. 2. Faster R-CNN network.
Fig. 3. Mask R-CNN network.
In our work, we propose a new model based on Faster RCNN for pedestrian
detection and bring in feature concatenation and hard negative mining to make a higher
recall and accuracy.
3.2 Mask
For the newly added mask branches, the output dimension for each ROI is the K * m * m
where m * m is the size of the mask and the K represents K categories, so a total of KKK
binary masks are generated for each ROI. This is the concept of class-specific mask
mentioned. After the prediction mask is obtained, the value of sigmoid function (so-called
per-pixel sigmoid) is calculated for each pixel value of the mask, and the result is taken as
one of the inputs of Lmask (cross entropy loss function). Sigmoid function is also called a
logistic function or squashing function that compresses the outputs of ‘neurons’ in the
multi-layer neural network. It influences the speed of the neural network training.
The ROI that is sent into the mask branch is actually the ROI that is only sent into the
positive sample. Only positive sample ROI will be used to calculate Lmask, and the
definition of positive sample is the same as that of target detection, which is defined as
positive sample with IOU greater than 0.5.
Mask-rcnn does not seek multiple classification loss of pixels on all channel
channels at once, but only seeks a sigmoid dichotomization for each pixel in the
category corresponding to each ROI. For the first ROI, we only dichotomized each
pixel in the corresponding category K = 3, that is, on the mask on the channel with
channel 3. For the second ROI, we only dichotomized each pixel in the category
corresponding to K = 8, that is, on the mask on the channel with channel 8 (see Fig. 4).
The defined Lmask allows the network to generate a mask for each category and only
calculate the loss on its own category (that is, channel) without competing with other
classes. After the loss is calculated for each pixel of the characteristic graph of each
fixed category, the average of all pixels is then calculated, which is why Lmask is
called average binary cross-entropy loss in the article.
Fig. 4. On the category corresponding to ROI, a sigmoid dichotomy is obtained for each pixel.
3.3 ROI Align

On the other hand, the ROI Align technology does not use quantization operation to
obtain feature maps of 7 7 fixed sized; it avoids quantization errors.
From Fig. 5, we can see that the blue dotted box represents the feature map
obtained after convolution. The black solid wireframe represents ROI feature. If the
final output is 2 2, the interpolation filter is used to estimate the pixel values cor-
responding to the locations of these blue points (virtual coordinate points, also known
as the grid points of bilinear interpolation) (jointly determined by the four real pixel
values around, such as the red pointer in the figure), and the corresponding output is
finally obtained. These blue points are ordinary random sampling points in the 2 2
cell, and the author points out that the number and location of these sampling points
does not have a significant effect on performance, which you can obtain in other ways.
Then carry out max pooling or average pooling operation in each orange area to obtain
the final output result of 2 2. Quantization operation is not used in the whole
process, and no error is introduced, that is, pixels in the original image and pixels in the
1520 W. Yu et al.
feature map are perfectly aligned without deviation, which will not only improve the
detection accuracy, but also facilitate the instance segmentation.
Fig. 5. ROI Align.
4 Experiments and Results
For the experiments, the Daimler dataset is used for training, and then the Caltech and
INRIA pedestrian datasets are used for testing.
4.1 Datasets
a) Caltech [5]: The Caltech ant Detection Benchmark is a Pedestrian Detection data
set that contains about 10 h of 640 480 resolution, 30 Hz video captured by an
on-board camera while driving around the city. A total of about 250,000 frames,
350,000 boxes, and 2,300 different pedestrians were tagged.
b) Daimler [6]: The Daimler pedestrian detection dataset Contains a number of
Daimler data sets; Daimler pedestrian segmentation benchmark data set; Daimler
pedestrian prediction benchmark data set; Daimler stereo pedestrian detection
baseline data set; Daimler one-way pedestrian detection data set;, etc.
c) INRIA [7]: The INRIA Person data set is used to detect upright pedestrians in
images and videos. The data set contains data in two formats, the first is the original
image and the corresponding upright pedestrian annotation. The second is the
orthostatic positive class normalized to 64 128 pixels and the negative class
image of the corresponding image.
4.2 Implementation Details

In training the Faster R-CNN, RoI and IoU in the truth box are observed. If it is bigger
than 0.5, it is a positive feature, otherwise it is negative. The mask loss function Lmask
is defined only on positive samples of RoI. The mask target is a mask for the inter-
section between RoI and its corresponding truth boxes.
In testing the Faster R-CNN, the mask branch can predict KK masks for each RoI,
but we only use the KK mask, where KK is the category predicted by the classification
branch. Then, one can adjust the size of the m mm m floating point number mask
output to the size of RoI; the threshold of 0.5 is used.
4.3 Pedestrian Detector Comparison

Through experiments, we compared the best performing methods on the datasets with
our own work. The results are shown in Fig. 6, we can see our miss rate is 24.27, is
better than others. When we used the INRIA datasets, we can see the miss rate is 10.31,
also better than others. So, the results show that our method is better than other
detection algorithms.
Fig. 6. Performance comparison of some good pedestrian detection methods using to (Caltech
pedestrian dataset on the left and bottom) INRIA pedestrian dataset.
1522 W. Yu et al.
5 Conclusions
In this work, we proposed a new method for pedestrian detection based on MASK
R-CNN. We learned about MASK R-CNN structure and training process and made
some improvements. The conclusion is drawn by comparing with several better
methods on the datasets: experimental results show that our work fits better than any
other method on these datasets. In the future we like to investigate an efficient way to
improve the detection accuracy when objects are in dynamic changes in scales.
Acknowledgment. This work has been supported partly by BK21 and Jeonbuk National
University of Korea.
References
1. Tang, Y.: Deep Learning using Linear Support Vector Machines. Computer Science, arXiv:
1306.0239 (2013)
2. Girshick, R.: Fast r-CNN. In: Proceedings of International Conference on Computer Vision,
pp. 1440–1448 (2015)
3. Shaoqing, K., Girshick, R.: Faster R-CNN: towards real-time object detection with region
proposal networks. In: Advances in Neural Information Processing Systems, pp. 91–99
(2015)
4. https://github.com/facebookresearch/Detectron
5. http://www.vision.caltech.edu/Image_Datasets/CaltechPedestrians/
6. http://www.gavrila.net/Datasets/Daimler_Pedestrian_Benchmark_D/daimler_pedestrian_
benchmark_d.html
7. http://pascal.inrialpes.fr/data/human/
A Revised Weighted Fuzzy C-Means
and Center of Gravity Algorithm
for Probabilistic Demand and Customer
Positions
Engin Bayturk1(&), Sakir Esnaf2, and Tarik Kucukdeniz2

1
American College of the Middle East,
Dasman 220, 15453 Kuwait City, Kuwait
[email protected]
2
Istanbul Cerrahpasa University, Avcilar, Istanbul, Turkey
{sesnaf,tkdeniz}@istanbul.edu.tr
Abstract. This study proposes four probabilistic fuzzy c-means algorithms

which include a probabilistic fuzzy c-means algorithm (Probabilistic FCM), a
probabilistic revised weighted fuzzy c-means algorithm (Probabilistic RWFCM)
and hybrid algorithms that combine these algorithms with the center of gravity
methods for the un-capacitated planar multi-facility location problem when
customer positions and customer demands are probabilistic with predetermined
service level. The performance of proposed algorithms was tested with 13 data
sets and compared with each other. Experimental results indicate that Proba-
bilistic RWFCM-COG algorithm performs better than other compared algo-
rithms in terms of cost minimization.
Keywords: Probabilistic demand and position Probabilistic fuzzy c-means

Center of gravity Multi-facility location problem
1 Introduction
One of the strategic decisions of a company is facility location selection. The correct
facility location selection will lead to less deviation in the cost estimates taken into
account in the investment decisions made over the years. In addition to the costs, order
delivery from suppliers to facilities and from facilities to customers at the desired time
will be directly affected by these decisions. It is one of the most frequently studied
subjects on the problem of facility location due to the important role outlined above.
The facility location problem was first proposed by [3]. This study has gained much
interest in the academic field, and then Kuenne and Soland [11], proposed the branch
and bound algorithm method for this particular problem. Murtagh and Niwattisyawong
[17] have developed a model on the problem of determination of facility location.
Megiddo and Supowit [14] conducted a study that proved that the facility location
problems were NP-Hard problems. Murray and Church’s [16] simulated annealing,
Ohlemuller’s [18] tabu search, Hansen et al.’s [9] p-median studies can be listed as
some of these studies. Hansen et al. [9] discuss the deterministic facility location
1524 E. Bayturk et al.
problems. In addition to these methods, solutions were suggested with a fuzzy

approach to facility location problems. Yang et al. [25] used the chance constraint
programming method for the capacitated facility location where the demand of cus-
tomers are fuzzy. Zhou and Liu [27] discussed three different facility location models
with fuzzy demand. Wen and Iwamura [21] developed the fuzzy cost model with the
Hurwicz criterion. Peidroa et al. [19] developed fuzzy supply chain models. Gao [7]
developed the uncertain shortest path problem. A facility location model with random
fuzzy demand proposed by Wen and Kang [22], while Wang and Watada [20] sug-
gested a model in a fuzzy random environment. Gao [8] developed a solution for the
uncertain single facility location problem.
Wen et al. [23] developed a mathematical model that determines the best possible
plant location with the mixed model in which the simplex method and genetic algo-
rithm are used together. Customer demands are determined by expert judgment. Wen
et al. [24] proposed a mathematical model for determining the capacity of a facility in a
problem where customer demands are not certain. The proposed model has a mixed
model with genetic algorithm, the simplex algorithm and Monte Carlo simulation.
Uncertain customer demands are determined by expert judgment. Markovic et al. [13]
discussed the problem of choosing a stochastic facility location where customer
demands are independent. Diabat et al. [4] developed a mathematical model in which
demand is not known, and lead time was determined jointly by plant location and stock
decisions.
In literature, some researchers also proposed a solution for the facility location
problem where the customers’ locations and demands are probabilistic. Zhou and Liu
[26] solved the facility location problem, where customer demands have random
parameters. Altinel et al. [1] developed solution for the capacitated facility location
problem where positions and demands of the customer are probabilistic. Mousavi et al.
[15] suggested a solution for the same problem using chance-constrained programming
and genetic algorithms.
Facility locations’ problems are NP-hard problems [14]. Continuous location-
allocation problem or Multisource Weber problem focuses on to determine locations of
c facilities to serve n customers with minimum total cost. In this particular problem,
customer positions are fixed. Revised Weighted Fuzzy C-Means (RWFCM) is first
discussed by [5]. This algorithm was shown as an improved version of Fuzzy C-Means
(FCM), which was first developed by [2].
Esnaf and Kucukdeniz [6] have improved the results by combining FCM and the
Center of Gravity (COG) methods. In this method; first, the FCM and the set of demand
points are divided into c clusters, after that the COG is calculated for each. These
centers are selected as new cluster centers. Kucukdeniz et al. [12] applied convex
programming and FCM methods for the capacitated facility problem. In their method;
cluster centers are determined using the FCM in the first step within the facility
capacity limit, then the centers of the new clusters are determined by convex pro-
gramming, which is treated as the problem of a single facility location selection.
The RWFCM method discussed by Esnaf and Kucukdeniz [5] used a different metric in
which the weight values are considered constant. It is calculated by multiplying the
Euclidean distance between the two points on the metric plane in the objective function
A Revised Weighted Fuzzy C-Means and Center of Gravity Algorithm 1525
by the demand quantity of the demand point. This method used the assumption that the
demands are fixed in the problem.
The literature review encouraged us to develop a solution with a successful fuzzy
clustering approach tried before to this problem.
In this paper, the probabilistic version of the RWFCM and a hybrid version of the
RWFCM and Center of Gravity methods are proposed, when customer demands and
positions are assumed to be distributed normally. Comparison of the proposed algo-
rithms are made in terms of total transportation cost. The probabilistic version of the
RWFCM and its hybrids for the multi-facility location problems have not been studied
in the literature yet according to the best of our knowledge.
The purpose of this paper to propose the probabilistic version of the RWFCM
algorithm combined with the center of gravity method and how to reach facility
locations that provide minimum total transportation costs when customer demands and
positions are probabilistic.
The rest of the paper is organized as follows; in the second section, problem
definition is provided. Third section, contains explanation of the proposed algorithms.
The results of the proposed algorithms using thirteen data sets are discussed in the
fourth section. The last section concludes with discussions about paper and future
studies.
It should be noted that this hybrid not yet used in this way.
It should be noted that this new hybrid approach presented for the first time and not
yet used before in this manner.
Multi-facility location problems (MFLP) focus on determining locations of c facilities

to serve n demand points with the minimum total cost when customer locations are
fixed. But if the locations and demands of the customers are variable, the MFLP can be
reformulated as follows:
Xk X
min ai 2Vj
wi d ai ; c j ð1Þ
c1 c2 ...::ck j¼1
Where, ai is the average location of customer i in a plane, i = 1, 2… n, wi is the

probabilistic demand of customer i, wi > 0, i = 1, 2… n, cj is the location of facility j,

Vj is the customer cluster that assigned to jth facility, and d ai ; cj is the probabilistic
distance (Euclidean) between the customer i and facility j.
3 Proposed Algorithms
In this section, Probabilistic FCM, Probabilistic RWFCM and hybridized versions with
the Center of Gravity method are discussed. Since demand and customer positions are
not fixed, the probabilistic approach is employed to find better facility locations to
reduce total costs with predetermined service level.
3.1 Probabilistic RWFCM

Esnaf and Kucukdeniz [5] developed RWFCM. In RWFCM, weights which are
symbolized as wi in equation applied as customer demand when customer demand is
fixed. Customer positions which are coordinates of demand points X, Y are fixed data
points in the set ai, and Euclidean distance is used. Unlike original RWFCM, demand
and coordinates in the Probabilistic RWFCM are not fixed values. Considering demand
values fit the Normal distribution and the weights for each customer are determined
according to the predetermined service level with the Eqs. (5) and (6). Distances
between demand points and facilities are calculated as in Eq. (7) and all remaining
calculations are the same as in the RWFCM.
Objective function is given as follows:
Xn Xk p
Jp ðU; cÞ ¼ i¼1 j¼1
w i u ij d
a i ; c j ð2Þ
The optimization problem is to minimize (2) under the constraint:

Xc
j¼1
uij ¼ 1; 8i ð3Þ
Membership values, uij, are calculated with the following equation:
1
uij ¼ p1 ð4Þ
Pk ai ci
2
j¼1 ai cj
Probabilistic RWFCM algorithm is given as follows in (5)–(8):

Step 1: At the beginning of the algorithm, the number of cluster, k, cluster centers of
subsets, fc1 ; c2 ; . . .; ck g, coefficient of fuzziness, p, and stopping criterion, e [ 0, are
determined.
Step 2: Probabilistic demand of the customers ðwi Þ are calculated.

wi wi
P z ¼T ð5Þ
rwi
i þ zrwi
wi ¼ w ð6Þ

Step 3: Probabilistic Euclidean distance of the customers d ai ; cj are calculated as [10]:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2ffi
a1i c1j
a2i c2j
d ai ; cj ¼ þ ð7Þ
s1i s2i
Step 4: Calculate membership values uij using Eq. (4).

Step 5: Cluster centers are calculated as follows:
Pn p
i¼1 wi uij
ai
c j ¼ Pn p ð8Þ
i¼1 wi uij
Step 6: If the difference of the calculation between consecutive cluster centers is bigger
than e, the algorithm continues starting from step 2, else, the algorithm is stopped.
Where P is the probability, wi is the average demand and rwi is the standard
deviation of demand for the customer i, T is the threshold value for customer service
level, ai is the average coordinates of customer i, and si is the variance of the coor-
dinates of customer i. The Probabilistic RWFCM becomes to the Probabilistic FCM in
the case that all weights are equal. Probabilistic FCM is the first proposed and used as a
benchmark algorithm in this study.
3.2 Probabilistic Center of Gravity

Y
Improving location of the cluster centers ðX; Þ by Probabilistic Center of Gravity
method is as follows in (9)–(12):
P
i wi
a1i
X¼ P ð9Þ
i wi
P
i wi a2i
Y ¼ P ð10Þ
i wi
P
wi a1i =d ai ; cj
¼ P
X i
ð11Þ
i wi =d ai ; c j
P

a2i =d ai ; cj
i wi
Y¼ P ð12Þ
i wi =d ai ; c j
3.3 Probabilistic RWFCM-COG

This method first matches plants and customers using Probabilistic RWFCM, after
which optimum plant locations within each created cluster are determined by a special
center of gravity approach.
Step 1: Probabilistic customer demands (wi) are calculated.

Step 2: Probabilistic Euclidean distance between customers and facilities d ai ; cj .
are calculated.
Step 3: Using Probabilistic RWFCM, customers are divided in to k number of
clusters (ck).
Step 3.1: Using Probabilistic RWFCM, k number of cluster center is determined.
Step 3.2: Customers are assigned to closest facilities. Sub clusters are created by
assigned customers.
Step 4: For every sub clusters, cluster centers are calculated by Probabilistic Center
of Gravity.
Step 5: Membership degrees (uij) are calculated.
Step 6: If the difference of the calculation between consecutive cluster centers is
bigger than e, algorithm continues starting from step 2, else, algorithm is stopped.
When algorithm stops, final versions of the cluster centers are determined and
customers are divided into sub clusters accordingly. Thereby, n number of customers
are divided in to k number of subsets with minimum total cost. If Probabilistic FCM
applied instead of Probabilistic RWFCM, the method called as Probabilistic FCM with
the Center of Gravity (Probabilistic FCM-COG).
4 Experimental Study
The performance of the Probabilistic RWFCM-COG algorithm against FCM-based

benchmark algorithms is conducted in this section.
Probabilistic FCM, Probabilistic FCM-COG and Probabilistic RWFCM are
methods employed for comparison. Equation (1) is used for calculating total trans-
portation costs and threshold value, T, is taken as 0.9 in all methods.
Probabilistic RWFCM-COG is compared with other proposed methods using 13
different datasets and 21 different trials. Data were obtained from a company in the
telecommunication sector in Turkey. Table 1 includes the details of the total trans-
portation costs of all methods for each trial.
Table 1. Total transportation costs of proposed algorithms.

Dataset Demand Clusters Probabilistic Probabilistic Probabilistic Probabilistic
Points FCM FCM-COG RWFCM RWFCM-COG
1 100 4 31242.66 31567.75 30383.11 30107.38
2 400 10 27421.6 27043.75 27140.48 27037.82
2 400 20 18108.51 17909.93 17716.87 17688.36
3 500 30 28118.66 27423.99 26657.18 26517
4 800 10 99675.04 98921.52 99166 98654.5
5 1200 100 30285.64 29128.2 28593.57 28298.05
5 1200 120 26637.7 25743.55 25917.26 25500.58
6 2000 120 169868 166142.2 165692.1 164029.2
(continued)
Dataset Demand Clusters Probabilistic Probabilistic Probabilistic Probabilistic
Points FCM FCM-COG RWFCM RWFCM-COG
6 2000 150 150134 146435.9 146749.1 146124.5
7 2500 200 36004.31 35123.45 34905.97 34611.42
8 2800 100 58190.81 57684.55 57464.25 57239.2
9 5000 50 117988.4 117812.9 117900.7 117798.6
9 5000 100 82070.04 81472.6 81587.69 81373.4
10 10000 50 272124.3 271137.8 268174.6 267673
10 10000 100 188013.1 187222.2 186299.8 185868.1
10 10000 200 131749.9 130582.1 130894.4 130449.8
11 20208 800 553645.9 527133.3 522225.2 519814
12 50111 200 2884595 2850871 2838437 2825892
12 50111 400 2014010 1986422 1986339 1964001
13 100000 100 9304092 9329676 9360277 9322158
13 100000 1000 1021425 1006551 959266 945384.6
5 Conclusion
In this paper, Probabilistic FCM, Probabilistic RWFCM, and hybrid algorithms with
the center of gravity methods are proposed and compared with each other using 13
different datasets and 21 different trials. As a result, in terms of total cost minimization,
the Probabilistic RWFCM-COG algorithm performs better than other proposed
algorithms.
In future studies, different probability distributions of demand and customer posi-
tions can be applied in various data sets. Instead of probabilistic customer location and
demand, stochastic versions of proposed algorithms will be used to solve the problem
and can be compared in terms of CPU time and total transportation cost. As a fine-
tuning method, the Nelder-Mead simplex search which is a derivative optimization
algorithm will be combined with a new RWFCM algorithm in both probabilistic and
stochastic telecommunication big data. Both the combination of the two methods and
its area applied will be novel if it is successful. Also, this approach can be used for
simulating the mobility of people infected by Covid-19 through their mobile phone-
data in future research.
References
1. Altinel, I.K., Durmaz, E., Aras, N., Ozkisacik, K.C.: A location–allocation heuristic for the
capacitated multi-facility Weber problem with probabilistic customer locations. Eur. J. Oper.
Res. 198, 790–799 (2009)
2. Bezdek, J.C.: Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum Press,
New York (1981)
3. Cooper, L.: Location-allocation problems. Oper. Res. 11, 331–343 (1963)

4. Diabat, A., Dehghani, E., Jabbarzadeh, A.: Incorporating location and inventory decisions
into a supply chain design problem with uncertain demands and lead times. J. Manuf. Syst.
43, 139–149 (2017)
5. Esnaf, S., Kucukdeniz, T.: Solving uncapacitated planar multi-facility location problems by a
revised weighted fuzzy c-means clustering algorithm. J. Multiple Val. Logic Soft Comput.
21, 147–164 (2013)
6. Esnaf, S., Kucukdeniz, T.: A fuzzy clustering-based hybrid method for a multifacility
location problem. J. Intell. Manuf. 20(2), 259–265 (2009)
7. Gao, Y.: Shortest path problem with uncertain arc lengths. Comput. Math Appl. 62, 2591–
2600 (2011)
8. Gao, Y.: Uncertain models for single facility location problems on networks. Appl. Math.
Model. 36, 2592–2599 (2012)
9. Hansen, P., Jaumard, B., Taillard, E.: Heuristic solution of the multisource weber problem as
a p-median problem. Oper. Res. Lett. 22, 55–62 (1998)
10. Johnson, R.A., Wichern, D.W.: Applied Multivariate Statistical Analysis, 6th edn. Pearson,
New York (2007)
11. Kuenne, R.E., Soland, R.M.: Exact and approximate solutions to the multisource Weber
problem. Math. Program. 3, 193–209 (1972)
12. Kucukdeniz, T., Baray, A., Ecerkale, K., Esnaf, S.: Integrated use of fuzzy c-means and
convex programming for capacitated multi-facility location problem. Expert Syst. Appl. 39
(4), 4306–4314 (2012)
13. Markovic, N., Ryzhov, I.O., Schonfield, P.: Evasive flow capture: a multi-period
probabilistic facility location problem with independent demand. Eur. J. Oper. Res. 257,
687–703 (2017)
14. Megiddo, N., Supowit, K.J.: On the complexity of some common geometric location
problems. SIAM J. Comput. 13, 182–196 (1984)
15. Mousavi, S.M., Niaki, S.T.A., Mehdizadeh, E., Tavarroth, M.R.: Capacitated multi-facility
location-allocation problem with probabilistic customer location and demand: two hybrid
meta-heuristic algorithms. Int. J. Syst. Sci. 44, 1897–1912 (2012)
16. Murray, A.T., Church, R.L.: Applying simulated annealing to location-planning models.
J. Heuristics 2, 31–53 (1996)
17. Murtagh, B.A., Niwattisyawong, S.R.: Efficient method for the multi-depot location dash
allocation problem. J. Oper. Res. Soc. 33, 629–634 (1982)
18. Ohlemuller, M.: Tabu search for large location–allocation problems. J. Oper. Res. Soc. 48
(7), 745–750 (1997)
19. Peidroa, D., Mula, J., Poler, R., Verdegay, J.L.: Fuzzy optimization for supply chain
planning under supply, demand and process uncertainties. Fuzzy Sets Syst. 160, 2640–2657
(2009)
20. Wang, S., Watada, J.: A hybrid modified PSO approach to VaR-based facility location
problems with variable capacity in fuzzy random uncertainty. Inf. Sci. 192, 3–18 (2012)
21. Wen, M., Iwamura, K.: Fuzzy facility location-allocation problem under the Hurwicz
criterion. Eur. J. Oper. Res. 184, 627–635 (2008)
22. Wen, M., Kang, R.: Some optimal models for facility location allocation problem with
random fuzzy demands. Appl. Soft Comput. 11, 1201–1207 (2011)
23. Wen, M., Qin, Z., Kang, R.: The a-cost minimization model for capacitated facility location-
allocation problem with uncertain demands. Fuzzy Optim. Decis. Making 13, 345–356
(2014)
24. Wen, M., Qin, Z., Kang, R., Yang, Y.: The capacitated facility location-allocation problem
under uncertain environment. J. Intell. Fuzzy Syst. 29, 2217–2226 (2015)
25. Yang, L.X., Ji, X.Y., Gao, Z.Y., Li, P.K.: Logistics distribution centers location problem and
algorithm under fuzzy environment. J. Comput. Appl. Math. 208, 303–315 (2007)
26. Zhou, J., Liu, B.: New probabilistic models for capacitated location-allocation problem.
Comput. Ind. Eng. 45, 111–125 (2003)
27. Zhou, J., Liu, B.: Modeling capacitated location-allocation problem with fuzzy demands.
Comput. Ind. Eng. 53, 454–468 (2007)
Mobile Application Based Automatic
Caption Generation for Visually Impaired
Özkan Çaylı1 , Burak Makav1 , Volkan Kılıç1(B) , and Aytuğ Onan2

1
Department of Electrical and Electronics Engineering,
Izmir Katip Celebi University, Izmir, Turkey
[email protected]
2
Department of Computer Engineering, Izmir Katip Celebi University, Izmir, Turkey
Abstract. Generating captions and text descriptions of images will

enable visually impaired extended accessibility to the real world, thus
reducing their social isolation, improving their well-being, employability
and education experience. In this paper, we propose a captioning sys-
tem based on a smartphone, which offers ultra-low-cost, portable and
user-friendly platform for visually impaired. Image captioning aims to
understand and describe a visual scene in terms of natural language
expressions, which requires an advanced level of image understanding.
In that sense, it brings computer vision and natural language process
(NLP) fields together to generate meaningful sentences. In the computer
vision side, more distinguished visual attributes need to be extracted
from the images to feed the NLP for caption generation. The proposed
captioning system based on a neural encoder-decoder framework that
uses inception-v3 deep learning architecture as the encoder to extract the
visual attributes. On the other hand, the decoder is based on the NLP
architecture, which consists of multi-layer recurrent neural networks, to
generate captions. Our captioning system is run on the remote server
integrated to our custom-designed Android application via a cloud sys-
tem after trained on MSCOCO 2017 captioning dataset. The user can
choose an image from the gallery or capture a new image to transfer the
image to the remote server. The caption is generated by the proposed
captioning system and transferred back to the Android application to be
displayed on the screen. The caption can be read out loudly using the
narrator option. Experimental results show that the proposed system
offers great advantages to the visually impaired in terms of portability,
simple operation and rapid response.
Keywords: Deep learning · Natural language processing · Android
1 Introduction
In recent years, there has been an increasing interest in automatic caption gen-
eration as it empowers human-machine interaction to another level. The image
captioning has found applications in various fields, including, visual search, vir-
tual reality, augmented reality, image retrieval and indexing.
Automatic Caption Generation for Visually Impaired 1533
Many approaches have been proposed by researchers for caption genera-

tion and one of the most popular is the recurrent neural networks (RNN) [20].
Mathews et al. [18] introduced a method, called SentiCap, to generate a caption
with positive or negative sentiments based on a novel switching RNN model. In
time steps, the model assesses the possibility of switching between two RNNs.
One generates a caption considering the descriptive words, and the other weighs
the words by sentiments. It, then, receives inputs from the hidden states of both
two RNNs during generating captions. Oruganti et al. [19] presented a fusion-
based model with three stages, i.e., image processing stage, language processing
stage and fusion stage. In the image processing stage, a convolutional neural
network (CNN) [11] has been utilized to process the images. In the language
processing phase, a long-short term memory network (LSTM) [8] has been uti-
lized to process sentences. Following that, the outputs of these two stages are
mapped into a joint space, where the fusion stage connects these two modalities
and makes predictions. Ma et al. [14] proposed to use structural words for auto-
matic caption generation. Their method consists of two phases, namely, struc-
tural word recognition, and sentence translation. In this scheme, a multi-layer
optimization technique has been initially implemented to form hierarchical con-
cepts, which represent images, as objects, attributes, activities and scenes. The
concepts act as structural words. Then, a machine translation model based on
LSTM has been employed to convert the structural words into a caption. Sugano
et al. [22] utilized the human gaze in conjunction with the attention mechanism
based deep neural network. The empirical analysis has been conducted on the
SILICON dataset [10]. Makav et al. [15] investigated the performance of the
neural encoder-decoder model for automatic caption generation using the Stan-
ford model [17]. The model has been evaluated on the MS COCO dataset [13]
in terms of performance metrics, such as BLEU [25] and CIDEr [2]. In addi-
tion, Makav et al. [16] proposed to employ VGG16 [21] CNN architecture and
the LSTM network for caption generation which was integrated with a smart-
phone application for visually impaired people. Tanti et al. [24] examined the
performance of gated recurrent units (GRU) on automatic caption generation
and reported that GRU yields comparable results with the init-inject method.
The earlier research contributions highlight the need for comprehensive nat-
ural language processing models for automatic caption generation. In this paper,
we propose an automatic caption generation model based on encoder-decoder.
We demonstrate that by using inception-v3 (for encoder) and multi-layer RNNs
(for decoder), can improve the performance for enhanced caption generation.
Additionally, the presented deep neural network architecture has been integrated
with a mobile application. In this way, the automatic caption generation model
can be utilized by visually impaired for their welfare.
The rest of this paper is organized as follows: the next section introduces
the proposed approach for caption generation, dataset and mobile application
(Virtual Eye). Section 3 presents performance metrics used in comparisons of
existing and proposed methods. Closing remarks with future directions are given
in Sect. 4.
1534 Ö. Çaylı et al.
2 Methods
In this section, we explain our approach for the image caption generation and
the mobile application.
2.1 Image Caption Generator

The proposed image captioning framework consists of two neural network archi-
tectures: encoder and decoder. The purpose of the encoder is to extract features
from a queried image. Then, these features have been fed into the decoder. After
that, the decoder processes the features and generates a caption word-by-word.
Encoder: An encoder is a CNN formed by convolution, pooling and fully con-

nected layers. Convolution layers leverage three crucial concepts: sparse inter-
actions, parameter sharing, and equivariant representations. Sparse interactions
mean the exposure of short and essential features by kernels. Parameter sharing
means utilizing the related parameter of the kernel at every location of the data.
Lastly, equivariant representations mean variations in data affect the result like-
wise [6]. Therefore, deep CNNs have a significant impact on feature extraction
of the images. Before the training, all image features were extracted using the
average pooling layer of the inception-v3 CNN [23], which is trained to perform
object recognition and returns a 2048-element vector. Throughout the training,
a dense layer reduces the feature vector into a smaller vector.
Decoder: A decoder consists of embedding layer, multi-layer RNN and dense

layers. Embedding layer describes words in the vocabulary as vectors before the
RNN and it begins with vectors that are randomly initialized at the beginning.
The embedding layer is trained as a chain of the decoder to acquire the most
useful descriptions of words for the image caption generator [24]. Given an input
sequence x = (x1 , ..., xT ), a standard RNN computes the hidden vector sequence
h = (h1 , ..., hT ) and output vector sequence y = (y1 , ..., yT ) by iterating the
following equations from t = 1 to T:
ht = H(Wxh xt + Whh ht−1 + bh ) (1)

yt = Why ht + by (2)
where W denotes weight matrices, b denotes bias vectors and H is the hidden
layer function [7]. RNN is a significant contribution to the development of an
image caption generator where captions are sequentially produced. The term
multi-layer RNN refers to stacking multiple RNNs into one neural network. A
major advantage of multi-layer RNN is that it allows the network to compute
more complex representations.
GRU was proposed to make each recurrent unit to capture dependencies of
different time scales adaptively [3]. The GRU has gating units that modulate
the flow of information [4]. The activation hjt of the GRU at time t is a linear
interpolation between the previous activation hjt−1 and the candidate activation
h̃jt is defined as
hjt = (1 − ztj )hjt−1 + ztj h̃jt (3)

ztj = σ(Wz xt + Uz ht−1 )j . (4)
where an update gate ztj decides how much the unit updates its activation or con-
tent. The candidate activation h̃jt is computed similarly to that of the traditional
recurrent unit
h̃jt = tanh(W xt + U (rt ht−1 ))j (5)
where rt is a set of reset gates and is an element-wise multiplication. The reset
gate rtj is computed similarly to the update gate:
rtj = σ(Wr xt + Ur ht−1 )j (6)
The reset mechanism helps the GRU to use the model capacity efficiently by
allowing it to reset whenever the detected feature is not necessary anymore [5].
Fig. 1. Flowchart of 3-layered image caption generator
A dense layer is used to predict the next word in the caption. The output of
the last recurrent layer is fed through a dense layer which computes the prob-
ability of each possible next word in the vocabulary. Based on this probability,
the next word is selected. Fig. 1 shows the overall image caption generator.
Dataset: In image captioning, there are two commonly used datasets: Flickr [9]
and MSCOCO [13]. Flickr contains approximately 30000 images while MSCOCO
includes more than 120 K images. Both datasets have at least 5 captions per
image. In addition, MSCOCO involves all Flickr images and their captions which
makes us to choose MSCOCO dataset.
Performance Metrics: Captions of the proposed approaches are evaluated

using BLEU, CIDEr and Rouge [12]. These are common metrics to compare dif-
ferent image captions as they are found to be similar to human judgment. BLEU
(Bilingual Evaluation Understudy) is an automatic machine translation assess-
ment process that is a delicate criterion and profoundly associated with human
assessment. CIDEr (Consensus-based Image Description Evaluation) makes it
possible to compare machine production approaches objectively according to
human resemblance without having to make random calls to content, grammar,
openness, etc. CIDEr has initially been revealed particularly to assess image cap-
tioning yet is also used in video captioning methods. ROUGE (Recall-Oriented
Understudy for Gisting Evaluation) concludes the quality of a caption by corre-
lating it with separate captions produced by people.
2.2 Smartphone Application: Virtual Eye
Virtual Eye consists of the integration of three systems: server setup, cloud com-
munication and smartphone application design. Server setup and cloud commu-
nication are needed as our image caption generator can not be embedded in
the smartphone application. The image caption generator is run on the server
to generate a caption for an image sent from the Virtual Eye. Communication
between Virtual Eye and server is provided by the Firebase cloud service. First,
Virtual Eye sends an image to the Firebase. Second, the server is invoked by a
recently uploaded image to get the image. Third, the server generates a caption
and returns it to the Firebase. In the end, Virtual Eye gets the caption and
presents it to the user. Virtual Eye smartphone application is developed for the
Android operating system. Virtual Eye allows users to capture a photo from the
camera or choose it from the gallery. After Virtual Eye gets the caption, users
can listen the generated caption by tapping the image.
In this study, we presented a new automatic caption generation approach based

on the inception-v3 deep learning architecture and multi-layer RNNs. The
proposed approach was evaluated with performance metrics Bleu, Rogue and
CIDEr. The results for performance metrics were listed in Table 1 where the
best scores were indicated bold. The proposed approach was employed with three
different multi-layer RNN and compared with existing methods [1,26,27]. The
performance of the proposed approach was improved by increasing the number
of layers. With three-layer RNN, our approach outperformed the existing meth-
ods in terms of Bleu-1, Bleu-2, Bleu-3, Bleu-4 and Rouge. It is worth mentioning
that the proposed approach was enhanced with a six-layer only in the CIDEr
metric. On the other hand, the proposed approach has been integrated with
the Android application (Virtual Eye) for visually impaired. Generated captions
have been depicted in Fig. 2.
Fig. 2. Captioning results in Virtual Eye. (a): a laptop computer sitting on a wooden
desk. (b): a wooden bench next to a tree. (c): a view of a busy city with a very tall
building.
Table 1. Performance metric results
Bleu-1 Bleu-2 Bleu-3 Bleu-4 Rogue CIDEr

Proposed 1-layer 0.435 0.241 0.130 0.070 0.329 0.238
Chen et al. [1] 0.505 0.308 0.191 0.121 – 0.600
You et al. [27] 0.510 0.322 0.207 0.136 – 0.654
Xu et al. [26] 0.671 0.488 0.343 0.239 – 0.733
Proposed 3-layer 0.675 0.492 0.350 0.248 0.491 0.786
Proposed 6-layer 0.666 0.483 0.341 0.240 0.489 0.793
4 Conclusion
In this paper, we present an image caption generation model based on
the inception-v3 and multi-layer RNNs. Experimental investigations on the
MSCOCO dataset show the advantage of our proposed method over the state-
of-art approaches with significant performance results. We demonstrate that by
using GRU layers we can improve the performance of the RNN for enhanced
caption generation. The proposed caption generator was integrated with our
custom-designed Android application, Virtual Eye, to improve the life quality
of visually impaired. The user can select an image from the gallery or capture a
new image using the smartphone camera, then send it to the server via the Fire-
base for caption generation. The generated caption comes back to the Virtual
Eye to display on the screen. The user can listen to the caption by tapping the
image. The application could be further improved by adding extra features such
as caption translation into other languages and running on IOS and Windows
platforms.
Acknowledgements. This research was supported by the Scientific and Technological

Research Council of Turkey (grant no. 1139B411900533).
References
1. Chen, T., Zhang, Z., You, Q., Fang, C., Wang, Z., Jin, H., Luo, J.: “Fac-
tual” or “emotional”: stylized image captioning with adaptive learning and atten-
tion. In: Proceedings of the European Conference on Computer Vision (ECCV),
pp. 519–535 (2018)
2. Chen, X., Fang, H., Lin, T.Y., Vedantam, R., Gupta, S., Dollár, P., Zitnick, C.L.:
Microsoft coco captions: Data collection and evaluation server. arXiv:1504.00325
(2015)
3. Cho, K., Van Merriënboer, B., Bahdanau, D., Bengio, Y.: On the properties of
neural machine translation: encoder-decoder approaches. arXiv:1409.1259 (2014)
4. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recur-
rent neural networks on sequence modeling. arXiv:1412.3555 (2014)
5. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Gated feedback recurrent neural
networks. In: International Conference on Machine Learning, vol. 37, pp. 2067–
2075 (2015)
6. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. MIT Press, Cambridge
(2016)
7. Graves, A., Mohamed, A.R., Hinton, G.: Speech recognition with deep recurrent
neural networks. In: international Conference on Acoustics, Speech and Signal
Processing, pp. 6645–6649. IEEE (2013)
8. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8),
1735–1780 (1997)
9. Hodosh, M., Young, P., Hockenmaier, J.: Framing image description as a ranking
task: data, models and evaluation metrics. J. Artif. Intell. Res. 47, 853–899 (2013)
10. Jiang, M., Huang, S., Duan, J., Zhao, Q.: SALICON: saliency in context. In: Pro-
ceedings of the IEEE Conference on Computer Vision and Pattern Recognition
(2015)
11. LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E., Hubbard, W.,
Jackel, L.D.: Backpropagation applied to handwritten zip code recognition. Neural
Comput. 1(4), 541–551 (1989)
12. Lin, C.Y.: ROUGE: a package for automatic evaluation of summaries. In: Pro-
ceedings of Workshop on Text Summarization Branches Out, Post2 Conference
Workshop of ACL (2004)
13. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P.,
Zitnick, C.L.: Microsoft COCO: common objects in context. In: European Confer-
ence on Computer Vision, pp. 740–755. Springer (2014)
14. Ma, S., Han, Y.: Describing images by feeding LSTM with structural words. In:
International Conference on Multimedia and Expo (ICME), pp. 1–6. IEEE (2016)
15. Makav, B., Kılıç, V.: A new image captioning approach for visually impaired peo-
ple. In: 11th International Conference on Electrical and Electronics Engineering
(ELECO), pp. 945–949. IEEE (2019)
16. Makav, B., Kılıç, V.: Smartphone-based image captioning for visually and hearing
impaired. In: 11th International Conference on Electrical and Electronics Engi-
neering (ELECO), pp. 950–953. IEEE (2019)
17. Manning, C.D., Surdeanu, M., Bauer, J., Finkel, J.R., Bethard, S., McClosky,
D.: The stanford corenlp natural language processing toolkit. In: Proceedings of
52nd Annual Meeting of the Association for Computational Linguistics: System
Demonstrations, pp. 55–60 (2014)
18. Mathews, A.P., Xie, L., He, X.: SentiCap: generating image descriptions with sen-
timents. In: Thirtieth AAAI Conference on Artificial Intelligence (2016)
19. Oruganti, R.M., Sah, S., Pillai, S., Ptucha, R.: Image description through fusion
based recurrent multi-modal learning. In: International Conference on Image Pro-
cessing (ICIP), pp. 3613–3617. IEEE (2016)
20. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by back-
propagating errors. Nature 323(6088), 533–536 (1986)
21. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale
image recognition. arXiv:1409.1556 (2014)
22. Sugano, Y., Bulling, A.: Seeing with humans: gaze-assisted neural image caption-
ing. arXiv:1608.05203 (2016)
23. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the incep-
tion architecture for computer vision. In: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp. 2818–2826 (2016)
24. Tanti, M., Gatt, A., Camilleri, K.P.: Where to put the image in an image caption
generator. Nat. Lang. Eng. 24(3), 467–489 (2018)
25. Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R.,
Bengio, Y.: Show, attend and tell: neural image caption generation with visual
attention. In: International Conference on Machine Learning, pp. 2048–2057 (2015)
26. Xu, N., Liu, A.A., Liu, J., Nie, W., Su, Y.: Scene graph captioner: image captioning
based on structural visual representation. J. Vis. Commun. Image Represent. 58,
477–485 (2019)
27. You, Q., Jin, H., Luo, J.: Image captioning at will: a versatile scheme for effectively
injecting sentiments into image descriptions. arXiv:1801.10121 (2018)
Chaos Control of BLDC Motor via Fuzzy
Based PID Controller
Alkım Gökçen(&), Mehmet Uğur Soydemir, and Savaş Şahin
Department of Electrical and Electronics Engineering, Faculty of Engineering

and Architecture, Izmir Katip Celebi University, Izmir, Turkey
[email protected],
Abstract. This paper presents a developed method which is a fuzzy-based

proportional integral derivative (PID) controller design for the chaos control of
brushless direct current (BLDC) motor model. The parameters of the BLDC
motor model are determined with a linear regression method via analyzing the
system dynamics in the MATLAB/Simulink simulation environment. After
finding the model parameters, the BLDC motor is tested under forcing the
chaotic behaviors range by using an external load torque input. The performance
analysis of the developed fuzzy-based PID controller is done for suppressing
chaotic behaviors of the BLDC motor dynamics that is to say the system
dynamics of the motor should provide their equilibrium or limit cycle solutions.
The performances of the developed fuzzy-based PID and conventional PID
controllers are compared to each other and their results are presented in terms of
the mean square error, and the suppression performance of the model’s chaotic
behaviors is tested by using the phase portraits and maximum Lyapunov
exponents.
Keywords: Fuzzy based PID controller BLDC motor Chaos control
1 Introduction
Chaos is a widely encountered phenomenon in the real-word applications including

biomedical engineering [1], chemical process control [2], mechanical system design [3]
and embedded systems in electronics engineering [4]. Within these systems, brushless
direct current (BLDC) motor is a commonly used as actuator and/or plant for chaos
control applications and chaotic behavior analysis due to its complex dynamics under
certain external load conditions [5–7]. Chaos control is a method that changes the
chaotic behavior of a system into asymptotical equilibrium point or limit cycle by using
a controller [8]. In this study, a three-time scales BLDC motor model is used [9], and
model parameters are adjusted to a real BLDC motor plant by estimating the param-
eters by real data in accordance with [10].
In this study, chaos analysis of the estimated model is examined according to
simulation studies that was realized with a simulation environment, and the load torque
conditions for chaotification of the model state variables are determined. A fuzzy based
proportional integral derivative (PID) controller is designed and implemented for
Chaos Control of BLDC Motor via Fuzzy Based PID Controller 1541
controlling chaotic behaviors of the BLDC motor dynamics. Also, a conventional PID
controller is used to compare performances of the developed fuzzy based PID con-
troller by taking into consideration the mean squared error (MSE) criteria. At the
controller performance test stage in the simulation, the controller is switched on after
the system is chaotified by an external load for 10 s. The purpose of the controller is to
get the chaotified system to the desired output called as the reference. According to the
obtained results of the experiments, phase portraits of the model state variables are
examined to observe suppression of the chaotic behaviors and maximum Laypunov
exponent is evaluated by using the results.
Remaining parts of the paper is organized as follows: In Sect. 2 three-time scales
BLDC motor model dynamics and its parameter estimation method are explained.
Section 3 presents the motor model parameters analyses for the chaotic operation and
the determination of the load torque model parameters. The developed fuzzy based PID
controller design is explained in Sect. 4. The results of the experiments and perfor-
mances are presented in Sect. 5, and the conclusion and future directions are given in
Sect. 6.
2 BLDC Motor Model
The electrical and mechanical model dynamics of a BLDC motor might be explained in
(1) given in [11, 12] in the following form
d 1
iq ¼ Riq nx Ld iq kt þ vq
dt Lq
d 1
id ¼ Rid þ nLq xiq þ vd ð1Þ
dt Ld
d 1
x ¼ ½T ðI; hÞ T‘ ðtÞ
dt j
where Ld and Lq are the inductance variables on direct-axis (d-axis) and quadrature-axis
(q-axis); n is the number of permanent magnet pole pairs; x represents the angular
speed of the rotor; R is the winding resistance; J represents the momentum of inertia; kt
represents the permanent magnet flux which is a constant; h is the angular displace-
ment; voltages and currents on d-axis and q-axis are as given as vd ,vq ,id and iq ;
T
I ¼ id iq , T‘ is the external load torque respectively. Electromagnetic torque TðI; h)
is given by

T iq ; id ¼ n kt iq þ Ld Lq iq id ð2Þ
Three time-scales model is a state space that is a transformed model of given state
space in (1) and (2) using multiple time constants s1 ,s2 and s3 . These time constants
can be described as given in [9].
1542 A. Gökçen et al.
Lq Ld JR
s1 ¼ ; s2 ¼ ; s3 ¼ 2 ð3Þ
R R kt
where s1 represents the first electrical time constant, s2 represents the second electrical
time constant, and s3 represents the mechanical time constant. After transforming the
given state space model using time constants, three-time scales state space model is
described by
s1 dtd x1 ¼ Vq x1 x2 x3 x3
s2 dtd x2 ¼ Vd þ x1 x3 x2 ð4Þ
s3 dt x3 ¼ rx1 þ qx1 x2 gx3 f
d
TL
where Vq , Vd , x1 and x2 are the q-axis and d-axis voltage current pairs; x3 represents the
angular speed, and f TL is the external load torque providing the chaotic system
dynamics behaviors. Model parameters r, q and g are described by
Rb Lq
r ¼ n2 ; q ¼ ð1 dÞn2 ; g ¼ ;d ¼ ð5Þ
kt2 Kd
where b is the damping coefficient of the viscosity, Rb is the winding resistance and kt
is the torque constant.
In this study, unknown parameters Ld and Lq of three time-scales BLDC motor state
space model are estimated to find time constants, and parameters r, q and g are chosen
from [11]. The values of the time constants are computed using linear regression
method [10], and these values are determined as s1 ¼ 0:0054, s2 ¼ 0:0050 and s3 ¼
0:0130 for the considered BLDC motor model.
3 Chaos in BLDC Motor Model
BLDC motors are an area that is commonly studied by the researchers due to its chaotic
behavior under spontaneous change conditions of load [9]. In this study, chaos analysis
of three time-scales BLDC motor model are performed to determine operating con-
ditions of the parameters. The homogeneous solution of the differential equations given
in three time-scales state space model is similar to the Lorenz system in terms of their
dynamical behaviors [13]. Phase portrait of the system with external load torque
f
TL ¼ 0, and initial conditions for states x ¼ ½100T is shown in Fig. 1. It can be seen
that behavior of the system model with estimated parameters is not chaotic for given
conditions.
For the chaotification of the BLDC motor model, f TL signal is applied as a square
wave with an amplitude 4Vpp and a frequency of 15Hz: Hence, the obtained phase
portrait of the BLDC motor model is depicted in Fig. 2 for the initial conditions of the
system with x ¼ ½000T . The chaotic behaviors of the motor model might be easily
observed with the external load applied. Maximum Lyapunov exponent, as a measure
of the divergence from initial conditions to a non-equilibrium point where its positive
Fig. 1. Phase portrait of the BLDC motor model without load
Fig. 2. Phase portrait of the BLDC motor model with load.
measurement describes the chaos of the considered system, of state variable x3 is

analyzed and computed with the MATLAB environment and it is found as
kx3 ¼ 0:5782.
4 Controller Design
In this study, the conventional PID controller and fuzzy based PID controller algo-
rithms are explained in the following subsections, respectively.
4.1 Conventional PID Controller

The PID controller is used to produce a control signal considering the proportional,
integral and derivative gains with the error signal so that the controlled output should
be approached the desired output of the closed loop system. The control signal standing
for uðtÞ in time domain [14] is described as follows
Z deðtÞ
uðtÞ ¼ Kp eðtÞ þ Ki eðtÞdt þ Kd ð6Þ
dt
where eðtÞ is the error signal in time domain, Kp is the proportional gain; Ki represents
the integral gain; Kd is the derivative gain; uðtÞ represents the control signal in time
domain. In this study, the PID controller parameters are tuned empirically by using
“PID Tuner” function of PID controller block of MATLAB/Simulink simulation
environment. The PID controller parameters are determined as Kp ¼ 0:38, Ki ¼ 65:96
and Kd ¼ 0:0004.
4.2 Fuzzy Logic Based PID Controller

Since fuzzy logic is a powerful tool to construct a relation between linguistic expres-
sions and the mathematical representations, the developed controller algorithms might
be formulated to control the open and/or closed-loop control system [15]. Herein,
input-output variables of the fuzzy based system might be defined by using the
membership functions that creates relation between all variables each other. These
membership functions might be defined as geometric based functions such as singleton,
trapezoidal, Gaussian and/or triangle.
The fuzzy logic controller might be designed with three stages [15]. These con-
secutive stages can be defined as fuzzification, inference, and defuzzification. The
process of defining the input-output relations system to be fuzzified and their mem-
bership functions can be named as the fuzzification stage. Then, the rule table is
designed in order to define the linguistic expression on the system which is called as
inference stage. The last stage is that the defuzzification that is used to transform the
fuzzified outputs into admissible values for real-world applications. In this study, the
developed fuzzy logic based PID controller is designed with gain scheduling approach
considering error signal E, and derivative of error CE (see Fig. 3).
Fig. 3. Membership functions for the developed controller inputs
5 Simulation and Results
Three time-scales BLDC motor model given in (4) is chaotified for 10 s by switching
the external load input on as square wave with the initial conditions x ¼ ½000T . During
the total simulation is 20 s, the controller is switched on for the next 10 s while the load
torque input remains switched on. When the external load torque input applied, the
chaotic behaviors can be seen in the BLDC motor model state variables and the time
evaluations of these state variables as x1 , x2 angular velocity are depicted in Fig. 4 and
Fig. 5 for the conventional PID controller and the developed fuzzy based PID con-
troller, respectively. The phase portraits are depicted for the comparisons of the con-
ventional PID and developed fuzzy based PID controllers’ performances for the
simulation duration. It is observed that fuzzy based PID controller cope with the
chaotic behaviors of the BLDC motor model in Fig. 6. As for the maximum Lyapunov
exponents evaluations, these values are computed as kPID ¼ 0:1855 and kfuzzy ¼ 0 for
both controllers to observe the suppressing chaos on dynamics. It is deduced that the
fuzzy-based PID controller suppression performance is better than the PID controller’s
performance. The chaos of the state x3 of the BLDC motor model is suppressed and it
behaves as a limit cycle. Moreover, the obtained MSE metrics are calculates as 0:6665
and 0:0501. As it can be seen from the MSE values are found as 0:6665 and 0:0501 for
the conventional PID and the developed fuzzy logic based PID controllers. The fuzzy
based controller performance ensures a better performance in terms of tracking error of
the closed loop system.
Fig. 4. Time evaluations of the BLDC motor model states with PID controller with applying the
load at 10 s
Fig. 5. Time evaluations of the BLDC motor model states with fuzzy based PID controller with
applying the load at 10 s
Fig. 6. Phase portraits for both PID and Fuzzy based PID controller
6 Conclusion
In this study, the developed fuzzy logic based PID controller is used to suppress the
chaotic behaviors of the BLDC motor model. In order to observe the chaos on BLDC
motor, an external load torque is applied and the three time-scales BLDC motor model
responses are analyzed for both developed and conventional PID controllers’ perfor-
mances in terms of MSE, maximum Lyapunov exponents and phase portraits of the
motor states. The maximum Lyapunov exponent of the motor state x3 which is com-
puted as 0 defining the chaotic behavior is suppressed and it can be seen limit cycle.
The obtained MSE result of the fuzzy logic based PID controller is found as 0:0501. In
the future direction, the reinforcement based controller design might be a new research
area for the chaos control applications.
References
1. Gokcen, A., Şahin, S.: Design of chaotic system based pacemaker on field programmable
analog array board. In: Medical Technologies National Congress, TIPTEKNO, pp. 1–2.
IEEE Xplore, Turkey (2019)
2. Epstein, I.R., Showalter, K.: Nonlinear chemical dynamics: oscillations, patterns, and chaos.
J. Phys. Chem. 100(31), 13132–13147 (1996)
3. Seth, S., Banerjee, S.: Electronic circuit equivalent of a mechanical impacting system.
Nonlinear Dyn. 99, 3113–3121 (2020)
4. Gokcen, A., Büyükçakır, B., Şahin, S.: Implementation and application of logistic map on
field programmable analog array board. In: 26th Signal Processing and Communications
Applications, SIU, IEEE, pp. 1–2. IEEE Xplore, Turkey (2018)
5. Das, M., Jadoun, V.K.: Stabilization of chaotic systems using robust optimal controller. In:
Kalam, A., Niazi, K., Soni, A., Siddiqui, S., Mundra, A. (eds.) Intelligent computing
techniques for smart energy systems. Lecture Notes in Electrical Engineering (2020), vol.
607, pp. 795–803. Springer, Singapore (2019)
6. Cheukem, A., Tsafack, A.S.K., Kingni, S.T., André, C.C., Pone, J.R.M.: Permanent magnet
synchronous motor: chaos control using single controller, synchronization and circuit
implementation. SN Appl. Sci. 2, 420 (2020)
7. Jouhet, G., González-Jiménez, L.E.: Model-based fault detection of permanent magnet
synchronous motors of drones using current sensors. In: Ghommam, J., Derbel, N., Zhu, Q.
(eds.) New trends in robot control. Studies in Systems, Decision and Control (2020), vol.
270, pp. 301–318. Springer, Singapore (2020)
8. Chen, Z., Chen, F.: Complex aperiodic mixed mode oscillations induced by crisis and
transient chaos in a nonlinear system with slow parametric excitation. Nonlinear Dyn. 100,
659–677 (2020)
9. Ge, Z.M., Cheng, J.W., Chen, Y.S.: Chaos anticontrol and synchronization of three time
scales brushless DC motor system. Chaos Solitions Fractals 24(2), 597–616 (2005)
10. Neugebauer, H.: Parameter identification of a permanent magnet synchronous motor. MS
thesis (2012)
11. Hemati, N.: Strange attractors in brushless DC motors, IEEE Trans. Circuits Syst. I: Fundam.
Theory Appl. 41, 40–45. IEEE (1994)
12. Zhaojun, M., Changzhi, S., Yuejun, A.: Chaos anti-control of permanent magnet
synchronous motor based on exact linearization via time-delayed state variable feedback.
Trans. China Electrotech. Soc. 22(3), 27–31 (2007)
13. Yang, S.K., Chen, C.L., Yau, H.T.: Control of chaos in Lorenz system. Chaos Solitions
Fractals 13(4), 767–780 (2002)
14. Büyükçakir, B., Elmaz, F., Sahin, S.: Stochastic optimization of PID parameters for twin
rotor system with multiple nonlinear regression. In: 6th International Conference on Control
Engineering & Information Technology, CEIT, pp. 2–3. IEEE Xplore, Turkey (2018)
15. Blanchett, T.P., Kember, G.C., Dubay, R.: PID gain scheduling using fuzzy logic. ISA
Trans. 39(3), 317–325 (2000)
Sustainability Oriented Scheduling Procedure
for Public Projects
Dorota Kuchta(&) and Ewa Marchwicka
Wroclaw University of Science and Technology,

Wyb. Wyspianskiego 27, 50-370 Wroclaw, Poland
[email protected]
Abstract. Public projects are the projects implemented by public administra-

tion units and with the involvement of funds from the budget of this adminis-
tration. Based on the literature analysis and the case studies of the completed
public projects in different countries, it can be concluded that public projects do
not always end on time and that many basic methods for project scheduling and
task duration estimation are not used. Public projects have a significant impact
on the future and consume huge amount of resources, thus their planning cannot
be detached from sustainability principles. Therefore, there is a need to recog-
nize scheduling of public projects as one of the key activities carried out by
project managers and to develop appropriate methods in this area. In the paper
we propose a procedure of scheduling public projects that allow to mitigate the
scheduling problems and help in applying sustainability principles where the
degree of sustainability is measured using fuzzy terms.
Keywords: Sustainability Project scheduling Fuzzy terms Public projects
1 Introduction
Administration are responsible for providing basic public services for citizens [1]. The
strategic investments [2] carried out by local government units are long term invest-
ments with the effects that can be observed after a few years. For this reason, local
government units by realizing strategic investments initiate changes and become
responsible for the future of citizens and the wealthiness of future generations. As all
other projects, public projects require project management. Recently, a lot of effort has
been made to change the way of determining project’s success factors. The classical
approach, that considers PM Iron Triangle, has been extended with a few new
dimensions to be considered. Project management becomes something more than the
process of managing cost, time and quality. Modern approaches include also other
important dimensions: economic, environmental and social dimensions. This is known
as the sustainable project management [3]. Sustainability can be defined as [4] ‘forms
of progress that meet the needs of the present without compromising the ability of
future generations to meet their needs’. There are 6 rules of sustainability [3]:
(1) balancing social, environmental and economic interests, (2) including both short
term and long term orientation, (3) including both global and local orientation
(4) consuming income, not capital, (5) providing transparency and accountability,
Sustainability Oriented Scheduling Procedure for Public Projects 1549
(6) cultivating social values and ethics. There are number of publications about sus-
tainable approaches in many different contexts, e.g. ecology [5], urbanization [6],
production and manufacturing [7], energy consumption [8], construction [9], supply
chain management [10] or education [11]. Sustainability in project management has
been widely analyzed [12–15] in many different use cases, including infrastructure
project management [16]. Also, many project management aspects, like corporate
strategy [17], project success factors/project evaluation [18], project stakeholders view
[19] have been discussed.
In this article we present the analysis of public projects realized by Polish local
government units in the context of scheduling. Scheduling is a crucial element of
providing transparency and accountability, which refers to the fifth rule of sustain-
ability and is very important in the social dimension of sustainability. Scheduling
enables communication to the stakeholders and to general public [3] about the list of
activities and the time of their realization. The time spent on communication campaigns
and information meetings should be included in the schedule as well [16]. In a broader
context of project planning, the process of scheduling, should be proceeded with other
sustainability-related activities, like risk assessment and cost analysis [3], which are
very important in the ecological dimension of sustainability. Risk assessment and
analysis can include the information about possible ecological disasters [3] and cost
analysis can estimate important cost components, like the impact of pollution [3],
which directly refers to the fourth rule of sustainability. In this work we describe public
projects, which have a strong influence on the future, involve the cooperation of many
stakeholders, consume huge amount of resources and often change the environment
[16], which refers to the first rule of sustainability: balancing social, environmental and
economic interests.
The work objective is to develop a procedure for public projects scheduling that
allows to mitigate the scheduling problems and to help in applying sustainability
principles, when the degree of sustainability is measured using fuzzy terms. The rest of
the paper is organized as follows. Section 2 describes the proposal of this procedure
Sect. 3 contains conclusions.
2 Proposal of Procedure
The procedure considers public projects’ schedule and is based on multi-criteria

heuristic procedure applied to the knowledge of domain experts. What is more, it
includes the uncertainty of domain experts’ decisions. There are three different sus-
tainability criteria are considered in the procedure:
• Human resources consumption (when human resources are overloaded with work,
we cannot assign any further activities to them; it is better to realize the activity
within a longer period, but considering a human factor; e.g. we can delegate the
activity to less efficient employees, but less tired and less appreciated. The activity
is expected to increase its duration, but its sustainability index will also increase.
Another possibility here is to delegate the activity to an external company, so that
employees are not overloaded with work.
1550 D. Kuchta and E. Marchwicka
• Materials consumption (some materials can be over-used or wasted, some materials

can be harmful to the environment, we suggest to use less scarce materials or less
environmental harmful materials, which possibly induces the need to add an extra task
of bringing the material to the company, but can also increase sustainability index).
• Negative influence on local environment and local communities (the good of the
citizens, for example, should be considered; if there are three renovations at once in a
given area, it can be too tiring for the local citizens; moving this activity from night to
day or from non-holidays time to holidays, can help citizens, e.g. renovation of a
tram line can be sifted to holidays when people use public transport less often).
Let us denote the three dimensions of sustainability listed above as SH (human
resources consumption), SM (material consumption) and SN (negative influence on local
community). For the purpose of the method a few assumptions are to be made:
• All three dimensions of sustainability are measurable at the time of scheduling, but
their values cannot be determined precisely; the dimensions are described as one of
possible values {LOW, MEDIUM, HIGH}.
• For the dimensions that are not measurable at the time of scheduling, e.g. dimension
SH can be non-measurable, because it can be hard to determine human consumption
prior to activity start, we omit them in the calculations. The values will be refined
during project progress and maybe rescheduling will be applied later on.
• The values {LOW, MEDIUM, HIGH} for each dimension are obtained by N inde-
pendent experts.
• Experts can use different criteria to assess their values for each dimension.
An expert assesses the value with the likelihood of decision’s correctness, which is
also from the set of three possible values {LOW, MEDIUM, HIGH}. Each expert
can give a set of possibilities, with several likelihoods associated to them.
• Expert decisions are mapped to trapezoidal fuzzy numbers, defined by (a, b, c, d).
Because different experts can use different criteria, the domain of the numbers is
normalized to [0, 1] interval, where 0 is the lowest possible value and 1 is the
highest possible value.
• Expert decisions can be mapped to fuzzy numbers using 3 different mappings. The
first mapping is {LOW, MEDIUM, HIGH} corresponding to {[0; 1/3], [1/3; 2/3],
[2/3; 1]}, with 10% of possible values overlapping with the values of adjacent
choice, i.e. 10% of the LOW values overlap with 10% of the MEDIUM values and
vice versa and 10% of the MEDIUM values overlap with 10% of the HIGH values
and vice versa. By overlapping we mean the situation when the membership
function is less than 1 and a few choices are possible. This is presented in the Fig. 1.
We denote this mapping as R1 = {LOW, MEDIUM, HIGH} => {[0; 1/3 ± 10%],
[1/3 ± 10%; 2/3 ± 10%], [2/3 ± 10%; 1]}. The second mapping is {LOW,
MEDIUM, HIGH} corresponding to {[0; 1/3], [1/3; 2/3], [2/3; 1]}, with 20% of
possible values overlapping with the values of adjacent choice. We denote this
mapping as R2 = {LOW, MEDIUM, HIGH} => {[0; 1/3 ± 20%], [1/3 ± 20%;
2/3 ± 20%], [2/3 ± 20%; 1]}. The third mapping is {LOW, MEDIUM, HIGH}
corresponding to {[0; 1/3], [1/3; 2/3], [2/3; 1]}, with 50% of possible values
overlapping with the values of adjacent choice (Fig. 2). We denote this mapping as
R3 = {LOW, MEDIUM, HIGH} => {[0; 1/3 ± 50%], [1/3 ± 50%; 2/3 ± 50%],
[2/3 ± 50%; 1]}.
Fig. 1. Experts choices mapped to trapezoidal fuzzy numbers. R1 = {LOW, MEDIUM,

HIGH} => {[0; 1/3 ± 10%], [1/3 ± 10%; 2/3 ± 10%], [2/3 ± 10%; 1]}, i.e. 10% of values
for each choice overlap with values of adjacent choice.
Fig. 2. Experts choices mapped to trapezoidal fuzzy numbers. R3 = {LOW, MEDIUM,

HIGH} => {[0; 1/3 ± 50%], [1/3 ± 50%; 2/3 ± 50%], [2/3 ± 50%; 1]}. i.e. 50% of values
for each choice overlap with values of adjacent choice.
• Define a set of possible mitigation actions, that allows to improve the activity
sustainability index. This set should be build based on the understanding of three
sustainability dimensions used and can look similarly to the example set of actions
listed above: {assign activity to different resources, outsource activity to an external
company, order different materials, move activity to holidays, etc.}.
The steps of the procedure used to determine the schedule are as follows:
1. Determine the activity that has not been scheduled yet, but its predecessors have
already been scheduled.
2. Place the activity at earliest possible time.
3. Determine the sustainability index of the activity. This index is calculated using the
following data: experts’ decisions (Table 1), non-deterministic factor (Monte Carlo
simulation), set of heuristic rules (Table 2). For i = 1 to M (number of simulations)
perform the following steps:
Table 1. Sustainability assessment for a single dimension. Sample results.

Expert Likelihood (LOW) Likelihood (MEDIUM) Likelihood (HIGH)
1 HIGH – –
2 HIGH MEDIUM –
…
N HIGH MEDIUM LOW
a. If it is possible to determine the sustainability index for the first dimension, cal-
culate the sustainability index of the first dimension as follows: for each set of
choices provided by the expert, randomly select only one choice according to the
likelihood distribution (the choice with higher likelihood is more likely to be
selected), then map the choice to fuzzy number using rules from Table 2, then
calculate the index as the average value of N experts SH (i) = (E1 (i) ⊕ E2 (i) ⊕
… ⊕ EN (i))/N, where E1 (i),…, EN (i) are expert choices and ⊕ is the summation
operator that is used with trapezoidal fuzzy numbers and is defined as follows [20]:
8 xða þ a Þ
>
>
1 2
; a1 þ a2 x b1 þ b2
< ðb1 þ b2 Þða1 þ a2 Þ
lAB ¼ 1; b1 þ b2 x b1 þ b2 ð1Þ
ðd1 þ d2 Þx
>
> ðd1 þ d2 Þðc1 þ c2 Þ ; c 1 þ c 2 x d1 þ d2
:
0; otherwise
where both A and B are trapezoidal fuzzy numbers defined as A = (a1, b1, c1, d1)
and B = (a2, b2, c2, d2).
b. If it is possible to determine the sustainability index for the second dimension,
calculate the sustainability index SM (i) of the second dimension similarly as in
step 3.1.
c. If it is possible to determine the sustainability index for the third dimension,
calculate the sustainability index SN (i) of the third dimension similarly as in step
3.1.
d. Determine the cumulative sustainability index for all dimensions as SH+M+N (i) =
(SH (i)+ SM (i)+ SN (i))/D, where D is the number of dimensions used.
e. Compare the cumulative sustainability index with LOW, MEDIUM and HIGH
fuzzy numbers (as defined in Figure 1), using the similarity measure of fuzzy
numbers. Such a similarity measure can be found in [21]: SLOW (i) = Similarity
(SH+M+N (i), LOW), SMED (i) = Similarity (SH+M+N (i), MEDIUM), SHIGH (i) =
Similarity (SH+M+N (i), HIGH).
4. Calculate the average similarity measures of all the simulations as

P
M P
M P
M
SLOW = M1 SLOW ðiÞ, SMED = M1 SMED ðiÞ, SHIGH = M1 SHIGH ðiÞ and pick the
i¼1 i¼1 i¼1
measure with the greatest similarity.
Table 2. Heuristic rules used for the procedure.

Rule name Description
Rule 1 If sustainability assessment contains only 1 choice
(e.g. Expert’s 1 choice) => map it using R1 mapping
Rule 2 If sustainability assessment contains 2 choices
(e.g. Expert’s 2 choices) => map it using R2 mapping
Rule 3 If sustainability assessment contains 3 choices
(e.g. Expert’s N choices) => map it using R3 mapping
5. If the measure with the greatest similarity is SHIGH, apply one or more mitigation
actions. If the measure is greatest for SLOW or SMED, don’t apply any mitigation
actions.
3 Conclusions
In this article we propose a method for scheduling infrastructure projects, that is aware
of sustainable aspects of project management. We suggested three dimensions of
sustainability to be considered: human resources consumption (SH), material con-
sumption (SM) and negative influence on local communities (SN). The procedure takes
advantage on the opinions of independent experts, includes heuristic rules applied to
their decisions and considers the uncertainty level of experts’ decisions. The process of
calculating activity sustainability index is based on fuzzy approach, where experts’
opinions are not precise, but fuzzy, and based on different criteria.
The real-project example analysis, as well as the analysis of different mapping rules
applied as well as the analysis of different sustainability dimensions can be subject to
future research.
This research was supported by the National Science Centre (Poland), under Grant
394311, 2017/27/B/HS4/01881: “Selected methods supporting project management,
taking into consideration various stakeholder groups and using type-2 fuzzy numbers”.
References
1. Witkowski, J., Kiba-Janiak, M.: The role of local governments in the development of city
logistics. Proc. Soc. Behav. Sci. 125, 373–385 (2014)
2. Terrados, T., Almonacid, G., Hontoria, L.: Regional energy planning through SWOT
analysis and strategic planning tools, Impact on renewables development. Renew. Sustain.
Energy Rev. 11(6), 1275–1287 (2007)
3. Silvius, G., Schipeer, R., Planko, J., van den Brink, J., Kohler, A.: Sustainability in Project
Management. Routledge, New York (2016)
4. World Commission on Environment and Development: Our Common Future. Oxford
University Press, Great Britain (1987)
5. Guerrero-Liquet, G.C., Sánchez-Lozano, J.M., García-Cascales, M.S., Lamata, M.T.,
Verdegay, J.L.: Decision-making for risk management in sustainable renewable energy
facilities: a case study in the dominican republic. Sustainability 8(5), 455 (2016)
6. Martek, I., Hosseini, R.M., Shrestha, A., Zavadskas, E.K., Seaton, S.: The sustainability
narrative in contemporary architecture: falling short of building a sustainable future.
Sustainability 10(4), 981 (2018)
7. Levy, D., Reinecke, J., Manning, S.: The political dynamics of sustainable coffee: contested
value regimes and the transformation of sustainability. J. Manage. Stud. 53(3), 364–401
(2015)
8. Hunter, G.W., Vettorato, D., Sagoe, G.: Creating smart energy cities for sustainability
through project implementation: a case study of Bolzano, Italy. Sustainability 10(7), 2167
(2018)
9. Wu, G., Qiang, G., Zuo, J., Zhao, X., Chang, R.: What are the key indicators of mega
sustainable construction projects? — a stakeholder-network perspective. Sustainability 10,
2939 (2018)
10. Quarshie, A.M., Salmi, A., Leuschner, R.: Sustainability and corporate social responsibility
in supply chains: The state of research in supply chain management and business ethics
journals. J. Purch. Supply Manage. 22(2), 82–97 (2016)
11. Thürer, M., Tomašević, I., Stevenson, M., Qu, T., Huisingh, D.A.: Systematic review of the
literature on integrating sustainability into engineering curricula. J. Clean. Prod. 181, 608–
617 (2017)
12. Köhler, A., van den Brink, J., Silvius, G.: The impact of sustainability on project
management. In: Linger, H., Owen, J. (eds.) The Project as a Social System: Asia-Pacific
Perspectives on Project Management. Monash University Publishing, Australia (2012)
13. Silvius, G., Schipper, R.: Sustainability in project management: A literature review and
impact analysis. Soc. Bus. 4(1), 63–96 (2014)
14. Dyllick, T., Muff, K.: Clarifying the meaning of sustainable business: introducing a typology
from business-as-usual to true business sustainability. Organ. Environ. 29(2), 156–174
(2015)
15. Morrison, R.L., Dillman, D.A., Christian, L.M.: Questionnaire design guidelines for
establishment surveys. J. Official Stat. 26(1), 43–85 (2010)
16. Kivilä, J., Martinsuo, M., Vuorinen, L.: Sustainable project management through project
control in infrastructure projects. Int. J. Project Manage. 35(6), 1167–1183 (2017)
17. Evans, S., Vladimirova, D., Holgado, M., van Fossen, K., Yang, M., Silva, E.A., Barlow, C.
Y.: Business model innovation for sustainability: towards a unified perspective for creation
of sustainable business models. Bus. Strategy Environ. 26(5), 597–608 (2017)
18. Khalilzadeh, M., Akbari, H., Foroughi, A.: Investigating the relationship of sustainability
factors with project management success. Ind. Eng. Manage. Syst. 15(4), 345–353 (2016)
19. Bellantuono, N., Pontrandolfo, P., Scozzi, B.: Capturing the stakeholders’ view in
sustainability reporting: a novel approach’. Sustainability 8(4), 1–12 (2016)
20. Rezvani, S., Molani, M.: Representation of trapezoidal fuzzy numbers with shape function.
Ann. Fuzzy Math. Inform. 8(1), 89–112 (2014)
21. Marchwicka E., Project risk management method. PhD dissertation, Wroclaw University of
Science and Management (2012)
Optimization of an Oil Refinery Valuation
System Through the Intuitionistic Fuzzy
InterCriteria Analysis
Velichka Traneva(B) and Stoyan Tranev
“Prof. Asen Zlatarov” University, “Prof. Yakimov” Blvd, 8000 Bourgas, Bulgaria
http://www.btu.bg
Abstract. In this paper, an extended form of three-dimensional intu-

itionistic fuzzy intercriteria analysis (3-D ICrA) is proposed and applied
to analyze a dataset gathered through the staff evaluations of a catalytic
oil refinery installation in Bulgaria over a three-year period. The 3-D
ICrA is applied in order to reduce the set of staff assessment criteria in
a refinery, taking into account the fact that their large number increases
the time and cost of the staff evaluation process. The ICrA is one of the
recently proposed approach for multicriteria decision making, combining
the classical correlation analysis and the fuzzy logic. The approach is
based on the concepts of index matrices (IMs) and intuitionistic fuzzy
sets (IFSs), defined by Atanassov.
The main contribution of the paper is that it illustrates the effec-
tiveness of the proposed method to optimize the staff rating system of
the catalytic oil refinery installation in Bulgaria. The open dependencies
between the evaluation criteria will lead to optimization of the com-
pany’s staff rating system by removing part of these criteria and will
support the decision-making process related to the motivation of human
resources. The results, obtained using this approach are compared with
those of the application of ICrA, Pearson’s (PCA), Spearman’s (SCA)
and Kendall’s (KCA) rank correlation analysis. The results of applying
these four statistical correlation analyzes to the actual data are almost
identical.
Keywords: Decision making · Index matrix · InterCriteria analysis ·

Intuitionistic fuzzy sets · Rating criteria · Oil refinery
1 Introduction
The ICrA method [6] was developed in 2014. It is based on the concepts of IFSs
and IMs. IFSs is first defined by Atanassov [1] as an extension of the concept
of fuzzy sets defined by Zadeh [19]. The concept of IMs is introduced in [2].
Supported by the Bulgarian National Scientific program “ICT in science, education
and security” (V. Traneva).
Later the ICrA has been extended in a theoretical aspect for its application to
two- and three-dimensional intuitionistic fuzzy data [10,16]. The approach has
discussed in a number of papers considering for example hybrid algorithms [13],
neural networks [14], economic investigations [9], etc.
In this paper, an extended form of 3-D ICrA [16] is introduced and applied
to analyze a dataset gathered through the staff evaluations of a catalytic oil
refinery installation in Bulgaria over a three-year period. Also we apply the
classic PCA, SCA, KCA [11] to the real data to show the advantages of the
proposed approach. The comparison of the obtained results illustrates that there
is no comparative difference between those, obtained from the ICrA and the other
three classic correlation analyzes, but only the ICrA can be applied to fuzzy data.
The originality of the paper comes from the proposed extended form of the 3-D
ICrA. The main contributions of the paper lie in its proposition for an extended
form of the 3-D ICrA over IF evaluations on the one hand, and its study of the
effectiveness of the method to optimize the rating system of the refinery. The
rest of the paper is structured as follows: Sect. 2 proposes an extension of the
3-D ICrA from [16]. In Sect. 3, we describes the assessment system in a refinery
installation and use the ICrA to optimize it. Finally, in Sect. 4, the obtained
results are compared with those obtained from PCA, SCA and KCA. Section 5
offers the conclusion and outlines aspects for future research.
2 An Extended Form of Three-Dimensional Intuitionistic

Fuzzy Intercriteria Analysis Method
In this section we extend the proposed form of 3-D ICrA in [16] following the
ideas of [7,18]. Let us a set of objects Op (1, ..., p, ..., n) is evaluated against a set
of criteria Cq (1, ..., q, ..., m) [6,16] in the index hg ∈ H for 1, ..., g, ...f , where H
is the third fixed scale and hg is its element. For example, index set H can be
interpreted as a time-scale. We obtain a 3-D IM [3]
⎧ ⎫
⎪
⎪ hg ∈ H O1 ··· Oi ··· Oj ··· On ⎪
⎪
⎪
⎪ .. .. . . . ⎪
⎪
⎪
⎪ . . . . . . . . . ⎪
⎪
⎪
⎪ . . . . . . . . ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎨ C k a C k ,O 1 ,hg
· · · aC k ,O i ,hg
· · · aC k ,O j ,hg
· · · aC k ,O n ,hg ⎪
⎬
A= .
. .
. . . .
. . . .
. . . .
. , (1)
⎪ . . . . . . . . ⎪
⎪
⎪ Cl ⎪
⎪
⎪
⎪ a · · · a · · · a · · · a ⎪
⎪
⎪
⎪
Cl ,O1 ,hg Cl ,Oi ,hg Cl ,Oj ,hg Cl ,On ,hg
⎪
⎪
⎪
⎪ .. .. .. .. .. .. .. .. ⎪
⎪
⎪
⎪ . . . . . . . . ⎪
⎪
⎩ ⎭
Cm aCm ,O1 ,hg · · · aCm ,Oi ,hg · · · aCm ,Oj ,hg · · · aCm ,On ,hg
where for every p, q (1 ≤ p ≤ m , 1 ≤ q ≤ n):
(1) Cp is an evaluation criterion; Oq is an evaluated object

(2) aCp ,Oq ,hg is the evaluation of the p-th object against the q-th criterion in
a time-moment hg , defined as a real number or an IF pair (IFP, [8]), that
is comparable about relation R with the other a-objects, so that for each
Optimization of an Oil Refinery Valuation System 1557
i, j, k, g: R(aCk ,Oi ,hg , aCk ,Oj ,hg ) is defined. Let R be the dual relation of R.
Let us suppose that some of the values of A are omitted (denoted by ⊥).
Following [7,18], let the numbers αo , β o , α⊥ , β ⊥ be fixed and let them satisfy
the inequalities
0 ≤ αo ≤ 1, 0 ≤ β o ≤ 1, 0 ≤ α⊥ ≤ 1,
0 ≤ β ≤ 1, 0 ≤ αo + β o ≤ 1, 0 ≤ α⊥ + β ⊥ ≤ 1.
⊥
Let for each index hg (1, ..., g, ..., f ):

μ
– Sk,l,g be the number of cases in which R(aCk ,Oi ,hg , aCk ,Oj ,hg ) and R(aCl ,Oi ,hg ,
aCl ,Oj ,hg ) are satisfied and the a-elements are different from ⊥;
ν
– Sk,l,g be the number of cases in which R(aCk ,Oi ,hg , aCk ,Oj ,hg ) and
R(aCl ,Oi ,hg , aCl ,Oj ,hg ), are satisfied and the a-elements are different from ⊥;
⊥
– Sk,l,g be the number of cases in which R(aCk ,Oi ,hg , aCk ,Oj ,hg ) and
R(aCl ,Oi ,hg , aCl ,Oj ,hg ) are satisfied, because at least one of the elements
aCk ,Oi ,hg and aCk ,Oj ,hg or aCl ,Oi ,hg and aCl ,Oj ,hg is equal to ⊥ (is empty);
o
– Sk,l,g be the number of cases in which aCk ,Oi ,hg = aCk ,Oj ,hg or aCl ,Oi ,hg =
aCl ,Oj ,hg and all a-elements are different from ⊥.
Let us define
π ⊥ o n(n − 1)
Sk,l,g = Sk,l,g + Sk,l,g ,N = , then
2
μ ν π
Sk,l,g + Sk,l,g + Sk,l,g = Np .
For every k, l, g, such that 1 ≤ k < l ≤ m, n ≥ 2 and g is fixed [16], we define
μ
Sk,l,g αo Sk,l,g
o
α⊥ Sk,l,g
⊥ ν
Sk,l,g β o Sk,l,g
o
β ⊥ Sk,l,g
⊥
μk,l,g = + + , νk,l,g = + + . (2)
Np Np Np Np Np Np
Hence,
μ
Sk,l,g ν o ⊥
Sk,l,g Sk,l,g Sk,l,g
μk,l,g + νk,l,g ≤ + + + ≤ 1.
Np Np Np Np
Follow the ideas

from [7], let us propose the following approach as follows:
let α, β ∈ 0, 14 be two fixed numbers, so that α + β ≤ 12 and let αo = α⊥ =
α, β o = β ⊥ = β. Then
μ
Sk,l,g o ⊥ ν o ⊥
αSk,l,g αSk,l,g Sk,l,g βSk,l,g βSk,l,g
μk,l,g = + + , νk,l,g = + + .
Np Np Np Np Np Np
Let us construct the IM R = [K, K, H, {aki ,lj ,hg }] that determines the degrees
of correspondence between criteria C1 , . . . , Cm and its elements aki ,lj ,hg =
μCi ,Cj ,hg , νCi ,Cj ,hg (1 ≤ i, j ≤ m, 1 ≤ g ≤ f ) are calculated according to (2).
Let us apply an aggregation operation α(H,#∗p ) (R, h0 ) = Ragg (1 ≤ p ≤ 10),
proposed in [17] to the 3-D IM R = [K, K, H, {aki ,lj ,hg }] (K, H ⊂ I ∗ ) and
let h0 ∈ / H. A set of the operations #p , (1 ≤ p ≤ 10) from [17] has used in

α(H,#∗p ) (R, h0 ) for generating a 10-element scale for which we can denote, e.g.,
#∗1 - super pessimistic forecast of the ICrA correlation coefficient values, etc.,
#∗10 - super optimistic forecast of the correlation coefficients. The last step of the
algorithm is to determine the degrees of correlation between groups of indicators
depending of the chosen thresholds for μ and ν. The ICrA pairs are calculated
μ
using the software described in [12]. The 3-D ICrA works with two IMs Ragg
ν μ
and Ragg , rather than with the IM Ragg . The IM Ragg contains as elements the
ν
first components of the IFPs of R, while Ragg contains the second components
of the IFPs of Ragg . The best ICrA correlations are selected according to (3):
x > y iff a > c and b < d; x ≥ y iff b ≤ d; x ≥2 y iff a ≥ c

(3)
x = y iff a = c and b = d; x ≥R y iff Ra,b ≤ Rc,d ,
where Ra,b = 0, 5(2 − a − b)0, 5(|1 − a| + |b| + |1 − a − b|) [15].

The set of criteria can be reduced by determining the degrees of correlations
between the criteria using the scale in [5] with values of γ = 0, 85 and δ = 0, 15.
3 An Application of the 3-D ICrA to the Refinery

Evaluation System
In this section, the proposed 3-D ICrA method is applied to data containing
values of the employees’s evaluation criteria at a catalytic oil refinery installation
over a three year period. The descriptions of the 13 evaluation criteria are:
Criterion C1 – knowledge of the technological schemes and materials;

Criterion C2 – proper use of the production and the technical equipment repair
equipment according to the instructions;
Criterion C3 – compliance with workplace instructions when preparing for start-
up and shutdown of the installation;
Criterion C4 – knowledge, keeping and observance of internal company docu-
mentation according to the requirements for the position occupied;
Criterion C5 – performing activities on the flooring during the repair work in
accordance with the established procedure;
Criterion C6 – observance of the rules of the occupational safety and protection
Criterion C7 – observe employment discipline and maintain cleanliness in the
workplace and surrounding area;
Criterion C8 – ability to respond quickly to critical situations; expeditiousness;
Criterion C9 – loyalty to the company/affection/;
Criterion C10 – teamwork and collaboration skills.
Criterion C11 – striving for career growth and self-improvement;
Criterion C12 – quality and efficiency of work;
Criterion C13 – knowledge and understanding of corporate norms and standards.
The assessments of the employees O = {O1 , O2 , . . . , On } by criteria from the

set C = {C1 , C2 , . . . , C13 } in a time-moment hg ∈ H for 1 ≤ g ≤ 3 are elements
of a 3-D IM A[C, O, H] with a structure such as (1). The results, obtained from
μ ν
the ICrA software [12], are in the form of two IMs Ragg and Ragg (see Fig. 1, 2),
containing, respectively, the membership and the non-membership parts of the
IFPs detected between each pair of criteria. The results of the ICrA are plotted
onto the IF triangle in Fig. 3. Table 1 describes the type of correlations, obtained
between the pairs of criteria follow the scale from [5]:
Fig. 1. Membership parts of the IFPs, giving the ICrA correlations.
Fig. 2. Non-membership parts of the IFPs, giving the ICrA correlations.

Fig. 3. The results of ICrA, plotted onto the IF interpretational triangle.
Table 1. Correlations between the pairs of criteria.
Type of correlation Pairs of criteria

Strong positive consonance [0, 95; 1] 2-3
Positive consonance [0, 85; 0, 95) –
Weak positive consonance [0, 75; 0, 85) 2-6, 3-6
Weak dissonance [0, 67; 0, 75) 5-7
Dissonance [0, 57; 0, 67) 1-7, 1-8, 4-12, 4-10, 4-8, 5-12, 6-7, 7-12
Strong dissonance [0, 43; 0, 57) 1-4, 1-5, 1-9, 1-10, 1-11, 1-12, 1-13, 2-4,
2-7, 3-4, 3-7, 3-9, 3-10, 3-13, 4-5, 4-6, 4-7,
4-9, 4-11, 5-6, 5-8, 5-9, 5-10, 5-13, 6-9,
6-10, 6-11, 6-13, 7-8, 7-9, 7-10, 7-11, 7-13,
8-9, 8-10, 8-11, 8-12, 8-13, 9-10, 9-12,
9-13, 10-11,10-12, 10-13, 11-13
Dissonance [0, 33; 0, 43) 1-13, 1-2, 1-3, 1-6, 2-5, 2-11, 2-12, 3-5,
3-11, 3-12, 4-13, 5-11, 6-8, 6-12, 9-11,
11-12, 12-13
Weak dissonance [0, 25; 0, 33) 2-8, 3-8
Weak negative consonance [0, 15; 0, 25) –
Negative consonance [0, 15; 0, 05) –
Strong negative consonance [0, 05; 0] –
The following conclusions can be obtained:
a) There is strong positive consonance between the criteria C2 “proper use of

production and technical equipment” and C3 “compliance with workplace
instructions” indicating that one of the two factors could be removed from
the refinery’s staff assessment system. It is more appropriate to remove cri-

terion C2 from the assessment system because it requires the proper use of
production and technical equipment that can not be realized without knowl-
edge of technological schemes, materials and products.
b) The indicators C2 and C3 are in a weak positive consonance with C6 “obser-
vance of the rules of occupational safety. The correlations between indicators
C5 and C7 are in a weak dissonance. The criterion C8 depends on C1 and C4 .
c) The criterion C9 “loyalty to the company” depends on C5 “performing activi-
ties on the flooring during the repair work in accordance with the established
procedure”, C8 “ability to respond quickly to critical situations” and C10
“teamwork and collaboration skills”. The criterion C12 “Quality and efficiency
of work” depends on C4 , C5 and C7 .
4 Comparative Analysis of the Results Obtained After

ICrA, PCA, SCA and KCA
In this section is performed a comparative analysis of the correlation dependen-

cies between the evaluation criteria in the refinery after application of ICrA,
PCA, SCA and KCA [11]. Let confidence interval for correlation is 95%. The
Table 2 describes the strongest correlations between the criterion pairs. It can
be seen from the Table 2 that the strongest correlation dependencies between
the criterion pairs, obtained after application of four statistical methods almost
coincide. The observed differences in the established correlation dependencies
between the criteria are due to high hesitancy degree π of the indicated intu-
itionistic fuzzy ICrA pairs:
– according to the ICrA, the criterion C11 is in the strongest correlation relation
with C1 3 (μ(C11 , C13 ), ν(C11 , C13 ) = 0, 52; 0, 09; according to the PCA,
SCA and KCA, the criterion C11 is in the strongest correlation relation 0,70
with C10 ; the correlation coefficient between C11 and C13 is equal respectively
to 0,16, 0,24 and 0,22 according to the PCA, SCA and KCA;
– according to the ICrA, the criterion C13 is in the strongest correlation rela-
tions with the criteria C2 and C3 , and μ(C13 , C2 ), ν(C13 , C2 ) = μ(C13 , C3 ),
ν(C13 , C3 ) =0, 54; 0; according to the PCA, SCA and KCA, C13 is in the
strongest correlation relation respectively 0,75; 0,56 and 0,63 according to
the PCA, SCA and KCA with C5 ; the correlation between C13 and C2 (C3 )
is equal respectively to 0,38; 0,37 and 0,38 according to the PCA, SCA and
KCA.
Table 2. The strongest correlations dependencies between the criterion pairs according
ICrA, PCA, SCA and KCA.
Criterion ICrA PCA SCA KCA

C1 C8 C8 C8 C8
C2 C3 C3 C3 C3
C3 C2 C2 C2 C2
C4 C10 C10 C10 C10
C5 C7 C7 C12 , C7 C7
C6 C3 C3 C3 C3
C7 C5 C5 C5 C5
C8 C1 C1 C1 C1
C9 C8 C8 C8 C5 , C8
C10 C4 C4 C4 C4
C11 C13 C10 C10 C10
C12 C7 C5 , C7 C5 , C7 C5 , C7
C13 C2 , C3 C5 , C4 C4 , C5 C4 , C5
5 Conclusion
In the paper the extended form of 3-D ICrA is introduced and applied to ana-
lyze a dataset gathered through the staff evaluations of a catalytic oil refinery
installation in Bulgaria over a three-year period, that aims to detect depen-
dencies between criteria. The comparative analysis of the results obtained after
the application of the classic statistical rank correlation analyzes to the real
data, gathered through the human resource assessments of a catalytic oil refin-
ery installation, illustrates that there is no comparative difference between those,
obtained from the ICrA and the other three classic correlation analyzes, but only
the ICrA can be applied to the fuzzy data. In the future study, the outlined app-
roach for ICrA, assisting the decision-making process, will be expanded and
applied to retrieve information to other types of multi-dimensional data [4,16].
References
1. Atanassov, K.: Intuitionistic fuzzy sets. In: Proceedings of VII ITKR’s Session,
Sofia (1983). (in Bulgarian)
2. Atanassov, K.: Generalized index matrices. C. R. Acad. Bulg. Sci. 40(11), 15–18
(1987)
3. Atanassov, K.: Index Matrices: Towards an Augmented Matrix Calculus. Studies
in Computational Intelligence, vol. 573, Springer, Cham (2014)
4. Atanassov, K.: n-Dimensional extended index matrices Part 1. Adv. Stud. Con-
temp. Math. 28(2), 245–259 (2018)
5. Atanassov, K., Atanassova, V., Gluhchev, G.: InterCriteria analysis: ideas and
problems. Notes Intuitionistic Fuzzy Sets 21(1), 81–88 (2015)
6. Atanassov, K., Mavrov, D., Atanassova, V.: InterCriteria decision making. A new
approach for multicriteria decision making, based on index matrices and intuition-
istic fuzzy sets. Issues IFSs Generalized Nets 11, 1–8 (2014)
7. Atanassov, K., Marinov, P., Atanassova, V.: InterCriteria analysis with Interval-
valued Intuitionistic fuzzy evaluations. In: Cuzzocrea, A., Greco, S., Larsen, H.,
Saccà, D., Andreasen, T., Christiansen, H. (eds.) FQAS 2019. LNCS, vol. 11529,
8. Atanassov, K., Szmidt, E., Kacprzyk, J.: On intuitionistic fuzzy pairs. Notes Intu-
itionistic Fuzzy Sets 19(3), 1–13 (2013)
9. Atanassova, V., Doukovska, L., Kacprzyk, J., Sotirova, E., Radeva, I., Vassilev, P.:
ICrA of the global competitiveness reports: from efficiency-to innovation-driven
economies. J. Multiple-Valued Logic Soft Comput. 31(5–6), 469–494 (2018)
10. Bureva, V., Sotirova, E., Atanassova, V., Angelova, N., Atanassov, K.: Intercriteria
Analysis over Intuitionistic Fuzzy Data. LNCS, vol. 10665, pp. 333–340. Springer,
Cham (2018)
11. Doane, D., Seward, L.: Applied Statistics in Business and Economics. McGraw-Hill
Education, New York (2016)
12. Ikonomov, N., Vassilev, P., Roeva, O.: ICrAData - software for InterCriteria anal-
ysis. Int. J. Bioautomation 22(1), 1–10 (2018). https://doi.org/10.7546/ijba.2018.
22.1.1-10
13. Roeva, O., Fidanova, S., Paprzycki, M.: InterCriteria analysis of ACO and GA
hybrid algorithms. Stud. Comput. Intell. 610, 107–126 (2016)
14. Sotirov, S., Atanassova, V., Sotirova, E., et al.: Application of the intuitionistic
fuzzy InterCriteria analysis method with triples to a neural network preprocessing
procedure. Comput. Intell. Neurosci. 2017, 1–9 (2017)
15. Szmidt, E., Kacprzyk, J.: Amount of information and its reliability in the ranking of
Atanassov intuitionistic fuzzy alternatives. In: Rakus-Andersson, E. (eds.) Recent
Advances in Decision Making. SCI, vol. 222, pp. 7–19. Springer, Heidelberg (2009)
16. Traneva, V., Tranev, S., Szmidt, E., Atanassov, K.: Three dimensional ICrA over IF
data. In: Kacprzyk, J. (ed.) Advances in Fuzzy Logic and Technology. Advances in
Intelligent Systems and Computing, vol. 643, pp. 442–449. Springer, Cham (2017)
17. Traneva, V., Tranev, S., Stoenchev, M., Atanassov, K.: Scaled aggregation opera-
tions over 2- and 3-dimensional IMs. Soft. Comput. 22(15), 5115–5120 (2018)
18. Traneva, V., Tranev, S.: A multidimensional intuitionistic fuzzy InterCriteria anal-
ysis in the restaurant. JIFS (2020, in press)
Double Edge–Vertex Domination
Bünyamin Şahin1(B) and Abdulgani Şahin2

1
Department of Mathematics, Faculty of Science,
Selçuk University, 42130 Konya, Turkey
[email protected]
2
Department of Mathematics, Faculty of Science and Letters,
Agri Ibrahim Cecen University, 04100 Ağrı, Turkey
[email protected]
Abstract. An edge e dominates a vertex v if e is incident to v or e is

incident to a vertex which is adjacent to v. A subset D ⊆ E is an edge-
vertex dominating set (in simply, ev-dominating set) of G, if every vertex
of a graph G is ev-dominated by at least one edge of D. The minimum
cardinality of an ev-dominating set is named as ev-domination num-
ber and denoted by γev (G). In this paper, we study double edge-vertex
domination where a subset D ⊆ E is a double edge-vertex dominating
set (in simply, double ev-dominating set) of G, if every vertex of V is
ev-dominated by at least two edges of D. The double ev-domination
number of a graph G is denoted by γdev (G) and it is equal to the min-
imum cardinality of a double ev-dominating set. We first obtain some
relations between double edge-vertex domination and other domination
parameters. We also investigate the relation between γdev (G) and γev (G).
Finally, we determine the double ev-domination number of paths and
cycles.
Keywords: Domination · Edge-vertex domination · Double

edge-vertex domination
1 Introduction
Domination is a growing topic in graph theory in the last years. For example,
it is used to model the RNA sequence [1], electric power networks [2], chemi-
cal materials which are used in drug chemistry [3] and distribution centers in
logistics [4]. Domination is also used to model the defence strategy of Roman
Empire which is called Roman domination problem [5]. Moreover domination is
a very useful tool in computer science for investigation of the complexity prob-
lem [6,7]. After the definition of the total version of the edge-vertex domination
in the recent [8], in this paper we study a new domination invariant which is
called double edge-vertex domination. We want to make a new contribution to
domination problem in graphs.
Let G = (V, E) be a simple, connected graph whose vertex set V and the
edge set E. For the open neighbourhood of a vertex v in a graph G, the notation
Double Edge–Vertex Domination 1565
NG (v) is used as NG (v) = {u|uv ∈ E(G)} and the closed neighborhood of v

is used as NG [v] = NG (v) ∪ {v}. The degree of a vertex v ∈ G is equal to the
number of vertices which are adjacent to this vertex and denoted by dG (v). If
the degree of a vertex is one, it is called a leaf. If a vertex is adjacent to a leaf, it
is called a support vertex. If a support vertex is adjacent to a leaf, it is called a
weak support vertex and if a support vertex is adjacent to at least two leaves, it
is called a strong support vertex. If an edge is incident to a leaf, it is named an
end edge. We denote a path, a cycle and a star of order n by Pn , Cn and S1,n−1
respectively. The diameter of a tree T is denoted by diam(T ).
A subset D ⊆ V is a dominating set, if every vertex in G either is an element
of D or is adjacent to at least one vertex in D. The domination number of a graph
G is denoted by γ(G) and it is equal to the minimum cardinality of a dominating
set in G. By a similar definition, a subset D ⊆ V is a total dominating set, if
every vertex of G has a neighbor in D. The total domination number of a graph
G is denoted by γt (G) and it is equal to the minimum cardinality of a total
dominating set in G [9]. A subset D ⊆ V is a 2-dominating set if every vertex of
V −D has at least two neighbors in D. The 2-domination number of a graph G is
denoted by γ2 (G) and it is equal to the minimum cardinality of a 2-dominating
set in G [10]. Fundamental notions of domination theory are outlined in the
books [10,11].
A vertex v ve-dominates an edge e which is incident to v and any edge which
is adjacent to e. A set D ⊆ V is a ve-dominating set if every edge of a graph
G is ve-dominated by at least one vertex of D. The minimum cardinality of a
ve-dominating set is called ve-domination number and denoted by γve (G).
An edge e dominates a vertex v if e is incident to v or e is incident to a vertex
which is adjacent to v. A subset D ⊆ E is an edge-vertex dominating set (in
simply, ev-dominating set) of G, if every vertex of a graph G is ev-dominated
by at least one edge of D. The minimum cardinality of an ev-dominating set is
named ev-domination number and denoted by γev (G).
Vertex-edge domination and edge-vertex domination were introduced by
Peters [12] and further studied by Lewis [13]. An improved upper bound of edge-
vertex domination number of a tree was obtained in [14] and trees with total
domination number equal to the edge-vertex domination number plus one were
studied in [15]. Double version of the vertex-edge domination was introduced and
studied by Krishnakumari et al. in [16]. A subset D ⊆ V is a double vertex-edge
dominating set (in simply, double ve-dominating set) of G, if every edge of E is
ve-dominated by at least two vertices of D. The double ve-domination number
of a graph G is denoted by γdve (G) and it is equal to the minimum cardinality
of a double ve-dominating set.
In this paper, we study double edge-vertex domination. A subset D ⊆ E is
a double edge-vertex dominating set (in simply, double ev-dominating set) of
G, if every vertex of V is ev-dominated by at least two edges of D. The double
ev-domination number of a graph G is denoted by γdev (G) and it is equal to the
minimum cardinality of a double ev-dominating set.
1566 B. Şahin and A. Şahin
In Sect. 2, firstly, it is shown that for every tree T , γt (T ) ≤ γdev (T ), γdve (T ) ≤

γdev (T ) and for every non-trivial, connected graph G, γ2 (G) ≤ γdev (G). In
Sect. 3, it is proved that for every simple, connected graph G, γdev (G) ≥
γev (G)+1. Moreover, the trees attaining the equations γdev (G) = γev (G)+1 and
γdev (G) = γev (G) + 2 are obtained. Finally, the double ev-domination number
of paths and cycles is determined and it is shown that γdve (Pn ) = γt (Pn ) for the
paths in the last part of Sect. 3.
2 Some Relations Between Domination Parameters
Observation 1 ([16]). For every connected graph G with diam(G) ≥ 3, there

is a γev (G)-set containing no end edge.
Observation 2 ([16]). Each support vertex of a graph G is contained in every

total dominating set.
Observation 3 ([16]). For every connected graph G with diam(G) ≥ 3, there

is a γt (G)-set containing no leaf.
Lemma 1 ([15]). For every triangle-free graph G without isolated vertices,

γdve (G) ≤ γt (G).
Lemma 2 ([16]). If T is a tree, then γev (T ) ≤ γt (T ).
It is indicated that the inequality in Lemma 2 is attained for bipartite graphs

in the following lemma.
Lemma 3. If G is a non-trivial, connected bipartite graph, then γev (G) ≤ γt (G).
Proof. Let P and S be the partite sets of G and let D be a total dominating set
of G. Assume that DP = D ∩ P and DS = D ∩ S. Since D is a total dominating
set, each vertex of S has a neighbor in DP and each vertex of P has a neighbor
in DS . The edges between these adjacent vertices are an ev-dominating set of G
and γev (G) ≤ γt (G).

There is no double edge-vertex dominating set in a graph which has one edge,
so we are interested the trees which have at least two edges.
Observation 4. Every end edge which is incident to a weak support vertex (with
degree two) of a tree T is contained in every γdev (T )-set.
Proof. Let D be a γdev (T )-set and uv be an end edge of T which has a weak
support vertex u with degree two. By Observation 1, there is a γev (T )-set which
contains no end edge. But in this case, the vertex v is ev-dominated once. In
order to doubly ev-dominate the vertex v, uv has to be contained in D.

Proposition 1. For every tree T , γt (T ) ≤ γdev (T ).

Proof. Let D be a γdev (T )-set. Let x be a vertex which is ev-dominated by the

edges xy, yz ∈ D. Clearly, x is ev-dominated by twice. Consider S as a γt (T )-set.
Furthermore, x is dominated by y and y is adjacent to z. It can be accepted that
the vertices y, z ∈ S. Thus, it is obtained that |S| ≤ |D|.

The relation in Proposition 1 is not extended to every graph. For example,

for the cycles G = Cn of order n, γdev (G) ≤ γt (G).
Proposition 2. For every tree T , γdve (T ) ≤ γdev (T ).
Proof. It is known that γdve (G) ≤ γt (G) by Lemma 1 and γt (T ) ≤ γdev (T ) by

Proposition 1. Thus, it is yielded that γdve (T ) ≤ γdev (T ).

In [15], the authors obtained a lower bound for double ve-domination number
of trees such that γdve (T ) ≥ (n − l − s + 4)/2 with order of n, l leaves and s
support vertices. Since γdve (T ) ≤ γdev (T ), this lower bound is a natural lower
bound for double ev-domination number of trees.
The differences γdev (T )−γt (T ) and γdev (T )−γdve (T ) can be arbitrarily large.
Let G be a graph obtained by adding a new vertex to every leaf of a star S1,n .
For this graph G, γt (G) = n + 1, γdve (G) = n + 1 and γdev (G) = 2n.
Proposition 3. For every triangle-free, connected graph G, γ2 (G) ≤ γdev (G).
Proof. Let D be a γdev (G)-set. It is clear that every vertex of G is ev-dominated

by at least two edges of D. This means that such a vertex has at least two
vertices on these edges. Since G is triangle-free, these neighbor vertices do not
lie on the same edge. Thus γ2 (G) ≤ γdev (G).

The bound of Proposition 3 is sharp for cycles.
3 Bounds
For every simple, connected graph with n ≥ 3 vertices, 2 ≤ γdev (G) ≤ n − 1. The
lower bound is attained for complete graphs and the upper bound is attained for
P3 , P4 , P5 . We will improve the lower bound for trees in the following subsection.
3.1 Lower Bounds
It is clear that γev (G) < γdev (G) for every graph. Note that an edge e ∈ H ⊆ E
has a private vertex v (to the subset H), if the vertex v is the unique vertex ev-
dominated by e. Now, a proposition will be stated for a minimal ev-dominating
set about the private vertex of an edge.
Proposition 4. Let D be an ev-dominating set of a simple, connected graph G.

Then D is a minimal ev-dominating set if and only if every edge e of D has at
least one private vertex with respect to D.
Proof. Assume that D is a minimal ev-dominating set of G. Then for every edge
e ∈ D, D \ {e} does not ev-dominate the graph G. Thus, there is a vertex v
which is not ev-dominated by D \ {e}. It means that v is a private vertex of e
with respect to D. For the reverse, assume that D is an ev-dominating set of G
such that every edge of D has a private vertex with respect to D. If D is not
minimal, then there is an edge e ∈ D such that D \ {e} ev-dominates the graph
G. It implies that every vertex of G is ev-dominated by D \ {e}, contradicting
the property of D. Thus, D is a minimal ev-dominating set.

Proposition 5. For every graph G which has at least two edges, γdev (G) ≥
γev (G) + 1.
Proof. Assume that γdev (G) = γev (G). Let D be a γdev (G)-set. D is an ev-
dominating set in the same time. By Proposition 4, every edge of D has a private
vertex with respect to D. But these private vertices are ev-dominated once by
D, is a contradiction to the fact that D doubly ev-dominates V (G).

The bound γdev (G) ≥ γev (G) + 1 is sharp for cycles C3 , C4 , C5 such that
γdev (C3 ) = 2, γev (C3 ) = 1; γdev (C4 ) = 2, γev (C4 ) = 1 and γdev (C5 ) = 3,
γev (C5 ) = 2.
Proposition 6. Let T be a tree. Then, γdev (T ) = γev (T ) + 1 if and only if T is

a star.
Proof. If T is a star, then γdev (T ) = 2 and γev (T ) = 1. Thus, γdev (T ) = γev (T )+1.
Prove the necessity. Let T be a tree with γdev (T ) = γev (T ) + 1. It will be shown
that the diameter of T is two. On the contrary assume that diam(T ) = d ≥ 3.
Firstly, consider a tree H with diameter three. Then, γev (T ) = 1 and γdev (T ) =
3. Thus, it can be assumed that d ≥ 4. Let x0 − x1 − x2 − · · · − xd−1 − xd be a
diametrical path in T . It can be seen easily that x0 and xd are leaves. In order to
doubly ev-dominate the vertex x0 , D contains the edges x0 x1 and x1 x2 . Similarly,
D contains the edges xd−2 xd−1 and xd−1 xd to doubly ev-dominate the vertex
xd . Since d ≥ 4 the edge xd−2 xd−1 ∈ / {x0 x1 , x1 x2 }. Thus D − {x0 x1 , xd−2 xd−1 }
is an ev-dominating set of T . It means that γev (T ) ≤ D − 2 = γdev (T ) − 2, a
contradiction. So T has diameter only two and therefore T is a star S1,n−1 .
By the last two propositions, the next corollary can be stated.
Corollary 1. If T is a tree with diam(T ) ≥ 3, then γdev (T ) ≥ γev (T ) + 2.
Now the trees attaining the equality γdev (T ) = γev (T ) + 2 are investigated.
The caterpillar graphs will be used for this objective. Let G be a caterpillar
graph with vertex set {v1 , v2 , . . . , vk } of a path and the number of the leaves
are denoted by m1 , m2 , . . . , mk which are attached to the vertices v1 , v2 , . . . , vk
respectively. G is denoted by G = C(m1 , m2 , . . . , mk ).
Proposition 7. If T is a tree with γdev (T ) = γev (T ) + 2 if and only if T =

C(m1 , m2 , . . . , mk ) where k ∈ {2, 3, 4, 5}, m1 , mk ≥ 1 and m2 = m3 = m4 = 0
if k = 5.
Proof. Let T be a tree attaining the equality γdev (T ) = γev (T ) + 2. By Corol-

lary 1, T has diameter at least three. Let diam(T ) ≥ 3 and x0 − x1 − x2 − · · · −
xd−1 − xd be a diametrical path in T . Thus x0 and xd are leaves and x1 and
xd−1 are their support vertices, respectively. Let D be a γdev (T )-set containing
a minimum number of leaves. D must contain the edges x0 x1 , x1 x2 , xd−2 xd−1
and xd−1 xd . Clearly, D − {x0 x1 , xd−2 xd−1 } is a minimum ev-dominating set of
T such that all leaves which are incident to x1 and xd−1 are ev-dominated by
x1 x2 and xd−2 xd−1 , respectively.
The aim is to show that every vertex (if exists) in V (T ) − V (P ) is adjacent
to a vertex of P . On the contrary assume that there are vertices in V (T ) − V (P )
not to be adjacent to vertices of P . Let uv ∈ E(T ) − E(P ) be a leaf at maximum
distance from P and vy be a parent edge of uv. Because of the choosing of
D, uv ∈ / D and vy ∈ D. To doubly ev-dominate u, D has another neighbor
of vy. Clearly, since P is a diametrical path, vx1 and vxd−1 ∈ / E(T ). Thus,
D − {x0 x1 , xd−1 xd , uv} is an ev-dominating set of T with size γdev (T ) − 3, a
contradiction. Consequently, it can be seen that every vertex (if exists) not in P
is a leaf and T is a caterpillar graph with the path x1 x2 . . . xd−2 xd−1 .

Assume that D = D − {x0 x1 , xd−1 xd } contains more than two edges. It

implies that D is a minimum ev-dominating set containing x1 x2 and xd−2 xd−1 .

By Proposition 4, every edge of D has a private vertex with respect to D .

If u is a private vertex of the edge xk−1 xk with respect to D , then u is ev-

dominated by D once, is a contradiction. It implies that D has two edges and
D = {x0 x1 , x1 x2 , xd−2 xd−1 , xd−1 xd }. It is deduced that T has diameter at most
six. If d ∈ {3, 4, 5}, then T = C(m1 , m2 , . . . , md−1 ) with m1 , mk ≥ 1 and mi ≥ 0
for i ∈
/ {1, d − 1}. Finally, assume that d = 6. If x3 is a support vertex, then the
leaf which is adjacent to x3 is not ev-dominated by x1 x2 and x4 x5 . Thus, x3 has
degree two. Furthermore, if x2 has a leaf u, u is ev-dominated by x1 x2 . But to
doubly ev-dominate u, the edge x2 u is contained by D, so the difference is that
γdev (T ) − γev (T ) > 2 in this case. Thus, x2 has degree two. The same approach is
current for x4 . Therefore, if d = 6, T = C(m1 , 0, 0, 0, m5 ) such that m1 , m5 ≥ 1.
The converse is trivial. This completes the proof.

The next Corollary is a direct result of Proposition 6 and Proposition 7.
Corollary 2. If T is a tree with diam(T ) ≥ 7, then γdev (T ) ≥ γev (T ) + 3.
Lemma 4 ([15]). For the paths Pn and cycles Cn with n vertex,

i)
n/2 + 1, n ≡ 2 (mod 4)
γdve (Pn ) = . (1)
n/2, otherwise
ii)
γdve (Cn ) = n/2. (2)
Lemma 5 ([9]). For the paths graphs Pn with n vertex,

n/2, n ≡ 0 (mod 4)
γt (Pn ) = . (3)
n/2 + 1, otherwise
Proposition 8. For the path graphs Pn with n vertex, γdve (Pn ) = γt (Pn ).
Proof. Lemma 4 and Lemma 5 will be used for proving and the proof will be
investigated three cases.
Case 1. Assume that n ≡ 2 (mod 4). Thus, it can be taken as n = 4k + 2 for
k ≥ 0. Therefore,
γdve (Pn ) = 2k + 2 = γt (Pn ). (4)
Case 2. Assume that n ≡ 0 (mod 4). Thus, it can be taken as n = 4k for k ≥ 1.

Therefore,
γdve (Pn ) = 2k = γt (Pn ). (5)
Case 3. Assume that n ≡ 1 or 3 (mod 4). Thus, n is an odd number in this

case. It can be taken as n = 4k ± 1.
Let n = 4k + 1 for k ≥ 1,
γdve (Pn ) = 2k + 1 = γt (Pn ). (6)
Let n = 4k − 1 for k ≥ 1,
γdve (Pn ) = 2k = γt (Pn ) (7)
and this completes the proof.

The double ev-domination number of paths and cycles is obtained in the next
proposition.
Proposition 9. For the paths Pn and cycles Cn with n vertex,
i)
n/2 + 1, if n ≡ 1 (mod 4)
γdev (Pn ) = . (8)
n/2 + 1, otherwise
ii)
γdev (Cn ) = n/2. (9)
Proof.
i) By using Proposition 2 and Lemma 4, it is obtained that γdve (Pn ) = n/2+1 ≤
γdev (Pn ) for the case n ≡ 2 (mod 4) and γdve (Pn ) = n/2 ≤ γdev (Pn ) for
the other cases. By Proposition 1 and Lemma 5, it is obtained that γt (Pn ) =
n/2 ≤ γdev (Pn ) for the case n ≡ 0 (mod 4) and γt (Pn ) = n/2+1 ≤ γdev (Pn )
for the other cases. Finally, γdev (Pn ) ≥ (n − l − s + 4)/2 = (n − 2 − 2 + 4)/2 =
n/2. Consequently, according to these results, it is obtained that γdev (Pn ) =
n/2 + 1 if n ≡ 1 (mod 4) and otherwise γdev (Pn ) = n/2 + 1.
ii) It is known that γdve (Cn ) = n/2 ≤ γdev (Cn ). Let G be a cycle with even
order and D be a γdev (G)-set. If the edges of D are chosen independently,
every vertex of G is a vertex of an edge e ∈ D. It means that every vertex
of G is ev-dominated twice. Thus, γdev (Cn ) = n/2, if n is even. Now assume
that G be a cycle with odd order. If the edges of D are chosen independently,
only a vertex u ∈ G is remaining at one distance from two elements of D.
Thus, each end vertices of these two edges is ev-dominated once and an edge
which is incident to u must be added to D. So γdev (Cn ) = n/2 and this
completes the proof.

4 Conclusion
The study deals with a graph theoretical problem of edge-vertex domination.
This kind of domination parameter is worth studying because of its connections
to the well studied total domination and the others. We conclude the paper
with some open problems. The trees T attaining the equalities γdev (T ) = γt (T ),
γdev (T ) = γ2 (T ), γdev (T ) = γdve (T ) can be characterized for future studies.
References
1. Haynes, T.-W., Knisley, D., Seier, E., Zou, Y.: A quantitive analysis of secondary
RNA structure using domination based parameters on trees. BMC Bioinform. 7,
Article number 108 (2006)
2. Haynes, T.-W., Hedetniemi, S.-M., Hedetniemi, S.-T., Henning, M.-A.: Domination
in graphs applied to electric power networks. SIAM J. Discret. Math. 15(4), 519–
529 (2002)
3. Ediz, S., Cancan, M.: On molecular topological properties of alkylating agents
based anticancer drug candidates via some ve-degree topological indices. Curr.
Comput. Aided Drug Des. 16(2), 190–195 (2020)
4. Desormeaux, W.-J., Haynes, T.-W., Hedetniemi, S.-T., Moore, C.: Distribution
centers in graphs. Discret. Appl. Math. 243, 286–293 (2018)
5. Cockayne, E.-J., Dreyer Jr., P.-A., Hedetniemi, S.-M., Hedetniemi, S.-T.: Roman
domination in graphs. Discret. Math. 278, 11–22 (2004)
6. Henning, M.-A., Pandey, A.: Alghoritmic aspects of semitotal domination in
graphs. Theoret. Comput. Sci. 766, 46–57 (2019)
7. Venkatakrishnan, Y.-B., Krishnakumari, B.: On the alghoritmic complexity of dou-
ble vertex-edge domination in graphs. In: Das, G.-K., et al. (eds.) WALCOM 2019.
LNCS, vol. 11355, pp. 188–198. Springer, Cham (2019)
8. Şahin, A., Şahin, B.: Total edge-vertex domination. RAIRO Theor. Inf. Appl.
54(1), 1–7 (2020)
9. Cockayne, E.-J., Dawes, R.-M., Hedetniemi, S.-T.: Total domination in graphs.
Networks 100, 211–219 (1980)
10. Haynes, T.-W., Hedetniemi, S.-T., Slater, P.-J.: Domination in Graphs: Advanced
Topics. Marcel Dekker, New York (1998)
11. Haynes, T.-W., Hedetniemi, S.-T., Slater, P.-J.: Fundamentals of Domination in
Graphs. Marcel Dekker, New York (1998)
12. Peters, J.-W.: Theoretical and algorithmic results on domination and connectivity.
Ph.D. thesis, Clemson University (1986)
13. Lewis, J.-R.: Vertex-edge and edge-vertex domination in graphs. Ph.D. thesis,
Clemson University (2007)
14. Venkatakrishnan, Y.-B., Krishnakumari, B.: An improved upper bound of edge-
vertex domination of a tree. Inf. Process. Lett. 134, 14–17 (2018)
15. Krishnakumari, B., Chellali, M., Venkatakrishnan, Y.-B.: Double vertex-edge dom-
ination. Discret. Math. Algoritmhs Appl. 9(4), 1750045, 11 p. (2017)
16. Krishnakumari, B., Venkatakrishnan, Y.-B., Krzywkowski, M.: On trees with total
domination number equal to edge-vertex domination number plus one. Proc. Indian
Acad. Sci.: Math. Sci. 126(2), 153–157 (2016)
Order Acceptance Scheduling on a Single
Machine with Energy Aspects
Mariam Bouzid(B) , Oussama Masmoudi, and Alice Yalaoui
ICD, LOSI, Université de technologie de Troyes,

12 Rue Marie Curie CS 42060, 10004 Troyes, France
{mariam.bouzid,oussama.masmoudi,alice.yalaoui}@utt.fr
Abstract. Order Acceptance Scheduling (OAS) is a two-fold decision

problem that consists in the sequencing of a subset of selected orders with
the objective of maximizing the total profit. In recent years, emphasis
focused on the reduction of energy consumption and carbon emissions.
Optimizing production schedules is a promising lever for this purpose.
In this context, we provide a mathematical formulation for the OAS
problem under energy aspects. To assess the performance of our model,
a comparative analysis is conducted. Our results clearly point out that
the arc-time-indexed model is better than the disjunctive one.
Keywords: Green manufacturing · Order Acceptance Scheduling ·

Energy
1 Introduction
With increasing concerns about environmental issues in the past decades − along
with stringent governments regulations (e.g. taxes) − the industrial sector has
reorganized its activities around sustainability [8]. These environmental and eco-
nomical aspects have been principally tackled by considering energy consump-
tion and greenhouse gases emissions (GHG) at every level of the supply chain [2].
Specifically, at the operational level − besides modernizing facilities − optimiz-
ing production schedules turns out to be the most cost-effective way for making
energy savings [10]. In this vein, recent works in green manufacturing take into
account the impact of various energy charging policies such as time-of-use (TOU)
tariffs [6].
This research presents an OAS problem on a single machine under electricity
TOU tariffs and taxed carbon emission periods. We propose an arc-time-indexed
formulation adapted from the article of Silva et al. [13] to solve the problem. A
comparative analysis for instances with different settings using the disjunctive
formulation in [4] and the proposed one is reported.
The remainder of this paper is organised as follows. Section 2 includes a
review on OAS problems and scheduling under electricity TOU tariffs. Section 3
states the problem and presents the solution approach. Section 4 is dedicated
to instances description and results. The last section concludes the paper and
draws some perspectives.
1574 M. Bouzid et al.
2 State of the Art

The standard OAS can be seen as an interrelated two-step decision problem. The
first step is to decide which orders should be accepted among a set of n orders.
At the second step, the selected orders are sequenced. As stated in [12], OAS
has gained attention from companies for the reason that integrating decision
making at operational level can improve the profits and resources use. Stern and
Avivi [14] proposed an OAS for a textile production where a subset of requested
fabric has to be selected in order to maximize the total profit while meeting the
due dates. In the literature, OAS problems are addressed for both single and
multi machines configuration. Furthermore, different jobs characteristics can be
integrated; among them, preemption, availability of the jobs and sequence depen-
dent setup-times. For instance, an OAS problem on a single machine with release
dates and deadlines is studied in [7]. Regarding our research, the immediately
related works are those presented in [3,11,13], which deal with a single machine
OAS with setup-times, release dates and deadlines.
To this day, energy aspects are increasingly taken into consideration in
scheduling problem for single machine and more complex systems. These aspects
are handled in multiple ways. Aghelinejad et al. [1] minimize energy cost by
exploiting on/off mechanisms on a single machine. Alternatively, for a job-shop
system, a power peak limitation is introduced while minimizing energy costs in
the work of Masmoudi et al. [9]. Pricing policies are also investigated to achieve
energy savings. TOU is a rate plan that charges electricity at higher cost during
predetermined intervals of the day, known as on-peak hours. Consequently, in
manufacturing companies, production times can be shifted from on-peak to off-
peak or mid-peak hours [15]. In the meantime, carbon policies are implemented
by government to pressure the companies to reduce their GHG emission. In this
research, the ToE (Tax on Emission) is deemed, which refers to a tax on the
amount of carbon emitted throughout the makespan [5]. In addition, carbon
emissions fluctuate during the day. Specifically, the latter are cheaper during
TOU on-peak periods and conversely during off-peak and mid-peak hours. This
leads to a trade-off between these two conflicting energetic aspects. To the best
of our knowledge, Chen et al. [4] are the first to examine a single machine OAS
along with TOU electricity tarrifs and CO2 emissions. Our work follows the
latter proposing a new formulation.
3 Problem Formulation and Solution Approach

The considered problem is a single non-preemptive OAS problem. Each order j
is defined by its processing time pj , due date dj , deadline d¯j , release date rj ,
revenue ej , power consumption Ωj and tardiness penalty wj . Besides, their setup-
times − denoted sij for each pair of orders i and j − are sequence-dependent,
i.e, they depend on the immediately previous sequenced order. The objective is
to maximize the total profit, knowing that exceeding the due date dj yields to
less incomes and that going beyond the deadline d¯j causes the order rejection.
This problem is known to be strongly N P-hard [3].
Order Acceptance Scheduling on a Single Machine with Energy Aspects 1575
Energy aspects are introduced in our formulation by the electricity TOU and
carbon emission periods as proposed in [4]. For this purpose, the time horizon
is divided into T = maxj {d¯j + pj } time slots which are characterized by an
electricity cost EC and an amount q of carbon emission per kWh. The latter
depends on a carbon tax per kg of emitted CO2 (T ax) and the amount of
electricity consumed. The environmental cost of orders are deducted from their
total profits. To solve the considered OAS problem, we propose an arc-time-
indexed formulation. This model relies on a discretization of time inspired in
the work of Silva et al. [13]. This approach allow us to represent the problem
as a directed graph in which nodes stand for specific instants of time t ∈ [0; T ]
at which orders start. The edges represent possible sequencing paths between
orders and their construction guarantee the satisfaction of problem constraints
along with their proper profits.
Formally, let J = {1, . . . , n} be the set of orders to be processed and J+ =
{0} ∪ J, the set of orders plus a dummy order 0. Let T = maxj∈J+ d¯j be the time
horizon. We define the binary decision variables xtij , for i, j ∈ J+ and t ∈ [0, T ]
that take the value one if the order j succeed directly to order i after its setup at
time t. Moreover, for each pair of orders i, j, the parameter pij takes the value
pi if i = j and 1 otherwise. The problem is formulated as a Mixed Integer Linear
Program (MILP), presented below.
d¯j −pj
sij pj

maximize xtij fjt − ct−t
j + ct+t
j (1)
j∈J i∈J+ t=ri +pi +sij t =1 t =1
i=j t−sij ≥rj
subject to
T

xt0j = 1 (2)
j∈J
t=rj +s0j
n d¯j −pj

xtij ≤ 1 (j ∈ J) (3)
i=0 t=ri +pi +sij
i=j t−sij ≥rj
t+pi +sij

xtji − xij − xt+1
ij = 0
j∈J+ j∈J+ j∈J+
t−pji −sji ≥rj j=i j=i
t+pi ≤d¯j
t+pi ≥rj t+1≥rj (4)
t+pi +sij +pj ≤d¯j t+1+pj ≤T
t−sji ≥ri
(i ∈ J, t ∈ [0; T ])

xtij ∈ {0, 1}, (i, j ∈ J+ , t ∈ ri + pi + sij ; d¯j − pj , t − sij ≥ rj ) (5)
The objective (1) is the maximization of the sum of the incomes of each order
j ∈ J, that is, the profit subtracted from the environmental cost. Constraint (2)
ensures that only one order starts the sequence. Constraints (3) specify that
each job can be processed at most once. Constraints (4) guarantee the flow
conservation for each order i ∈ J at any time t. Finally, constraint (5) gives the
domains of definition of the decision variables.
The profit of an order j ∈ J at starting time t is calculated as follows.

fjt = ej − wj max (0, (t + pj ) − dj ) (j ∈ J, t ∈ rj ; d¯j − pj )
For each order j ∈ J, we introduced the term ctj to calculate the environ-
mental cost at ending time t where 1x takes value 1 when condition x holds and
0 otherwise. Moreover, bk∈{0,...,m−1} and gl∈{0,...,h−1} correspond respectively to
the starting times of TOU and carbon emission periods:

m
h Ωj
ctj = Ωj ECk 1t≥bk−1 + T ax ql 1t≥gl−1 (j ∈ J, t ∈ rj ; d¯j − pj , Ωj = 60
)
k=1 t<bk l=1 t<gl
At last, the dummy order takes special values: p0 = r0 = 0, d¯0 = maxj∈J d¯j ,
d0 = maxj∈J dj , and ∀j ∈ J, sj0 = 0.
To conclude this section, the Fig. 1 presents an illustration of the impact
of introducing energy aspects (TOU) in OAS problem. In this example, three
orders have to be processed with p = [0, 4, 2, 2], r = [0, 0, 8, 6], d = [11, 4, 11, 9],
d¯ = [12, 7, 12, 10] and e = [0, 3, 2, 1]. We suppose negligible setup. In addition,
Ω = [1, 1, 1], b = [0, 5, 11, 15] and EC = [0.0, 1.0, 0.0, 0.0].
σ∗ 1 3 2
0 5 10 12
σ ∗ 1 2
Fig. 1. An example of optimal solutions for the OAS problem with (σ ∗ ) and without
(σ ∗ ) energy considerations. In the sequencing σ ∗ , order 3 is rejected because the TOU
tarrif at period 2 is higher than its profits and order 2 is shifted to a low-priced period.
4 Instances Description and Computational Results
The evaluated instances are provided in Chen et al. [4]. Instances are built with
diverse number of orders n = 10, 15, 20, 25, 50, 100. Moreover, orders parameters
are randomly generated. Processing times, setup-times and revenues are gen-
erated from the uniform distribution pj , ej ∼ U(0, 20) and sij ∼ U(1, 10). To
determine the other parameters, two coefficients are employed by [4]. The first
one is the tardiness factor τ ∈ {0.1, 0.5, 0.9}. The last one is the due date range
R ∈ {0.1, 0.5, 0.9}.
The values of the release dates are determined by rj = U(0, τ pT ) with pT =
j∈J pj . The due dates are defined as dj = rj +maxi∈J ; sij +max{slack, pi } with

slack ∼ U pT 1 − τ − R2 , pT 1 − τ + R2 . Finally, deadlines are computed
ej
with d¯j = dj + Rpj and tardiness penalties with wj = d¯j −d j
. Finally, to suit
the specificity of the problem, Chen et al. [4] incorporate the information of the
TOU tarrifs and carbon emissions.
We have implemented and solved the formulations in a commercial solver
(IBM CPLEX Optimization Studio v 12.9) on a desktop computer with processor
Intel i5 2 GHz CPU with 4 GB RAM. Solving time is limited to 3600 s.
Table 1 resumes the results obtained with various number of orders and values
of τ . For each values of the latter, there are 3 corresponding instances that match
with the values of due dates range R ∈ {0.1, 0.5, 0.9}. For all formulations, the
table presents the results in term of number of feasible (#fea) and optimal (#opt)
solutions found, average CPU time (cpu in seconds) and average gap (gap in %).
Note that the values of the gap are retrieved from the solver relative gap, i.e.
bestobj −bestint
10−1 +|bestint | , where bestobj is the upper bound found by the solver and bestint
the objective value of the best integer solution.
Table 1. Results of the comparative analysis for 45 instances.
n τ Disjunctive formulation Arc-time-indexed formulation

#fea #opt cpu gap #fea #opt cpu gap
10 0.1 3 3 111 0 3 3 1.1 0
0.5 3 3 3.9 0 3 3 0.39 0
0.9 3 3 0.20 0 3 3 0.16 0
15 0.1 3 2 1235 1.4 3 2 1215 0.28
0.5 3 3 845 0 3 3 3.0 0
0.9 3 3 0.32 0 3 3 0.33 0
25 0.1 3 0 3600 4.2 3 2 1357 0.21
0.5 3 0 3600 9.0 3 3 92.8 0
0.9 3 3 16.1 0 3 3 2.1 0
50 0.1 3 0 3600 15.7 3 0 3600 37.4
0.5 3 0 3600 16.4 3 0 3600 85.8
0.9 3 0 3600 15.0 3 3 79.9 0
100 0.1 3 0 3600 84 0 0 – –
0.5 3 0 3600 283.1 3 0 3600 98.8
0.9 3 0 3600 43.3 3 1 3567 12.7
Total 45 20 1957 27.7 42 29 1222 16.8
Our arc-time indexed model can find optimal solution up to n = 100 orders at
acceptable computational time, in particular when τ > 0.5. Besides, we note that
this formulation attains more often optimality (29 out of 42 solutions, meaning
69% instead of 44% for the disjunctive one), whereas the disjunctive formulation
is more effective to find good feasible solutions.
Table 2 presents the computational results for a set of instances with τ = 0.5,
R = 0.5 and different settings − with or without release date (rj ) or setup-times
(sij ). This is justified by the fact that a variety of OAS problems are considered
in the literature and thus, this comparative analysis assesses the genericity of
our approach.
Table 2. Comparision of formulation performances in terms of CPU time (s) and Gap
(%) for a set of instances with τ = 0.5 and R = 0.5 and different settings
n rj sij Disjunctive formulation Arc-time indexed formulation

CPU (s) Gap (%) CPU (s) Gap (%)
10 – – 0.13 0 0.11 0
– 0.08 0 0.11 0
5 0 0.25 0
– 23.2 0 0.39 0
15 – – 0.36 0 0.28 0
– 0.14 0 0.22 0
2015 0 1 0
– 3600 7.5 2 0
20 – – 24.8 0 6.1 0
– 3600 1.4 5.7 0
3600 2.1 1.6 0
– 3600 10.7 1.6 0
25 – – 3600 1.9 92.5 0
– 3600 2.3 21.6 0
3600 10.4 27.5 0
– 3600 13.7 227.2 0
50 – – 3600 5.1 3600 11.5
– 3600 5.5 3600 10.4
3600 12 3600 63.8
– 3600 16.5 3600 66.2
100 – – 3600 35.1 3600 83.6
– 3600 75 3600 74.3
3600 171 3600 116.5
– 3600 46.2 – –
The results indicate that our approach can be more effective than the dis-
junctive one for small to medium size instances with or without release dates.
However, the arc-time indexed formulation has a pseudo-polynomial number of
variables and constraints. As a consequence, within an hour the model cannot
find some feasible solution when n = 100.
Table 3 gives the number of variables and constraints of each formulation.

As can be seen, the arc-time-indexed formulation is disadvantaged by its huge
number of variables. Indeed, the number of orders n coupled with the great
granularity of the time horizon T can affect the computational time.
Table 3. Spatial complexity in the worst case for each formulation
#variables #constraints
Disjunctive formulation O(n2 ) + O(nm) + O(nh) O(n2 ) + O(nm) + O(nh)
Arc-time-indexed O(n2 T ) O(nT )
formulation
5 Conclusion and Perspectives

In this work we have used an arc-time-indexed formulation for the order accep-
tance scheduling problem considering energy aspects. For small instances, the
formulation seems to lead to less computational effort. Besides, this is a pur-
poseful choice for this problem since the objective is the maximization of a time-
dependent profit (i.e. subject to tardiness penalties and electricity cost varying
at each period) with sequence dependent setup-times.
Our undergoing work is focused on improving the evaluated model, and devel-
oping exact and heuristic methods for larger instances. For instance, it could
consist in the generation of good initial solutions and incorporate them into the
arc-time-indexed model. Branch and Cut is also an interesting approach, how-
ever, even the relaxation of variables is computationally arduous due to their
number.
References
1. Aghelinejad, M., Ouazene, Y., Yalaoui, A.: Production scheduling optimisation
with machine state and time-dependent energy costs. Int. J. Prod. Res. 56(16),
5558–5575 (2018)
2. Cai, W., Lai, K.H., Liu, C., Wei, F., Ma, M., Jia, S., Jiang, Z., Lv, L.: Promot-
ing sustainability of manufacturing industry through the lean energy-saving and
emission-reduction strategy. Sci. Total Environ. 665, 23–32 (2019)
3. Cesaret, B., Oǧuz, C., Sibel Salman, F.: A tabu search algorithm for order accep-
tance and scheduling. Comput. Oper. Res. 39(6), 1197–1205 (2012). https://doi.
org/10.1016/j.cor.2010.09.018
4. Chen, S.H., Liou, Y.C., Chen, Y.H., Wang, K.C.: Order acceptance and scheduling
problem with carbon emission reduction and electricity tariffs on a single machine.
Sustainability 11(19) (2019). https://doi.org/10.3390/su11195432
5. Foumani, M., Smith-Miles, K.: The impact of various carbon reduction policies on
green flowshop scheduling. Appl. Energy 249, 300–315 (2019)
6. Gahm, C., Denz, F., Dirr, M., Tuma, A.: Energy-efficient scheduling in manufac-
turing companies: a review and research framework. Eur. J. Oper. Res. 248(3),
744–757 (2016). https://doi.org/10.1016/j.ejor.2015.07.017
7. Garcia, C.: Resource-constrained scheduling with hard due windows and rejection
penalties. Eng. Optim. 48(9), 1515–1528 (2016)
8. Herrmann, I.T., Hauschild, M.Z.: Effects of globalisation on carbon footprints of
products. CIRP Ann. 58(1), 13–16 (2009)
9. Masmoudi, O., Delorme, X., Gianessi, P.: Job-shop scheduling problem with energy
consideration. Int. J. Prod. Econ. 216, 12–22 (2019)
10. Mouzon, G., Yildirim, M.: A framework to minimise total energy consumption
and total tardiness on a single machine. Int. J. Sustain. Eng. 1(2), 105–116 (2008).
https://doi.org/10.1080/19397030802257236
11. Og, C., Salman, F.S., Yalçın, Z.B., et al.: Order acceptance and scheduling decisions
in make-to-order systems. Int. J. Prod. Econ. 125(1), 200–211 (2010)
12. Palakiti, V.P., Mohan, U., Ganesan, V.K.: Order acceptance and scheduling:
overview and complexity results. Int. J. Oper. Res. 34(3), 369–386 (2019)
13. Silva, Y., Subramanian, A., Pessoa, A.: Exact and heuristic algorithms for order
acceptance and scheduling with sequence-dependent setup times. Comput. Oper.
Res. 90, 142–160 (2018). https://doi.org/10.1016/j.cor.2017.09.006
14. Stern, H.I., Avivi, Z.: The selection and scheduling of textile orders with due dates.
Eur. J. Oper. Res. 44(1), 11–16 (1990)
15. Zhang, H., Zhao, F., Fang, K., Sutherland, J.W.: Energy-conscious flow shop
scheduling under time-of-use electricity tariffs. CIRP Ann. 63(1), 37–40 (2014)
The Analysis of Human Oriented System
of Weighted Fuzzy Petri Nets for the Passenger
Transport Logistics Problem
Zbigniew Suraj1(&), Oksana Olar2(&), and Yurii Bloshko1(&)

1
University of Rzeszów, 35310 Rzeszów, Poland
2
Yuriy Fedkovych Chernivtsi National University, 58012 Chernivtsi, Ukraine
[email protected]
Abstract. This paper describes an experiment, the goal of which is to prove

that the use of classical triple of functions is better and more precise over an
optimized one for the problem of the passenger transport logistics. It includes a
detailed description of different triples of function which can be applied in this
field of research and their application, where different combinations lead to
different results. The experiment presented in this paper includes a group of
triples of functions that are located between classical and optimized one. They
give results in the range from the middle to the highest possible mathematical
value. In this manner, it is possible to verify the effectiveness of the application
of classical triple of functions based on the majority of similar decisions
achieved from the other triples. Moreover, this paper presents the detailed
description of wFPN model with its mathematical tools. An explanation of the
higher effectiveness of the classical triple compared to the other triples con-
sidered in the paper for a given problem was obtained.
Keywords: Knowledge representation Production rules Decision-support

system Optimized fuzzy Petri net Weighted fuzzy Petri net Transport
logistic task
1 Introduction
Transport logistics problem consists of several levels of tasks that are interconnected
between each other [1]. It gives a possibility to structure the knowledge for the deeper
understanding of the problem as well as simplifies the search for the best and the most
detailed answer for this question. Since the problem of transport logistics can be
described in terms of dynamic discrete events, the use of fuzzy Petri nets (FPN) is
considered to be a relevant mathematical tool for the research of these tasks [2].
This paper aims to continue the research on the application of weighted FPN
(wFPN) for the passenger transport logistics problem. Specifically, there were already
applied different triples of functions in the numerous experiments. The goal is to find
the triple that is most likely to be applied for the user’s needs [3, 4]. In other words,
every decision-support system is oriented on the user’s preferences. Therefore, it is
compulsory to do a transformation of the mathematical language into the human
1582 Z. Suraj et al.
language in the easiest possible way. Thus, the goal is to find some triple of functions
which can be applied by any user and will precisely reflect the results in accordance to
the user’s preferences.
The methodology used in the paper is based on the knowledge representation in
tables of type “Object-property”, production rules, triangular norms (t-norms and
s-norms), weighted fuzzy Petri net, optimized fuzzy Petri net [3–9]. These techniques
and tools implement the idea of a decision-support system in the relevant transport
logistics problem. In order to show the application of this theory in the experiment, all
input information was set to the PNeS software. The network describes the structure of
connections and information which leads to some numerical output results (decisions)
[10]. All calculations and conclusions are included.
This paper consists of five sections: introduction, problem formalization, triangular
norms for wFPN, net representation model of passenger transportation and conclusion.
2 Problem Formalization
In order to describe the problem of transport logistics, there was introduced a four-level
scheme with concrete tasks (and subtasks) and their relationships between each other
[1]. Figure 1 presents the structure of transport logistics problem, where each level
contains some specific knowledge.
Fig. 1. Scheme description of the transport logistics problem.
The general idea is to divide the main goal on the smaller tasks to simplify the
search for the solution. Each element of the structure contains some knowledge about
the problem. The elements, in turn, are combined into a logical whole that represents
The Analysis of Human Oriented System 1583
the full description of the problem. It leads to the possibility of representing the
knowledge in the tabular form of type “Object-property”.
Production rule of type IF ri THEN dj is the most suitable approach for this kind of
situation since there are connections at the internal level: between properties and object
as well as at the outer level: between tables. Also, production rules can be extended
with the use of logical AND, OR: IF ri1 AND (OR)… AND (OR) rin THEN dj, where
rik (k = 1, …, n) – property, and dj – object. The next step is to create wFPN model
based on these production rules and set there an appropriate triple of functions as well
as betta b(t) and gamma c(t).
3 Triangular Norms for wFPN
Every transition consists of three mathematical elements: betta b(t), gamma c(t) and
triple of operators (In, Out1, Out2). Each triple (In, Out1, Out2) consists of the following
elements: (a) In – the input function which takes as an input markings from the input
places and connects them with the transition t; (b) Out1 - the first output function,
which takes as its first argument the value of the input function In and for the second -
the value of the betta function b(t); (c) Out2 - the second output function, which takes
as its first argument the value of the first output function Out1, and its second argument
is the marking value taken from the output place to which the transition t is connected.
Every element from the triple is being changed with some t-norm or s-norm [7]. The
following condition must be satisfied to fire a transition:
Inðwi1 Mðpi1 Þ; wi2 Mðpi2 Þ; . . .; wik Mðpik ÞÞ cðtÞ [ 0 ð1Þ
where: (1) In is an input function of the triple of operators that is changed with some
function; (2) wij (j = 1,…,k) is a weight that describes a strength of connection between
input place and a transition in the net. Also, it describes a strength of connection
between a property and an object in the knowledge table. This weight is set at the
intersection of the corresponding object and property; (3) M(pij) is a marking of a place;
(4) c(t) is gamma that is set by experts in the corresponding field of studies.
Another element of the triple is betta b(t) and it is calculated in the following way:
b ¼ k=ðk þ 1Þ ð2Þ
where k is a number of input places connected to this transition.

In addition, there is a need to give a detailed description of the mathematical tool of
wFPN. Figure 2 is a representation of triples of functions for logical AND. There are
125 possible combinations in total. The structures which describe the top-vertex and
bottom-vertex triples of function for production rules that apply logical AND and OR
were presented in paper [9]. The further disclosure will be dedicated to the production
rules which apply logical AND, because it is the most-often used logical operator in the
experiment of the current paper.
Fig. 2. A cube with all possible triples of functions for the FPN (case of logical AND).
A triple of functions (ZtN, GtN, ZsN) is known as a classical triple, because it is the
most commonly used triple in the calculations for the production rules which apply
logical AND. Also, this cube presents the bottom-vertex (a so-called minimal) triple of
functions: (LtN, LtN, ZsN) and the top-vertex (a so-called maximal) triple of functions:
(ZtN, ZtN, LsN). The group of triples of functions that is in between the minimal triple
and the classical one is highlighted as a green (horizontal) rectangle in this cube, while
the group of functions that is in between classical and maximal triples is highlighted as
a vertical rectangle with blue color.
The mathematical description of these functions is as follows: ZsN (Zadeh s –
norm): ZsN(a, b) = max (a, b); ZtN (Zadeh t – norm): ZtN(a, b) = min (a, b); GtN
(Goguen t – norm): GtN(a, b) = ab; LtN (Lukasiewicz t – norm): LtN(a, b) = max (0,
a + b − 1); LsN (Lukasiewicz s – norm): LsN(a, b) = min (1, a + b).
Since, the paper [3] is dedicated to the analysis and comparison of application of
classical and optimized triples of functions, the area of interest is highlighted in a blue
rectangle of the cube presented in Fig. 2.
As it can be mentioned in this rectangle, there are 15 different combinations of
triples. It gives is a good possibility to prove which type of triples (classical or opti-
mized one) is the most suitable for the problem of passenger transport logistics. In
order to give an answer for this question, there is a proposal to test 13 triples which are
in between classical and optimized triples. This approach allows to find out what these
triples have in common with these boundary functions, i.e. the initial and final ones.
It is worth noting the specification of each wFPN model for the given problem:
every level of output places for the corresponding level of transitions is always empty
before firing of these transitions. In other words, only the first level of places is filled
with the corresponding markings at the beginning, while other levels of places are
empty. Therefore, the third operator of the triple doesn’t make an influence on the
output results, because output place is always empty before firing a transition. Thus, the
second input value for the third operator Out2 is always equal to 0. As a result, it is
possible to consider only vertical lines of the blue rectangle presented in Fig. 2 as it
describes the second operator Out1 which can be interpreted as one of the following
functions: GtN, HtN, ZtN.
The description of these functions is as follows – GtN, ZtN as above and HtN
(Hamacher t – norm):

0 for a ¼ b ¼ 0
HtN ða; bÞ ¼ ab ð3Þ
a þ bab otherwise
Additionally, the first operator In is the same for any triple of functions presented in
the blue rectangle in Fig. 2 as it is on the one plane (side) of the cube.
4 Net Representation Model of Passenger Transportation
This paper presents two different input sets of numerical values aiming to analyze the
results of the equal models as well as to make some conclusions on the change of triple
of functions. Moreover, this experiment aims to give an answer which triple of function
is better to use for the passenger transport logistics problem by applying triples of
functions that are presented in the blue rectangle in Fig. 2.
In order to answer these questions, the first level of the scheme presented in Fig. 1
is taken into consideration. The aim is to find the best transport mean (type of trans-
port). There were created equal wFPN models with different triples of functions which
applied two sets of inputs.
The model presented in Fig. 3 consists of 4 levels of places and 3 levels of tran-
sitions, where the first level of places is dedicated to properties of user’s preferences for
the choice of transport type and the last, fourth level, is dedicated to the objects as the
decision: “Airplane”, “Automobile vehicle” or “Train”. In addition, there are two
transit levels of places that help narrow down the answer to the problem with a wide
range of options. The network was created from the production rules that were
extracted from the knowledge tables of type “Object-property”.
This type of table has a feature to make a switch between objects and properties in
tables that are interconnected between each other. In this manner, output places of the
previous level of transitions become input places for the following level of transitions
at the outer level, while the transition connects some number of properties (input
places) to the object (output place) at the internal level.
Table 1 presents two sets of input values for the model presented in Fig. 3.
Table 1. Table of sets of input values.

Set 1 0.95 0.8 0.75 0.8 0.65 0.8 0.85 0.7 0.55 0.7 0.75 0.8
Set 2 0.95 0.7 0.85 0.9 0.8 0.9 0.85 0.75 0.6 0.7 0.9 0.8
Table 2 presents the resulting values at output places of the wFPN model for 2 sets
of inputs and two different triples of functions.
Table 2. Resulting values at output places for two sets of inputs with two different triples of
functions.
Objects
Airplane Automobile vehicle Train
Set 1 (ZtN, GtN, ZsN) 0.2 0.1 0.13
(ZtN, ZtN, LsN) 0.44 0.24 0.42
Set 2 (ZtN, GtN, ZsN) 0.22 0.1 0.16
(ZtN, ZtN, LsN) 0.48 0.24 0.5
Triples of functions
As it can be noted from the results for both sets, the theory presented in papers [3, 9]
is proved: optimized triple of functions gives a higher output values. Additionally, the
resulting decisions should be carefully observed: they are the same for the input set 1,
while they are different for the input set 2. Here the situation arises where the optimized
triple of function achieved higher output results, but the change of triples of functions
leaded to another decision. So, there is a need to apply additional triple of functions that
is in between classical and optimized triples of functions presented in a blue rectangle in
Fig. 2. Since, ZtN and GtN were already applied, only HtN as the second operator can
be selected. The following triple of functions (ZtN, HtN, ZsN) is formed for the next test
of the wFPN model on the best type of transport.
Fig. 3. An example of the wFPN model for the best type of transport that applies triple of
functions (ZtN, GtN, ZsN).
Table 3 presents the results for the input set 2 including additional triple of func-
tions (ZtN, HtN, ZsN). The use of (ZtN, HtN, ZsN) suggests on the same decision as
the classical triple (ZtN, GtN, ZsN). With such an approach, the decision on the
“Airplane” as the best type of transport is more probable. Additionally, it is worth
mentioning that the triple of functions (ZtN, HtN, ZsN) resulted in the values that are in
between values received from classical and optimized triples for every object. As it was
expected from the Fig. 5, the following correlation is satisfied (ZtN, GtN, ZsN) (
ZtN, HtN, ZsN) (ZtN, ZtN, LsN) for every object: Airplane: 0.22 < 0.33 < 0.48;
Automobile vehicle: 0.1 < 0.19 < 0.24; Train: 0.16 < 0.29 < 0.5.
Table 3. Resulting values at output places for the input set 2 with three different triples of
functions.
Objects
Airplane Automobile vehicle Train
Set 2 (ZtN, GtN, ZsN) 0.22 0.1 0.16
(ZtN, HtN, ZsN) 0.33 0.19 0.29
(ZtN, ZtN, LsN) 0.48 0.24 0.5
Triples of functions
Another approach which can be also tested is the average mathematical value
calculated for each object (output place) that was conducted by every triple of func-
tions. If the results for each object in Table 3 were substituted into the formula:
ResðZtN; GtN; ZsN ÞObji þ . . . þ ResðZtN; HtN; ZsN ÞObji þ . . . þ ResðZtN; ZtN; LsN ÞObji
ResðObji Þ ¼
Num of triples
ð4Þ
then the following average values for each object were obtained:
– Airplane: 0:22 þ 0:33
3
þ 0:48
¼ 0:34ð3Þ.
– Automobile vehicle: 0:1 þ 0:19 3
þ 0:24
¼ 0:17ð6Þ.
0:16 þ 0:29 þ 0:5
– Train: 3 ¼ 0:31ð6Þ.
The highest average value which is equal to 0.34 belongs to the object (decision)
“Airplane”. It is the same decision that was achieved by two triples of functions: (ZtN,
GtN, ZsN) and (ZtN, HtN, ZsN). Thus, the average mathematical value is another
approach to confirm the achieved decision.
5 Conclusion
The experiment consisted of two parts for two different sets of inputs which were tested
by classical and optimized triple of functions. According to the results, optimized triple
of functions always leads to the higher numerical values at the output, but the decisions
made by classical triple and optimized one are not always the same. Therefore, wFPN
model was tested with additional triple (ZtN, HtN, ZsN) that is in between classical and
optimized one. The obtained decisions achieved by triples (ZtN, HtN, ZsN) and (ZtN,
GtN, ZsN) were the same. Thus, additional model gave a possibility to confirm the
truth probability of the decision achieved by the classical triple (ZtN, GtN, ZsN).
Another approach that confirmed this theory was the calculation of the arithmetic
mean for each object taken from each equal wFPN model that used different triples of
functions. Again, the highest resulting value was dedicated to the object which was also
achieved by the classical triple of functions.
Therefore, this paper suggests to use classical triple of functions (ZtN, GtN, ZsN)
for the problem of the passenger transport logistics as far as this triple of function is the
most used in the wFPN calculations and the decision was confirmed by majority of
triples as well as by the average value.
The next step of the research is to apply other triples of functions presented in the
cube in Fig. 2 in order to confirm the effectiveness of the classical triple of functions or
to find some alternative.
References
1. Suraj, Z., Olar, O., Bloshko, Y.: Conception of fuzzy petri net to solve transport logistics
problems. Current Research in Mathematical and Computer Sciences II, Publisher UWM,
Olsztyn, pp. 303–313 (2018)
2. Liu, H.C., You, J.X., Li, Z.W., Tian, G.: Fuzzy petri nets for knowledge representation and
reasoning: a literature review. Eng. Appl. Artif. Intell. 60, 45–56 (2017)
3. Suraj, Z., Olar, O., Bloshko, Y.: Optimized fuzzy petri nets and their application for transport
logistics problem. In: Proceedings of International Workshop on Concurrency, Specification
and Programming (CS&P 2019), Olsztyn (2019)
4. Suraj, Z., Olar, O., Bloshko, Y.: Hierarchical weighted fuzzy petri nets and their application
for transport logistics problem. Submitted to the 15th International Conference on Intelligent
Systems and Knowledge Engineering (ISKE 2020), Cologne (2020)
5. Lyashkevych, V., Olar, O., Lyashkevych, M.: Software ontology subject domain intelligence
diagnostics of computer means. In: Proceedings of the 7th IEEE International Conference on
Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applica-
tions (IDAACS-2013), Berlin (2013)
6. Lokazyuk, V.: Software for creating knowledge base of intelligent systems of diagnosing
process. In: Lokazyuk, V., Olar, O., Lyaskevych, V. (eds.) Advanced Computer System and
Networks: Design and Application, Lviv (2009)
7. Klement, E.P., Mesiar, R., Pap, E.: Triangular Norms. Kluwer, Dordrecht (2000)
8. Chen, S.M.: Weighted fuzzy reasoning using weighted fuzzy petri nets. IEEE Trans. Knowl.
Data Eng. 14(2), 386–397 (2002)
9. Suraj, Z.: Toward optimization of reasoning using generalized fuzzy petri nets. In:
Proceedings of the IJCRS 2018, Vietnam (2018)
10. Suraj, Z., Grochowalski, P.: Petri nets and PNeS in modeling and analysis of concurrent
systems. In: Proceedings of the International Workshop on Concurrency, Specification and
Programming (CS&P 2017), Warsaw (2017)
Concepts and Methods of “Digital Twins”
Models Creation in Industrial Asset
Performance Management Systems
Nodirbek Yusupbekov1, Fakhritdin Abdurasulov2, Farukh Adilov3,

and Arsen Ivanyan3(&)
1
Tashkent State Technical University, Tashkent, Uzbekistan
[email protected]
2
Oltin Yo’l GTL Company, Tashkent, Republic of Uzbekistan
[email protected]
3
LLC Ximavtomatika, Tashkent, Republic of Uzbekistan
{farukh.Adilov,arsen.Ivanyan}@himavtomatika.uz
Abstract. The paper describes new solutions and methods for computer
modelling of different aspects and layers of industrial assets. In this paper not
only new software solutions for creating of model itself are considered but also
new solutions for creating of software environment for further models’ inter-
action with real industrial plant in real-time mode are described. This paper
concentrates mostly on “digital twin” models for asset performance monitoring
and effectiveness increase, however some aspects of other types of “digital twin”
models for industrial applications are raised also. Despite that some fundamental
physic model types like first-principle model or thermodynamic model are well-
known in science some other modelling methods like creation of fault-symptom
model or data-driven model are new approaches related to industrial “digital
twins”. The motivation of the study for this paper is to find the scientific-
technical way to raise the performance and efficiency of industrial enterprise to
absolutely new level which can be established only on the applying of real-time
based digitized intelligent infrastructure.
Keywords: Asset Performance Management (APM) Digital twin

Performance models Analytic models Health & fault models Original
equipment manufacturers (OEM) Asset Performance Library Advanced
pattern recognition (APR) Data-driven model Key performance indicator
(KPI)
1 Introduction
In today’s competitive environment, process plants are under more pressure than ever
to deliver improved operating performance with fewer skilled staff. Personnel are often
unaware that their processes are under-performing. In fact, many plants do not even
monitor process performance in real-time. Production, efficiency, quality, and cost are
all measured after-the-fact. Industrial organizations need up to date advanced analytics
to make decisions in near real-time to run their production processes more efficiently.
1590 N. Yusupbekov et al.
Traditionally, plant operators and engineers have relied on periodic, manual,

spreadsheet-based tools to calculate critical variables and key performance indicators
(KPIs).
By taking advantage of advanced “digital twin” technology enabling the most
advanced monitoring, analytical, and predictive capabilities, process engineers can
implement around-the-clock monitoring of plant data, and ongoing operational health
checks and recommendations to close performance gaps [1].
As the example of advanced “digital twin” technology for industrial plant opti-
mization and production effectiveness increase we will consider special cloud and
software infrastructure called Asset Performance Management (APM) system. Asset
Performance Management technology unifies asset and process data through integra-
tion of all sensory sources, control systems, process historians, and machine condition
monitoring systems. The ability to reveal new levels of untapped plant productivity
begins with the creation of this unified process-asset dataset.
The unified dataset is securely pushed to the cloud, where digital twins perform
performance calculations and analytic data analysis [2]. Drawing from a broad library
of embedded asset models, machine efficiency (performance) is calculated. This pro-
vides a quantifiable measure of excessive energy usage, as well as leading indication of
process or asset health issues. The models also combine the most modern cloud ana-
lytic techniques. The analytics and machine learning actually compliment the perfor-
mance metric [3].
Performance models excel at quantifying work done, while data driven analytics
excel at recognizing irregularities early – revealing root cause, and providing proba-
bility estimates for future outcomes [4]. This complementary approach to digital twins
provides the earliest insight available to recognize impending issues and is the key to
finding hidden improvement opportunities that otherwise would go unnoticed.
Scalable across the extents of an organization’s enterprise and functionally scalable
to monitor any asset-process KPI, APM drives team collaboration. De-localizing access
to machinery and process performance reveals insight to all plant equipment stake-
holders, regardless of where they sit globally.
The organization of this paper consists of Abstract and Introduction sections where
actual problem statements and motivation to this study is described, Methods and
results section where authors describe in details fundamental methodology of stated
problems resolving, as well as basic results achieved by solutions replied, Conclusion
section with study results finalized and suggestions for future researches, References
section with references used in the paper.
2 Methods and Results
Innovation idea of industrial facility optimization based on APM system considers

incorporation of library of wide range of asset models into IT-infrastructure APM.
The library combines three crucial asset modeling methods into asset-specific
templates. By combining modeling methods, the library provides the most in-depth
view of overall asset performance.
Concepts and Methods of “Digital Twins” Models Creation 1591
1. Performance Models:
– Best substitute when original equipment manufacturer’s (OEM) factory per-
formance curves are not available
– Physics-based (thermodynamic) first principle equations of multiple perfor-
mance measurement types (e.g. the pressure and flow generated by a pump or
compressor) [5]
– Performance models excel at accurately measuring the work done by a machine.
They quantify efficiency by relating the energy consumed vs. the work done
– Performance models are good at recognizing effects of changing process con-
ditions, machine fouling, and process leaks
– Performance models may be slow to recognize minor variation in measured
parameters
– Performance models are poorly suited for identifying mechanical health of
bearings, couplings, lube oil conditions, shaft misalignment and unbalance, etc.
2. Analytic Models:
– Based on statistical calculation rather than thermodynamic formulas
– Analytic models excel at early detection of variation in process conditions, as
well as machine mechanical parameters such as bearing vibration
– Analytic models are adept at recognizing issues with sensors, including cali-
bration drift and erratic measurements
– Analytic models do not quantify any physical parameter. However, by calcu-
lating rates of change, they can serve to predict a timeframe for RUL (remaining
useful life) of a machine.
3. Health & Fault Models:
– Health models allow for traditional health monitoring relative to established
alarm limits
– Health models monitor and record standard KPI’s such as uptime/downtime,
machine availability, running status or status of needed maintenance, etc.
– Fault models provide a ready form of fault tree for identifying where in a system
a fault or alarm is creating impact. It also assists in drilling to root cause of
faults.
The Asset Performance Library is a continually evolving set of templates for
quickly creating digital twins for a variety of industrial machinery types. The library
incorporates the latest expertise and knowledge that machinery subject matter experts
have gathered from the deployment of asset performance management (APM) solutions
with clients across industries. The models within each template represent the evolution
of experience gained through implementing automation systems, and modeling and
simulation of numerous industrial processes.
Use of these templates provides access to this embedded knowledge and greatly
reduces the time needed to configure APM to elicit insight and intelligence from plant
data.
A plant typically has thousands of points and control tags where data is continu-
ously collected, yet they may have few means to process that data effectively to elicit
actionable information. Likewise, model building is time consuming and has proven to
be a large reason for reluctance in implementing an APM solution. APM solves these
issues with asset templates library. The models within each asset template turn asset
and process data into actionable insight by applying three distinct areas of monitoring:
– Performance-based Monitoring
– Analytics-based Monitoring
– Health Monitoring & Fault Models [6]. Figure 1 illustrates this concept.
Fig. 1. Asset performance library concept of APM
The performance models calculate any variation between baseline performance and
the actual performance of a machine in real-time. This provides a means to find latent
inefficiency in a machine’s operation and to quantify costs associated with such inef-
ficiencies [5].
The performance models contained within the library are focused on recognizing
two primary areas of energy loss of monitored machinery: operational loss and design
loss.
Operational loss: loss caused by equipment degradation and/or the machine being
operated sub-optimally due to non-standard process conditions or upsets.
Design loss: loss caused by equipment operating off its ideal operating point during
normal process operating conditions. (e.g. loss from oversized pumps and motors,
undersized valves, etc.)
Derived insight from performance:
– Continuously reveals latent inefficiencies
– Identifies asset performance improvement opportunities
– Accurately quantifies energy loss
– Provides an objective basis for maintenance planning
– Eliminates offline, manual, and periodic calculation methods
Fig. 2. Performance models’ category of APM digital twins
Figure 2 illustrates performance models’ category of digital twins.

Performance models coupled with predictive machine learning data analytics is the
most effective means for early recognition of asset issues.
Industrial APM utilizes machine learning and advanced pattern recognition
(APR) methods in its data analytics models. Each analytic model learns a machine’s
normal operating patterns by drawing from the plant data stream. Once trained, a model
continuously compares current data patterns to the learned and expected patterns. The
changes in the patterns are used to alert early on developing problems.
As with any data driven approach, the usefulness of the results depends on careful
selection of both the model inputs and the training datasets. The first consideration is
addressed by each embedded asset template, the second depends on the available data.
Figure 3 illustrates analytic models’ category of digital twins.
Fig. 3. Analytic models’ category of APM digital twins

The third category of APM digital twins is health and fault models. Health mon-
itoring is based on direct sensory data, such as vibration, pressures and temperatures.
Users can quickly configure health monitoring and associated alarm levels for both the
primary asset and its auxiliary systems. For example, the configuration template for
centrifugal compressors provides a template for configuring a compressor’s health
model – as well as the configuration of health of the compressor auxiliaries, such as a
lube oil system, a seal oil system, or an applicable gearbox. Together these monitored
inputs tie to the templated fault models.
Augmenting the health and fault models is a pre-configured template for standard
metrics. Every asset type in the Asset Performance Library is pre-configured to cal-
culate the:
– Asset Index = Availability * Performance/100
– Availability – percent of time available
– Metrics Start Date – time stamp on all metrics
– Performance – percent of “no faults” time
– Status (Faulted) – in/out of fault; potential fault
– Status (Maintenance) – in/out & planned/unplanned
– Status (Running) – running/not running
– Status (Uptime) – available/unavailable
– Time (Downtime)
– Time (Faulted)
– Time (Maintenance)
– Time (Running Cumulative)
– Time (Running)
– Time (Uptime)
Figure 4 illustrates health and fault models’ category of digital twins.
Fig. 4. Health and fault models’ category of APM digital twins

3 Conclusion
APM is a powerful, scalable solution for managing all industrial assets across an
enterprise. APM provides the latest capabilities for serving a company’s asset perfor-
mance management initiatives – whether that initiative is to embark on a new asset
management program (e.g. IIoT Digital Transformation and cloud hosted delivery
model) or expand and transform an established condition monitoring system.
APM integrates all plant data into a single database. By integrating process, asset
and operational data, and constructing digital twins over the unified dataset,
Forge APM delivers increased performance and reliability to all processes and assets –
not just critical machines.
APM goes beyond traditional machine monitoring and data gathering. By merging
together decades of machine and process modeling experience with modern cloud
analytics, APM digital twins predict machinery availability, drill to the root cause of
inefficient machine operation, and bring order to reliability and maintenance planning.
There is also further research activity going on related to the integration of APM
infrastructure together with another aspects of optimization in management for
industrial enterprise like process performance management, smart methodology of
people management, control and profit performance management [7].
References
1. Honeywell Process Solutions: Real-Time Process Performance Monitoring: Powerful
Analytic Tool Enables Smart Operations. Whitepaper, WP-18-02-ENG. Honeywell Process
Solutions, Houston (2018)
2. Grieves, M.: Digital Twin: Manufacturing Excellence through Virtual Factory Replication.
Digital Twin White Paper. Oracle, Redwood Shores (2014)
3. Zhiqiang, G., Zhihuan, S., Steven, X.D., Biao, H.: Data mining and analytics in the process
industry: the role of machine learning. IEEE Access 5(12), 20590–20616 (2017)
4. Data’s Big Impact on Manufacturing: A Study of Executive Opinions. Study by Honeywell
and KRC Research Inc. Honeywell Process Solutions, Houston (2016)
5. Patrakhin, V.: Proactive maintenance of equipment as practical implementation of GE digital
twin concept. World Autom. 2(4), 64–68 (2017)
6. Yusupbekov, N., Somakumaran, S., Narwadkar, A., Abdurasulov, F., Adilov, F., Ivanyan, A.:
Advantages of analytic solutions of sentience cloud platform. Chem. Technol. Control
Manag. 4–5(12), 9–12 (2018)
7. Yusupbekov, N., Abdurasulov, F., Adilov, F., Ivanyan, A.: Increase of effectiveness of
industrial enterprise based on forge software-analytical platform. In: 5th International
Proceedings on Hybrid and Synergetic Intelligent Systems, HSSS 2020, pp. 23–30. BFU,
Zelenogradsk (2020)
Finding the Optimal Features Reduct,
a Hybrid Model of Rough Set and Polar
Bear Optimization
Amer Mirkhan1 and Numan Çelebi2(&)

1
Computer Engineering Department, Sakarya University, Sakarya, Turkey
2
Information Systems Engineering Department,
Sakarya University, Sakarya, Turkey
[email protected]
Abstract. The target of this research is reducing the size of a dataset which is
usually needed before starting the data analysis in scientific research, this can be
done by removing attributes that do not affect the accuracy of the dataset, this
process will enhance the performance of data analysis and will result in more
accurate results and decisions. The rough set theory introduced by Pawlak
provides a powerful technique to measure the influence of each attribute in the
dataset and the effect of excluding an attribute on the accuracy of the dataset.
However, evaluating all possible combinations of features is an NP-Hard
problem and is usually not possible when dealing with datasets having large
number of attributes. To solve this kind of problems the heuristics algorithms
can play a crucial role to avoid scanning all possible combinations. Polar Bear
Optimization Algorithm PBO, is a pretty new meta heuristic algorithm has an
advantage over other heuristic algorithms of solving such kind of problems
using dynamic population with a flexible production and death mechanism,
which results in finding optimal solution quickly by keep producing good
solutions out of hopeful candidates and keep removing unpromising ones. Our
proposed algorithm could find the optimal reduct in better performance com-
paring to other algorithms in terms of execution time, population size and
number of iterations.
Keywords: Rough set Heuristic algorithms Features selection

Optimization
1 Introduction
In scientific research, usually the process starts by preparing data in order to start
analyzing this data and discover rules and insights. Before starting analyzing datasets,
especially in huge ones, only relevant features -attributes- should be selected. However,
deciding if a feature is necessary or not is not a straight forward decision, mainly when
the research is being performed by someone out of the domain. Features needed to
make the decision will be called “Minimal Reduct”.
The rough set theory was first introduced by Pawlak in 1982 [1, 2], provides a very
powerful mathematical based technique to deal with inconsistent data and generate
Finding the Optimal Features Reduct, a Hybrid Model of Rough Set 1597
uncertain rules. In the field of machine learning, several techniques exist to find the
relation between input attributes and the output attribute(s), however, those techniques
do not provide a solution when dataset is fuzzy and definite rules do not exist.
However, when dealing with big datasets with high number of attributes, rough set
techniques will require evaluating all possible combinations of attributes to find the
minimal reduct which can be turned into NP-hard problem. To avoid exploring all
possible reducts, heuristics algorithms can support this process by only evaluating
limited number of solutions.
Polar Bear Optimization PBO, is a new meta-heuristic algorithm proposed by
Dawid Polap in 2017 [3]. The new technique that PBO proposes is the dynamic birth
and death mechanism which allows investing more in good potential solutions and
keep removing hopeless ones.
Our proposed technique is to convert the functions of PBO to deal with binary
representations of solutions instead of solutions with coordinates (x, y, z, …). This will
include re-writing the local search, global search and fitness functions, and then adopt
this algorithm along with rough-set to find the minimal reduct of a dataset via exploring
limited number of solutions defined as an input for the algorithms in terms of popu-
lation size and maximum number of iterations.
The rest of this paper is organized as: Sect. 2, is a literature review for related work,
Sect. 3, will be an introduction to rough set theory and an explanation of the PBO
algorithm, Sect. 4 will present the details of our proposed algorithm, Sect. 5 will
summarize our experimental results, Sect. 6 concludes this research in addition to
future work.
2 Related Work
A considerable amount of literature just focused either on implementing the rough set
techniques to find minimal reducts in various research areas, or just used the heuristic
algorithms to solve NP-hard problems. However, in the summary below only research
implemented the heuristic algorithms with rough set theory were mentioned:
Chen, Zhu and Xu proposed a novel Rough Set based method to feature selection
using fish swarm algorithm [11]. Later, a similar research done by Su and Guo pro-
posed a novel search strategy for minimal attribute reduction based on rough set theory
(RST) and fish swarm algorithm (FSA) to identify the core attributes of a dataset [7].
Both have proved that using the FSA algorithm can improve the accuracy of finding the
minimal reduct in addition to efficiency of convergence rate. Another implementation
of a population-based algorithm was proposed by Zouache and Ben Abdelaziz by
presenting a new cooperative swarm intelligence algorithm for feature selection based
on quantum computation and a combination of Firefly Algorithm (FA) and Particle
Swarm Optimization (PSO) and obtained a better rate of feature reduction and a high
accuracy classification [6]. Rodríguez-Diez, Martínez-Trinidad, Carrasco-Ochoa,
LazoCortésc and Olvera-Lópeza proposed a new algorithm for computing all the
shortest reducts based on binary cumulative operations over a pair-wise comparison
matrix, and a fast candidate evaluation process, the result of their analysis showed that
their technique could find the minimum reduct faster than similar algorithms [4].
1598 A. Mirkhan and N. Çelebi
Another heuristic approach was proposed by Alweshah, Alzubi and Alaqeel (2016)
who have introduced a combination between wrapper approach and genetic pro-
gramming algorithm, Wrapper Genetic Programming (WGP) to find the most infor-
mative attributes [5].
3 Background
In this section the theoretical concepts associated with of the rough set will be
explained. After that the basic characteristics of heuristic algorithms will be introduced
in general. Finally, the details of the PBO algorithm will be described.
3.1 Rough Set Theory

Pawlak (1982) [1, 2] has introduced the rough set theory as a technique for knowledge
discovery to deal with fuzzy datasets where exact and certain rules cannot be obtained
using the traditional classification algorithms. Below the main definitions of rough set
theory are presented:
Information System: An information system can be considered as table of rows and
columns. Rows will be called objects, while columns will be called attributes or fea-
tures. Formally, the decision table can be written as I ¼ ðU; AÞ, where U is the uni-
verse, a finite non-empty set of objects, A is a finite non-empty set of features.
Dependency of Features: The dependency between a set of condition attributes B and
a set of decision features R is given by the Eq. (1):
cB ðDÞ ¼ jPOSjUBjðDÞj 1 ð1Þ
POSB ðDÞ is the positive region [2] of B, so here we divide number of elements on
the lower approximation by total number of objects.
When the value of cB ðDÞ = 0, this means B is independent of D, and if D is fully
dependent on B the value will be 1.
3.2 Polar Bear Optimization

Polar bear optimization is a swarm-based optimization algorithm introduced by Połap
and Woźniak in 2017 [3], a simulation to the way that polar bear hunting in harsh
condition. Here we will list the general characteristics of PBO.
Population Generation: The algorithm starts by specifying number of bears which
will be looking for the target, increasing number of bears will increase the chance of
finding the global optimal solution, but having too many bears will have negative
impact on the execution time of the algorithm.
Local Search: At every iteration all bears will try to make one step toward the tget,
before moving the new solution is evaluated, if the new position is better than the
current one then the bear will move, otherwise it will stay at current location.
Global Search: After completing each loop, where all bears have tried to make a step
and improve their locations, one of the best solutions will be selected and all other
bears will try to make a step toward, however actual movement will not take place until
the new solution is evaluated and shows a better position.
Dynamic Control of Population: In each iteration and based on a random number, a
decision is made to produce a new member or remove one. If the decision is to
reproduce, then two of the best bears generate a new one by combining the two
solutions, here we assume that combining two good solutions will produce another
good solution.
4 Proposed Model, Polar Bear Optimization for Feature

Selection PBO-FS
The original version of PBO deals with solutions as spatial objects, each solution can
be represented as a set of coordinates ðx1 ; x2 ; . . .xn Þ and all its functions (local search
global search, moving functions, …), consider this kind of objects. However, solutions
in feature selection problems can be considered as a set of selected and not selected
features. Those solutions can be represented as a binary array, where “1” indicates that
a feature is selected while “0” means not selected. In one dataset all arrays will have the
same length. PBO cannot be applied directly to solve RST problems as both algorithms
are speaking different languages. Here we will list our modifications to the original
PBO to adopt RST.
Local Search: To generate a new solution based on the current one, our proposal is to
switch s number of attributes, where s is calculated according to the Eq. (3):
s ¼ 4 4 maxc
i
ð2Þ
Here i is the current loop counter, and maxc is the total number of loops. The idea
here is to start the loop by changing four features and then gradually decreasing number
of changed features to one when solutions become more stable. This behavior was
inspired from Simulated Annealing Algorithm [9].
Global Search: After all bears finish their local search, we randomly select two of the
best bears in order to produce and new one. Only features exist in one of those two
solutions (XOR operator) will be included in the new solution.
Fitness Function: To evaluate each solution, the dependency should be calculated
according to Eq. (1), however, this equation does not give any attention to the number
of the selected attributes. We needed a mechanism that gives weighting the quality of
the proposed solutions in addition to the number of the selected feates.ere our fitness
function will be the Eq. (3) proposed by Wang [10]:
jC j jRj
Fitness ¼ a cR ðDÞ þ b ð3Þ
jC j
The first part of the equation a cR ðDÞ is the classification quality as defined in
Eq. (1), while |C| is the total number of features and |R| is the number of selected
features. This equation also yields a kind of tolerance to the quality calculation, where
a and b can be considered as tolerance parameters. Increasing b parameter will increase
the tolerance and will allow accepting solutions with fewer attributes. Those parameters
can be adjusted according to the nature of data. According to Wang [10], ideal values
for a, b can be 0.9 and 0.1 respectively.
PBO-FS Algorithm
Input: decision table, number of bears, number of itera-

tions and tolerance parameters α and β
Output: convergence list of bears
(1) Initialize variables
(2) Create 75% of the bears and calculate fitness using
Equation (3)
(3) while (iter <= maxc)
(4) {
(5) Calculate “step” according to Equation (2)
(6) for each bear current population
(7) {
(8) next = randomly change “step” number of attributes
(9) If next position is better than current
(10) Move the bear to the new position
(11) if the new position is better than global_optimal
(12) update global_fitness
(13) update global_min
(14) add current bear to convergence
(15) }
(16) Select one of the top 10% and make one more move
(17) Choose random value κ between 0 and 1
(18) If k < 0.25
(19) Select two of the top 10% and produce a new one
(20) else
(21) if number of current bears > bears * 0.5
(22) remove the worst bear from population
(23) iter = iter + 1
(24) }
(25) Output convergence
5 Experimental Analysis
In order to evaluate the efficiency of our proposed algorithm, three datasets from the
UCI Machine Learning Repository [8] were selected. To be able to compare the
performance of our algorithm with similar algorithms we have selected the same
datasets used and evaluated by [6]. Our analysis was performed using a personal
computer with 1.8 GHz CPU and 16 GB RAM running Windows 10. Since our pro-
posed approach is fully stochastic based process, each execution might give different
result, so each dataset was evaluated 10 times and the optimal results were written in
this analysis.
Table 1 contains the used datasets with along with number of records and number
of features, also includes the minimum reduct obtained using brute force technique by
examining all possible solutions. Then the table has results for the optimal reduct
obtained by three heuristic algorithms, RSAR [10], RSAR-Entropy [10] and QCISA-
FS [6]. Last column of the table is the result of our algorithm.
Table 1. Best solution obtained by each heuristic algorithm

Dataset Instances Features Min RSAR RSAR-Entropy QCISA-FS PBO-FS
Lymphography 148 18 6 6 8 7 6
Mushroom 8124 22 4 5 5 4 4
Zoo 101 16 5 5 10 5 5
From Table 1 we can compare the result obtained by our algorithm with three other
algorithms in terms of minimal reduct found by each algorithm. As we can notice in
this table, our algorithm was always able to find a reduct similar or better than other
algorithms. Especially for the Mushroom dataset, which is the most complicated
database taking into consideration number of features and number of samples.
In Fig. 1, we can see convergence diagram for all datasets, depending on the
complexity of the dataset, the figure represents how our algorithm is trying to improve
the fitness of each dataset. We notice that all datasets start at fitness 0.9, the fitness is
calculated according to Eq. (3), the algorithm starts by calculating the fitness when all
features are included, in this case the right operand of the “+” sign in Eq. (3) is zero,
while the left side should be 0.9 when the original dataset is crisp [2].
To decide the needed population size for each dataset, we used several values of
bears across all datasets, we noticed that the optimal number of bears needed to find the
minimal reduct is same number of features. Selecting the population size correctly has
an important impact on the performance of the algorithm. Selecting small number of
bears might not allow the algorithm finding good results specially when number of
features is high. On the other hand, selecting high number of populations might result
in finding good results but will have negative impact on the execution time. From
Fig. 2 we can notice that at certain threshold increasing the number of bears will not
produce better solutions anymore as the optimal solution is already found. Another
important parameter is number of iterations, this parameter also affects the ability of
finding the optimal solution and also has impact on the execution time. According to
1
Lymphography
0.95
Fitness
Mushroom
0.9 Zoo
0.85
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Convergence
Fig. 1. Fitness value per dataset (convergence chart)
our analysis and as we can see in Fig. 3, the optimal solution is found at a certain
threshold, which is double number of features, after this threshold, the algorithm will
keep looping but no more solutions are found.
0.985
Fitness
0.98
0.975
0 10 20 30 40 50
Number of bears
Fig. 2. Mushroom dataset classification fitness per population size
1
Fitness
0.98
0.96
0 10 20 30 40 50 60 70
Number of iteraƟons
Fig. 3. Mushroom dataset fitness per number of iterations
The uniqueness of our algorithm is dynamic population, to prove this we have

included additional information to each solution includes the iteration number which
produced this solution and the number of steps made in local search until finding the
optimal solution, from Table 2 we can notice that in most cases the optimal solution
was generated using the re-production mechanism and not by the initial population.
Table 2. Analysis for the obtained solution for all used datasets
Dataset Features Iteration count Best solution found at Changed at Duration(s)
Lymphography 18 36 12 18 <1
Mushroom 22 44 30 33 55
Zoo 16 32 7 15 3
Because the original version of the polar bear algorithm deals with spatial coordinates
while each solution in rough set is list of selected features of the dataset, we had to
make some amendments to the original PBO to be make it compatible with rough set.
Our analysis showed that the dynamic population of our algorithm found the minimal
reduct efficiently in comparing with similar algorithms. In our proposal we imple-
mented hamming distance using XOR operator to reproduce new solutions out of two
good solutions, the results showed that this was a promising technique, however, we
believe that implementing other binary operators might even produce better results, this
could be a subject for future research to be evaluated.
References
1. Pawlak, Z.: Rough sets. Int. J. Comput. Inform. Sci. 11, 341–356 (1982)
2. Pawlak, Z.: Some Issues on Rough Sets. Lecture Notes in Computer Science (2004)
3. Połap, D., Woźniak, M.: Polar bear optimization algorithm: meta-heuristic with fast
population movement and dynamic birth and death mechanism. Symmetry 9, 203 (2017)
4. Lazo-Cortés, M.S., Martínez-Trinidad, J.F., Carrasco-Ochoa, J.A., Diaz, G.S.: A new
algorithm for computing reducts based on the binary discernibility matrix. Intell. Data Anal.
20(2), 317–337 (2016)
5. Alweshah, M., Alzubi, O.A., Alzubi, J.A., Alaqeel, S.: Solving attribute reduction problem
using wrapper genetic programming. Int. J. Comput. Sci. Netw. Secur. (IJCSNS) 16(5), 77
(2016)
6. Zouache, D., Abdelaziz, F.B.: A cooperative swarm intelligence algorithm based on
quantum-inspired and rough sets for feature selection. Comput. Ind. Eng. 115, 26–36 (2018)
7. Su, Y., Guo, J.: A novel strategy for minimum attribute reduction based on rough set theory
and fish swarm algorithm. Comput. Intell. Neurosci. 2017, 1–7 (2017)
8. Bache, K., Lichman, M.: UCI machine learning repository (2013)
9. Van Laarhoven, P.J., Aarts, E.H.: Simulated annealing. In: Simulated Annealing: Theory
and Applications, pp. 7–15. Springer, Dordrecht (1987)
10. Wang, X., Yang, J., Teng, X., Xia, W., Jensen, R.: Feature selection based on rough sets and
particle swarm optimization. Pattern Recogn. Lett. 28(4), 459–471 (2007)
11. Chen, Y., Zhu, Q., Xu, H.: Finding rough set reducts with fish swarm algorithm. Knowl.-
Based Syst. 81, 22–29 (2015)
Design and Manufacture a Vehicle Auxiliary
System Model Controlled by Using
Smartphone
Nguyen Phu Thuong Luu(&)
Automotive Engineering Department, Ho Chi Minh City University

of Technology (HUTECH), Ho Chi Minh City, Vietnam
[email protected]
Abstract. This paper presents a design and manufacture of an automotive

auxiliary electrical system (AAES) model which was controlled by using a
smartphone. The auxiliary electrical systems are included security systems,
comfort and safety systems, lighting systems and information systems. This
research focused on the lighting systems (front lighting, rear lighting, com-
partment lighting, and signalization lighting) and comfort and safety systems
(windshield wiper and window cleaning). Those systems are connected to an
Arduino module then an android smartphone device was used to control the
systems. An application for the android device was built by using MIT App
inventor. In this study, a user interface was set on the android device and the
systems via a Bluetooth channel. The experiments for control systems were
confirmed that we can apply for the modern vehicle control system.
Keywords: Auxiliary system Control system Arduino Smartphone
1 Introduction
With the increasing development of science and technology, AVR microcontrollers and
PIC microcontrollers are becoming more and more common and used world wide, but
it can be said that the arrival of Arduino in 2005 in Italy has opened a new direction for
microcontroller [1, 2]. The arrival of Arduino has helped humans a lot in programming
and design, especially for those who started to learn about microcontrollers without too
much knowledge, deep understanding of physics and electronic [3]. The hardware of
the device has integrated many basic functions and open source [4].
The Java-based programming language is extremely easy to use, compatible with
the C language, and the library is rich and free to share. Because of these reasons,
Arduino is currently gaining popularity and is growing more and more powerful
worldwide [5–7]. In this research, the author uses Arduino Uno R3 to play a key role in
receiving and processing signals to control the system’s modes. The transmitter and
receiver distance of Bluetooth Module HC06 within 20 m is responsible for receiving
control signals via phone, then sending that signal to Arduino to control the system.
This model can be used for training and vocational education in automotive
engineering.
Design and Manufacture a Vehicle Auxiliary System Model 1605
2 Design and Build the Model System

2.1 Control Blocks System
Power Source
The power supply unit creates a constant current and voltage that provides safety for
the circuit. Because the Arduino circuit uses 5 V power, we use the additional DC low
voltage circuit XL4015 (5A) with Current Adjustment used to reduce the DC voltage as
the usual pulse voltage reducing circuit, the circuit integrated IC Opamp compared at
the output and Variable current limiting resistor helps to ensure safety when using, in
case of short circuit or overcurrent circuit will alert (Fig. 1).
Fig. 1. A block diagram of auxilary monitor device
Android Mobile Interfaces

All phones running android can install our mobile application, which is written based
on the MIT App Inventor 2 platform to create an intuitive interface, easy to use for
control and supervise electrical appliances. The ability to control and monitor multiple
devices was shown in [4] (Fig. 2).
Fig. 2. Configuration menu on the mobile application software
Block Module Bluetooth

The HC-06 module was used to receive Bluetooth broadcast from the Central control
circuit connected via Android phones, using the HC-06 Module as shown in Fig. 3.
1606 N. P. T. Luu
Fig. 3. HC-06 module Bluetooth
Block Module Arduino

The Arduino receives signals from the Bluetooth module and sensor signals, then
processes the signals and controls the system’s modes through the built-in program as
shown in Fig. 4.
Fig. 4. Arduino module
Block Relay
Relay is a simple remote control switch; it uses a small current to control a large current
so it is used to protect the switch, so it is also considered a protective device. A typical
relay controls the circuit and source control. The relay structure consists of an iron core,
a magnetic coil, and a contact.
Relay is responsible for receiving control signals from Arduino to shut down,
power the device depending on the mode controlled on the phone as shown in Fig. 5.
Fig. 5. Relay module
2.2 Principle of Operation

First, launch the Android software “Device Control”. The software interface appears to
select Bluetooth connection. The user can click on the icon to turn on the Bluetooth and
make an automatic connection to the Bluetooth module HC-06 the software an
automatically takes the address of the Bluetooth Module (input when programming).
When connected successfully, the software will display a notice that it is connected
successfully. The software on the application includes 2 control folders for the lighting
system and wiper system as shown in Fig. 6.
Fig. 6. The complete software control interface unit for auxiliary system
ON Function. When clicking on a button, the android software will send the packet
(containing device opening command) via the device’s Bluetooth device to the Blue-
tooth Module HC-06. At this time, the Bluetooth HC-06 module will receive the packet
and transmit it to the Arduino module to give the signal as a power supply for the Relay
to close.
OFF Function. When you click once more on a button, the android software will send
the packet (containing the device close command) via the phone’s Bluetooth device to
the Bluetooth Module HC-06. At this time, the Bluetooth HC-06 module will receive
the packet and transmit it to the Arduino module to give the signal a power outage for
the open relays.
2.3 Algorithm Flowchart
Program Algorithm for the Lighting System. The input signals from the Bluetooth
module HC-06 and the light sensor which show in Fig. 9 input signals to the Arduino
will process the signal and control the opening and closing of the relay depending on
the mode installed on the phone App. The lighting circuit schematic of the hardware
interface unit was shown in Fig. 8 (Fig. 7).
Program Algorithm for the Wiper System. The input signals from Bluetooth
module HC-06 and water sensor which show in Fig. 11 that input signals into Arduino
will process signals and control opening and closing depending on each mode installed
on the mobile App. The flowchart for the control system was shown in Fig. 10.
1608 N. P. T. Luu
Fig. 7. Microcontroller program flowchart for the lighting control system
Fig. 8. The lighting circuit schematic of the hardware interface unit
Fig. 9. The lighting sensor

Fig. 10. Microcontroller program flowchart for the Fig. 11. Water sensor module
wiper control system
2.4 Mechanical Design

for the Auxiliary Model
Fig. 12. The 3D design of the model
The complete model of the system consists of the parts shown in Fig. 12, 13 and 14.
1610 N. P. T. Luu
Fig. 13. The complete circuit schematic of the hardware interface unit
THE VEHICLE AUXILIARY SYSTEM MODEL

CONTROLLED BY USING SMARTPHONE
Fig. 14. The auxiliary model

3 Conclusion
This is a comprehensive topic including circuit construction design and programming

techniques. Although, we have learned the operation of the hardware features as well as
the Arduino module programming software, capturing the outstanding features of this
control module in the application of measurement monitoring and control. We were
able to implement a reliable low-budget auxiliary vehicle system prototype of a
wireless portable monitoring device that can control the systems in the vehicle. The
main aspect of our system design is the utilization of a smartphone and its built-in
components to control our device and to display adaptive results on the model. Custom
made mobile application software with the practical interface was also developed to
simplify user operation. This proof of concept has demonstrated its applicability to
mobile applications. This monitoring system can be deployed in many settings such as
the starting system, AC system, and other systems in the vehicle for its versatility and
convenience. This research also can apply for technical training and vocational edu-
cation teaching.
Acknowledgment. This work was supported in part by the research program 2020 of Ho Chi
Minh City University of Technology (HUTECH) University in Viet Nam.
References
1. Tahat, A., Said, A., Jaouni, F., Qadamani, W.: Android-based universal vehicle diagnostic
and tracking system. In: 16th International Symposium on Consumer Electronics, pp. 137–
143, Harrisburg. IEEE (2012)
2. Press Release Canalys. Android takes almost 50% share of worldwide Smart phone market,
Singapore and Reading (UK), Palo Alto (2011)
3. Tahat, A., Sacca, A., Kheetan, Y.: Design of an integrated mobile system to measure blood
pressure. In: 18th Symposium on Communications and Vehicular Technology (SCVT), pp. 1–
6. IEEE (2011)
4. AppInventor. http://ai2.appinventor.mit.edu/
5. Tahat, A., Khalaf, M., Elmuhesen, O.: A solar energy water heater remote monitoring and
control system. In: International Conference on Electronic Devices, Systems, and Applica-
tions (ICEDSA), pp. 98–103 (2011)
6. Audi. Intelligent connectivity and communication using smartphone. https://www.audi-
mediacenter.com/en/technology-lexicon-7180/infotainment-7183
7. Carter, S.: Nissan Releases Android App for Leaf Vehicle (2019)
A Study on Vehicle Air Condition System
Model Controlled by Using Smartphone
Nguyen Phu Thuong Luu(&)
Automotive Engineering Department, Ho Chi Minh City University

of Technology (HUTECH), Ho Chi Minh City, Vietnam
[email protected]
Abstract. This paper presents the design and manufacture of an air condition
system (ACS) model which was controlled by using a smartphone. The ACS
model included condenser, fan, electric motor, dryer, evaporator, power supply,
control panel, and compressor. This paper focused on controlling the fan,
compressor and control panel. Those components are installed into the body
frame model, which are connected to an Arduino module then an android
smartphone device was used to control the ACS model. An application for the
android device was built by using MIT App inventor. In this study, a user
interface was set on the android device and the systems via a Bluetooth channel.
The experiments for control ACS were confirmed that we can apply for the
modern vehicle control system.
Keywords: AC system Control system Arduino Smartphone
1 Introduction
Normally, the car air conditioner can not be controlled by phone but only by the button
which was set in the vehicle. Controlling the car air conditioner by mobile phone help
people to easily control it even when they outside the car. Due to combining
mechanical and electrical control, we created the “model of automatic air conditioning
controlled by phone”. Those make more potential for developing and enhance the
convenience of the vehicle.
Nowadays, only a luxurious car has the ability to control the air conditioner by
phone [1, 2]. Our country is a developing country so in the automotive engineering
field, the control of the air conditioner by phone is still a new thing. That is the reason
why our group did the project of research and creating a model of automatic air
conditioning controlled by phone. The Java-based programming language is extremely
easy to use, compatible with the C language, and the library is rich and free to share.
Because of these reasons, Arduino is currently gaining popularity and is growing more
and more powerful worldwide and also was used in this research [3].
A Study on Vehicle Air Condition System Model 1613
2 Design and Build the Model System

2.1 3D Modeling Designs
The 3D modeling was designed in this article by using AutoCAD software as shown in
Fig. 1.
1. Condenser
2. Fan
3. Electric motor
4. Dryer
5. Evaporator
6. Power supply
7. Compressor
Fig. 1. Drawing template of AC system model
All the parts in this model are using the real components in the vehicle systems that
can be bought in the market and have been design in the AutoCAD to make a 3D
model for simulation. The model components are included seven main part as shown in
Figs. 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 and 15. The technical drawings also were
displayed for model designing and the optimal part for the system. The function of each
component as the same in the real AC system in the vehicle such as the condenser, fan,
electric motor (this component is the power for the model as the same engine in the
vehicle and using the AC power source), dryer, evaporator, power supply (use for
control unit and compressor) and compressor.
Fig. 2. Condenser Fig. 3. Condenser technical drawing

1614 N. P. T. Luu
Fig. 4. Fan Fig. 5. Fan technical drawing
Fig. 6. Electric motor Fig. 7. Electric motor technical drawing
Fig. 8. Dryer Fig. 9. Dryer technical drawing

Fig. 10. Evaporator Fig. 11. Evaporator technical drawing
Fig. 12. Power supply Fig. 13. Power supply technical drawing
Fig. 14. Compressor Fig. 15. Compressor technical drawing
2.2 Mechanical Design for the Air-Condition System Model

The purpose of this study is to make the model for technical training and vocational
education. So the convenience for moving is the first bound condition for design. The
mechanical is strong enough for all components assembly on its which is shown in
Fig. 16, 17 and easy for the visible system as shown in Fig. 18.
1616 N. P. T. Luu
Fig. 16. Model frame Fig. 17. Model frame technical drawing
Fig. 18. The complete air-condition vehicle system model
2.3 Algorithm Flowchart

The control app was designed on the website called “http://ai2.appinventor.mit.edu/”
and coded in Arduino IDE software [4].
MIT App Inventor for Android is an open code app provided by Google and
maintained by MIT. These help the creator to create an application for Android OS. By
using friendly display allow the creator to easily pull and drop blocks for creating
applicant. The input signals from Bluetooth module HC-06 and a temperature sensor
which shows in Fig. 19 that input signals into Arduino will process signals and control
opening and closing depending on each mode installed on the mobile App. The
flowchart for the control system was shown in Fig. 19. The complete circuit schematic
of the hardware interface unit and the complete components of the hardware interface
unit as shown in Fig. 20 and Fig. 21.
Fig. 19. Microcontroller program flow- Fig. 20. The complete circuit schematic of
chart for the AC control system the hardware interface unit
2.4 Android Mobile Interface

To communicate with the hardware, the android mobile interface was needed to control
the system. First, connect to the Android software “Device Control”. The software
interface appears to select Bluetooth connection. Click on the icon to turn on the
Bluetooth and make an automatic connection to the Bluetooth module HC-06 the
software automatically takes the address of the Bluetooth Module (input when pro-
gramming). When connected successfully, the software will display a notice that it is
connected successfully. The software on the application as the same in the vehicle air
condition system control as shown in Fig. 22. The control system is including an
ON/OFF function, Fan speed, and AC
Fig. 21. The complete components of the hardware Fig. 22. The complete software control
interface unit interface unit for the AC system
1618 N. P. T. Luu
Fig. 23. The complete circuit Schematic of the hardware interface unit for simulation
status.
The applicant for communicating with Arduino was used by Bluetooth through
module HC-06. When using, the applicant sent the signal to the HC-06 and Arduino
receives the signal and processes that signal to control relay in a sequence that has been
programmed before in Arduino. All phones running android can install our mobile
application, which is written based on the MIT App Inventor 2 platform to create an
intuitive interface, easy to use for control and supervise electrical appliances. The
ability to control and monitor multiple devices was shown in [4]. The complete circuit
Schematic of the hardware interface unit for simulation as shown in Fig. 23.
3 Conclusion
After researching, building the model, this project has successfully created the model of
automatic air conditioning controlled by phone to create a material of car air condi-
tioner for researchers and local car brand to refer.
This project also is an important document in teaching in college or university. It
can help to improve the teaching method more realistic and help the student to
understand about car air conditioner better.
Although the article still has a few limits such as reliability or the build quality is
not good enough. But it can apply in a small vehicle or vehicle that has a simple air
conditioner.
Acknowledgment. This work was supported in part by the research program 2020 of Ho Chi
Minh City University of Technology (HUTECH) University in Viet Nam.
References
1. Cheng, C.C., Lee, D.: Smart sensors enable smart air conditioning control. Sensors 14(6),
11179–11203 (2014)
2. Carter, S.: Nissan Releases Android App for Leaf Vehicle (2019)
3. Tahat, A., Khalaf, M., Elmuhesen, O.: A solar energy water heater remote monitoring and
control system. In: International Conference on Electronic Devices, Systems, and Applica-
tions, pp. 98–103 (2011)
4. AppInventor. http://ai2.appinventor.mit.edu
Multi-shift Single-Vehicle Routing
Problem Under Fuzzy Uncertainty
F. Nucci(B)
Department of Engineering for Innovation, University of Salento, Lecce, Italy

[email protected]
Abstract. This research considers the single-vehicle routing problem

(VRP) with multi-shift and fuzzy uncertainty. In such a problem, a
company constantly uses one vehicle to fulfill demand over a scheduling
period of several work shifts. In our case, a crew executes maintenance
jobs in different sites. The working team runs during different work shifts,
but recurrently returns to the depot by the end of the shift (overtime
avoidance). The goal consists in minimizing the number of work shifts
(makespan). We observe the impact of uncertainty in travel and main-
tenance processing time on the overtime avoidance constraint. We real-
ize an Artificial Immune Heuristic to get optimal solutions considering
both makespan and overtime avoidance. First, we present a Pareto-based
framework to evaluate the uncertainty influence. Then, we show a numer-
ical real case study to survey the problem. In particular, a case study
scenario has been created on the basis of the environmental changes
in travel and processing times observed in Italy during the Covid-19
lockdown period (started on March 9, 2020). Results present important
improvements are obtained with the proposed approach.
Keywords: Single-vehicle routing problem · Scheduling under

uncertainty · Artificial Immune Heuristic
1 Introduction
Vehicle routing problem (VRP) consists in determining a set of routes to visit
a set customers, in order to minimize the path length. Different versions of the
VRP exist. If customers are only available in a time windows, a VRP with Time
Windows (VRPTW) is considered. Basic variants of the VRP consider the route
planning for a vehicle fleet in a single period (shift). In that case, the vehicles
return to the depot before the end of the shift. This problem originates from a
healthcare routing issue [1–3]. When the health care company ships products to
medical sites, if overtime is allowed, performance could be significantly improved.
For example, if a location scheduled for the next shift is on the current return
route to the depot, a limited overtime allows the vehicle to serve it. This can
significantly reduce the workload of the next shift. On one hand, overtime reduces
the total number of shifts necessary to complete the work. On the other, the
Multi-shift Single-Vehicle Routing Problem Under Fuzzy Uncertainty 1621
continuous use of overtime can lead to crew health problems [4]. Since overtime
increases the chance of micro-sleep in car drivers [5], company manager could be
accused of vehicle collision due to wrong workload scheduling [6]. Indeed, shift
planning should maintain shift duration as constant as possible.
Uncertainty on travel and processing time can lead to unexpected overtime
and performance reduction. Frequently, in optimization problems, data are sup-
posed to be known with certainty. However, in practice this is infrequent. More
generally, real data are dependent on uncertainty due to their irregular nature.
Since the solution of the optimization problem typically shows a great tendency
to data disruption, overlooking the ambiguity of the data can lead to non-optimal
or unrealistic solutions for a real case. Robust Optimization is a significant
technique to address optimization problems subject to uncertainty [7]. In this
case, a methodology is needed to analyze the trade-off between performance and
robustness.
On one hand, Stochastic VRP (SVRP) was introduced in [8] when uncer-
tainty is statistically known. See [9,10] for a complete review of SVRP. On the
other, Fuzzy set theory is a useful approach to handle non-stochastic uncertainty
[11]. Fuzzy sets theory is widely adopted for studying the influence of uncertain
factors on VRP [12–20]. In these works, Fuzzy VRP is analyzed and fuzzy set
theory is adopted to manage such uncertain data.
In this paper, we examine a multi-shift VRP with travel and processing time
modeled as Fuzzy Numbers. The objective consists in reducing both overtime
and makespan. The question we considered is derived from a routing problem
in maintenance activities as reported in [21–25]. A maintenance team performs
jobs in different sites using a vehicle for movements. A crew works in shifts and
should come back to the depot before the shift ends. The goal is completing the
maintenance activities in various places reducing both overtime and makespan.
We investigate the influence of the uncertainty of driving and job processing
time on the objective.
The originality of the paper consists in the meta-heuristic approach adopted
to solve the problem. Indeed, our meta-heuristic uses a 2-factor ranking method,
based on overtime and makespan, to sort the solution set at each step. Conse-
quently, a Pareto set of optimal solutions exists. Considering the papers already
examined and two additional review articles [26,27], it can be established that
no such an approach exists.
The body of this paper is structured as follows. In Sect. 2, we report the
problem formulation. In Sect. 3, we propose the Artificial Immune Heuristic to
solve the problem. We present in Sect. 4 a case study solved with the proposed
approach, considering two scenarios: before and after Covid-19 lockdown in Italy
(March 9, 2020). We give concluding remarks in Sect. 5.
2 Problem Definition
In the VRP, we assume a horizon of P shifts (periods), and let the set P =
{1, 2, ..., P } index the shifts of the planning horizon. Each shift duration is L.
1622 F. Nucci
shift 1-2 shift 1 shift 1-2

shift 1 shift 2 shift 2
origin 6 origin 7 8 destination
6 7 8 destination
depot
depot depot depot
2 3 5 1 4 2 3 5 1 4
1 2 3 4 5 6 1 2 3 4 5 6
6 2 3 5 7 1 4 8 6 2 3 5 1 7 4 8
(a) (b)
Fig. 1. Example of path and synthetic representation for two solutions
On a given planning, we seek to serve a set of N customers. Associated with

each customer is a task i to be executed, i ∈ N = {1, 2, ..., N }. Task i processing
time is qi . For each shift, the crew departs from and returns to the depot. A
travel time between task i and task j locations is defined as di,j . We assume
that triangular inequality is valid for driving times. Any task can be executed in
a shift. Our problem objective is to minimize the latest task completions time
(makespan). The problem can be represented as a directed graph G = (V, E),
where V = N ∪ {N + 1, . . . , N + P + 1}. We create P + 1 depot copies represented
by node set {N + 1, . . . , N + P + 1}. Node N + 1 is the origin depot of shift 1.
Node N + h represents the destination depot of shift h − 1 together with the
origin depot of shift h, with h ∈ {2, . . . , P }. Node N + P + 1 stands for the
destination depot of last shift P .
A solution problem can be represented as a path in the graph G. Since any
path starts at node N + 1 and ends at node N + P + 1, a solution can be
synthetically represented by variables ωi ∈ N ∪ {N + 2, . . . , N + P } for i =
1, . . . , N + P − 1. Variable ωi stands for the visiting sequence of location task
nodes N and shift break nodes {N + 2, . . . , N + P }. Shift duration is equal to
the arrival time at node N + 1 + h and is represented by σh for h ∈ P.
Considering N = 5 tasks and P = 2 shifts, in the solution reported in Fig. 1a,
crew executes task 2, 3 and 5 on first shift, whereas task 1 and 4 are allocated
in the second shift. Instead in Fig. 1b solution, task 1 is anticipated to shift 1.
In solution b, overtime occurs in shift 1 although makespan decreases (shift 2
duration is minimal).
We model the uncertainty on driving and working times with Triangular
Fuzzy Numbers (TFN), see Fig. 2. In particular, fuzzy travel time is indicated
with dij = (dA ij , dij , dij ) ∀i, j ∈ V, whereas fuzzy job processing time is referred
B C
as qi = (qi , qi , qi ) ∀i ∈ N . For this reason, variables related to shift duration

A B C
h become fuzzy and have be analyzed to determine whether overtime occurs.

σ
In order to assess whether fuzzy shift duration σ h is lower than maximum
shift duration L, we adopt the function Φ( σh ≤ L) defined in (1). Such a function
calculates the possibility a fuzzy number ( σh ) is lower than a crisp value (L). It is
equal to the portion of the area under the TFN membership function on the left
of crisp value. For TFN, if σhC ≤ L, we have Φ( σh ≤ L) = 1, so condition σ h ≤ L

t = (tA , tB , tC )
possibility

t = (tA , tB , tC )
possibility 1
1
0
µt (t) t2 tA tB t3 tC t1
tC ≤ t1 →
t ≤ t1 is true
t2 ≤ tA →
t ≤ t2 is false
t
tA < t3 < tC →
t ≤ t3 is uncertain
0
tA tB tC
Fig. 2. Triangular Fuzzy Number

t = (tA , tB , tC ) and comparison with a crisp value
is certainly true and no-overtime occurs for shift h. Whereas if L ≤ σhA , we have
Φ(σh ≤ L) = 0, so condition σ h ≤ L is certainly false and overtime occurs for
shift h. Finally, if σhA < L < σhC , then 0 < Φ( σh ≤ L) < 1 and overtime may
occur ; no-overtime possibility is equal to Φ( σh ≤ L).
t0
A μ (τ )dτ

Φ(t ≤ t0 ) = ttC t̃ ∈ [0, 1] (1)
tA
μt̃ (τ )dτ
3 Solution Method: Artificial Immune Heuristic

Artificial Immune Algorithm (AIA) is a meta-heuristic based on animal immune
system [28–30]. This paper proposes a fuzzy AIA to find optimal solutions for
the considered problem. In short, AIA stands on behavior of animal immune
system that protects against foreign pathogens. Immune system reacts to germs
and improves the activity of recognizing and eliminating pathogens by using two
principles: clonal selection and affinity maturation. Clonal selection creates new
immune cells. Such cells encounter high rate of mutations, along with a selection
process.
As reported in Fig. 1, a solution is encoded as a string by using a fixed-length
integer code, providing the order in which nodes are reached. For a solution ψ,
we adopt an innovative 2-factor affinity. Indeed, a solution ψ is described by the
node visiting sequence ωiψ with i = 1, . . . , N + P − 1 and for each shift the cor-
responding fuzzy duration σ h is computed. Then, the 2-factor solution affinity
is determined as (λ, ρ), where λ represents the makespan and ρ is the possibility
no-overtime occurs; λ should be minimized whereas ρ should be maximized. In
(2), L·(P −1) is the maximum length of P −1 shifts and σPB is the crisp duration
of last shift. In (3), the minimum of no-overtime possibilities is calculated over
shift set. Considering the two solutions a and b reported in Fig. 1, the corre-
sponding affinities are the two pairs (λa , ρa ) and (λb , ρb ). Performing the Pareto
comparison of solution a and b, we have λa > λb and ρa > ρb (supposing ρa = 1
and ρb > 0), consequently no solution is better than the other.
λ = L · (P − 1) + σPB (2)
1624 F. Nucci
Table 1. Proposed algorithm
Step Description
1 Initialization
1.1 Fix the population popsize, the No. generations ng, the No. clones nc, the
mutation rate mr, the No. mutations nm, the No. exchangeable antibodies nea
1.2 Create popsize/2 initial solutions by Rule1 and produce popsize/2 initial
solutions by Rule2
2 Affinity Computation
2.1 Calculate the 2-factor affinity (λ, ρ) for each antibody
2.2 Determine Pareto optimal antibodies
3 Generate Next Population
3.1 Copy the Pareto optimal antibodies to the next population
3.2 Select nc antibodies randomly and generate nc clones of the selected antibodies
3.3 Choose nm antibodies, randomly, from nc clones and use mutation to create
nm extra antibodies. Apply each mutation operator with the probability 50%
3.4 Include the nm extra antibodies to the next population
3.5 Add ne brand new solutions (see Step 2) to the next population
3.6 Copy solutions from current to the next population to reach popsize solutions
4 Check Stop Condition
4.1 If ng populations have been generated return the Pareto optimal antibodies
4.2 otherwise go to Step 2
ρ= σh ≤ L)
min Φ( (3)
h=1,...,P
The proposed AIA is described in Table 1. At Step 1.2, Rule1 is the full
random rule: random selection of node ωi ∈ N ∪ {N + 2, . . . , N + P } with
i = 1, . . . , N + P − 1. While in Rule2 we choose nodes using a probability
that is inversely proportional to the distance between the current node and
each candidate node. Mutation operator randomly selects two solution indexes
i, j = 1, . . . , N + P − 1 and swaps their content ωi and ωj . If the depot node
sequence N + 2, . . . , N + P is unfeasible, mutation is cancelled. Considering the
ordinary AIA approach, the innovation of this work relies in the step 2 and 3.1.
Indeed, step 2 is used to determine the new antibody affinity, whereas step 3.1
preserves the entire Pareto set in the next population.
4 Numerical Results
We validated our approach, described in Sect. 3 and we set AIA parameters as

follows: population size popsize = 200, No. generations ng = 10000, No. clones
nc = 20, mutation rate mr = 0.75, mutation number per generation nm = 40,
No. exchangeable antibodies nea = 20.
A real case study, in the field of elevator maintenance and repair, is consid-
ered. Since data obtained from the company are protected from disclosure, we
Pareto Optimal Solutions Pareto Optimal Solutions

100% 100%
Possibility of no-overtime
Possibility of no-overtime
80% 80%
60% 60%
40% 40%
AIA solutions AIA solutions
20% Manual solutions 20%
Manual solutions
0% 0%
1920 2040 2160 2280 2400 2400 2520 2640 2760 2880
Makespan (min) Makespan (min)
(a) (b)
Fig. 3. Results for base scenario (a) and lockdown scenario (b)
report only summary data. Company and its customers are located in Salento,
in the southeast region of Italy. Uncertainty affects driving and working times,
inferred from empirical data. Maximum shift duration is set to L = 480 min.
No. jobs is equal to N = 22, whereas No. shifts is P = 5.
A base scenario is considered with crisp job processing time qiB of 40 to
80 min. Processing time uncertainty is 20% of crisp value, that is qiC − qiA =
0.2 · qiB , so q A = 0.9 · qiB and q C = 1.1 · qiB . Crisp driving time dB
i range from 20
to 50 min and driving time uncertainty is also 20% of crisp value: dA = 0.9 · dB i
and dC = 1.1 · dB i . The Algorithm described in Sect. 3 produces the Pareto
set reported in Fig. 3a along with two manual solutions designed by company
experts. Company experts analyzed the eight AIA solutions that dominate their
own solutions. Since rightmost AIA solution (λ, ρ) = (2300, 100%) is very conser-
vative, managers are unlikely to accept such a high safety margin. Experts pre-
ferred solution (2241, 95%) because makespan decreases by almost one hour with
5% risk. Also, solution (2166, 84%) is remarkable because of the good makespan
compared to the significant possibility of 84% to avoid overtime. Managers dis-
carded solutions having ρ < 0.5 because of the high risk of overtime.
Another scenario called lockdown was analyzed. Because of the environmen-
tal changes in travel and processing times during the Italian Covid-19 lock-
down period (started on March 9, 2020), maintenance planning was completely
redesigned. From one hand, new activities were introduced in the tasks such as
cleaning of surfaces using appropriate disinfection methods and wearing personal
protective equipment. Crisp working time increased by 8% plus 10 min. More-
over, processing time uncertainty reached 30% of crisp value. From the other,
road traffic decreased significantly. Crisp driving times were reduced by 25%.
In lockdown scenario, Fig. 3b shows AIA Pareto optimal solutions. Managers
experienced difficulties in designing good planning. Note that P = 6 shifts are
necessary to complete the previous job set. Because of the high uncertainty only
two Pareto optimal solutions were found with ρ > 0.5. Significant difference
exists between AIA and Manual solutions in the lockdown scenario: performing
tasks in the same area may not be the best strategy because crew may overrun
1626 F. Nucci
the shift. On the other hand, a change of zone may lead to a better fit of the
tasks in the shift, as travel time is shorter than usual.
5 Conclusion
This study presents the single-vehicle routing problem with multi-shift when
fuzzy uncertainty is introduced in driving and job processing times. The objec-
tive consists of minimizing both the system makespan and shift overtime occur-
rence. We provide optimal solutions for the decision-maker considering a 2-factor
comparison. Our approach was adopted in a real company case study. During
the Italian Covid-19 lockdown period, a new robust maintenance planning was
rapidly issued. In the future, the possibility of copying only a subset of Pareto
optimal solutions in the next population will be investigated.
References
1. Zangeneh-Khamooshi, S., Zabinsky, Z.B., Heim, J.A.: A multi-shift vehicle routing
problem with windows and cycle times. Optim. Lett. 7(6), 1215–1225 (2013)
2. Nucamendi-Guillén, S., Martı́nez-Salazar, I., Angel-Bello, F., Moreno-Vega, J.M.:
A mixed integer formulation and an efficient metaheuristic procedure for the k-
Travelling Repairmen Problem. J. Oper. Res. Soc. 67(8), 1121–1134 (2016)
3. Onder, G., Kara, I., Derya, T.: New integer programming formulation for multiple
traveling repairmen problem. Transp. Res. Procedia 22, 355–361 (2017)
4. Sparks, K., Cooper, C.L., Fried, Y., Shirom, A.: The effects of working hours on
health: a meta-analytic review. In: From Stress to Wellbeing, vol. 1, pp. 292–314.
Palgrave Macmillan UK, London (2013)
5. Caruso, C.C.: Negative impacts of shiftwork and long work hours. Rehabil. Nursing
39(1), 16–25 (2014)
6. Costa, G.: Shift work and health: current problems and preventive actions. Saf.
Health Work 1(2), 112–123 (2010)
7. Lee, C., Lee, K., Park, S.: Robust vehicle routing problem with deadlines and travel
time/demand uncertainty. J. Oper. Res. Soc. 63(9), 1294–1306 (2012)
8. Cook, T.M., Russell, R.A.: A simulation and statistical analysis of stochastic vehi-
cle routing with timing constraints. Decis. Sci. 9(4), 673–687 (1978)
9. Gendreau, M., Laporte, G., Séguin, R.: Stochastic vehicle routing. Eur. J. Oper.
Res. 88(1), 3–12 (1996)
10. Yaohuang, G., Binglei, X., Qiang, G.: Overview of stochastic vehicle routing prob-
lems. J. Southwest Jiaotong Univ. 10(2), 113–121 (2002)
11. Xu, J., Yan, F., Li, S.: Vehicle routing optimization with soft time windows in
a fuzzy random environment. Transp. Res. Part E: Logist. Transp. Rev. 47(6),
1075–1091 (2011)
12. Teodorović, D., Pavković, G.: The fuzzy set theory approach to the vehicle routing
problem when demand at nodes is uncertain. Fuzzy Sets Syst. 82(3), 307–317
(1996)
13. de C.T. Gomes, L., Von Zuben, F.J.: Multiple criteria optimization based on unsu-
pervised learning and fuzzy inference applied to the vehicle routing problem. J.
Intell. Fuzzy Syst.: Appl. Eng. Technol. 13(2–4), 143–154 (2002)
14. He, Y., Xu, J.: A class of random fuzzy programming model and its application to
vehicle routing problem. World J. Model. Simul. 1(1), 3–11 (2005)
15. Sáez, D., Cortés, C.E., Núñez, A.: Hybrid adaptive predictive control for the multi-
vehicle dynamic pick-up and delivery problem based on genetic algorithms and
fuzzy clustering. Comput. Oper. Res. 35(11), 3412–3438 (2008)
16. Ghannadpour, S.F., Noori, S., Tavakkoli-Moghaddam, R., Ghoseiri, K.: A multi-
objective dynamic vehicle routing problem with fuzzy time windows: model, solu-
tion and application. Appl. Soft Comput. 14, 504–527 (2014)
17. Novaes, A.G., Bez, E.T., Burin, P.J., Aragão, D.P.: Dynamic milk-run OEM opera-
tions in over-congested traffic conditions. Comput. Ind. Eng. 88(C), 326–340 (2015)
18. Muñoz-Carpintero, D., Sáez, D., Cortés, C.E., Núñez, A.: A methodology based
on evolutionary algorithms to solve a dynamic pickup and delivery problem under
a hybrid predictive control approach. Transp. Sci. 49(2), 239–253 (2015)
19. Ewbank, H., Wanke, P., Hadi-Vencheh, A.: An unsupervised fuzzy clustering app-
roach to the capacitated vehicle routing problem. Neural Comput. Appl. 27(4),
857–867 (2016)
20. Zhu, Z., Xiao, J., He, S., Ji, Z., Sun, Y.: A multi-objective memetic algorithm based
on locality-sensitive hashing for one-to-many-to-one dynamic pickup-and-delivery
problem. Inf. Sci. 329(C), 73–89 (2016)
21. Avci, M., Avci, M.G.: A grasp with iterated local search for the traveling repairman
problem with profits. Comput. Ind. Eng. 113, 323–332 (2017)
22. Zamorano, E., Stolletz, R.: Branch-and-price approaches for the multiperiod tech-
nician routing and scheduling problem. Eur. J. Oper. Res. 257(1), 55–68 (2017)
23. Chen, X., Hewitt, M., Thomas, B.W.: An approximate dynamic programming
method for the multi-period technician scheduling problem with experience-based
service times and stochastic customers. Int. J. Prod. Econ. 196, 122–134 (2018)
24. Miranda, D.M., Conceição, S.V.: The vehicle routing problem with hard time win-
dows and stochastic travel and service time. Expert Syst. Appl. 64, 104–116 (2016)
25. Chen, X., Thomas, B.W., Hewitt, M.: The technician routing problem with
experience-based service times. Omega 61, 49–61 (2016)
26. Braekers, K., Ramaekers, K., Nieuwenhuyse, I.V.: The vehicle routing problem:
state of the art classification and review. Comput. Ind. Eng. 99, 300–313 (2016)
27. Khalfay, A., Crispin, A., Crockett, K.: A review of technician and task scheduling
problems, datasets and solution approaches. In: 2017 Intelligent Systems Confer-
ence (IntelliSys), pp. 288–296 (2017)
28. Al-Enezi, J., Abbod, M., Alsharhan, S.: Artificial immune systems - models, algo-
rithms and applications. Int. J. Res. Rev. Appl. Sci. 3(2), 118–131 (2010)
29. Bagheri, A., Zandieh, M., Mahdavi, I., Yazdani, M.: An artificial immune algorithm
for the flexible job-shop scheduling problem. Future Gener. Comput. Syst. 26(4),
533–541 (2010)
30. Mobini, M., Mobini, Z., Rabbani, M.: An artificial immune algorithm for the
project scheduling problem under resource constraints. Appl. Soft Comput. 11(2),
1975–1982 (2011)
On Merrifield-Simmons Index of Trees
Bünyamin Şahin(B)
Department of Mathematics, Faculty of Science, Selçuk University,

42130 Konya, Turkey
[email protected]
Abstract. Classification of graphs with respect to a parameter is

an important optimization problem in graph theory. The Merrifield-
Simmons index is defined as the total number of independent sets. This
index can be calculated by recurrence relation in terms of Fibonacci num-
bers. In this paper, we characterize the trees whose Merrifield-Simmons
indices are calculated by recurrence relation with k steps. Finally, we
obtain the extremal values of trees whose Merrifield-Simmons indices
are calculated recursively in k steps.
Keywords: Independent sets · Merrifield-Simmons index · Recurrence

relation
1 Introduction
Graph theory has many applications in physics, chemistry and related fields.
The carbon skeletons of chemical materials are modelled by molecular graphs.
The carbon atoms are represented by vertices and bonds between these atoms
are represented by the edges. A molecular descriptor is a closed formula obtained
from the molecular graphs of chemical materials. They are very useful tools in
providing correlations about some physical and chemical properties of chemical
materials for example, boiling point, entropy etc. [1].
The Merrifield-Simmons index (or σ-index) is a well studied molecular
descriptor in graph theory. This index was introduced by Richard Merrifield
and Howard Simmons [1]. More details can be found in the excellent survey [2].
This molecular descriptor is investigated for trees with given number of pendant
vertices [3], for trees with a given diameter [4], for trees with bounded degree [5],
for a class of trees [6], for trees with two branching vertices [7], for thorn regular
caterpillars [8], for unicyclic graphs [9,10]. Moreover the graphs, the Merrifield-
Simmons index of which is calculated in a single step by recurrence relation, are
characterized in [11].
Let G = (V, E) be a simple graph which is consisted of the vertex set V and
the edge set E. The open neighbourhood of a vertex v is defined as NG (v) =
{u|uv ∈ E(G)} and the closed neighborhood of v is defined as NG [v] = NG (v) ∪
{v}.
The degree of a vertex v ∈ G is equal to the number of edges which are
incident to v and denoted by degG (v). If the degree of a vertex is one, it is called
On Merrifield-Simmons Index of Trees 1629
a leaf. If a vertex is adjacent to a leaf, it is called a support vertex. A path and a

star graph with n vertices are denoted by Pn and Sn , respectively. If A ⊆ V and
W ⊆ V , then the subgraph of G which is obtained by deleting the vertices of A
is denoted by G − A. If the set A = {u}, we prefer the notation G − u instead
of G − {u}.
A subset of V (G) is called independent if no two vertices are adjacent.
The number of independent sets with k vertices is denoted by σ(G, k), for
0 ≤ k ≤ |V (G)|/2. By definition σ(G, 0) = 1 for any graph G. Furthermore,
the Merrifield-Simmons index of a graph G is denoted by σ(G) and it is equal
to the total number of independent sets of G such that in the following equation
|V (G)|

2
σ(G) = σ(G, k) (1)

k=0
We use Fibonacci number sequence with initial terms F1 = F2 = 1 and

Fn = Fn−1 + Fn−2 for n ≥ 3.
The following lemma is very important for our investigations. It is used in
computing the σ-index of molecular graphs (the graphs whose vertices have
degree at most four). For example, the σ-index of triphenylene molecule is cal-
culated in [11].
Lemma 1 [12]. Let G be a graph. Then,
If G1 , G2 , . . . , Gm are the components of the graph G, then σ(G) =

i)
m
k=1 σ(Gk ),
ii) If u ∈ V (G), then σ(G) = σ(G − u) + σ(G − NG [u]).
iii) σ(Pn ) = Fn+2 and σ(Sn ) = 2n−1 + 1.
In this paper, we generalize the results in [11] and we present a general

condition not only for molecular graphs but also for every graphs. Moreover, we
characterize the graphs whose Merrifield-Simmons index is calculated in exactly
k steps (k ≥ 1) and we obtain the trees which has maximal Merrifield-Simmons
index in this tree family. Finally, we obtain the extremal trees whose Merrifield-
Simmons index is calculated in two steps.
2 Main Results
In [11] Gutman et al. obtained the molecular trees, unicyclic, bicyclic and tri-
cyclic molecular graphs whose Merrifield-Simmons index can be calculated in a
single step by using recurrence relation which is given Lemma 1. The molecular
trees mentioned above are depicted in Fig. 1.
1630 B. Şahin
Fig. 1. The molecular trees T1 , T2 and T3 whose Merrifield-Simmons index is calculated

in a single step.
Thus σ index of T1 , T2 and T3 is as follows [4].
σ(T1 ) = σ(Pa1 +a2 +1 )σ(Pb1 +b2 +1 ) + σ(Pa1 )σ(Pa2 )σ(Pb1 )σ(Pb2 ) (2)
σ(T2 ) = σ(Pa1 +a2 +1 )σ(Pb1 +b2 +1 )σ(Pc1 +c2 +1 )

+σ(Pa1 )σ(Pa2 )σ(Pb1 )σ(Pb2 )σ(Pc1 )σ(Pc2 )
(3)
σ(T3 ) = σ(Pa1 +a2 +1 )σ(Pb1 +b2 +1 )σ(Pc1 +c2 +1 )σ(Pd1 +d2 +1 )

+σ(Pa1 )σ(Pa2 )σ(Pb1 )σ(Pb2 )σ(Pc1 )σ(Pc2 )σ(Pd1 )σ(Pd2 ).
(4)
Now we give an essential theorem which is characterize the graphs whose

σ-index can be calculated in a single step.
Theorem 1. The Merrifield-Simmons index of a graph G can be calculated in a

single step if and only if G contains a vertex v whose deletion yields only paths.
Proof. We assume that the σ index of G is calculated in a single step. For a

vertex v ∈ V (G) we use Nv instead of NG [v] as in the following expression
σ(G) = σ(G − v) + σ(G − Nv ) (5)
It implies that operation is finished and each component of G − v and G − Nv

is consisted in terms of known Merrifield-Simmons values such that paths. In this
case if G − v contains the components Pa1 , Pa2 , . . . , Par , then G − Nv contains
the paths Pa1 −1 , Pa2 −1 , . . . , Par −1 . Therefore
σ(G) = σ(Pa1 )σ(Pa2 ) . . . σ(Par ) + σ(Pa1 −1 )σ(Pa2 −1 ) . . . σ(Par −1 ). (6)
The converse is clear and we omit the details.
In the following theorem, we generalize the result in Theorem 1 and we

characterize the graphs whose Merrifield-Simmons index is calculated in exactly
l steps.
Theorem 2. Let l ≥ 1 be a positive integer and G be a connected graph. Then

the Merrifield-Simmons index of G can be calculated in exactly l steps if and
only if there exists v ∈ V (G) such that the σ index of each component of G − v
can be computed in at most l − 1 steps and moreover, the σ index of at least one
of the component of G − v can be computed in exactly l − 1 steps.
Proof. We prove the necessary condition. Thus G can be constructed in the

following recursive way as depicted in Fig. 2.
Denote by Sq the set of all connected graphs whose Merrifield-Simmons index
can be computed in exactly q ≥ 0 steps. Then S0 is the set of all paths.
Fig. 2. The graph G whose Merrifield-Simmons index is calculated in exactly l steps.
Gl−1
1 ∈ Sl−1 and for k ≥ 1, m1 , m2 , . . . , mk ≥ 1
l−1

G≤l−1
2 , G≤l−1
3 , . . . , G≤l−1
k ∈ Sq .
q=0
We use contraposition and assume that σ(G) is not be computed in l steps

by recurrence relation. Thus, for v ∈ V (G)
σ(G) = σ(G1 )σ(G2 ) . . . σ(Gk ) + σ(G1 − Nv )σ(G2 − Nv ) . . . σ(Gk − Nv ). (7)
We continue this operation untill the every component is reduced to paths.

By our assumption G1 , G2 , . . . , Gk can be reduced to paths at maximum l − 2
steps, is a contradiction. Thus, there is a component of G − v, say G1 , whose
Merrifield-Simmons index is computed in exactly l − 1 steps. The step number
of other components in G − v can be changed such that at maximum l − 1 steps.
The converse is trivial and we finish the proof.

1632 B. Şahin
Lemma 2. Let T1 and T2 be trees different from a path and G be a graph which
is consisted of T1 , T2 and a path between these trees such that in Fig. 3. If σ(G)
is computed at least two steps, then distance between T1 and T2 has to be at least
three.
Proof. If the Lemma 1(ii) is applied to u or t, then
σ(G) = σ(G − u) + σ(G − Nu ) (8)
σ(G) = σ(T1 − u)σ(T2 + vw) + σ(T1 − Nu )σ(T2 + w). (9)

Clearly if T2 is a star, then σ(T2 + vw) and σ(T2 + w) can be computed in a
single step and the result is obtained. If the recurrence relation is applied to v
or w, then
σ(G) = σ(G − v) + σ(G − Nv ) (10)
σ(G) = σ(T1 )σ(T2 + w) + σ(T1 − u)σ(T2 ). (11)
Consequently, if T2 is a minimal tree different from a path such that a star,
then σ(T2 + w) is computed in a single step and the proof is completed.
Fig. 3. The graph G which is used for Lemma 2.
Now we define a tree family Tn,k containing all trees with n vertices whose
σ-index is calculated in exactly k steps.
Theorem 3. Among all trees in family Tn,k , T4 has maximal Merrifield-

Simmons index which is depicted in Fig. 4.
Proof. We know from [3] if the number of leaves adjacent to a vertex increases,
Merrifield-Simmons index increases too. Thus, we add two vertices between Ki
for i = 1, 2, . . . , k by Lemma 2. In order to attain k recurrence relations, we use
k − 1 stars S4 and K1 which has n − [2(k − 1) + 4(k − 1)] = n − 6k + 6 vertices
and K1 = Sn−6k+6 .
As a result of Theorem 3, we can express:
Corollary 1. By Theorem 3 among all trees in family Tn,2 , T5 has maximal

Merrifield-Simmons index which is depicted in Fig. 5.
Fig. 4. The tree T4 .
Corollary 2. Among all trees in family Tn,2 , T6 has minimal Merrifield-

Simmons index which is depicted in Fig. 6. Because this problem is similar to
minimal values of the Merrifield-Simmons index of trees with two branch vertices
in [7].

1634 B. Şahin
3 Conclusion
In this study, we characterize the trees whose Merrifield-Simmons index can be
computed in exactly k steps and we obtain some extremal trees with respect to
this characterization. The Hosoya index (which is equal to the total number of
matchings of a graph G) of a graph G is computed with recurrence relation and
an other generalization can be obtained for Hosoya index.
Acknowledgement. The author thanks Professor Audace Dossou-Olory for his con-
tributions in this paper.
References
1. Merrifield, R.-E., Simmons, H.-E.: Topological Methods in Chemistry. Wiley, New
York (1989)
2. Wagner, S., Gutman, I.: Maxima and minima of the Hosoya index and the
Merrifield-Simmons index. Acta Appl. Math. 112, 323–346 (2010)
3. Yu, A., Lv, X.: The Merrifield-Simmons indices and Hosoya indices of trees with
k pendant vertices. J. Math. Chem. 41(1), 33–43 (2007)
4. Li, X., Zhao, H., Gutman, I.: On the Merrifield-Simmons index of trees. MATCH
Commun. Math. Comput. Chem. 54, 389–402 (2005)
5. Heuberger, C., Wagner, S.: Maximizing the number of independent subsets over
trees with bounded degree. J. Graph Theory 44, 49–68 (2008)
6. Tian, W., Zhao, F., Sun, Z., Mei, X., Chen, G.: Ordering of a class of trees with
respect to the Merrifield-Simmons index and the Hosoya index. J. Comb. Optim.
38(2), 1286–1295 (2019)
7. Cruz, R., Marin, C.-A., Rada, J.-P.: Extremal values of Merrifield-Simmons index
for trees with two branching vertices. Kragujevac J. Math. 42(1), 97–106 (2018)
8. Boutiche, M.-A., Belbachir, H., Gutman, I.: Merrifield-Simmons and Hosoya index
of thorn-regular graphs. Sci. Publ. State Univ. Novi Pazar. Ser. A: Appl. Math.
Inform. Mech. 8, 113–129 (2016)
9. Li, S., Zhu, Z.: The number of independent sets in unicyclic graphs with a given
diameter. Discret. Appl. Math. 155, 2294–2302 (2007)
10. Pedersen, A.-S., Vestergaard, P.-D.: The number of independent sets in unicyclic
graphs. Discret. Appl. Math. 152, 246–256 (2005)
11. Gutman, I., Gültekin, İ., Şahin, B.: On Merrifield-Simmons index of molecular
graphs. Kragujevac J. Sci. 38, 83–95 (2016)
12. Gutman, I., Polansky, O.-E.: Mathematical Concepts in Organic Chemistry.
Springer, Berlin (1986)
A Fuzzy Goal Programming Model to Evaluate
Green Energy Alternatives
Beyzanur Cayir Ervural(&)

Konya Food and Agriculture University, Konya, Turkey
[email protected]
Abstract. Many practical and multi-faceted decision problems may not be

exactly identified in an uncertain environment due to a vague understanding of
the system. Fuzzy goal programming is one of the most prominent optimization
tools for complex energy planning problems. Green energy sources, which are
alternative to fossil energy sources, are strongly favoured because of their
sustainable, effective and reliable structures. However, it is necessary to develop
optimal investment plans by analyzing and evaluating the existing green energy
resources from a social, environmental and economic perspective. In this study,
we propose a practical decision model-fuzzy goal programming approach-
which combines the optimal allocation of green power plants to obtain social,
environmental and economic components according to Turkey’s 2023 energy
goals under the fuzzy environment. The proposed model makes strategically
important decisions to meet long-term energy needs on the basis of multiple
objectives.
Keywords: Fuzzy goal programming Green energy alternatives Energy

planning
1 Introduction
Uncertainty arises when a decision can have multiple consequences, including multi-
dimensional domains found in complex systems such as energy planning, resource
allocation, and energy network design [1]. Optimal and correct management of
renewable energy sources is a critical issue that is followed all over the world. Effective
assessment of renewable energy potentials is among the energy policies published by
the Ministry of Energy and Natural Resources (MENR) in Turkey [2].
In general, energy planning has a multidimensional and complex structure because
of the multiple and conflicting objectives. Energy plans include multifaced decisions
regarding the optimal mix of the various power plant types that each facility needs to be
installed, and capacity expansion decisions for each facility over a period of time.
Multi-criteria decision-making methods are widely applied in such complex energy
planning problems.
Fuzziness is a kind of uncertainty that takes place in the decision-making process,
principally expressing in linguistic terms, which cannot be statistically described in
nature and cannot be characterized by a classical probability distribution. Fuzzy logic
1636 B. C. Ervural
adaptation with a mathematical programming approach is an attractive and successful

solution alternative used by scientists and practitioners. When the literature is analyzed,
it is seen that energy decision-making problems are mostly based on linear program-
ming models and multi-objective decision-making methods, and also fuzzy models are
very limited [3–6]. The main disadvantage of the traditional goal programming
approach is that all the determinants must be exactly known in the decision-making
process. Especially in real life problems such as energy problems, deterministic data are
often not available due to the changing and uncertain nature of energy market condi-
tions. A fuzzy goal programming model provides an effective and rational solution to
achieve specific targets for indefinite conditions.
In this study, a fuzzy goal programming model is proposed for a renewable energy
investment planning in Turkey based on twenty-one regions, at a country level, which
includes fuzzy objective functions under several constraints. In the model, three
objectives are considered, which minimize the investment cost, maximize the utiliza-
tion rate of each renewable energy potential and try to achieve the renewable energy
generation goals. The study aims to make contributions in the following ways. It is the
first application of a fuzzy goal programming model in a renewable energy planning
problem at a country level, in Turkey. In addition to its methodological approach and
strategic results, it also offers the opportunity to evaluate the real-life situation at the
country level in terms of unique constraints and fuzzy objective functions. And with
adjusting aspiration levels according to decision maker’s expectations, different and
more convenient results/scenarios can be obtained.
The rest of the paper is organized as follows. In the next section, the fuzzy goal
programming approach is explained in detailed. In the third section, the proposed fuzzy
goal programming model is clarified and the obtained results are discussed in the fourth
section. Finally, conclusions and future directions are provided in the fifth section.
2 Background Information
2.1 Forming of Membership Function

Let blt be the lower tolerance level for the success of the aspiration level at of the tth
fuzzy goal. The membership function lt ð X Þ; and the fuzzy goal Ft ð X Þ can be specified
as follows [7]: lt ð X Þ describes the following formulation under % the type of
restriction
8
<1 if Ft ð X Þ at ;
Ft ð X Þ ðat blt Þ
lt ð X Þ ¼ if at blt Ft ð X Þ \ at ; ð1Þ
: blt
0 if Ft ð X Þ \ at blt ;
2.2 Fuzzy Goal Programming Formulation

The weighted fuzzy goal programming model consists of a T number of goals with
weight wt membership is to be maximized [8, 9].
A Fuzzy Goal Programming Model 1637
Maximize k
k Ft ð xÞ; t ¼ 1; . . .; T ð2Þ
Cx ct
x0
where
X
T X
T
k¼ wt lt for wt ¼ 1
t¼1 t¼1
where wt is the relative importance of fuzzy goals and lt is a membership function of

each objective function.
3 The Proposed Fuzzy Goal Programming Model
3.1 Definition of Variables and Parameters:

The energy planning model is a classical multi-criteria decision-making problem. In
this study, the following index, decision variable and parameters are considered.
Index
i : type of renewable power plant I : fWind; Solar; Hydro; Biomass; Geotermalg

j; t : regions J : f1; . . .; 21g
Decision variables

1 if power plant i will be installed in region j
xij ¼
0 otherwise
yijt ¼ The amount of energy generated in region j using

renewable power plant i to be used in region t
Parameters
Ginv The goal for investment of each renewable power plant ($)
Gpowp
i The goal for power production of power plant i (GWh)
Gpot The goal for the utilization of renewable energy potentials of each region (%)
Dt Energy demand of region t (GWh)
Kij Capacity of renewable power plant i in region j (GWh)
Oi Initial investment cost (fixed cost) of renewable power plant i ($ per year)
ci Variable cost per unit of electricity for renewable power plant i ($ per year)
e Inflation rate, N Number of periods (20 years)
1638 B. C. Ervural
xij is a binary variable and if power plant i is installed in region j it equals to 1, yijt is a
continuous variable and it means that the amount of energy generated in region j using
renewable power plant i to be used in region t.
3.2 Fuzzy Goal Descriptions

The evaluation parameters are obtained from MENR reports and the published inter-
national energy reports. In the study, the power production goals of each renewable
power plant are calculated using Turkey’s 2023 Strategic Energy Plan. In detail, the
energy demand of each region is formed by considering the population ratio of the
regions, according to energy demand projections calculated by MENR [2]. The power
production goal of power plants determined by Turkey’s investment plans, and depends
on the capacity building rates of the Turkish Electricity Transmission Corporation. The
data of the renewable energy potential is obtained from the Ministry of Energy.
!!
X X X ð 1 þ eÞ N 1 X X
y ð ci Þ
t ijt
þ y O
j ijt i
. Ginv ð3Þ
i j e ð1 þ eÞN i
Investment cost consists of initial investment cost and the unit cost of generating
electricity per unit of GWh., and which is also known as levelized cost of energy. The
cost of generating per unit electricity includes operating, maintenance, rehabilitation
and fuel costs. Constraint 3 gives a fuzzy investment objective and it should be cost-
effective for an economic energy investment plan. The utilization of renewable energy
potentials should be maximized in order to achieve the best use of available resources,
and Constraint 4 presents the energy potential objective. The power production goal of
each type of renewable energy resource should be ensured according to energy policy
goals and Constraint 5 indicates power generation goals for each type power plant.
Constraint 6 provides the capacity of each power plant in regions and Constraint 7
gives the energy demands of each region. Constraint 8 provides sign restrictions related
to the decision variables. The objective function in Eq. 9 aims to maximize all of the
determined goals with regarded weighted memberships. The goals have been consid-
ered according to different importance levels based on expert opinions from the energy
sector (w1 ¼ 0:50; w2 ¼ 0:35; w3 ¼ 0:15).
X X X
i j
y EPij J Gpot
t ijt
ð4Þ
X X
j
y
t ijt
J Gpowp
i ð5Þ
X
y
t ijt
Kij xij 8i 2 I; j 2 J ð6Þ
X X
i
y
j ijt
Dt 8t 2 J ð7Þ
xij 2 ½0; 1 yijt 0 l1 ; l2; l3 2 ½0; 1 ð8Þ
Maxðw1 l1 þ w2 l2 þ w3 l3 Þ ð9Þ
In the next section, the membership functions of the identified fuzzy goals and the
exact fuzzy goal programming model of the problem are illustrated with a real case
study of renewable energy planning in Turkey.
Table 1. Goals for investment and energy potentials.

Goal Direction Value Unit
Investment Ginv 8000000406.8 1000*$
Renewable Gpot Wind Solar Hydro Biomass Geothermal %
energy 0.0949 0.4287 0.1195 0.2054 0.1514
potential
Table 2. Goals for power production of each renewable power plants.

Renewable power plant Goal Direction Value (GW)
Wind (W) Gpowp
wind 35948.185
Solar (S) Gpowp
solar
6499.976
Hydro (H) Gpowp
hydro ¼ 4420.014
Biomass (B) Gpowp
bio ¼ 4606.502
Geothermal (G) Gpowp
geo 810.982
Table 3. Data descriptions of the fuzzy goals.

Goal Aspiration level Tolerance limit
Lower Upper
Investment (1000*$) 8000000406.8 9500000000.8
Renewable energy potential (%)
Wind 0.0948 0.0785
Solar 0.4287 0.3012
Hydro 0.1195 0.0802
Biomass 0.2054 0.1215
Geothermal 0.1514 0.1210
Power production (GW)
Wind 35948.185 32948.152
Solar 6499.976 6300.875
Hydro 4420.014 3090.892
Biomass 4606.502 4501.463
Geothermal 810.982 700.602
1640 B. C. Ervural
All required data has been presented in the following tables (Table 1–Table 3). The
data for the aspiration levels and tolerance limits of fuzzy goals are given in Table 3.
The data in Table 1 and Table 2, and the membership functions of the identified
fuzzy goals can be built by using Eq. 1–Eq. 2. The membership function of the fuzzy
investment cost for the upper tolerance limit constructed as follows:
Investment
!!
XXX ð 1 þ eÞ N 1 XX
yijt ðci Þ N þ yijt Oi . 8000000406:8
i j t e ð1 þ eÞ i j
P P P P P
ð1 þ eÞN 1
9500000000:8 i j t yijt ð ci Þ eð1 þ eÞN
þ i j yijt Oi
l1 ¼
1499999594:0
Other membership functions of fuzzy goals in the modal can be defined with same
systematic way. The membership goals of the determined membership functions are
given as follows:
Energy Potential
The energy potential of each energy types is given as follows with an example:
X
21 X
21
y1jt EP1j J 0:0948
j t
The constructed membership values of the fuzzy energy potential goal is given as
follows with an example:
P P
21 21
j t y1jt EP1j 0:0785
l2 ¼
0:0163
Likewise, the membership functions can be defined for each fuzzy goal in the
model, and then the membership goals of the characterized membership functions are
created as follows:
Power Production
The amount of each type power generation and the related membership goal is given
with an example as follows:
P P
XX ð j t y1jt Þ 32948:152
y1jt J 35948:185 l3 ¼
j t
3;000:033
4 Results and Discussions
GAMS software (Version 23.2) was performed, and the defined fuzzy goals are exe-
cuted according to their significance levels. The obtained results of the decision vari-
ables are given in Table 4, and the achievement degree of fuzzy goals acquired as
highly important (kmax ¼ 0:9376). Table 4 presents the amount of renewable energy
production and the region where each plant should be installed. The analysis results
with details can be discussed as follows: Membership function values are obtained as
follows 0.876, 0.998, 0.998, 0.998, 0.998, 0.998, 1.00, 1.00, 1.00, 1.00, 1.00. In more
detail, the investment cost is not satisfied with the desired goal due to expectation of all
multiple conflicting fuzzy objectives at the same time. When the obtained results are
examined in more detail, it is seen that the most installed renewable energy type is the
wind power. The main reason for allocating more wind power plant is that the country
has higher wind energy potential and the investment cost of wind energy is lower than
the others.
If the developed model results do not meet the experts expectation, the minimum
aspirations can be changed and the model can be resolved under the determined goals.
Moreover, the weights of the goals can be changed for different cases in order to
analyze behaviour of the model with various situations. The developed model provides
high flexibility for energy investment decision model when the goals are not clearly
defined.
Table 4. The amount of energy production and regions to install the power plant.
Power Region Region Yijt Power Region Region Yijt
plant (j) (t) plant (j) (t)
type (i) type (i)
1 1 19 1335.33 2 3 16 177.48
1 2 18 1170.94 2 4 18 155.91
1 3 17 232.69 2 7 18 116.71
1 4 17 608.44 2 9 16 909.59
1 5 17 607.18 2 11 11 66.13
1 6 17 2745.86 2 14 12 786.77
1 7 16 112.85 2 16 6 17.9
1 8 16 654.93 2 17 19 976.14
1 9 16 1400.55 2 18 6 151.19
1 10 1 2390.77 2 19 11 165.29
1 10 14 2025.78 2 20 10 1490.03
1 10 15 819.73 2 20 16 1238.73
1 11 10 1709.35 3 3 17 1337.19
1 12 10 388.42 3 6 11 525.94
1 13 5 3621.39 3 9 13 839.57
1 14 5 1258.42 3 16 11 60.56
(continued)
1642 B. C. Ervural
Power Region Region Yijt Power Region Region Yijt
plant (j) (t) plant (j) (t)
type (i) type (i)
1 15 11 6.97 3 19 21 331.76
1 16 5 320.21 4 2 16 281.69
1 17 5 1496 4 3 18 386.58
1 17 19 1473.49 4 5 18 281.69
1 18 4 4869.51 4 12 6 397.39
1 18 15 359.21 4 16 21 36.58
1 19 2 3872.68 4 17 19 616.79
1 19 15 1293.48 4 18 18 758.68
1 19 20 2.84 4 20 14 593.89
1 20 13 56.92 4 21 5 418.77
1 21 10 1764.03 4 21 13 788.42
2 1 1 19.82 5 10 5 700.98
2 2 5 34.28
5 Conclusion
Uncertainty and complexity in the nature of energy problems require the use of fuzzy
approaches. In an imprecise environment, fuzzy goal programming models help to
solve serious, critical situations rationally since the working principle of the models
allow the model to relax with tolerance limits and aspiration levels.
Renewable energy investments are long-term strategic plans with their high added
values. In this study, renewable energy investment planning is developed with a fuzzy
goal programming approach. Five different renewable power plants are installed in
twenty-one regions based on the determined fuzzy goals. According to the obtained
results, the most installed renewable energy type is the wind power due to its cost and
wind energy potential. The proposed model shows that fuzzy goal programming is a
helpful, practical and efficient decision-making tool for complex problems.
For future studies, the model can be expanded by adding new constraints and goals
according to various importance levels. Adapting other multi-objective approaches to
the model may also provide more comprehensive compromise solutions.
References
1. Sadeghi, A., Larimian, T., Molabashi, A.: Evaluation of renewable energy sources for
generating electricity in province of Yazd: a fuzzy MCDM approach. Procedia - Soc. Behav.
Sci. 62, 1095–1099 (2012)
2. MENR, Republic of Turkey Ministry of Energy and Natural Resources. https://www.enerji.
gov.tr/en-US/Mainpage. Accessed 17 Oct 2019
3. Sadeghi, M., Mirshojaeian Hosseini, H.: Energy supply planning in Iran by using fuzzy linear
programming approach (regarding uncertainties of investment costs). Energy Policy 34(9),
993–1003 (2006)
4. Huang, Z., Yu, H., Chu, X., Peng, Z.: A goal programming based model system for
community energy plan. Energy 134, 893–901 (2017)
5. Bitar, S.D.B., Costa Junior, C.T., Barreiros, J.A.L., Neto J.C.L.: Expansion of isolated
electrical systems in the Amazon: an approach using fuzzy multi-objective mathematical
programming. Energy Policy 37(10), 3899–3905 (2009)
6. Bilbao-Terol, A., Arenas-Parra, M., Cañal-Fernández, V.: A fuzzy multi-objective approach
for sustainable investments. Expert Syst. Appl. 39(12), 10904–10915 (2012)
7. Pal, B.B., Moitra, B.N., Maulik, U.: A goal programming procedure for fuzzy multiobjective
linear fractional programming problem. Fuzzy Sets Syst. 139(2), 395–405 (2003)
8. Lin, C.C.: A weighted max-min model for fuzzy goal programming. Fuzzy Sets Syst. 142(3),
407–420 (2004)
9. Yang, T., Ignizio, J.P., Kim, H.J.: Fuzzy programming with nonlinear membership functions:
Piecewise linear approximation. Fuzzy Sets Syst. 41(1), 39–53 (1991)
Application of Fuzzy Logic Model for Correct
Lighting in Computer Aided
Interior Design Areas
Rahib Imamguluyev(&)
Odlar Yurdu University, Baku, Azerbaijan

[email protected]
Abstract. Proper lighting is one of the important things in architectural design

areas. Designers use a variety of software to design the interior of any area (e.g.
AutoCAD, 3DsMax, SketchUp, Revit, etc.). Typically, designers determine the
location lighting system and the number of bulbs prepared by the program.
Problems arise when the design space provided by the designer is properly
illuminated. To avoid these problems, it is necessary to pay attention to the
following: 1. LUX definition by location, 2. Calculation of reflection coefficients
according to color tones, 3. Identification of chandeliers and lamps, 4. Calcu-
lation of light flow (Lumen), 5. Calculation of bright intensity, 6. Calculation of
the room index, 7. Lighting account. This article explores how to properly and
efficiently illuminate daylight and artificial illumination in the interior design of
3Ds MAX according to a fuzzy logic model based on color tones. Based on the
values obtained, calculations were made in the Fuzzy Toolbox section of Matlab
and the results were obtained with lighting suitable for the interior design
developed by 3DsMax.
Keywords: Fuzzy logic Lighting Daylighting Color Interior design

CAD 3DsMax
1 Introduction
As is known, light is a form of energy that affects our eyes and creates a sense of vision.
This energy is defined by the theories developed by two independent conceptions [1–4].
• Light sources are classified in different systems in various publications. Light
sources in terms of light production are as follows.
• Primary light sources; are objects that can emit light on their own (sun, candle,
incandescent lamp, etc.)
• Secondary light sources; they are objects that emit light by reflecting or passing the
light they receive from primary light sources, (moon, atmosphere, window, wall
surface, etc.) Another classification is made according to the geometric shapes of
the light sources;
• Point light sources,
• Linear light sources,
• They are listed as superficial light sources.
Application of Fuzzy Logic Model 1645
When light sources are classified according to the source of light;

• Natural light sources, sun, sky, windows, etc.
• Artificial light sources, candle, incandescent lamp, etc. We see that it meets in two
main groups.
• The eyes’ eyesight increases (visual acuity, eyesight increases).
• Eye health is protected, visual impairments are prevented.
• As the visual performance will increase, the efficiency of the work will increase,
thus providing economic benefits.
• Visual comfort is also provided in psychological terms. The beneficiary feels
happier in his /her environment.
• Accidents that may arise from inability to see well or to be blind are reduced.
• A sense of security is provided.
Fig. 1. Sample lighting.
1.1 Indoor Lighting
Light Flux Definition

Luminous flux Lumen (lm); is the total amount of light that a light source gives in
every direction. It is the part of the electrical energy supplied to the light source that
turns into light energy [6, 7]. We can also say the efficiency of the armature used. The
light flow is denoted by the letter i.
Definition of Luminous Intensity
The sum of the luminous flux per unit surface is called luminous intensity. It indicates
the level of light given by a light source in every direction. The unit of luminous
intensity is luxury.
1646 R. Imamguluyev
1.2 Lighting Types
• Natural Lighting: It can be defined as the lighting system designed to meet the
visual comfort needs of daylight, the main source of which is the sun.
• Artificial Lighting: It can be defined as the lighting system designed to meet the
visual comfort needs of the light produced from artificial light sources,
• Integrated Lighting: It can be defined as the lighting system in which light is used as
a supplement in meeting the visual comfort requirements in cases where daylight is
insufficient.
• In the classification according to the place of illumination, we see that lighting is
handled in two types, we can see this classification in many sources written on this
subject.
• Indoor Lighting: It deals with the lighting system of the interior spaces, separated
from the external environment by various structural elements.
• Outdoor Lighting: Building outside the building is the subject of the illumination
system of various sizes.
2 Calculation of Indoor Lighting

2.1 Lighting Account
Various data and calculations are used for correct lighting. The tables given to you here
are in the form of sampling. With these data, you can make more specific applications.
Illumination calculations and factors of a place are explained step by step below [1, 6].
2.2 Reflection Coefficients of Important Substances

In the table below, the reflection coefficients of some materials according to their color
status are given. You can use these coefficients in the calculation [3, 7] (Table 1).
Table 1. Reflection coefficients of some materials and wall colors.

Material % Wall paints %
Dark 0.10–0.20 Oak light color 0.25–0.35
Brown 0.60–0.70 Chipboard cream color 0.50–0.60
Light yellow 0.45–0.55 Plaster 0.90
Light green 0.30–0.50 Anodized aluminum 0.85
Light red 0.35–0.45 Concrete 0.10–0.50
Sky blue 0.70–0.90 Glass-silver-moon 0.85–0.90
White 0.45–0.55 Granite 0.20–0.25
Pink 0.40–0.60 White marble 0.60–0.65
Light grey 0.20–0.30 Lime whitewash 0.40–0.45
Table 2. Room lighting efficiency according to k values µ.

Ceiling 0.80 0.50 0.30
Wall 0.50 0.30 0.50 0.30 0.10 0.30
Floor 0.30 0.10 0.30 0.10 0.30 0.10 0.30 0.10 0.10 0.10
Room index Room lighting efficiency factor (µ)
k ¼ h*a*b
ða þ bÞ
0.60 0.24 0.23 0.13 0.18 0.20 0.19 0.15 0.15 0.12 0.15
0.80 0.31 0.29 0.24 0.23 0.25 0.24 0.20 0.19 0.16 0.17
1.00 0.36 0.33 0.29 0.28 0.29 0.28 0.24 0.23 0.20 0.20
1.25 0.41 0.38 0.34 0.32 0.33 0.31 0.28 0.27 0.24 0.20
1.50 0.45 0.41 0.38 0.36 0.36 0.34 0.32 0.30 0.27 0.26
2.00 0.51 0.46 0.45 0.41 0.41 0.38 0.37 0.35 0.31 0.30
2.50 0.56 0.49 0.50 9.45 0.45 0.41 0.41 0.38 0.35 0.34
3.00 0.59 0.52 0.54 0.48 0.47 0.43 0.43 0.40 0.38 0.36
4.00 0.63 0.55 0.58 0.51 0.50 0.45 0.47 0.44 0.41 0.39
5.00 0.66 0.57 0.62 0.54 0.53 0.48 0.50 0.46 0.44 0.40
2.3 Lighting Account Formulas

k = room index
a = short edge length of the room
b = long edge of the room
H = height between armature and working surface
ab
k¼ ð1Þ
H ða þ bÞ
Using these values, the room index k value is calculated. When each luminaire is
installed in place, the lamp is usually 20 to 60 cm below the ceiling. B type luminaires
do not have a distance, but for example for chandeliers, rod length can be taken as
60 cm. The working plane is generally considered to be 70–80 cm from the table
length. In the sitting plane, 50–60 cm can be considered.
3 Color Temperature Selection
In fact, the color temperature expressed as Kelvin has nothing to do with physical heat
for the user [3]. On the contrary, in summer time when the bright blue lights of the sun
are effective, the color temperature of daylight is defined as cold white. We call it
orange-yellow or amber-colored light, which is given by candlelight, as warm white. In
artificial light sources, the color temperature is generally in the range of 2500 to
8500 K degrees.
1648 R. Imamguluyev
3.1 Warm White Light (2500–3300 K)

Color temperatures of 2500–3300 K are considered as “warm light”. Red, orange,
yellow and brown are generally earthy colors used in homes, warm light colors are
suitable for these spaces. Warm light generally provides a softer, better vision and a
comfortable atmosphere than bright cold light. For this, it is used in bedrooms and
living areas.
3.2 Natural White Light (4000–4500 K)

Although warm light is generally accepted in homes, some people may prefer natural
white light. Natural white light is at 4000–4500 K color temperature. Natural white
light is preferred in environments decorated with green, white and blue tones. It is
preferred in offices and home offices since it creates an energetic and refreshing mood
for people and employees.
At the same time, light in the color temperature of 4000–4500 K supports us to see
detail and makes it easier to see imperfections in objects. For this reason, the places
where cleaning is important is beneficial for the lighting in the makeup rooms.
3.3 Cool White Light (5000–6500 K)

Technically, 5000–6500 K light is the degree of kelvin of summer sun. So with such a
light, you imitate the sun in lighting. Due to the excess blue light in 6500 K light, risk
begins to occur for our eyes. For this reason, it is necessary not to stay in these
environments for too long. In working environments, cool white light can increase
work efficiency, it is used in offices, garages, jewelery stores, leather and shoe stores, in
displaying silver objects, in sports areas, outside of residences.
It is known that psychologically warm white light feels one to two degrees higher in
the ambient temperature. Accordingly, according to sales statistics, preferences for the
warm white light in the northern countries and the cold white light increase as they
approach the equator region. One final note, if you are having insomnia, do not prefer
cool white light in the evening, even if you love it, be illuminated with warm white
light. Thus, each designer can select the light source needed when designing the
interior as 3dsmax program and change the Color temperature value in the intensity
section on the right (see Fig. 2).
Fig. 2. Color temperature selection in 3ds max program.

4 Fuzzy Logic and Lighting Control
Daylight are a dynamic source of lighting and changes in daylight, season, location or
latitude and cloudiness. Different levels of skylight can be found in the same sunlight,
and even when the sky pattern remains, the range of solar lighting range of an instant
turbidity filter or scatter is above the sun. As a result, the system must be flexible to
allow any prediction multivariate system changes and skylight that characterize the
combination of sunlight [7, 9]. In recent years, control technology has been well
developed, and the vehicles in the most successful industry. However, due to the above
mentioned aspects, traditional control systems, mathematical models, daylight energy
management controls its limits [2, 5, 6, 10]. Taking into account the potentially existing
random pattern, rapid change of daylight and artificial illumination properties, fuzzy
control has proven to be a more suitable solution.
In our case, taking into account the windows head height, the pattern of the daylight
is presented in (see Fig. 3); accordingly, four control zones parallel to the short side of
the room have been identified.
Fig. 3. Daylighting of the room.
5 Practical Application of Fuzzy Technologies
When both inputs and outputs are used with membership functions, the conventional
expressions between them are constructed as follows: If (input UF x), then (output UF y).
For example, “if the temperature is high, the value of success is very low” Fig. 1. In the
general approach, the model can have more than one input parameter and must be
combined with AND, OR, NOT. Parameters and rules are unstable and can be changed
under different conditions, which increases the use of the model. In a fuzzy process, each
rule is evaluated to determine the output, and as a result, the Fuzzy Information System is
the average of all outcomes [8]. In this way, each design work is done to a certain extent,
which shows that it is extremely successful in terms of time and quality. This value is in
the range [0–8000]. When the process is complete, a list is created and can be used
visually.
1650 R. Imamguluyev
Every linguistic variable has five fuzzy values with triangular or trapezoid mem-
bership functions, as follows:
• For input variables – Fig. 4: VH-veryhot; H – hot; N – natural; C – Cold; VC –
verycold
• For output variables – Fig. 5: VL – very low; L – low; M –medium; H – high; VH –
very high.
Fig. 4. Input variables fuzzification.
Fig. 5. Fuzzification of output variables.
The knowledge base used by the control system is given in Table 3, where li
(i = 1…4) represents the membership functions corresponding to the four control
zones.
Table 3. Membership functions.

IF THEN
A B C l1 l2 l3 l4
VH VH VH V H V H VH VH
VH VH H V H V H VH VH
VH VH N V H V H VH VH
… … … … … … …
L L L VL VH VL VL
The processing stage invokes each appropriate rule and generates a result for each
of them, then combines the results of the rules; this mechanism was implemented by
the max-min inference method.
6 Conclusion
The evaluation of the alternatives given in Table 2 in the MATLAB\Fuzzy Develop-

ment Systems note described above allowed the following results to be obtained:
Alternatively, in the 3dsMax program, we get the color temperature more accurately by
taking the color temperature value in the range 0–8000 instead of 3500, 6500, 8000.
The results obtained are more rhetorical, showing that it is effective to estimate the
determination of a color temperature based on a fuzzy set theory under conditions of
uncertainty.
On the other hand, it is useful to use fuzzy set theory when determining lighting
instead of a simple goal method based on perfect interior design, taking into account
other simple but vague and difficult shaping criteria. and making decisions based on
them is reasonable and economically feasible.
References
1. Cziker, A., Chindris, M., Miron, A.: Fuzzy controller for indoor lighting system with
daylighting contribution. In: ELECO 2007 5th International Conference on Electrical and
Electronics Engineering, Turkey (2007)
2. Cao, C., Zhu, X.: Energy management using optimal fuzzy logic control in wireless sensor
network. Int. J. Online Biomed. Eng. 14(9), 35–52 (2018)
3. Rea, M.S.: IESNA Lighting Handbook, 9th edn. In: Illuminating Engineering (2000)
4. Zhang, L., Yu, Y., Ma, H., Zhang, Y., Cao, P.: Design of photovoltaic power supply MPPT
circuit for WSN node based on current observation. Int. J. Online Eng. 14(7), 45–61 (2018)
5. Cziker, A., Chindris, M., Miron, A: implementation of fuzzy logic in daylighting control. In:
11th International Conference on Intelligent Engineering Systems, INES, pp. 195–200
(2007)
6. Görgülü, S., Ekren, N.: Energy saving in lighting system with fuzzy logic controller which
uses light-pipe and dimmable ballast. Energy Build. 61, 172–176 (2013)
7. Lighting Controls Association, Introduction to Lighting Automation. http://lightingcontrol-
sassociation.org/2005/02/14/introduction-to-lighting-automation-for-the-home/. Accessed 30
Apr 2020
8. Zadeh, L.A.: Fuzzy logic, neural networks, and soft computing. Commun. ACM 37, 77–84
(1994)
9. IEA Task 21, Application Guide for Daylight Responsive Lighting Control. https://www.iea-
shc.org/Data/Sites/1/publications/8–8-1%20Application%20Guide.pdf. Accessed 30 Apr
2020
10. Daylighting Control, Design and Application Guide. https://www.legrand.us/-/media/brands/
wattstopper/resources/application-guide/ws-appguide-daylightingcontrol.ashx. Accessed 30
Apr 2020
Review and Discussion Papers
Extensions of Ordinary Fuzzy Sets:
A Comparative Literature Review
Cengiz Kahraman1(&), Basar Oztaysi1, Irem Otay2,

and Sezi Cevik Onar1
1
Macka, Besiktas, 34367 Istanbul, Turkey
[email protected]
2
Department of Industrial Engineering, Istanbul Bilgi University,
Istanbul, Turkey
Abstract. Fuzzy sets extensions have been often used in the modeling of
problems including vagueness and impreciseness in order to better define the
membership functions together with the hesitancy of decision makers. More
than 20 different extensions of ordinary fuzzy sets have appeared in the literature
after the introductions of ordinary fuzzy sets by Zadeh (1965). These sets
involve interval-type fuzzy sets, type-2 fuzzy sets, hesitant fuzzy sets, intu-
itionistic fuzzy sets, Pythagorean fuzzy sets, q-rung orthopair fuzzy sets,
spherical fuzzy sets, picture fuzzy sets, fermatean fuzzy sets, etc. Mainly, these
extensions can be divided into two classes: extensions with two independent
membership parameters and extensions with three independent membership
parameters. In this paper, we briefly classify these extensions and present some
comparative graphical illustrations.
Keywords: Fuzzy sets Extensions Intuitionistic fuzzy sets Pythagorean

fuzzy sets Picture fuzzy sets q-rung orthopair fuzzy sets Spherical fuzzy sets
1 Introduction
It is about 55 years journey from the introduction of ordinary fuzzy sets to today
including more than 20 extensions. Each of these extensions has been introduced to
correctly define the thoughts of decision makers on membership functions with more
details than the others. We can divide into two major groups these extensions: the
extensions defining the hesitancy of experts depending on membership and non-
membership degrees, and the extensions defining the refusal degree of experts
depending on membership, non-membership, and hesitancy degrees. Thus, the first
group is based on two independently assigned degrees whereas the second group is
based on three independently assigned degrees. There are other types of fuzzy set
extensions which are generally based on ordinary fuzzy set principles such as hesitant
fuzzy sets, type-2 fuzzy sets, fuzzy multi-sets, non-stationary fuzzy sets, etc.
The first group of extensions involves intuitionistic fuzzy sets and their secondary
products while the second group of extensions is based on neutrosophic sets and their
secondary products. The secondary products of the first group are Pythagorean fuzzy
sets, fermatean fuzzy sets, and q-rung orthopair fuzzy sets. The secondary products of
the second group are picture fuzzy sets, spherical fuzzy sets, and t-spherical fuzzy sets.
Figure 1 illustrates the extensions of fuzzy sets with a historical order.
Fig. 1. Extension of ordinary fuzzy sets
The rest of the paper is organized as follows. Section 2 summarizes the basic
definitions of these sets together with graphical illustrations. These graphs provide a
comparative analysis among the extensions. Section 3 concludes the paper with future
directions and suggestions.
2 Extensions of Fuzzy Sets: Definitions and Literature

Review
2.1 Group 1: Hesitancy Depending on Membership and Non-

membership
2.1.1 Intuitionistic Fuzzy Sets
Intuitionistic fuzzy set: Let X 6¼ ; be a given set. An intuitionistic fuzzy set in X is an
object A~ given by

~¼
A x; lA~ ð xÞ; vA~ ð xÞ ; x 2 X ð1Þ
where lA~ : X ! ½0; 1 and vA~ : X ! ½0; 1 satisfy the condition 0 lA~ ð xÞ þ vA~ ð xÞ 1;
for every x 2 X. The hesitancy of a decision maker is calculated as pA~ ¼ 1 lA~ ð xÞ
vA~ ð xÞ. Figure 2 shows the frequencies of IFS publications with respect to the years while
Fig. 3 illustrates the distribution of subject areas of IFS publications.
Extensions of Ordinary Fuzzy Sets 1657
Fig. 2. Frequencies of IFS publications by years
Fig. 3. Subject areas of IFS publications
2.1.2 Pythagorean Fuzzy Sets (PFS)

Yager (2013) introduced the Pythagorean fuzzy sets (PFS). Let X be a universe set.
~ is an object having the form:
A single-valued PFS P

~¼
P x; lP~ ð xÞ; vP~ ð xÞ ; x 2 X ð2Þ
where the function lp : X ! ½0; 1 defines the degree of membership and vp : X !

½0; 1 defines the degree of non-membership of the element x 2 X to P, respectively,
and for every x 2 X, it holds that:
0 ðlP ð xÞÞ2 þ ðvP ð xÞÞ2 1 ð3Þ
The degree of hesitancy is given by Eq. (4).

pP ð xÞ ¼ 1 ðlP ð xÞÞ2 ðvP ð xÞÞ2 ð4Þ
Figure 4 shows the frequencies of PFS publications with respect to the years while
Fig. 5 illustrates the distribution of subject areas of PFS publications.
Fig. 4. Frequencies of PFS publications by years
Fig. 5. Subject areas of PFS publications
2.1.3 Fermatean Fuzzy Sets

A single-valued Fermatean Fuzzy Set (FFS) is an object having the form:
~ ¼ f\x; lF ð xÞ; vF ð xÞ; pF ð xÞjx 2 X g

F ð5Þ
where the function lF ð xÞ : X ! ½0; 1 defines the degree of membership and vF ð xÞ :

X ! ½0; 1 defines the degree of non-membership of the element x 2 X to F, respec-
tively, and for every x 2 X, it holds that:
0 ðlF ð xÞÞ3 þ ðvF ð xÞÞ3 1 ð6Þ

pF ð xÞ ¼ 1 ðlF ð xÞÞ3 ðvF ð xÞÞ3 ð7Þ
There are only six papers published on fermatean fuzzy sets: 5 papers in 2019 and 1
paper in 2020. Computers and mathematics are the most used subject areas of FFS.
2.1.4 q-Rung Orthopair Fuzzy Sets

q-Rung Orthopair Fuzzy Sets (q-ROFS), introduced by Yager (2018), are an extension
of Pythagorean fuzzy sets. Their important characteristic is that the sum of the qth
powers of the membership degree and non-membership degree is equal to or less than 1.
A single-valued q-rung orthopair fuzzy set (q-ROFS) is an object having the form:
~ ¼ f\x; lO ð xÞ; vO ð xÞ; pO ð xÞjx 2 X g

O ð8Þ
where the function lO ð xÞ : X ! ½0; 1 defines the degree of membership and vO ð xÞ :

X ! ½0; 1 defines the degree of non-membership of the element x 2 X to O, respec-
tively, and for every x 2 X, it holds that:
0 ðlO ð xÞÞq þ ðvO ð xÞÞq 1 ð9Þ

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pO ð xÞ ¼ 1 ðlO ð xÞÞq ðvO ð xÞÞq ð10Þ
Fig. 6. Frequencies of q-ROFS publications by years
Fig. 7. Subject areas of q-ROFS publications

Figure 6 shows the frequencies of q-ROFS publications with respect to the years
while Fig. 7 illustrates the distribution of subject areas of q-ROFS publications.
2.2 Group 2: Refusal Degree Depending on Membership,

Non-membership, and Hesitancy
2.2.1 Neutrosophic Sets
Smarandache introduced neutrosophic sets in 1999. A neutrosophic set helps to handle
uncertainty and convert it into the fuzzy set theory. Smarandache (1999) defined
neutrosophic sets as (T, I, F) where, percentage of truth in a subset T, the percentage of
indeterminacy in a subset I, and the percentage of falsity in a subset F. (Ye 2014)
introduces a single valued neutrosophic set to handle uncertainty while imprecise,
incomplete, and inconsistent information is very common in the data set.
Let U be a universe of discourse. Neutrosophic set A ~ in U is an object having the
form,

~¼
A u; TA~ ðuÞ; IA~ ðuÞ; FA~ ðuÞ u 2 U ð11Þ
Fig. 8. Frequencies of NS publications by years
Fig. 9. Subject areas of NS publications

where, TA~ is the truth-membership function, IA~ is the indeterminacy-membership

function and FA~ is the falsity-membership function satisfying that 0 TA~ ðuÞ þ IA~ ðuÞ þ
FA~ ðuÞ 3.
Figure 8 shows the frequencies of NS publications with respect to the years while
Fig. 9 illustrates the distribution of subject areas of NS publications.
2.2.2 Picture Fuzzy Sets

Cuong (2014) proposed picture fuzzy sets (PiFS) which are composed of three
parameters with degree of positive membership, neutral membership, and negative
membership.
A picture fuzzy set A on a universe X is an object in the form of
A ¼ fðx; lA ð xÞ; gA ð xÞ; vA ð xÞÞjx 2 X g ð12Þ
Fig. 10. Frequencies of PiFS publications by years
Fig. 11. Subject areas of PiFS publications

where lA ð xÞ 2 ½0; 1 is the degree of positive membership; gA ð xÞ 2 ½0; 1 is the degree

of neutral membership; and vA ð xÞ 2 ½0; 1 is the degree of negative member-
ship. lA ð xÞ; gA ð xÞ; and vA ð xÞ satisfy the following condition:
0 l A ð x Þ þ gA ð x Þ þ v A ð x Þ 1 ð13Þ
The refusal membership degree is equal to 1 lA ð xÞ vA ð xÞ gA ð xÞ.

Figure 10 shows the frequencies of PiFS publications with respect to the years
while Fig. 11 illustrates the distribution of subject areas of PiFS publications.
2.2.3 Spherical Fuzzy Sets
Fig. 12. Frequencies of SFS publications by years
Fig. 13. Subject areas of SFS publications

Membership lS ðxÞ, non-membership vS ðxÞ and hesitancy degrees pS ðxÞ in spherical

fuzzy sets (SFS) are assigned independently satisfying Eq. (14). The definition space is
1/8 of a unite sphere (Fatma Gundogdu and Kahraman, 2019a).
0 l2S ð xÞ þ v2S ð xÞ þ p2S ð xÞ 1 ð14Þ
The degree of refusal is given by Eq. (15).

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
rS ð xÞ ¼ 1 ðlS ð xÞÞ2 ðvS ð xÞÞ2 ðpS ð xÞÞ2 ð15Þ
Figure 12 shows the frequencies of SFS publications with respect to the years while
Fig. 13 illustrates the distribution of subject areas of SFS publications.
2.2.4 t-Spherical Fuzzy Sets

Membership ltS ðxÞ, non-membership vtS ðxÞ and hesitancy degrees ptS ðxÞ in spherical
fuzzy sets are assigned independently satisfying Eq. (14).
0 ltS ð xÞ þ vtS ð xÞ þ ptS ð xÞ 1 ð16Þ
The degree of refusal is given by Eq. (15).

r S ð xÞ ¼ 1 ðlS ð xÞÞt ðvS ð xÞÞt ðpS ð xÞÞt ð17Þ
Figure 14 shows the frequencies of t-SFS publications with respect to the years
while Fig. 15 illustrates the distribution of subject areas of t-SFS publications.
Fig. 14. Frequencies of t-SFS publications by years

Fig. 15. Subject areas of t-SFS publications
3 Conclusions
Fuzzy set extensions try to define membership functions better than the previous
definitions. Hence, it seems to be that new extensions will continue to appear in the
literature in the future. The most popular extension is intuitionistic fuzzy sets and their
secondary products such as Pythagorean fuzzy sets. The second most popular extension
is neutrosophic sets and picture fuzzy sets and their secondary products such as
spherical fuzzy sets, and t-spherical fuzzy sets. Multicriteria decision making appli-
cations such as DEA, AHP, SWOT analysis, and axiomatic design analysis (Otay et al.
2017; Kahraman et al. 2006; Kahraman et al. 2008; Cebi and Kahraman 2010) and
statistical analyses such as process capability analyses and hypothesis testing (Kaya
and Kahraman 2010; Kaya and Kahraman 2009) have been handled by new extensions
of fuzzy sets. In the future, the possible direction of fuzzy set extensions might be the
fuzzification of the existing parameters in the membership functions, making the
process more complex.
References
Atanassov, K.T.: Intuitionistic fuzzy sets. Fuzzy Sets Syst. 20(1), 87–96 (1986)
Atanassov, K.T.: On intuitionistic fuzzy sets and their applications in: actual problems of
sciences, Sofia. Bulg. Acad. Sci. 53, 1 (1989)
Cebi, S., Kahraman, C.: Extension of axiomatic design principles under fuzzy environment.
Expert Syst. Appl. 37(3), 2682–2689 (2010)
Cuong, B.C.: Picture fuzzy sets. J. Comput. Sci. Cybern. 30(4), 409–420 (2014)
Garibaldi, J.M., Ozen, T.: Uncertain fuzzy reasoning: A case study in modelling expert decision
Grattan-Guinness, I.: Fuzzy membership mapped onto intervals and many-valued quantities.
Zeitsehr Math. Logik und Grundlagen Math. 22, 149–160 (1976)
Jahn, K.U.: Intervall-wertige Mengen. Math. Nach. 68, 115–132 (1975)
Kahraman, C., Demirel, N.Ç., Demirel, T., Ateş, N.Y.: A SWOT-AHP application using fuzzy
concept: e-government in Turkey. In: Kahraman, C. (ed.) Fuzzy Multi-Criteria Decision
Making, pp. 85–117. Springer (2008). https://doi.org/10.1007/978-0-387-76813-7_4
Kahraman, C., Gülbay, M., Kabak, Ö.: Applications of fuzzy sets in industrial engineering: a
topical classification. In: Kahraman, C. (ed.) Fuzzy Applications in Industrial Engineering,
pp. 1–55. Springer (2006). https://doi.org/10.1007/3-540-33517-X_1
Kaya, I., Kahraman, C.: Fuzzy robust process capability indices for risk assessment of air
pollution. Stoch. Env. Res. Risk Assess. 23(4), 529–541 (2009)
Kaya, I., Kahraman, C.: Development of fuzzy process accuracy index for decision making
problems. Inf. Sci. 180(6), 861–872 (2010)
Kutlu, G.F., Kahraman, C.: Spherical fuzzy sets and spherical fuzzy TOPSIS method. J. Intell.
Fuzzy Syst. 36(1), 337–352 (2019)
Otay, I., Oztaysi, B., Cevik Onar, S., Kahraman, C.: Multi-expert performance evaluation of
Knowl.-Based Syst. 133, 90–106 (2017)
Smarandache, F.: A Unifying Field in Logics. Neutrosophy: Neutrosophic Probability, Set and
Logic, Rehoboth. American Research Press, Santa Fe (1999)
Torra, V.: Hesitant fuzzy sets. Int. J. Intell. Syst. 25(6), 529–539 (2010)
Yager, R.R.: On the theory of bags. Int. J. Gen. Syst. 13(1), 23–37 (1986)
Yager, R.R.: Pythagorean fuzzy subsets. In: Joint IFSA World Congress and NAFIPS Annual
Meeting (IFSA/NAFIPS), Edmonton, AB, pp. 57–61 (2013)
Yager, R.R.: Generalized orthopair fuzzy sets. IEEE Trans. Fuzzy Syst. 26(5), 1222–1230 (2018)
Ye, J.: Single valued neutrosophic cross-entropy for multicriteria decision making problems. 38
(3), 1170–1175 (2014)
Zadeh, L.A.: Fuzzy sets. Information. Control 8(3), 338–353 (1965)
Zadeh, L.A.: The concept of a linguistic variable and its application to approximate reasoning.
Inf. Sci. 8, 199–249 (1975)
Systematic Literature Review for Work
Sampling with IOT Technology in Retail Store
Operations
Gizem Omeroglu(&) , Abdullah Ertug ,

and Aysegul Ozkavukcu
Migros Ticaret A.Ş., 34758 Istanbul, Turkey

[email protected]
Abstract. In store logistics operations are not standard in every store due to
different physical condition. For that reason, it is difficult to ensure standard
operations and it is even more difficult to measure that operations. Work study is
a technique for using measurement of standard operations. However, work study
method in retail is difficult than manufacturing. This is because in store logistics
operations can be differentiated by store to store, labor to labor and physical
conditions of stores. For instance, a store can be located in 2 floors. If a store has
2 floors, the employee must use elevator etc. This directly effects the method of
operations and results of variance. We prefer to use “Work Sampling” technique
to handle with variation. In this study, publications are searched about work
sampling and IOT technology in the retail sector. Within this scope, literature
has reviewed, and 26 study examined. Whom to authors best knowledge, there
is no study that includes both grocery and work sampling techniques. For the
future research, big data based analysis conducted in order to visualize dyna-
mism of retail operations.
Keywords: Work sampling Retail In-store logistic IOT Efficiency
1 Introduction
Compared with the manufacturing, using work study technique is highly difficult in
retail. Because, in manufacturing sector production are supplied from few production
facilities with the standard process. In retail sector, there are thousands of groceries and
thousands of employees. Operations can be changed by employee to employee, store to
store. For instance, stores can be located in different center such as hospital, school,
university, bus station etc. Physical condition of stores can also change by structural
factors. For this reason, methods of the operations are not standard.
In this study, work sampling method preferred to handle with this variation. Work
sampling is one of the work measurement techniques. Work sampling is a method that
creates outcomes regarding the process by making observations at random times of a
work or process performed as standard. Various statistical tools use in work sampling
method [1]. Work sampling method is very costly for companies which serving in the
retail sector. Because companies in the retail sector have many stores and many
Systematic Literature Review for Work Sampling 1667
employees. For this reason, it is very difficult to establish a standard process like
production companies.
It is quite difficult and costly to perform traditional work sampling method and
ensure continuity it in all stores. In order to solve this problem, advanced technology
tools such as beacon and smart id cards are used in work sampling method for real time
and continuous work measurement.
To the best of authors knowledge, there is no study that includes both grocery and
work sampling techniques with IOT technology. In this study, work sampling method
will be discussed according to new technological tools such as beacon, IOT, in-store
logistic, work study, big data and big data analysis tools. This study is the starting step
of an on-going project. All peer-reviewed publications are reviewed and reported
through systematically.
This paper is organized as follows. Section 2 explains the details of the literature
review. Section 3 presents the results of the literature review with research opportu-
nities that may be a guide for future studies.
2 Literature Review
In this section, publications will be described in four categories. These are work study,
work sampling, IOT in retail and in store logistics. The publications which includes
these four keywords are subjected to systematic research. Within the scope of the study,
26 publications were examined. According to the results of the literature review, tables
were created by keywords and sectors (Tables 1 and 2).
Table 1. Paper number of keywords

Keyword Paper number
Work study 4
Work sampling 9
IOT in retail 9
In store logistics 2
The papers are group by the keywords and all the publications examine according to
Mathematical Model, using IOT tools and Case Study. Table 3 is indicating the methods
and the subjects of the articles which are reviewed within the scope of this study.
Gregory, Roberts and Hader aimed to increase efficiency by used work sampling
method in 1958 [26]. In 1996, Tsai used work sampling method for how this method
use to evaluate effort spent on activities [25]. In other study published in 1996,
Buchholz, Paquet, Punnet, Lee and Moir purposed developing characterize the ergo-
nomic threat of construction by using work sampling approach [24]. In 1997, Everhart
examined about work sampling application. In the study, Everhart aimed to prove that
the work sampling method is effective and reliable method for collecting data [23].
Jenkins and Orth provided a guide for work sampling technique and implementation in
1668 G. Omeroglu et al.
Table 2. Paper number of keywords and sector

Keywords/Sectors Manufacturing Construction Retail Service Energy Logistic
Work sampling 4 4 – – – 1
IOT 1 3 4 – 1 –
Work study 1 1 – 1 – –
In store logistic – – 2 – – –
2004 [22]. In other study, Bircan and Iskender used work sampling method for com-
paring performance level of department at the Cumhuriyet University Research
Hospital in 2005 [21]. In 2008, Brown and Dant made a literature research about
scientific method for retail sector [20]. In other study, Guner, Unal and Arikan used
work sampling method to show the effect of productivity on layout plan of apparel
company in 2009 [1]. In 2010, Hengqin, Lihua and Ting suggested in store logistic
activity to increase productivity [19]. Condea, Thiesse and Fleisch aimed to analyze
RFID based shelf replenishment in retail stores and compare with standard techniques
in 2011 [18]. In other study, De Marco, Cogliano, Nervo and Rafele aimed to show
impact of item level RFID technology in retail sector in 2011 [17]. In 2011, Cheng,
Venugopal, Teizer, Vela evaluated the UWB system to record tasks in the field [16]. In
other study, Migliaccio, Gatti and also Cheng and Teizer evaluated a new way of self-
acting labor productivity in 2012 [15]. Olfa, Allard C.R., and Janjaap made a study
which offer suggestion to increase effectiveness in store logistic operations in 2013
[14]. In 2013, Metzger, Thiesse, Gershwin, and Fleisch used RFID technology to
provide high product availability at minimum operating costs [13]. Loera, Espinosa,
Enríquez, and Rodriguez made a study in terms of applying the work sampling and it is
proposed to get the extent of time that representatives are performing beneficial work in
2013 [12]. In 2013 Degirmenci and Celik examined the efficiency of a yarn processing
in a factory via using Work Sampling methods [11]. Durak and Unverdi used time
study method for decreasing the logistic cost of an enterprise in the year 2014 [10]. In
2014, Thiesse and Buckel used RFID technology for improving shelf replenishment
process [9]. Akhavian and Behzadan made a study about classify collecting data by
using smart phones that embedded accelerometer and gyroscope sensors in 2016 [8]. In
the year 2017, Mou, Robb, and DeHoratius made a systematic literature review with
the numbers of 255 papers on retail store operations from 32 operations research,
management science, retailing and general management journals over the period 2008–
2016 [7]. Garza, Hernández, and Antonyan made a study which aim to improve the
management of construction projects and their associated by using work study method
in 2017 [6]. Lou used work sampling method in the study of activity recognition
method. It receives surveillance videos as input and produces diverse and continuous
activity labels of individual workers in the field of view in 2018 [5]. In the year 2019,
Wahyuni, Budiman, Sembiring and Panjaitan used work sampling method in the
productivity analysis of the employees in small medium size enterprises [4]. Gong, Ma,
Zheng and Song developed a method for recognizing and analyzing the activities of the
worker in a complex environment [3]. Kim and Cho examined about the safety and
productivity of the construction workers in the year 2020. In this study, motion sensors
are located the worker’s body and using LSTM model for evaluation [2].
Table 3. Literature review
No Title of study Authors Year of Is there a/an Methods used Subject
publication Mathematical IOT Case
model? tools? study?
1 Effective inertial sensor quantity [2] 2020 x LSTM Model In this study, in order to ensure the safety
and locations on a body for deep and productivity of the construction
learning-based worker’s motion workers, motion sensors are located the
recognition worker’s body and using LSTM model
for evaluation
2 A deep model method for [3] 2020 x x MCPM Algorithm In this study, a method is developed for
recognizing activities of workers on the recognizing and analyzing the
offshore drilling platform by activities of the worker in a complex
multistage convolutional pose envirenment
machine
3 Analysis of workload and number [4] 2019 x Work Sampling The research has made the productivity
of workers in brick SMEs analysis of the employees with a work
sample method in small and medium
sized enterprises. Thus, optimization of
the number of employees has been tried to
determined
4 Towards efficient and objective [5] 2018 x x x Work Sampling An activity recognition method, which
work sampling: recognize workers Labor Productivity receives surveillance videos as input and
activities in site surveillance videos produces diverse and continuous activity
with two-stream convolutional labels of individual workers in the field of
networks view
5 Increase of productivity through the [6] 2017 x Work Study Aim of this study is to improve the
study of work activities in the Work Sampling management of construction projects and
construction sector their associated
6 Retail store operations: Literature [7] 2017 Literature Review This study review 255 papers on retail
Systematic Literature Review for Work Sampling
review and research directions store operations

(continued)
1669
1670

7 Smartphone-based construction [8] 2016 x x x Machine Learning Classify by collecting data using smart
workers’ activity recognition and algorithms phones embedded accelerometer and
classification Activity recognition gyroscope sensors
8 A comparison of RFID-based shelf [9] 2014 x x RFID Technology The goal of the study is improving shelf
replenishment policies in retail replenishment process
G. Omeroglu et al.
stores under suboptimal read rates

9 The Use of Cost Information in [10] 2014 x Time Study This study compares the using outsource
Frozen Food Logistics or insource in order to decrease the
operational and logistic cost of enterprise
due to operational flow chart
10 Application of Work Study Method [11] 2013 x Work Study Evaluates the efficiency of a yarn
in Yarn Spinning processing in a factory via using Work
Sampling methods
11 Productivity in Construction and [12] 2013 x Work Sampling Apply work sampling and proposed to get
Industrial Maintenance time that representatives are performing
beneficial work
12 The impact of false-negative reads [13] 2013 x x RFID Technology The goal is provide high product
on the performance of RFID-based availability at minimum operating costs
shelf inventory control policies
13 Managing in-store logistics: a fresh [14] 2013 Confirmator Factory Offer a suggestion to increase in store
perspective on retail service Analysis, logistic operations
Average Variance
Extracted
14 Automated task-level activity [15] 2012 x x The automated data Presents a unique way to evaluate self-
analysis through fusion of real time analysis methodology acting labor efficiency
location sensors and worker’s
thoracic posture data
(continued)
15 Performance evaluation of ultra [16] 2011 x x Automated work sampling Evaluate the ability of a commercially
wideband technology for Signal synchronization available Ultra Broadband (UWB) system
construction resource location to record its tasks in the field
tracking in harsh environments
16 Using system dynamics to assess [17] 2011 x x x System Dynamics Assist in the process of assessing the
the impact of RFID technology Simulation, economic impact of item level RFID
RFID Technology technology on retail sector
17 RFID-enabled shelf replenishment [18] 2011 x x x In Store Logistic This study aimed to analyze the
with backroom monitoring in retail Management characteristic of RFID-based shelf
stores replenishment policies in retail stores and
to compare them with the traditional
procedure
18 Research of Supermarket Store [19] 2010 x x x Operational Research Product counting, placement and queue
Logistic and Solutions Techniques problems are exemined and suggestions
are made to increase efficiency
19 A Layout Planning by Using [1] 2009 x x Work Sampling Work sampling method was used. With
Closeness Ratings Procedure Work the software developed, the effects of the
Sampling Method In Apparel changes made on the floor plan of the
Industry garment company on productivity were
examined
20 Scientific method and retailing [20] 2008 Literature Research Literature review
research: a retrospective
21 Application of Time Study among [21] 2005 x Time Study Used the work measurement technique
The Work Measurement for comparing the present performance
Techniques on Endoscopic Data level versus the potential performance
Systematic Literature Review for Work Sampling
level at the department of General

Surgery
(continued)
1671
1672
22 Productivity Improvement through [22] 2004 Work Sampling Provide a guide that develop work
G. Omeroglu et al.
Work Sampling sampling method and implementation

23 Work Sampling: The Application [23] 1997 x Work Sampling This study demonstrated that work
of an Industrial Research Technique sampling is an effective and reliable way
to School Library Media Centers of collecting data in school library media
center
24 Path: A work sampling-based [24] 1996 x Work Analysis Method Aim of this study is using a work
approach to ergonomic job analysis sampling-based approach for developing
for construction and other non- to characterize the ergonomic hazards of
repetitive work and other non-repetitive work
25 A Technical note on using work [25] 1996 x x Work Sampling Method Explain using of work sampling to
sampling to estimate the effort on estimate the effort spent on activities
activities under activity based under activity based costing and indicate
costing percentage of nonvalue added activities
26 Work-sampling as a technique in [26] 1958 x Work Sampling This study used work sampling method
determining labour utilization of for comparing present status and known
dairy plants standard labor operation to increase the
efficiency of labour in operation
Sum 26 11 6 22 26
Systematic Literature Review for Work Sampling 1673
3 Conclusion
To conclude, traditional work sampling is a powerful method for analyzing standard

process such as manufacturing operation. On the other hand, retail sector and in store
operations are not standard as manufacturing process. It can change stores to stores,
employee to employee and location to location. It is inevitable to bring a new per-
spective to work sampling applications in retail industry.
Stores are consisting of several department such as warehouse, office, selling space,
fruits and vegetables etc. In the on-going project, the stores divide according to these
departments. For achieving the accurate data each receiver located to each zone. This
receiver determines the employee activity by smart id card. By combining IOT tech-
nology and work sampling method, it can be possible to know the employee activity,
place and other data about working schedule for real time and continuously. For the
future studies, big data based analysis conducted in order to visualize dynamism of
retail operations.
Work sampling is still a powerful method which can be combined with industry 4.0
trends like big data and IOT approach. For the future research, combining industry 4.0
trends and work sampling method create a value on the literature as an innovation in
the retail sector.
References
1. Guner, M., Unal, C., Arikan, C.: A layout planning by using closeness ratings procedure and
work sampling method in apparel. Tekstil ve Konfeksiyon 20(2), 172–177 (2010)
2. Kim, K., Cho, Y.K.: Effective inertial sensor quantity and locations on a body for deep
learning-based worker’s motion recognition. Autom. Constr. 113, 103126 (2020)
3. Gong, F., Ma, Y., Zheng, P., Song, T.: A deep model method for recognizing activities of
workers on offshore drilling platform by multistage convolutional pose machine. J. Loss
Prev. Process Ind. 64, 104043 (2020)
4. Wahyuni, D., Budiman, I., Sembiring, M.T., Panjaitan, N.: Analysis of workload and
number of workers in brick SMEs. In: IOP Conference Series Material Science and
Engineering, vol. 505, no. 1, p. 012027 (2019)
5. Luo, X., Li, H., Cao, D., Yu, Y., Yang, X., Huang, T.: Towards efficient and objective work
sampling: recognizing workers’ activities in site surveillance videos with two-stream
convolutional networks. Autom. Constr. 94, 360–370 (2018)
6. Espinosa-Garza, G., Loera-Hernández, I., Antonyan, N.: Increase of productivity through the
study of work activities in the construction sector. Proc. Manuf. 13, 1003–1010 (2017)
7. Mou, S., Robb, D.J., DeHoratius, N.: Retail store operations: literature review and research
directions. Eur. J. Oper. Res. 265(2), 399–422 (2018)
8. Akhavian, R., Behzadan, A.H.: Smartphone-based construction workers’ activity recognition
and classification. Autom. Constr. 71, 198–209 (2016)
9. Thiesse, F., Buckel, T.: A comparison of RFID-based shelf replenishment policies in retail
stores under suboptimal read rates. Int. J. Prod. Econ. 159, 126–136 (2014)
10. Durak, M.G., Ünverdi, İ.: The use of cost information in frozen food logistics. Bus. Econ.
Res. J. 5(4), 19–41 (2014)
1674 G. Omeroglu et al.
11. Deǧirmenci, Z., Çelik, N.: Application of work study method in yarn spinning mill. Tekst. ve
Muhendis 20(92), 49–58 (2013)
12. Loera, I., Espinosa, G., Enríquez, C., Rodriguez, J.: Productivity in construction and
industrial maintenance. Proc. Eng. 63, 947–955 (2013)
13. Metzger, C., Thiesse, F., Gershwin, S., Fleisch, E.: The impact of false-negative reads on the
performance of RFID-based shelf inventory control policies. Comput. Oper. Res. 40(7),
1864–1873 (2013)
14. Bouzaabia, O., van Riel, A.C., Semeijn, J.: Managing in-store logistics: a fresh perspective
on retail service. J. Serv. Manage. 24(2), 112–129 (2013)
15. Cheng, T., Teizer, J., Migliaccio, G.C., Gatti, U.C.: Automated task-level activity analysis
through fusion of real time location sensors and worker’s thoracic posture data. Autom.
Constr. 29, 24–39 (2013)
16. Cheng, T., Venugopal, M., Teizer, J., Vela, P.A.: Performance evaluation of ultra wideband
technology for construction resource location tracking in harsh environments. Autom.
Constr. 20(8), 1173–1184 (2011)
17. De Marco, A., Cagliano, A.C., Nervo, M.L., Rafele, C.: Using system dynamics to assess the
impact of RFID technology on retail operations. Int. J. Prod. Econ. 135(1), 333–344 (2012)
18. Condea, C., Thiesse, F., Fleisch, E.: RFID-enabled shelf replenishment with backroom
monitoring in retail stores. Decis. Support Syst. 52(4), 839–849 (2012)
19. Hengqin, X., Lihua, A., Ting, BY.. Research of supermarket store logistics and solutions. In:
2010 International Conference on Logistics Systems and Intelligent Management (ICLSIM),
vol. 3, pp. 1569–1572. IEEE (2010)
20. Brown, J.R., Dant, R.P.: Scientific method and retailing research: a retrospective. J. Retail.
84(1), 1–13 (2008)
21. Bircan, H., İskender, G.: İş Ölçümü Tekniklerinden Zaman Etüdü Üzerine Bir Uygulama.
İktisadi ve İdari Bilimler Dergisi 2, 199–217 (2005)
22. Jenkins, J.L., Orth, D.L.: Productivity improvement through work sampling. Cost Eng. 46
(3), 27 (2004)
23. Everhart, N.: Work sampling: the application of an industrial research technique to school
library media centers. Libr. Inf. Sci. Res. 19(1), 53–69 (1997)
24. Buchholz, B., Paquet, V., Punnett, L., Lee, D., Moir, S.: PATH: a work sampling-based
approach to ergonomic job analysis for construction and other non-repetitive work. Appl.
Ergon. 27(3), 177–187 (1996)
25. Tsai, W.H.: A technical note on using work sampling to estimate the effort on activities
under activity-based costing. Int. J. Prod. Econ. 43(1), 11–16 (1996)
26. Gregory, M.E., Roberts, W.M., Hader, R.J.: Work-sampling as a technique in determining
labor utilization of dairy plants. J. Dairy Sci. 41(12), 1820–1822 (1958)
Engineering Economics Using Fuzzy Sets:
A Literature Review
Eda Boltürk(&)

Macka, 34367 Istanbul, Turkey
[email protected]
Abstract. Estimating the values of investment parameters in engineering

economics is a fundamental issue which is based on human thoughts and
judgments. It can be quietly difficult to make an investment decision especially
in vague and indefinite environments. When a model is based on human
thoughts, fuzzy logic offers excellent solutions in an environment of uncertainty.
In addition, fuzzy logic can be an agreeable tool when no probabilities are
accessible for states of nature, and when decisions are given under uncertain-
ness. In this paper, the studies on engineering economics employing the
extensions of fuzzy sets are summarized with a historical order. The future
possible studies are suggested in the conclusion.
Keywords: Engineering economics Fuzzy sets Literature review
1 Introduction
Selecting the best beneficial investment alternative with economic analysis is very
significant issue for firms. Investment analysis is a subject of engineering economics
and investments decisions should be based on the techniques of engineering eco-
nomics. The investment parameters are difficult to estimate and there are some
approaches developed in the literature. Engineering economics uses mathematical
techniques in order to compare investment alternatives simply and deals with time
value of money. Engineering economic decision making is very significant in limited
environmental conditions and requires detailed analysis. In order to make the best
decision, the quality and the amount of the data gain importance. When the data are
sufficient, probabilistic analyses are used. While probability analysis is used when there
is enough data, probability approaches are used in situations where there is not enough
data in order to make future estimates. The main analyses of engineering economics are
present worth analysis, annual cash flow analysis, rate of return analysis, incremental
analysis, cost estimating, etc. These tools are used in order to make better decisions in
investment analysis. In real life, these tools can be employed under non-deterministic
and complex environment conditions. Fuzzy logic was introduced by Zadeh in 1965
[1]. It allows decision expert on decision-making with estimated values under or
uncertain incomplete information. Fuzzy set approaches are suitable to use when the
modeling of human knowledge is necessary and when human evaluations are needed.
In order to model the vagueness, incompleteness and uncertainty better, the fuzzy logic
1676 E. Boltürk
is expanded to its various extension such as type-2 fuzzy sets [2], intuitionistic fuzzy
sets [3], fuzzy multisets [4], hesitant fuzzy sets [5], Pythagorean fuzzy sets [6], neu-
trosophic sets [7], intuitionistic fuzzy sets of second type [6, 8], picture fuzzy sets [9],
Q-rung orthopair fuzzy sets [10], spherical fuzzy sets [11], fermatean fuzzy sets [12].
These fuzzy sets let decision makers make their decisions when no probabilities are
accessible for states of nature that means uncertainty. In these times, decision makers
give their decision based on a linguistic scales such as very good, good, normal, bad,
too bad, and etc. Because of vague, imprecise, and insufficient data in investment
analysis, decision makers use fuzzy logic to make evaluations correctly under uncer-
tainty. Vagueness has been represented by probabilities in engineering economics. The
main purpose of this study is to show the utilization of fuzzy sets in engineering
economics problems and to see the gaps in the literature for future studies. The paper is
organized as follows: In Sect. 2, the literature review is given and a detailed literature
analyses is given. In Sect. 3, the conclusion is given and future suggestions are
proposed.
2 Literature Review
In this section, we review the publications on engineering economics with fuzzy sets.
Life-cycle cost analysis [13–17], cost estimating [18–32], investment analysis [33–40],
present worth analysis [41–46], engineering economic analysis [47–55], annual worth
analysis [56, 57], rate of return [58, 59], pricing [60], capital rationing [61], real options
valuation [62], inflation [63], depreciation and income tax considerations [64], cost-
benefit analysis [65–80], present value analysis [81, 82], net present value [83, 84],
cash flow analysis [85–92], and replacement analysis [93] are issued in fuzzy engi-
neering economics papers. Figure 1 shows the distribution of fuzzy engineering eco-
nomics papers with respect to years and the most of the papers have been published in
2008 with a rate of 27%. Figure 2 shows the percentage of area of fuzzy engineering
economics. Cost-benefit analyses and cost estimating are the most published areas
among others.
Fig. 1. Percentage of fuzzy engineering economics papers with respect to years.

Engineering Economics Using Fuzzy Sets Review 1677
Fig. 2. Percentage of year of fuzzy engineering economics publications.
The document types on fuzzy engineering economics are three types: articles with a
percentage 50% and book chapters with a percentage of 30 and conference papers with
a percentage of 20% (see Fig. 3).
Fig. 3. Distribution of fuzzy engineering economics with respect to the type of publications.
1678 E. Boltürk
3 Conclusion
In this paper, we summarize the papers in the literature on fuzzy engineering eco-
nomics. We observe that there are a lot of publications on engineering economics with
fuzzy sets. However, the fuzzy set extensions have not been used exhaustively. In the
future study, other types of fuzzy sets such as spherical fuzzy sets, neutrosophic sets
can be used and comparative analyses can be added for usefulness. In addition, we
think that the usage of the other fuzzy sets especially Pythagorean and spherical fuzzy
sets can be effective in decision making in investment analysis.
References
3. Atanassov, K.: Intuitionistic fuzzy sets. Fuzzy Sets Syst. 20(1), 87–96 (1986)
4. Yager, R.R.: On the theory of bags. Int. J. Gen. Syst. 13(1), 23–37 (1986)
6. Atanassov, K.T.: More on intuitionistic fuzzy sets. Fuzzy Sets Syst. 33(1), 37–45 (1989)
7. Smarandache, F.: A unifying field in logics. Neutrosophy: Neutrosophic probability, set and
logic. (1999)
9. Cuong, B.C.: Picture fuzzy sets. J. Comput. Sci. Cybern. 30(4), 409 (2014)
(2016)
11. Kutlu Gündoğdu, F., Kahraman, C.: Spherical fuzzy sets and spherical fuzzy topsis method.
12. Senapati, T., Yager, R.R.: Fermatean fuzzy sets. J. Ambient Intell. Humaniz. Comput. 11,
663–674 (2020)
13. Goh, K.C., Goh, H.H., Chong, H.: Integration model of fuzzy AHP and life-cycle cost
analysis for evaluating highway infrastructure investments. J. Infrastruct. Syst. 25(1),
04018045 (2019)
14. Bagdatli, M.E.C.: Fuzzy logic-based life-cycle cost analysis of road pavements. J. Stomatol.
144(4), 04018050 (2018)
15. Heravi, G., Esmaeeli, A.N.: Fuzzy multicriteria decision-making approach for pavement
project evaluation using life-cycle cost/performance analysis. J. Infrastruct. Syst. 20(2),
04014002 (2014)
16. Chen, C., Flintsch, G.W.: Fuzzy logic pavement maintenance and rehabilitation triggering
approach for probabilistic life-cycle cost analysis. Transp. Res. Rec. 1990, 80–91 (2007)
17. Liang, Q., Song, B., Jiang, J.: Fuzzy-grey ideal point method for estimating the life cycle
cost and effectiveness of torpedo. Xi Tong Gong Cheng Yu Dian Zi Ji Shu/Syst. Eng.
Electron. 28(9), 1375–1377+1397 (2006)
18. Plebankiewicz, E., Meszek, W., Zima, K., Wieczorek, D.: Probabilistic and fuzzy
approaches for estimating the life cycle costs of buildings under conditions of exposure to
risk. Sustainability 12(1), 226 (2020)
19. Chen, T.: Estimating unit cost using agent-based fuzzy collaborative intelligence approach
with entropy-consensus. Appl. Soft Comput. J. 73, 884–897 (2018)
20. Alshibani, A., Hassanain, M.A.: Estimating facilities maintenance cost using post-occupancy
evaluation and fuzzy set theory. J. Qual. Maintenance Eng. 24(4), 449–467 (2018)
21. Kasie, F.M., Bright, G., Walker, A.: Estimating cost of new products using fuzzy case-based
reasoning and fuzzy analytic hierarchy process. Adv. Transdisc. Eng. 5, 969–976 (2017)
22. Muñoz, M., Miranda, E., Sánchez, P.J.: A fuzzy system for estimating premium cost of
option exchange using mamdani inference: derivatives market of mexico. Int. J. Comput.
Intell. Syst. 10(1), 153–164 (2017)
23. Muñoz, M., Miranda, E.: A fuzzy system for estimating premium cost of option exchange
using mamdani inference: derivates market of Mexico. In: Paper presented at the 2016 IEEE
International Conference on Fuzzy Systems (FUZZ-IEEE 2016), pp. 888–895 (2016)
24. Zima, K.: Fuzzy case based reasoning in sports facilities unit cost estimating. Paper
presented at the AIP Conference Proceedings, vol. 1738 (2016)
25. Zima, K.: The use of fuzzy case-based reasoning in estimating costs in the early phase of the
construction project. In: AIP Conference Proceedings, vol. 1648 (2015)
26. Yi, T., Zhang, Y., Cui, S.: Cost estimating method of power engineering overhead line based
on fuzzy analogy. Energy Educ. Sci. Technol. Part A: Energy Sci. Res. 32(6), 6571–6580
(2014)
27. Fayek, A.R., Rodriguez Flores, J.R.: Application of fuzzy logic to quality assessment of
infrastructure projects at conceptual cost estimating stage. Can. J. Civ. Eng. 37(8), 1137–
1147 (2010)
28. Cocodia, E.: Risk based fuzzy modeling of cost estimating relationships for floating
structures. In: Proceedings of the International Conference on Offshore Mechanics and
Arctic Engineering - OMAE 1, pp. 129–142 (2008)
29. El-Cheikh, M., Chan, A.H.C., Lamb, J.: Fuzzy applications in cost estimating and
construction planning. In: Proceedings of the 11th International Conference on Civil,
Structural and Environmental Engineering Computing, Civil-Comp 2007 (2007)
30. Shaheen, A.A., Fayek, A.R., Abourizk, S.M.: Fuzzy numbers in cost range estimating.
J. Constr. Eng. Manage. 133(4), 325–334 (2007)
31. Mason, A.K., Kahn Sr., D.J.: Estimating costs with fuzzy logic. AACE Int. Trans. 12, 122
(1997)
32. Petley, G.J., Edwards, D.W.: Further developments in chemical plant cost estimating using
fuzzy matching. Comput. Chem. Eng. 19(SUPPL.1), 675–680 (1995)
33. Kahraman, C., Çevik Onar, S., Oztaysi, B., Sarı, I.U., Ilbahar, E.: Wind energy investment
analyses based on fuzzy sets. In: Kahraman, C., Kayakutu, G. (eds.) Energy Management—
Collective and Computational Intelligence with Theory and Applications. Studies in
Systems, Decision and Control, vol. 149, pp. 141–166. Springer, Cham (2018)
34. Xiaoping, B., Jia, F.: A new optimization model about enterprise equipment investment
project based on integrated entropy weight and fuzzy matter element. Res. J. Appl. Sci., Eng.
Technol. 4(20), 3885–3890 (2012)
35. Sheen, J.N.: Information security investment decision by fuzzy economics. In: Paper
presented at the Proceedings - 3rd International Conference on Information Sciences and
Interaction Sciences, ICIS 2010, pp. 648–653 (2010)
36. Sheen, J.N.: Applying fuzzy engineering economics to evaluate project investment
feasibility of wind generation. WSEAS Trans. Syst. 8(4), 501–510 (2009)
37. Majlender, P.: Soft decision support systems for evaluating real and financial investments.
In: Kahraman, C. (ed.) Fuzzy Engineering Economics with Applications, pp. 307–338
(2008)
1680 E. Boltürk
38. Kahraman, C.: Investment analyses using fuzzy decision trees, In: Kahraman, C. (ed.) Fuzzy
Engineering Economics with Applications, pp. 231–242 (2008)
39. Dymova, L., Sevastjanov, P.: Fuzzy multiobjective evaluation of investments with
applications, In: Kahraman, C. (ed.) Fuzzy Engineering Economics with Applications,
pp. 243–288 (2008)
40. Araque, F., Salguero, A., Carrasco, R., Martinez, L.: Using fuzzy multi-attribute data mining
in stock market analysis for supporting investment decisions. In: Kahraman, C. (ed.) Fuzzy
41. Dimitrovski, A., Matos, M.: Fuzzy present worth analysis with correlated and uncorrelated
cash flows. In: Kahraman, C. (ed.) Fuzzy Engineering Economics with Applications, pp. 11–
42 (2008)
42. Kuchta, D.: Optimization with fuzzy present worth analysis and applications. In: Kahraman,
C. (ed.) Fuzzy Engineering Economics with Applications, pp. 43–70 (2008)
43. Kahraman, C., Onar, S.C., Oztaysi, B.: Present worth analysis using pythagorean fuzzy sets
(2018)
44. Sarı, I.U., Kahraman, C.: Economic analysis of municipal solid waste collection systems
using type-2 fuzzy net present worth analysis (2017)
45. Beskese, A., Kahraman, C., Irani, Z.: Quantification of flexibility in advanced manufacturing
systems using fuzzy concept. Int. J. Prod. Econ. 89(1), 45–56 (2004)
46. Kahraman, C., Cevik Onar, S., Oztaysi, B.: Engineering economic analyses using
47. Çoban, V., Onar, S.C.: Pythagorean fuzzy engineering economic analysis of solar power
plants. Soft. Comput. 22(15), 5007–5020 (2018)
48. Kahraman, C., Sarı, I.U., Onar, S.C., Oztaysi, B.: Fuzzy economic analysis methods for
environmental economics. In: Intelligent Systems Reference Library, vol. 113, pp. 315–346
(2017)
49. Kahraman, C.: Future directions in fuzzy engineering economics. In: Kahraman, C. (ed.)
Fuzzy Engineering Economics with Applications. Studies in Fuzziness and Soft Computing,
vol. 233, pp. 381–382. Springer, Heidelberg (2008)
50. Kahraman, C.: Fuzzy Engineering Economics with Applications. Springer, Heidelberg
(2008)
51. Kahraman C.: Fuzzy sets in engineering economic decision-making. In: Kahraman, C. (ed.)
Fuzzy Engineering Economics with Applications, pp. 1–10 (2008)
52. Dimitrovski, A.D., Matos, M.A.: Fuzzy engineering economic analysis. IEEE Trans. Power
Syst. 15(1), 283–289 (2000)
53. Kaino, T., Hirota K., Pedrycz, W.X.: Fuzzy sensitivity analysis and its application. In:
Kahraman, C. (ed.) Fuzzy Engineering Economics with Applications, pp. 183–216 (2008)
54. Yoon, K.P.: A probabilistic approach to fuzzy engineering economic analysis. In: Kahraman,
C. (ed.) Fuzzy Engineering Economics with Applications, pp. 217–230 (2008)
55. Nguyen, H.T., Sriboonchitta, S., Wu, B.: A statistical basis for fuzzy engineering economics.
Int. J. Fuzzy Syst. 17(1), 1–11 (2015)
56. Kahraman, C., Kaya, I.: Fuzzy equivalent annual-worth analysis and applications. In:
Kahraman C (ed) Fuzzy Engineering Economics with Applications, pp. 71–82 (2008)
57. Matos, M., Dimitrovski, A.: Case studies using fuzzy equivalent annual worth analysis. In:
58. Kuchta, D.: Fuzzy rate of return analysis and applications. In: Kahraman, C. (ed.) Fuzzy
59. Sewastjanow, P., Dymowa, L.: On the fuzzy internal rate of return. In: Kahraman, C. (ed.)
60. Buckley, J.J., Eslami, E.: Pricing options, forwards and futures using fuzzy set theory. In:
61. Kahraman, C., Bas, E.: Fuzzy capital rationing models. In: Kahraman, C. (ed.) Fuzzy
62. Ucal, I.: Cengiz Kahraman: Fuzzy Real Options Valuation for Oil Investments. J. Technol.
Econ. Devel. Econ. 15(4), 646–669 (2009)
63. Kahraman, C., Demirel, T., Demirel, N.: Effects of inflation under fuzziness and some
applications. In: Kahraman, C. (ed.) Fuzzy Engineering Economics with Applications,
pp. 173–182, (2008)
64. Kahraman, C., Kaya, I.: Depreciation and income tax considerations under fuzziness. In:
65. Wicaksono, F.D., Arshad, Y., Sihombing, H., Baihaqi, I.: Gaussian fuzzy analytic hierarchy
process for the evaluation of benefit, cost, and risk analysis in the indonesian oil and gas
processing area. J. Teknol. 81(6), 25–37 (2019)
66. Kahraman, C., Kaya, I.: Fuzzy benefit/cost analysis and applications. In: Kahraman, C. (ed.)
67. Bağdatlı, M.E.C., Akbıyıklı, R., Papageorgiou, E.I.: A fuzzy cognitive map approach
applied in Cost–Benefit analysis for highway projects. Int. J. Fuzzy Syst. 19(5), 1512–1527
(2017)
68. Verma, NK., Ghosh, A., Dixit, S., Salour, A.: Cost-benefit and reliability analysis of
prognostic health management systems using fuzzy rules. In: 2015 IEEE Workshop on
Computational Intelligence: Theories, Applications and Future Directions, WCI 2015 (2016)
69. Rabbani, M., Soufi, H.R., Torabi, S.A.: Developing a two-step fuzzy cost-benefit analysis for
strategies to continuity management and disaster recovery. Saf. Sci. 85, 9–22 (2016)
70. Kahraman, C., Onar, S.C., Oztaysi, B.: A comparison of wind energy investment alternatives
using interval-valued intuitionistic fuzzy benefit/cost analysis. Sustainability 8(2), 118
(2016)
71. Fulari, H., Fegade, V., Loharkar, P.K.: The fuzzy cost benefit analysis of design for product
development process with perspective of remanufacturing. Int. J. Appl. Eng. Res. 10(11),
10114–10118 (2015)
72. Thengane, S.K., Hoadley, A., Bhattacharya, S., Mitra, S., Bandyopadyay, S.: Cost-benefit
analysis of different hydrogen production technologies using AHP and fuzzy AHP. Int.
J. Hydrogen Energy 39(28), 15293–15306 (2014)
73. Xu, C.W., Yao, J., Li, J.: A hybrid approach based on the fuzzy AHP and benefit cost
analysis to evaluating cutting blade alternatives. Appl. Mech. Mater. 364, 513–518 (2013)
74. Carmen, T., Ramona, C., Daniela, B.: Fuzzy uncertainty analysis and multi-period cost-
benefit analysis. innovation and sustainable competitive advantage: from regional develop-
ment to world economies. In: Proceedings of the 18th International Business Information
Management Association Conference, vol. 3, pp. 1490–1498 (2012)
75. Kahraman, C., Ateş, N.Y., Çevik, S., Gülbay, M.: Fuzzy multi-attribute cost-benefit analysis
of e-services. Int. J. Intell. Syst. 22(5), 547–565 (2007)
76. Kahraman, C., Tolga, E., Ulukan, Z.: Justification of manufacturing technologies using
fuzzy benefit/cost ratio analysis. Int. J. Prod. Econ. 66(1), 45–52 (2000)
77. Dompere, K.K.: Cost-benefit analysis, benefit accounting and fuzzy decisions. (II).
A hypothetical application to mental illness. Fuzzy Sets Syst. 100(1–3), 101–116 (1998)
78. Dompere, K.K.: The theory of approximate prices: Analytical foundations of experimental
cost-benefit analysis in a fuzzy-decision space. Fuzzy Sets Syst. 87(1), 1–26 (1997)
79. Wang, M., Liang, G.: Benefit/cost analysis using fuzzy concept. Eng. Econ. 40(4), 359–376
(1995)
1682 E. Boltürk
80. Xu, LD.: Fuzzy multiobjective mathematical programming’s application. In: Cost-Benefit
Analysis. Proceedings of the 1986 IEEE International Conference on Systems, Man and
Cybernetics; Atlanta, GA, pp. 638–643 (1986)
81. Aydin, S., Kahraman, C., Kabak, M.: Evaluation of investment alternatives using present
value analysis with simplified neutrosophic sets. Eng. Econ. 29(3), 254–263 (2018)
82. Ustundag, A., Cevikcan, E.: A fuzzy inference system for net present value analysis of
construction projects. Paper presented at the World Scientific Proceedings Series on
Computer Engineering and Information Science 7; Uncertainty Modeling in Knowledge
Engineering and Decision Making - Proceedings of the 10th International FLINS Conf,
pp. 783–787 (2012)
83. Sari, I.U., Kuchta, D.: Fuzzy global sensitivity analysis of fuzzy net present value. Control
and Cybern. 41(2), 481–496 (2012)
84. Shahriari, M.: Mapping fuzzy approach in engineering economics. Int. Res. J. Finance Econ.
81, 6–12 (2011)
85. Komolavanij, S., Rogers, K.J.: Performing cash flow analysis with fuzzy data. J. Eng.
Valuation Cost Anal. 2(6), 425–435 (2000)
86. Kahraman, C., Beskese, A., Ruan, D.: Measuring flexibility of computer integrated
manufacturing systems using fuzzy cash flow analysis. Inf. Sci. 168(1–4), 77–94 (2004)
87. Çetin, N., Kahraman, C.: Fuzzy future value and annual cash flow analyses. In: Zhong, N.,
Skowron, A., Ohsuga, S. (eds.) New Directions in Rough Sets, Data Mining, and Granular-
Soft Computing. RSFDGrC 1999. LNCS, vol. 1711, pp. 266–270. Springer, Heidelberg
(1999)
88. Boussabaine, A.H., Elhag, T.: Applying fuzzy techniques to cash flow analysis. Constr.
Manage. Econ. 17(6), 745–755 (1999)
89. Karsak, E.: Measures of liquidity risk supplementing fuzzy discounted cash flow analysis.
Eng. Econ. 43(4), 331–344 (1998)
90. Chiu, C., Park, C.S.: Fuzzy cash flow analysis using present worth criterion. Eng. Econ. 39
(2), 113–138 (1994)
91. Ward, T.L.: Discounted fuzzy cash flow analysis. In: 1985 Fall Industrial Engineering
Conference Proceedings, pp. 476–481 (1985)
92. Kahraman, C., Oztaysi, B., Onar, S.C.: Annual cash flow analysis using hesitant fuzzy sets.
In: 2015 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE) (2015)
93. Kahraman, C., Demircan, M.L.: Fuzzy replacement analysis. In: Kahraman C (ed) Fuzzy
Scientific Cooperation in the Field
of Economics in Selected European Countries
Paweł Lula(&) , Urszula Cieraszewska , and Monika Hamerska
Cracow University of Economics, Rakowicka 27, 31-510 Kraków, Poland

[email protected]
Abstract. The paper presents the results describing the research productivity in
the area of economics in 36 European countries in the period 2011–2020.
During the first stage of the research, the publications’ distributions over the
structure of authors’ teams, over the contribution of foreign co-authors and over
main research topics mentioned in papers were analyzed. Also the impact of
mentioned above features on the number of citations was studied. The authors
tested the possibility of building a recommendation system indicating the best
way of preparing scientific publications with respect to the number of citations.
The results show that the prediction of the exact number of citation is rather
impossible, but the main determinants may be identified. The set of the most
important factors having a huge impact on the citation measures involves the
research experience of authors confirmed by previous publications, the structure
of authors’ teams, the degree of their internationalization and proper selection of
research topics belong to the group of main success factors in publication
activity.
Keywords: Research productivity in economics Analysis of research

publications Success factors in publication activity
1 Introduction
1.1 Scientific Cooperation

From the beginning of the twentieth century, one can observe a phenomenon affecting
the life of the academic community, in line with the philosophy of “publish or perish”
[1]. Research shows that one of the determinants affecting the level of scientific pro-
ductivity expressed by the number of publications is scientific cooperation and the
intensity of cooperation increases with increasing productivity [2]. Sonnenwald defined
scientific collaboration as interaction taking place within a social context among two or
more scientists that facilitates the sharing of meaning and completion of tasks with
respect to a mutually shared, superordinate goal [3]. During the last few decades the
scientific collaboration, both at intra-country and international levels, has increased
rapidly in diverse research areas [4]. Literature of the subject point to some features
characteristic for the development of scientific cooperation:
• Factors based on cultural, linguistic and institutional differences cause additional
obstacles to long-distance cooperation [5];
1684 P. Lula et al.
• There is a greater tendency to cooperate with scientists with a higher level of

scientific productivity [2];
• International cooperation is lower in countries with a lower level of science and
technology [6];
• Co-authored works have a better chance of being accepted for publication than
works by a single author [7];
The network of international co-authorship relations has been dominated by certain
European nations and the USA, but this network is rapidly expanding at the global
level [8]. Scientific collaboration is an important mechanism that enables the integra-
tion of the least developed countries into research activities [9].
The authors of the paper established the following list of research goals:
• analysis the distribution of publications over the number of authors,
• analysis of the internationalization of the authors’ teams,
• identification and analysis of research topics covered by publications,
• the analysis of citations,
• checking the possibility of building a recommendation system indicating the way of
preparing scientific publications guaranteeing the highest number of citations.
1.2 Economics and Management Fields

Trend of growing collaboration is also visible in economic sciences, research showed
that post-war economics literature has been characterized by a marked trend toward co-
authored articles and especially in 20th century publication in economics has expanded
rapidly [10]. Numerous studies have been carried out to examine the production of
scientific knowledge in the field of economics and managements, co-authorship pat-
terns, and the development of co-authorship in some areas of the economy and man-
agement or major economic journals [11, 12]. Research show that the percentage of co-
authored papers grew in journals from this area [11]. If we concerning about co-
authoring in economics and management [13, 14] its seems that authors exploiting the
gains from specialization within increasingly specialized fields, hedging against the
risks of rejection or delayed review, and/or changing the trade-off between quantity and
quality [13].
In research it suggests that, although publication patterns differ between disciplines
of the social science and humanities (SSH) group, these patterns are rather similar
within SSH disciplines across different countries [15].
Research shown not only that cooperation is increasing, but also that the number of
co-authors [11]. This pattern is not only visible in country with the long tradition of
international cooperation but also in Poland as in other post-communist countries [16].
In literature we can find that the subject specific differences in citation patterns arise
for the following reasons like: - different numbers of journals indexed depending on
subject in bibliometric databases [17]; - different cultures of citation and authorship
practices among fields; - different production functions depending on subject [18]; -
numbers of researchers among fields [19].
Scientific Cooperation in the Field of Economics in Selected European Countries 1685
2 Methodology
2.1 The Scope of the Analysis
The analysis covers the research productivity in the area of economics in selected
European countries. The study was performed with the use of data concerning research
papers in the field of economics prepared by authors from 36 European countries and
registered in the Scopus database from 2011 to 2020 year. The total number of papers
taken into account in the analysis was 124460. The distribution of papers over selected
countries in presented in the Table 1.
Table 1. The distribution of research papers in the area of economics over European countries
in the period 2011–2020.
Country N Country N Country N Country N
Albania 602 Finland 2841 Lithuania 1351 Romania 2344
Austria 3405 France 14448 Luxembourg 776 Serbia 925
Belgium 5141 Germany 18837 Malta 146 Slovakia 1605
Bulgaria 358 Greece 3420 Montenegro 164 Slovenia 858
Croatia 1150 Hungary 1400 Netherlands 9591 Spain 13435
Cyprus 849 Iceland 225 North 161 Sweden 5454
Macedonia
Czech 3948 Ireland 2369 Norway 3609 Switzerland 6393
Republic
Denmark 3660 Italy 11724 Poland 4069 Turkey 4643
Estonia 343 Latvia 267 Portugal 3247 United 23977
Kingdom
2.2 Research Methodology

During the analysis several methods were used. Data sets were retrieved from the
Scopus database in the CSV format. Data preprocessing was conducted in R language.
All analysis of papers’ distributions (over number of authors, number of countries of
authors’ affiliation, number of citations) were carried out with standard R tools. For
cluster analysis the Ward’s algorithm was used with the Bhattacharyya distance (due to
the fact that objects were described by distributions).
The identification of research topics mentioned in paper’s abstracts was carried out
with the use of ontology-based system design by the authors in implemented in R
language. The detailed description of this tool is presented in [20]. The system cal-
culates measures of similarity between phrases derived from abstracts with a set of
above 7000 patterns assigned to different concepts existing in the JEL ontology. To
measure the similarity, the modified version of Jaccard coefficient was used. Aggre-
gation of measures of similarities within every main JEL concept allowed to calculate
the contribution of every first-level JEL class in a given document.
1686 P. Lula et al.
Two version of recommendation systems were tested. The first one was based on a
regression model using linear regression approach. In the second approach the clas-
sification model build with the XGBoost algorithm was used.
3 Analysis of Research Productivity in the Area of Economics

in Selected European Countries
3.1 Analysis the Distribution of Publications Over the Number

of Authors
The analysis of the number of members in authors’ teams was the main goal of this
stage of the analysis. Aggregated information for all countries concerning the number
of papers divided into classes with respect to the number of authors is presented in the
Table 2.
Table 2 The distribution of research papers in the area of economics over the number of
members of authors’ teams (aggregated data for 36 European countries in the period 2011–2020).
1 2 3 4 5 6 or more
29931 (24.0%) 42906 (34.5%) 32493 (26.1%) 11824 (9.5%) 3588 (2.9%) 3718 (3.0%)
The results for individual countries are presented in Fig. 1.
Albania 53.3% 26.2% 13.1% 4.7% 2.2% 0.5%

Austria 19% 33% 27% 11.4% 3.8% 5.8%
Belgium 18.1% 32.7% 28.3% 11.4% 3.9% 5.6%
Bulgaria 38.5% 28.8% 11.7% 3.9% 6.4% 10.6%
Croatia 25.2% 34.6% 33.8% 2.3% 1% 3.1%
Cyprus 25% 28% 28.3% 13.8% 3.2% 1.8%
Czech_Republic 27.2% 35.4% 20.7% 10% 3.4% 3.3%
Denmark 19.6% 32.1% 28.1% 11.2% 3.9% 5.1%
Estonia 24.2% 32.9% 23% 7.6% 2.6% 9.6%
Finland 23.1% 31.4% 24.5% 11.6% 4.2% 5.3%
France 24.6% 33% 26% 10% 2.7% 3.8%
Germany 17.5% 38.5% 27.8% 9.8% 2.9% 3.5%
Greece 17.6% 35.6% 29.2% 11.1% 2.9% 3.6%
Hungary 31.9% 31.9% 16.7% 8.6% 3.6% 7.3% Value [%]
Iceland 16% 32.9% 21.3% 13.8% 4% 12%
Ireland 23% 32.4% 26.3% 10.3% 3.8% 4.3% 50
Italy 15.9% 32.6% 29.7% 12% 4.1% 5.6%
Country
40
Latvia 26.2% 38.2% 19.1% 8.6% 3% 4.9%
Lithuania 10% 29.5% 28.7% 18.4% 7.7% 5.7% 30
Luxembourg 28.7% 31.3% 27.2% 8.1% 2.8% 1.8% 20
Malta 16.4% 26% 30.1% 13.7% 4.1% 9.6%
Montenegro 36% 18.9% 23.8% 14.6% 3.7% 3% 10
Netherlands 15.4% 33.5% 30.4% 12.4% 3.6% 4.7%
North_Macedonia 19.3% 37.3% 25.5% 8.7% 5% 4.3%
Norway 20.4% 34.2% 26.9% 10.9% 3% 4.5%
Poland 37.9% 30.1% 17.2% 7.6% 3% 4.2%
Portugal 14.3% 33% 31.1% 11.2% 4% 6.4%
Romania 21% 27.6% 25% 17.5% 5.2% 3.7%
Serbia 17.5% 27.6% 33.6% 10.6% 5.4% 5.3%
Slovakia 17.8% 31.5% 26% 15.1% 5.2% 4.4%
Slovenia 17.6% 37.4% 27.9% 9.2% 2.7% 5.2%
Spain 14.1% 33.8% 34.8% 10.7% 2.6% 3.8%
Sweden 21.2% 33.4% 25.8% 10.9% 3.9% 4.8%
Switzerland 20.3% 35.2% 28.8% 10.4% 2.3% 3.1%
Turkey 25% 37.2% 24% 8.3% 2.6% 2.9%
United_Kingdom 18.5% 33.1% 29.5% 11.7% 3.4% 3.7%
1 author 2 authors 3 authors 4 authors 5 authors 6+ authors
Number of authors
Fig. 1. The distribution of research papers in the area of economics over the number of members
of authors’ teams in selected European countries.
The results presented in Fig. 1 allowed to perform cluster analysis showing the
similarities among European countries with respect to the number of members in
authors teams. The dendrogram obtained as a result is presented in Fig. 2.
Similarity of European countries with respect to the number of members in authors' teams
0.12
0.08
0.04
Height
0.00
Belgium
Italy
Germany
Turkey
Malta
Cyprus
Greece
Hungary
Spain
Denmark
Norway
Austria
Latvia
France
Estonia
United_Kingdom
Poland
Ireland
Albania
Bulgaria
Iceland
Slovakia
Serbia
Romania
Slovenia
Sweden
Croatia
Lithuania
Netherlands
Finland
Switzerland
Montenegro
Luxembourg
Portugal
North_Macedonia
Czech_Republic
Ward's method, Bhattacharyya distance
Fig. 2. The similarity of European countries with respect of the number of members in authors’
teams preparing research papers in the area of economics.
3.2 The Analysis of the Internationalization of the Authors’ Teams

Using the information concerning authors’ affiliation, the distribution of research
papers over countries in which authors were affiliated was performed. First, to evaluate
the degree of internationalization, for every country the percentage of papers prepared
with foreign co-author(s) were expressed. The results are presented in Fig. 3.
Data describing the structure of authors’ teams has allowed to build a matrix of
cooperation between selected European countries. It is a square matrix in which rows
and columns correspond to countries and an element on position (i, j) indicates how
many times authors from countries represented by i-th and j-th row (column) appear
together on list of authors of the same publication. Elements on positions (i, i) show
cases in which at least two authors from the same i-th country worked together on the
same paper. The matrix of cooperation is presented in Fig. 4.
Also cluster analysis of countries with respect to the strength of their research
cooperation with other countries was performed. The results are shown in Fig. 5.
3.3 The Identification and Analysis of Research Topics Covered

by Publications
The ontology-based system for identification of topics mentioned in paper’s abstracts
have allowed to measure the importance of topics represented by consecutive JEL
concepts existing on the first level of this classification system. The contribution of
topics represented by main JEL classes in presented in Fig. 6.
1688 P. Lula et al.
Fig. 3. The percentage of papers prepared as a result of international cooperation in the area of
economics.
Albania
Austria
Belgium
Bulgaria
Croatia
Cyprus
Czech_Republic
Denmark
Estonia
Finland
France
Germany
Greece
Hungary
Iceland
Ireland
Italy
Latvia
Lithuania
Luxembourg
Malta
Montenegro
Netherlands
North_Macedonia
Norw ay
Poland
Portugal
Romania
Serbia
Slovakia
Slovenia
Spain
Sw eden
Sw itzerland
Turkey
United_Kingdom
Portugal
Belgium
Malta
Latvia
Serbia
Spain
Albania
Austria
Bulgaria
Estonia
Croatia
Finland
Iceland
Ireland
Lithuania
Poland
Romania
Slovakia
Slovenia
Sweden
France
Greece
United_Kingdom
Montenegro
Switzerland
Luxembourg
North_Macedonia
Italy
Cyprus
Denmark
Germany
Hungary
Norway
Turkey
Netherlands
Czech_Republic
Fig. 4. Matrix of cooperation between European countries.

The similarity of European countries with respect to patterns of cooperation
1.5
1.0
0.5
Height
0.0
Belgium
Italy
Cyprus
Turkey
Germany
Latvia
Greece
Norway
Hungary
Denmark
Malta
France
Serbia
United_Kingdom
Austria
Spain
Croatia
Slovakia
Poland
Iceland
Sweden
Ireland
Estonia
Albania
Slovenia
Romania
Bulgaria
Finland
Netherlands
Montenegro
Lithuania
Luxembourg
Switzerland
Czech_Republic
North_Macedonia
Portugal
Ward's method, Bhattacharyya distance
Fig. 5. The similarity of European countries with respect to their pattern of cooperation with
foreign partners.
The contribution of the main JEL concepts in research papers

JEL A
JEL B
JEL C
JEL D
JEL E
JEL F
JEL G
JEL H
JEL I
JEL J
JEL K
JEL L
JEL M
JEL N
JEL O
JEL P
JEL Q
JEL R
JEL Y
JEL Z
0.00
0.02
0.04
0.06
0.08
0.10
The significance
Fig. 6. The contribution of main JEL concepts in publications in the area of economics.
3.4 The Analysis of Citations

In the next step of analysis, the number of citations of papers were studied. The
distribution of papers over the number of citation is presented in the Table 3. The range
for classes was defined in a way which assure the similar number of elements in every
group.
Table 3. The distribution of research papers in the area of economics over the number of
citations (aggregated data for 36 European countries in the period 2011–2020).
Number of citations 0 1–2 3–5 6–12 13–2251
Number of papers 32392 28206 20572 20532 22758
1690 P. Lula et al.
3.5 Recommendation System for Choosing the Best Way of Preparing

Scientific Publications Guaranteeing the Highest Number of Citations
The final step of the research process was related to the problem of building a rec-
ommendation system indicating the optimal way of preparing scientific publication
guaranteeing the highest number of citations. It was assumed that the system should be
built of two main components:
• modelling component allowing to predict the number of citations using a set of
selected variables,
• control component calculating the optimal values of input variables to maximize the
number of citations.
For a dataset describing all analyzed papers, a set of potential input variables
included:
• a year of paper’s publication,
• number of authors in author’s team,
• number of countries in which members of author’s team were affiliated,
• an average number of citations for papers published in previous years for the author
with highest citation,
• an average number of citations for papers published in previous years for all
members of author’s team,
• information about the structure of author’s team with respect to countries in which
they were affiliated (for every paper the proportion of authors from selected 36
European countries extended by United States of America, Japan, Russia and
China),
• contribution of every main JEL concept in the abstract of a given paper (a vector
with elements related to main JEL classes with measures of contribution in a given
paper).
For every paper two potential output variables were studied:
• a number of citation – used for regression models,
• a label of class describing the number of citations (classes described in the section
“The analysis of citations” were used) – for classification models.
For regression models different version of linear models were used and for clas-
sification the XGBoost models were tested. Unfortunately quality measures (R2 for
regression models and accuracy for classification models) indicate the impossibility of
building a formal model predicting an exact number of citations for a given paper. But
statistical significance of coefficients calculated for regression models confirms use-
fulness of obtained models for describing general relationships existing between input
variables and a number of citations. Using the methods proposed in [21] and in [22] the
importance of input variables was estimated. The list of variables which have the
strongest impact on the number of citations is presented in Fig. 7.
year
max.prev.citation
USA
n.authors
JEL_Q
n.country
United_Kingdom
JEL_M
JEL_O
avg.prev.citation
0.0
0.1
0.2
0.3
0.4
Fig. 7. The contribution of main JEL concepts in publications in the area of economics.
The analysis of regression coefficients indicates that a variable “year” has negative
impact on the number of citations and remaining variables have positive influence.
4 Conclusions
The analysis allowed to describe main rules describing research productivity in the area
of economics in 36 European countries.
The research results show that the prediction of the exact number of citations for a
given paper is very difficult. But the main determinants were identified. The findings
confirm that the following factors have a crucial influence on the citation measures:
• time (the number of quotations increases with time),
• involvement of an author with high number of citation for his/her previous publi-
cations in an authors’ team,
• working in multi-author and multinational teams,
• participation of authors from United States of America or from United Kingdom,
• contents of paper related to areas represented by Q, O or M concepts from JEL
classification system.
Acknowledgements. The research has been carried out as part of a research initiative financed
by the Ministry of Science and Higher Education within “Regional Initiative of Excellence”
Programme for 2019–2022. Project no.: 021/RID/2018/19. Total financing: 11 897 131,40 PLN.
References
1. Fanelli, D., Larivière, V.: Researchers’ individual publication rate has not increased in a
century. PLoS ONE 11(3), 1–12 (2016)
2. Pravdic, N., Oluic-Vukovic, V.: Dual approach to multiple authorship in the study of
collaborator/scientific output relationship. Scientometrics 10(5–6), 259–280 (1986)
3. Sonnenwald, D.H.: Scientific collaboration. Annu. Rev. Inf. Sci. Technol. 41(1), 643–681
(2007)
1692 P. Lula et al.
4. Luukkonen, T.: The measurement of international scientific collaboration. Scientometrics

28(1), 15–36 (1993)
5. Okubo, Y., Zitt, M.: Searching for research integration across Europe: a closer look at
international and inter-regional collaboration in france. Sci. Public Policy. 31(3), 213–226
(2004)
6. Wagner, C., Brahmakulam, I., Jackson, B., Wong, A., Yoda, T.: Science and Technology
Collaboration: Building Capacity in Developing Countries (2001)
7. Laband, D.N., Tollison, R.D.: Intellectual collaboration. J. Polit. Econ. 108, 632–662 (2000)
8. Leydesdorff, L., Wagner, C., Park, H.W., Adams, J.: International collaboration in science:
the global map and the network. El Prof. Inf. 22(1), 87–94 (2013)
9. González-Alcaide, G., Park, J., Huamaní, C., Ramos, J.M.: Dominance and leadership in
research activities: collaboration between countries of differing human development is
reflected through authorship order and designation as corresponding authors in scientific
publications. PLoS ONE 12(8), 1–30 (2017)
10. Schymura, M., Löschel, A.: Investigating JEEM empirically: a story of coauthorship and
collaboration. ZEW Discussion Paper, 12-29 (2012)
11. Goyal, S., Leij, V.D.M.J., González, J.L.M.: Economics: an emerging small world. J. Polit.
Econ. 114(2), 403–412 (2006)
12. Acedo, F.J., Barroso, C., Casanueva, C., Gala, J.: Co-authorship in management and
organizational studies: an empirical and network analysis. J. Manag. Stud. 43(5), 957–983
(2006)
13. Hudson, J.: Trends in multi-authored papers in economics. J. Econ. Perspect. 10, 153–158
(1996)
14. Ong, D., Chan, F.H., Torgler, B., Yang, Y.A.: Collaboration incentives: endogenous
selection into single and coauthorships by surname initial in economics and management.
J. Econ. Behav. Organ. 147, 41–57 (2018)
15. Sivertsen, G.: Patterns of internationalization and criteria for research assessment in the
social sciences and humanities. Scientometrics 107(2), 357–368 (2016)
16. Bukowska, G., Fałkowski, J., Łopaciuk-Gonczaryk, B.: Teaming up or writing alone -
authorship strategies in leading Polish economic journals. Working Papers, University of
Warsaw, Faculty of Economic Sciences, no. 29/146, pp. 1–26 (2014)
17. Marx, W., Bornmann, L.: On the causes of subject-specific citation rates in web of science.
Scientometrics 102(2), 1823–1827 (2015)
18. McAllister, P.R., Narin, F., Corrigan, J.G.: Programmatic evaluation and comparison based
on standardized citation scores. IEEE Trans. Eng. Manag. 30(4), 205–211 (1983)
19. Kostoff, R.N.: Citation analysis of research performer quality. Scientometrics 53(1), 49–71
(2002)
20. Kovaleva, A.S., Lula, P., Tuchowski, J.: Ontology-based measurement of study program
innovativeness in the area of economics and management. In: Babić, V., Nedelko, Z.,
Hershey, P.A. (eds.) Handbook of Research on Enhancing Innovation in Higher Education
Institutions, pp. 381–407. IGI Global, Hershey (2020)
21. Pratt, J.: Dividing the indivisible: using simple symmetry to partition variance explained. In:
Pukkila, T., Puntanen, S. (eds.) Proceedings of Second Tampere Conference in Statistics,
pp. 245–260. University of Tampere, Tampere (1987)
22. Grömping, U.: Relative importance for linear regression in R: the package relaimpo. J. Stat.
Softw. 17(1), 1–12 (2006)
Sentiment Analysis on Students’ Evaluation
of Higher Educational Institutions
Mansur Alp Toçoğlu1 and Aytuğ Onan2(&)

1
Department of Software Engineering, Faculty of Technology, Manisa Celal
Bayar University, 45400 Manisa, Turkey
[email protected]
2
Department of Computer Engineering, Faculty of Engineering
and Architecture, İzmir Katip Çelebi University, 35620 İzmir, Turkey
[email protected]
Abstract. Sentiment analysis is the method of identifying and classifying the

views of users from text documents into different sentiments, such as positive,
negative or neutral. Sentiment analysis can be employed to extract structured
and informative knowledge from unstructured text pieces. This knowledge can
serve as an important source of information for decision support systems and
individual decision makers. Sentiment analysis plays an important role in many
fields like education, where student input is crucial for assessing the effective-
ness of educational institutions. In this paper, we present a machine learning
based approach to sentiment analysis on students’ evaluation of higher educa-
tional institutions. We analyze a corpus containing approximately 700 student
reviews written in Turkish, with the use of conventional text representation
schemes and machine learning classifiers. In the experimental analysis, three
conventional text representation schemes (i.e., term-presence, term-frequency,
TF-IDF scheme) and three N-gram models (1-gram, 2-gram and 3-gram) have
been considered in conjunction with four classifiers (i.e., support vector
machines, Naïve Bayes, logistic regression, and random forest algorithm). The
predictive performance of four ensemble learners (i.e., AdaBoost, Bagging and
Random Subspace and voting algorithm) have been also evaluated. The
empirical results indicate that machine learning based approach yields promising
results on students’ evaluation of higher educational institutions.
Keywords: Sentiment analysis Machine learning Text mining Ensemble

learning
1 Introduction
The enormous quantity of user-generated information was shared on the web with the
developments in information and communication technology. The user-generated text
documents include product reviews, film reviews and hotel reviews. Sentiment analysis
is the process of identifying and classifying the views of users from text documents into
different sentiments, such as, positive, negative and neutral [1]. Sentiment analysis can
be employed to extract structured and informative knowledge from unstructured text
pieces. This knowledge can serve as an important source of information for decision
1694 M. A. Toçoğlu and A. Onan
support systems and individual decision makers [2]. Sentiment analysis on educational
data can be utilized to obtain feedback on learning contents and resources, which can
provide useful insights for improving the quality of learning material and recognizing
student-learning behaviors. Sentiment analysis can play an important role in education,
where student input is crucial for assessing the effectiveness of higher educational
institutions. Sentiment analysis methods can be broadly divided into two groups, as
lexicon-based methods and machine-learning based methods [3]. For the lexicon-based
methods, the semantic orientation of words has been identified based on a dictionary of
positive and negative emotion values. In this way, sentiment of text document has been
identified. For the machine-learning based methods, sentiment analysis has been
modelled as a classification problem, in which a labelled set of text document has been
utilized in conjunction with supervised learning algorithms, such as, Naïve Bayes,
support vector machines, k-nearest neighbour algorithm and random forest classifier
[4]. In addition to conventional supervised learning algorithms, ensemble learning
methods may be employed to enhance the predictive performance. In this paper, we
present a machine learning based approach to sentiment analysis on students’ evalu-
ation of higher educational institutions. We analyse a corpus containing approximately
700 student reviews written in Turkish, with the use of conventional text representation
schemes and machine learning classifiers. In the experimental analysis, three conven-
tional text representation schemes (i.e., term-presence, term-frequency, TF-IDF
scheme) and three N-gram models (1-gram, 2-gram and 3-gram) have been consid-
ered in conjunction with four classifiers (i.e., support vector machines, Naïve Bayes,
logistic regression, and random forest algorithm). The predictive performance of four
ensemble learners (i.e., AdaBoost, Bagging and Random Subspace and voting algo-
rithm) have been also evaluated. The empirical results indicate that machine learning
based approach yields promising results on students’ evaluation of higher educational
institutions.
The rest of this paper is structured as follows: In Sect. 2, earlier research contri-
butions in sentiment analysis on education are presented. In Sect. 3, the methodology
of the study (i.e., the process of data set collection, feature extraction schemes, clas-
sification algorithms and ensemble learning methods) is presented. In Sect. 4, experi-
mental procedure and the empirical results are presented. In Sect. 5, the concluding
remarks of the study have been given.
2 Related Work
Feedback on learning contents and tools can be utilized to collect sentiment analysis
data on education domain, which can provide potentially useful insights to improve the
quality of learning materials. Adamopoulos [5] employed sentiment analysis on mas-
sive online open course reviews to model factors on student retention, such as course,
platform and university. In another study, Valakunde and Patwardhan [6] presented
sentiment analysis on student evaluation reviews. In this paper, text documents have
been represented by TF-IDF weighting scheme. The predictive performances of two
supervised learning algorithms (namely, Naïve Bayes and support vector machines)
have been evaluated. In a similar way, Wen et al. [7] employed sentiment analysis on
Sentiment Analysis on Students’ Evaluation of Higher Educational Institutions 1695
MOOC forum posts to model the characteristics of drop out on students. The analysis
indicated that there is a significant correlation between the sentiments of forum posts
and the completion rates for MOOC discussion forum posts. In a similar way, Altra-
bsheh et al. [8] presented a machine learning based sentiment analysis scheme on text
feedbacks provided by students. In this scheme, feedbacks of students on several
courses, such as, calculus, communication skills, database and engineering have been
evaluated. Text documents have been represented with three N-gram based models
(i.e., unigram, bigram and trigram). The predictive performances of feature extraction
schemes have been evaluated in conjunction with four classification algorithms (i.e.,
Naïve Bayes algorithm, support vector machines, maximum entropy classifier and
random forest algorithm). In another study, Bogdan [9] employed sentiment analysis on
course reviews on MOOCs. Similarly, Abdi et al. [10] evaluated students’ comments
about performance of instructors. In the presented scheme, two supervised learning
algorithms (namely, support vector machines and random forest algorithm) have been
employed. Moreno-Marcos et al. [11] examined the predictive performance of lexicon-
based and machine learning based schemes for sentiment analysis on MOOC forum
posts. In this scheme, several supervised learning methods, such as, logistic regression,
support vector machines, decision trees, random forest and Naïve Bayes algorithm have
been employed. Recently, Santos [12] presented a text mining and sentiment analysis
based scheme to identify the drivers of online success of higher educational
institutions.
3 Methodology
This section presents the methodology of the study. Namely, the dataset collection and
pre-processing, feature extraction, classification algorithms and ensemble methods
have been briefly presented.
3.1 Dataset Collection and Preprocessing

In this study, we created a dataset for sentiment analysis on students’ evaluation of
higher educational institutions. To do so, we have collected student reviews from
several microblogging platforms. The sentiment annotation for the forum posts have
been manually conducted as positive and negative. The dataset utilized in the empirical
analysis contains 372 negative and 326 positive text messages regarding the university
evaluations.
3.2 Feature Extraction Schemes

For natural language processing and language modelling, one conventional scheme is
N-gram based modelling. An n-gram is an adjacent sequence of n items from a given
text document. Those artifacts may be phonemes, syllables, letters, words, or characters
in this scheme. Word-based n-grams and character n-grams were widely used in the
study of the natural language tasks. N-grams of size were referred to as unigrams,
N-grams of size 2 were referred to as bigrams and N-grams of size 3 were referred to as
trigrams. To represent text documents, we used word-based n-gram models, which

considered unigrams, bigrams and trigrams. We considered three different schemes in
the vector space model (VSM) to represent text documents, namely, the term-presence
based representation, term frequency-based representation and TF-IDF based repre-
sentation. In terms of frequency-based representation, the number of occurrence of
words in the documents have been counted, namely, each document has been repre-
sented by an equal length vector with the corresponding word counts. Let t denote a
phrase in document d. The term frequency of t has been specified as TF(t, d). In term
presence-based representation, presence or absence of a word in a given document has
been utilized to represent text documents, such that a particular word t is represented as
1 if it is present on a particular document d and zero, otherwise.
3.3 Classification Algorithms

In the experimental analysis, three conventional text representation schemes (i.e., term-
presence, term-frequency, TF-IDF scheme) and three N-gram models (1-gram, 2-gram
and 3-gram) have been considered in conjunction with four classifiers (i.e., Naïve
Bayes, support vector machines, random forest and logistic regression). Naïve Bayes
(NB) algorithm is a statistical supervised learning algorithm based on Bayes’ theorem
and conditional independence assumption [13]. Support vector machines (SVM) are
linear supervised learning algorithms that may be utilized for classification and
regression tasks. SVM finds a hyperplane in the higher dimensional space to separate
instances of different classes [14]. Logistic regression (LR) is a linear classification
algorithm, which provides a scheme to apply linear regression to classification prob-
lems. In this scheme, a linear regression model and transformed target variables have
been employed to obtain a linear classification scheme [15]. Random forest (RF) al-
gorithm is an ensemble of bagging algorithm and random subspace algorithm. In this
algorithm, decision trees have been employed as the base learner. Each tree has been
built based on bootstrap samples of the training data. The diversity among the base
learners have been provided by a random feature selection. In this way, the model
yields satisfactory results in the existence of noisy or irrelevant data [16].
3.4 Ensemble Methods

Ensemble learning (also, known as, multiple classifier systems) is the process of
combining the predictions of multiple learning algorithms and combining their pre-
dictions by regarding them as a committee of decision makers. Ensemble learning aims
to identify a learning model with higher predictive performance [17]. The algorithms
have been briefly described: AdaBoost is a boosting based ensemble learning algo-
rithm, in which a more robust classification model has been obtained by focusing on
the instances that are harder to learn [18]. Bagging (also known as, Bootstrap aggre-
gating) is another ensemble algorithm, which combines base learning algorithms
trained on different training subsets obtained from the original training set by the
bootstrap sampling [19]. Random Subspace algorithm is another ensemble learning
method, which achieves diversity among the members of the ensemble with feature-
space based partition [20].
4 Experimental Procedure and Results
In this section, evaluation measures, experimental procedure and experimental results

have been presented. To evaluate the performance of classification algorithms, clas-
sification accuracy has been utilized. In the empirical analysis, we used 10-fold cross
validation. In this scheme, the original dataset is randomly divided into ten mutually
exclusive folds. Training and testing process are repeated ten times and each part is
tested and trained ten times. The results reported in this section are the average results
for 10-folds. In the empirical analysis, three different feature extraction methods
(namely, term-presence based representation, term-frequency based representation and
TF-IDF weighting scheme) have been taken into consideration. In addition, different
sizes of n-gram (ranging from n = 1 to n = 3) have been evaluated. In addition, the
combination of unigram, bigram and trigram models have been also considered.
In the empirical analysis, we have considered three different text representation
schemes (namely, term presence, term frequency, TF-IDF) with three different stem-
ming schemes (i.e., F5, raw and SnowBall stemming). In Table 1, the classification
accuracy values obtained by different text representation schemes for feature size of
2000 have been presented. Regarding the predictive performance of different text
representation and stemming schemes listed in Table 1, raw stemming scheme (i.e.,
representation without any stemming process) generally outperforms the other stem-
ming schemes. For the compared configurations, TF-IDF scheme generally outper-
forms the two other weighting schemes (i.e., term presence and term frequency based
representation). Among all the compared configurations, the highest predictive per-
formance has been achieved by TF-IDF based text representation with raw stemming in
conjunction with support vector machines classifier.
Table 1. Classification accuracy values for conventional classifiers and representation schemes.
Classifier Text representation Feature size F5 Raw SnowBall
SVM TF-IDF 2000 0,726501035 0,73494824 0,7004762
SVM TF 2000 0,672049689 0,706356108 0,6575155
SVM TP 2000 0,67194617 0,691821946 0,6690476
NB TF-IDF 2000 0,664803313 0,673188406 0,6719255
NB TF 2000 0,650289855 0,679026915 0,6674948
NB TP 2000 0,654720497 0,684720497 0,659089
LR TF-IDF 2000 0,700393375 0,707784679 0,6992133
LR TF 2000 0,703540373 0,703436853 0,6777019
LR TP 2000 0,69057971 0,713416149 0,7005383
RF TF-IDF 2000 0,674906832 0,658923395 0,639089
RF TF 2000 0,664824017 0,677846791 0,6690683
RF TP 2000 0,692008282 0,673312629 0,6533126
In Table 2, the classification accuracy values for different N-gram models and
conventional classifiers have been presented. Regarding the three weighting schemes,
TF-IDF scheme generally outperforms the two other weighting schemes. For support
vector machines and random forest classifier, unigram model outperforms the other N-
gram models. For Naïve Bayes and logistic regression classifiers, the utilization of
unigram and bigram feature sets yields higher predictive performance. In Table 3,
classification accuracy values for classification algorithms and ensemble methods have
been given for feature size of 2000 and raw stemming scheme. As it can be observed
from the results listed in Table 3, we cannot obtain any significant performance
enhancement in classification accuracy values with the use of ensemble methods.
Table 2. Classification accuracy values for different N-gram models and classifiers.
Classifier N-model Feature size TF-IDF TF TP
SVM unigrams 2000 0,73494824 0,706356108 0,6918219
SVM (uni-bi) grams 2000 0,723540373 0,683374741 0,6760663
SVM bigrams 2000 0,633167702 0,630310559 0,6320083
SVM (uni-bi-tri) grams 2000 0,739275362 0,680559006 0,7077019
SVM trigrams 2000 0,575983437 0,571863354 0,5701449
NB unigrams 2000 0,680559006 0,67621118 0,667764
NB (uni-bi) grams 2000 0,693519669 0,683354037 0,6863561
NB bigrams 2000 0,541573499 0,540082816 0,5416149
NB (uni-bi-tri) grams 2000 0,693436853 0,68184265 0,6834576
NB trigrams 2000 0,488467909 0,487329193 0,4898551
LR unigrams 2000 0,720703934 0,716376812 0,7134161
LR (uni-bi) grams 2000 0,726376812 0,720621118 0,7219669
LR bigrams 2000 0,636149068 0,634554865 0,6275155
LR (uni-bi-tri) grams 2000 0,729026915 0,706273292 0,7034783
LR trigrams 2000 0,571532091 0,565548654 0,5687164
RF unigrams 2000 0,676459627 0,677846791 0,6733126
RF (uni-bi) grams 2000 0,670517598 0,673312629 0,6705176
RF bigrams 2000 0,624616977 0,608881988 0,6146998
RF (uni-bi-tri) grams 2000 0,677515528 0,672153209 0,6604348
RF trigrams 2000 0,571656315 0,560227743 0,5615114
Table 3. Classification accuracy values for different N-gram models and classifiers.
N-model AdaBoost Bagging Random subspace Voting
SVM unigrams 0,532898551 0,704865424 0,707784679 0,707743271
SVM (uni-bi) grams 0,532753623 0,714741201 0,687826087 0,729233954
SVM bigrams 0,532857143 0,62610766 0,583043478 0,631884058
SVM (uni-bi-tri) grams 0,532898551 0,707681159 0,703540373 0,72068323
SVM trigrams 0,532981366 0,561614907 0,545817805 0,57447205
NB unigrams 0,670455487 0,670455487 0,619026915
NB (uni-bi) grams 0,695093168 0,686190476 0,630372671
NB bigrams 0,62447205 0,531635611 0,500124224
NB (uni-bi-tri) grams 0,674679089 0,693395445 0,680414079
NB trigrams 0,57010352 0,482774327 0,487184265
LR unigrams 0,532732919 0,704782609 0,689026915
LR (uni-bi) grams 0,532960663 0,704968944 0,699109731
LR bigrams 0,532857143 0,607287785 0,593250518
LR (uni-bi-tri) grams 0,533022774 0,719296066 0,693478261
LR trigrams 0,533064182 0,563022774 0,554430642
RF unigrams 0,691821946 0,681966874 0,690538302
RF (uni-bi) grams 0,684596273 0,659026915 0,673333333
RF bigrams 0,620393375 0,604720497 0,617515528
RF (uni-bi-tri) grams 0,670559006 0,656086957 0,666149068
RF trigrams 0,575797101 0,552981366 0,567494824
5 Conclusion
Sentiment analysis (also known as, opinion mining) is the process of identifying the
sentiment orientation of text documents, as positive, negative and neutral. Sentiment
analysis is an important application field to obtain structured information from text
documents. In this paper, we present a comprehensive analysis of conventional clas-
sifiers, ensemble methods and text representation schemes for sentiment analysis on
students’ evaluation of higher educational institutions. The empirical results indicate
that the predictive performance of classifiers enhance as the size of feature set increases.
Regarding the different term weighting schemes taken into consideration TF-IDF
scheme generally outperforms the two other schemes. The empirical results indicate
that machine learning based approach yields promising results on students’ evaluation
of higher educational institutions.
References
1. Pang, P., Lee, L.: Opinion mining and sentiment analysis. Found. Trends Inf. Retrieval 2, 1–
135 (2008)
2. Fersini, E., Messina, E., Pozzi, F.A.: Sentiment analysis: Bayesian ensemble learning. Decis.
Support Syst. 68, 26–38 (2014)
3. Medhat, W., Hassan, A., Korashy, H.: Sentiment analysis algorithms and applications: a
survey. Ain Shams Eng. J. 5(4), 1093–1113 (2014)
4. Onan, A., Korukoğlu, S., Bulut, H.: Ensemble of keyword extraction methods and classifiers
in text classification. Expert Syst. Appl. 57, 232–242 (2016)
5. Adamopoulos, P.: What makes a great MOOC? An interdisciplinary analysis of student
retention in online course. In: Proceedings of ICIS 2013, pp. 1–21. ACM, New York (2013)
6. Valakunde, N.D., Patwardhan, M.S.: Multi-aspect and multi-class based document sentiment
analysis of educational data catering accreditation process. In: Proceedings of International
Conference on Cloud and Ubiquitous Computing and Emerging Technologies, pp. 188–192,
IEEE, New York (2013)
7. Wen, M., Yang, D., Rose, C.: Sentiment analysis in MOOC discussion forums: what does it
tell us? In: Proceedings of 7th International Conference on Educational Data Mining 2014,
pp. 130–137, ACM, New York (2014)
8. Altrabsheh, N., Cocea, M., Fallahkhair, S.: Predicting learning-related emotions from
students’ textual classroom feedback via Twitter. In: 8th International Conference on
Educational Data Mining 2015, pp. 436–439, ACM, New York (2015)
9. Bogdan, R.: Sentiment analysis on embedded systems blended courses. Broad Res. Artif.
Intell. Neurosci. 8(1), 17–23 (2017)
10. Abdi, A., Shamsudding, S.M., Aliguliyev, R.M.: QMOS: query-based multi-documents
opinion-oriented summarization. Inf. Process. Manag. 54(2), 318–338 (2018)
11. Moreno-Marcos, P.M., Alario-Hoyos, C., Munoz-Merino, P.J., Kloos, C.: Sentiment
analysis in MOOCs: a case study. In: IEEE Global Engineering Education Conference 2018,
pp. 1489–1496. IEEE, New York (2018)
12. Santos, C.L., Rita, P., Guerreiro, J.: Improving international attractiveness of higher
education institutions based on text mining and sentiment analysis. Int. J. Educ. Manag. 32
(3), 431–447 (2018)
13. Lewis, D.D.: Naïve (Bayes) at forty: the independence assumption in information retrieval.
In: European Conference on Machine Learning 1998, pp. 4–15, Springer, Berlin (1998)
14. Vapnik, V.: The Nature of Statistical Learning Theory. Springer, Heidelberg (2013)
15. Hastie, T., Tibsharani, R., Friedman, J.: Springer series in statistics. Elements Math. Intell.
27, 83–85 (2009)
16. Breiman, L.: Random forests. Mach. Learn. 45(1), 5–32 (2001)
17. Freund, Y., Schapire, R.E.: Experiments with a new boosting algorithm. ICML 96, 148–156
(1996)
18. Breiman, L.: Bagging predictors. Mach. Learn. 24(2), 123–140 (1996)
19. Ho, T.K.: The random subspace method for constructing decision forests. IEEE Trans.
PAMI 20(8), 832–844 (1998)
20. Wolpert, H.: Stacked generalization. Neural Netw. 5(2), 241–259 (1992)
Author Index
A Aydin, Nizamettin, 1431

Abdollahi, Saranaz, 1164 Aydın, Serhat, 293
Abdollahzadeh, Sohrab, 484 Aydogdu, Ebru, 1267
Abdurasulov, Fakhritdin, 1589 Aygülen, Berna Tektaş, 1407
Abiyev, Rahib H., 964 Aygün, Halis, 1258, 1267
Acarbay, Algı, 616 Aytore, Can, 582
Adalı, Esra Aytaç, 733, 742
Adilov, Farukh, 1589 B
Ahn, Youngshin, 1100 Babbar, Anuradha Verma, 903
Akbaş, Muhammet Fatih, 1156 Bacanin, Nebojsa, 718, 955
Akdemir, Halid, 1242 Bagriyanik, Selami, 171
Akın, Özgün, 817 Bahçevan, Cenk Anıl, 833
Aklan, Sami Can, 1116 Bakioglu, Gozde, 503
Aktas, Ahmet, 265 Barranco, Manuel J., 894
Alaybeyoğlu, Ayşegül, 1242 Basar, Murat, 534
Albayrak, Inci, 1483 Bayturk, Engin, 1523
Aldemir, Başak, 1267 Belyakov, Stanislav, 10, 18, 1475
Alem, Sarper, 557 Bereketli, Ilke, 608
Alkan, Nurşah, 884, 1373, 1447 Besri, Zineb, 331
Alkhanafseh, Yousef, 833 Bezdan, Timea, 718, 955
Alptekin, Emre, 347 Bharatraj, Janani, 372
Ari, Ergun, 642 Biçer, Cenker, 1125
Arıoğlu, Mahmure Övül, 1008 Bilgiç, Emrah, 625
Armağan, Ersin, 1068 Bloshko, Yurii, 1581
Arman, Kevser, 658 Bolat, Bersam, 1363
Arslan, Armagan, 1092 Boltürk, Eda, 110, 1675
Arslanoğlu, Yasin, 1149 Boulmakoul, Azedine, 331, 779, 825, 933,
Asan, Umut, 1172 1059, 1183
Asrin, Fauzan, 807 Bouzid, Mariam, 1573
Atahan, Ali Osman, 503 Bozhenyuk, Alexander, 10, 18, 1475
Atak, Üstün, 1149 Bozov, Hristo, 910
Atanassov, Krassimir, 26 Bunyak, Rabia Burcu, 833
Atik, Erdem, 1133 Bureva, Veselina, 674
Atik, Serhat, 591 Buyuk, Aysu Melis, 151
Aydemir, Salih Berkan, 440 Büyüközkan, Kadir, 1440
https://doi.org/10.1007/978-3-030-51156-2
1702 Author Index
C Erez, Arif Sercan, 195

Çağlıyor, Sandy, 1077 Ergin, Fatih, 1015
Çağrı Budak, M., 257 Erkal, Gizem, 381
Cagri Tolga, A., 525 Ertug, Abdullah, 1666
Çakır, Altan, 246, 817, 833 Ervural, Beyzanur Cayir, 512, 1635
Çakır, Esra, 302, 311 Esen, M. Fevzi, 625
Cakmakci, Mehmet, 195 Eskiocak, Defne İdil, 180
Campión, María Jesús, 1301 Esnaf, Sakir, 1523
Cao, Zhangjun, 990
Castro, Jorge, 894 F
Çaylı, Gül Deniz, 1233 Fan, Ching-Lung, 788
Çaylı, Özkan, 1532 Farrokhizadeh, Elmira, 432, 457, 466, 475, 484
Çebi, Ferhan, 213, 1116 Fazekas, Z., 1059
Cebi, Selcuk, 769, 1327, 1345
Çelebi, Numan, 1596 G
Çevik Onar, Sezi, 257 Galar, Mikel, 1023
Chen, Fei, 1100, 1499, 1515 Gáspár, P., 1059
Cherradi, Ghyzlane, 933, 1059 Gedikli, Tolga, 512
Choi, Jaeho, 1100, 1499, 1515 Georgescu, Irina, 1292
Cieraszewska, Urszula, 1683 Georgieva, Penka V., 1216
Cinaroglu, Songul, 54 Gerasimenko, Evgeniya, 981
Çoban, Veysel, 274, 415 Gökçen, Alkım, 1540
Collan, Mikael, 1292 Guan, Bing, 1499
Gundoğdu, Fatma Kutlu, 101, 110
D Gündüz, Sevcan Yilmaz, 440
Dagdeviren, Metin, 87 Güner, Elif, 1258, 1267
Daissaoui, Abdellah, 1183 Güngör, Cengiz, 1156
Dal, Mehmet Berkin, 1084 Gurcan, Omer Faruk, 875
Deliktas, Derya, 1423
Demir, Ezgi, 666 H
Demircan, Murat Levent, 616 Haktanır, Elif, 1385, 1396
Deniz, Halil Faruk, 817 Hamal, Serhan, 1008
Denizci, Aykut, 1250 Hamerska, Monika, 1683
Denizhan, Berrin, 634 Hasab, Mohanad Abdulkareem Hasan, 390
Digalwar, Abhijeet, 649 Hasgul, Zeynep, 582
Dizbay, İkbal Ece, 682 Hemam, Sofiane Mounine, 63
Dogan, Onur, 39, 47, 875 Henge, Santosh Kumar, 221, 903
Doğruak, Utku, 71, 79 Hioual, Ouassila, 63
Dominguez-Catena, Iris, 1023 Hioual, Ouided, 63
Donyatalab, Yaser, 432, 457, 466, 475, 484 Hiziroglu, Abdulkadir, 39
Durak, Gul, 525 Hung, Nguyen Nhut, 397
Dursun, Mehtap, 642
I
E İçer, Çağlar, 246
Emin Baysal, M., 319, 1440 Idemen, Bengi Tugcu, 1284
Engin, Orhan, 319, 1440 İkizceli, Türkan, 1125
Erbay, Hasan, 949, 1125 Ilbahar, Esra, 769, 1327
Ercan-Tekşen, Hatice, 750 Imamguluyev, Rahib, 1644
Erdem, Yusra, 769 Imrenci, Rasit, 195
Erdogan, Melike, 550 Induráin, Esteban, 1301
Author Index 1703
Israfilova, Elvira, 1092 Küçükali, İrem, 941

Ivanyan, Arsen, 1589 Kucukdeniz, Tarik, 1523
Kuliev, Elmar, 842, 1491
J Kumam, Poom, 355
Jaller, Miguel, 118 Kumar, Anish, 567
Kumar, M. A. Vinod, 238
K Kumar, Pradeep, 567
Kabak, Mehmet, 87, 265, 293 Kumbasar, Tufan, 1068
Kacprzyk, Janusz, 10, 18 Kureichik, Vladimir, 981
Kadan, Merve, 576, 1199 Kursitys, Ilona, 842, 1491
Kahraman, Cengiz, 127, 134, 142, 339, 494, Kurubaş, Erhan, 833
701, 884, 916, 1316, 1327, 1385, 1396,
1447, 1655 L
Kalender, Zeynep Tugce, 381 Labella, Á., 1506
Karaarslan, Enis, 1156 Lbath, Ahmed, 779, 825, 933, 1183
Karabıyık, Esra, 833 Lula, Paweł, 1683
Karaçavuş, Seyhan, 1125 Luu, Nguyen Phu Thuong, 1604, 1612
Karadeniz, Sinan, 1250
Karakul, Aygülen Kayahan, 1048, 1355 M
Karakul, Kayahan, 1407 Maden, Ayça, 347
Karasan, Ali, 550, 701 Makav, Burak, 1532
Karayazi, Ferhat, 608 Mangla, Sachin Kumar, 567
Karim, Lamia, 779, 825, 933, 1059, 1183 Marchwicka, Ewa, 1548
Kaska, Melis, 543 Martínez, L., 1506
Katrancı, Ali, 658 Mashinchi, Mashaallah, 916
Kavi, Mert, 1225 Masmoudi, Oussama, 1573
Kaya, Tolga, 71, 79, 213, 850, 1092, 1149 Mehr, Ali Danandeh, 1242
Kayikci, Yasanur, 567 Memari, Nogol, 1164
Keivanpour, Samira, 690, 797 Mercan, Öykü Berfin, 1276
Kemik, Hasan, 1084 Mert, Buse, 180
Khalil, Shuker Mahmood, 390 Mieszkowicz-Rolka, Alicja, 758
Khalilpoor, Parisa, 1309 Milošević, Dušan, 599
Khan, Muhammad Jabir, 355 Milošević, Mimica, 599
Khodabakhsh, Sonia, 1164 Mirkhan, Amer, 1596
Kilic, Huseyin Selcuk, 381 Mizrahi, Rozi, 1048, 1355, 1407
Kılıç, Volkan, 1276, 1532 Moghbel, Mehrdad, 1164
Kim, Jong-Myon, 858, 1107 Mohamed, Ali, 1031
Kim, Sumi, 1100, 1515 Murat, Miraç, 1172
Kinnunen, Jani, 1292
Kizi, Isgandarova Ilhama Tarlan, 998 N
Kizi, Mikayilova Rena Nuru, 973 Nahri, M., 779
Kızıltan, Alp, 817 Namlı, Özge H., 941
Knyazeva, Margarita, 18, 1475 Natskevich, Alexandr, 1491
Koca, Gözde, 666 Navarro-Arribas, Guillermo, 3
Kocken, Hale, 1483 Nazlı Günesen, S., 850
Köklü, Murat, 204 Nefis, Doğukan, 817
Komurcu, Yesim, 1464 Nouri, Faranak, 941
Kondratenko, Galyna, 1456 Nucci, F., 1620
Kondratenko, Yuriy, 1456
Korobov, Alexander, 709 O
Kosenko, Olesiya, 1475 Oglu, Alekperov Ramiz Balashirin, 228, 998
Koyuncu, Yusuf Mertkan, 941 Oglu, Salahli Vuqar Mamadali, 228
Kravchenko, Yury, 842, 1491 Olar, Oksana, 1581
Kuchta, Dorota, 726, 1206, 1548 Omeroglu, Gizem, 1666
1704 Author Index
Onan, Aytuğ, 1133, 1532, 1693 Şahin, Emir, 634

Onar, Sezi Çevik, 127, 134, 142, 171, 274, Şahin, Savaş, 1540
339, 415, 494, 1655 Saide, Saide, 807
Oniz, Yesim, 1084 Sanchez, Pedro J., 894
Organ, Arzu, 658 Saraswat, S. K., 649
Otay, Irem, 118, 339, 423, 591, 1655 Sari, Irem Ucal, 448, 1316
Ozceylan, Eren, 87 Sarucan, Ahmet, 319, 1440
Özçil, Abdullah, 733, 742 Seifi, Seyyed Hadi, 457, 466
Ozdemir, Cagatay, 171 Seker, Sukran, 1336
Özdemir, Mehmet Hilmi, 576, 1199 Sen, Durmus Tayyar, 512
Özgün, Kağan, 1116 Şengör, N. Serap, 941
Ozkan, Baris, 87 Sennaroglu, Bahar, 1008
Özkan, Gökhan, 576, 1199 Sergi, Duygu, 448
Ozkavukcu, Aysegul, 1666 Seyfi Shishavan, Seyed Amin, 432, 457, 475
Özok, Ahmet Fahri, 32 Seymen, Omer Faruk, 39
Öztaş, Gülin Zeynep, 733, 742 Sezer, Emine, 1284
Öztaş, Tayfun, 733, 742 Sezer, Emrah, 867
Öztayşi, Basar, 127, 134, 142, 339, 494, 1077, Sezer, Hülya Başeğmez, 867
1225, 1655 Shah, Vipul A., 1039
Öztürk, Işıl, 180 Shishavan, Seyed Amin Seyfi, 466, 484
Öztürkoğlu, Ömer, 682 Sidenko, Ievgen, 1456
Sindhu, M. Sarwar, 282
P Singh, Bhupinder, 221
Pala, İrem, 634 Sivri, Mustafa, 1483
Parchami, Abbas, 916, 924, 1309, 1415 Sotirov, Sotir, 26, 910
Parlak, Ismail Burak, 1015 Sotirova, Evdokia, 26, 910
Parmaksız, Deniz, 246 Soydemir, Mehmet Uğur, 1540
Pashaei, Elham, 1431 Stanojević, Ana, 599
Pashaei, Elnaz, 1431 Strumberger, Ivana, 718, 955
Patel, Himanshukumar R., 1039 Suraj, Zbigniew, 1581
Paternain, Daniel, 1023
Phu, Nguyen Dinh, 397 T
Piltan, Farzin, 858, 1107 Tan, Fatma Gülşah, 187
Poleshchuk, Olga M., 1191 Taranov, Mykyta, 1456
Taşdemir, Sakir, 204
Q Tekin, Ahmet Tezcan, 213, 1116
Qian, Ji, 1140 Tektaş, Berna, 1048, 1355
Quynh, Le Thi Ngoc, 397 Temelcan, Gizem, 1483
Temur, Gul Tekin, 151, 160
R Toçoğlu, Mansur Alp, 1693
Ran, Meng, 1140 Tolga, A. Cagri, 534, 543
Rashid, Tabasam, 282 Torra, Vicenç, 3
Ratna, Silvia, 807 Tranev, Stoyan, 363, 1555
Raval, Sejal, 1039 Traneva, Velichka, 363, 1555
Raventós-Pujol, Armajac, 1301 Tuba, Eva, 718, 955
Rayis, Osama, 1031 Tuba, Milan, 718, 955
Rezaei, Saeideh, 1164 Tunç, Ali, 204
Rodríguez, R. M., 1506 Tuş, Ayşegül, 733, 742
Rolka, Leszek, 758 Tuzkaya, Gulfem, 381
Rozenberg, Igor, 10
U
S Ulu, Cenk, 1250
Şahin, Abdulgani, 1564 Ulukan, Ziya, 302, 311
Şahin, Bünyamin, 1564, 1628 Umutcan Ay, H., 850
Author Index 1705
Unal, Yagmur, 160 Yali, Rao, 1140

Unalir, Murat Osman, 1284 Yaman, Tutku Tuncalı, 625, 867
Urhan, Mustafa, 1423 Yanık, Seda, 941, 1464
Yaşlı, Fatma, 1363
V Yatsalo, Boris, 709
Vasilev, Valentin, 910 Yera, Raciel, 894
Vasilev, Valetin, 26 Yiğit, Ahmet Talha, 71, 79
Yildirim, Nihan, 1092
W Yıldız, Tuğba, 423
Wadii, Basmi, 825 Yörükoğlu, Mehmet, 293
Wang, Dong, 990 Yu, Wenjun, 1515
Wang, Lachun, 990 Yüksel, Asım Sinan, 187
Wei, Shen, 1140 Yurttakal, Ahmet Haşim, 949, 1125
Wenda, Alex, 807 Yusupbekov, Nodirbek, 1589
Y
Yadav, S. S., 649 Z
Yalaoui, Alice, 1573 Zabor, Adam, 1206
Yalçın, Adnan, 1133 Zaporozhets, Dmitry, 842
Yalcin, Ahmet Selcuk, 381 Zivkovic, Miodrag, 718, 955

Intelligent and Fuzzy Techniques: Smart and Innovative Solutions

Uploaded by

Copyright:

Available Formats

Intelligent and Fuzzy Techniques: Smart and Innovative Solutions

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Intelligent and Fuzzy Techniques: Smart and Innovative Solutions

Uploaded by

Copyright:

Available Formats

Advances in Intelligent Systems and Computing 1197

Cengiz Kahraman · Sezi Cevik Onar ·

More information about this series at http://www.springer.com/series/11156

Basar Oztaysi Irem Ucal Sari

Selcuk Cebi A. Cagri Tolga

Intelligent and Fuzzy

Basar Oztaysi Irem Ucal Sari

Selcuk Cebi A. Cagri Tolga

ISSN 2194-5357 ISSN 2194-5365 (electronic)

INFUS is an acronym for Intelligent and Fuzzy Systems. It is a well-established

Invited Speakers’ Papers

A Hybrid Approach for Clustering and Selecting of Cloud Services

Picture Fuzzy Sets and Spherical Fuzzy Sets

Estimation and Prediction

Hesitant Fuzzy Sets

Solar Energy Power Plant Investment Selection with Unbalanced

Intuitionistic Fuzzy Sets

Interval Valued Intuitionistic Fuzzy Gaussian Membership Function:

Pythagorean Fuzzy Sets and Q-Rung Orthopair Fuzzy Sets

Hamacher Aggregation Operators Based on Interval-Valued q-Rung

Multicriteria Decision Making-Applications

Investigating Enablers to Improve Transparency in Sustainable Food

InterCriteria Analysis Applied to Emerging Europe and Central Asia

Multicriteria Decision Making-Theory

Intelligent Statistical Analyses

Realising Newspaper Sales by Using Statistic Methods . . . . . . . . . . . . . 875

Comprehensive Research of the Commodity Market: Theoretical,

Predicting Movie Ratings with Machine Learning Algorithms . . . . . . . . 1077

A Compromise-Based New Approach to Learning Fuzzy

Center-of-Gravity Real Options Method Based on Interval-Valued

Defects Control Charts Using Interval-Valued Penthagorean

A Bioinspired Algorithm for Improving the Effectiveness

Finding the Optimal Features Reduct, a Hybrid Model of Rough Set

Review and Discussion Papers

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1701

Vicenç Torra1,2(B) and Guillermo Navarro-Arribas3

Abstract. The amount of information currently available is a threat to

Keywords: Data privacy · Fuzzy sets and systems · Fuzzy meets

2 Data Protection Methods Based on Fuzzy Techniques

Given a database X, a masking method is a function ρ that applied to X returns

– Microaggregation is a very eﬀective method for data protection. Eﬀective

Microaggregation proceeds with the construction of small clusters of data,

3 Information Loss Measures Based on Fuzzy Techniques

4 Disclosure Risk Measures Based on Aggregation

5 Conclusion and Future Work

Acknowledgements. Partially supported by Vetenskapsrådet project: “Disclosure

Stanislav Belyakov1 , Alexander Bozhenyuk1(&) ,

Keywords: Intellectual geo-information systems Context Analysis

2 Review of Related Publications

Modern GIS implement the basic principles of cartographic visualization [1] in

3 Management of Analysis Process

5. The recommender subsystem calls the function of perception control.

4 Conceptual Model of Knowledge

reliable decisions in new, previously unstudied conditions. Indication of permissible

5 Optimal Number of Contexts

Consider the issue of the applicability of context-based planning related to their

It is easy to verify that the minimum cost is achieved when

This expression (2) allows us to draw the following conclusions:

Alexander Bozhenyuk1(&) , Stanislav Belyakov1 ,

ð8xi 2 VÞ½xi 2 Ba _ ð9xj 2 Ba jlðxj ; xi Þ la & mðxj ; xi Þ ma Þ: ð7Þ

proposition ((xj,xi),m(xj,xi)) a. Passing from the quantiﬁer form of proposition (7) to

a _ a & b ¼ a; a & b _ a & b ¼ a; n1 & a _ n2 & a & b ¼ n1 & aif n1 n2 : ð9Þ

ha; bi hc; d i if and only if a c and b d:

hl1 ; m1 i hl2 ; m2 i . . . hln ; mn i and b1 b2 . . . bn

hold. When there exists i, 1 i n 1 for which