Advances in Intelligent Systems and Computing 634

Valentina Emilia Balas

Lakhmi C. Jain
Marius Mircea Balas Editors

Soft Computing
Applications
Proceedings of the 7th International
Workshop Soft Computing Applications
(SOFA 2016), Volume 2
Advances in Intelligent Systems and Computing

Volume 634

Series editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
e-mail: [email protected]
About this Series
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.
The publications within “Advances in Intelligent Systems and Computing” are primarily
textbooks and 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.
Advisory Board
Chairman
Nikhil R. Pal, Indian Statistical Institute, Kolkata, India
e-mail: [email protected]
Members
Rafael Bello Perez, Universidad Central “Marta Abreu” de Las Villas, Santa Clara, Cuba
e-mail: [email protected]
Emilio S. Corchado, University of Salamanca, Salamanca, Spain
e-mail: [email protected]
Hani Hagras, University of Essex, Colchester, UK
e-mail: [email protected]
László T. Kóczy, Széchenyi István University, Győr, Hungary
e-mail: [email protected]
Vladik Kreinovich, University of Texas at El Paso, El Paso, USA
e-mail: [email protected]
Chin-Teng Lin, National Chiao Tung University, Hsinchu, Taiwan
e-mail: [email protected]
Jie Lu, University of Technology, Sydney, Australia
e-mail: [email protected]
Patricia Melin, Tijuana Institute of Technology, Tijuana, Mexico
e-mail: [email protected]
Nadia Nedjah, State University of Rio de Janeiro, Rio de Janeiro, Brazil
e-mail: [email protected]
Ngoc Thanh Nguyen, Wroclaw University of Technology, Wroclaw, Poland
e-mail: [email protected]
Jun Wang, The Chinese University of Hong Kong, Shatin, Hong Kong
e-mail: [email protected]

More information about this series at http://www.springer.com/series/11156

Valentina Emilia Balas Lakhmi C. Jain
•

Marius Mircea Balas

Editors

Soft Computing
Applications
Proceedings of the 7th International
Workshop Soft Computing Applications
(SOFA 2016), Volume 2

123
Editors
Valentina Emilia Balas Marius Mircea Balas
Department of Automatics and Applied Universitatea Aurel Vlaicu Din Arad
Informatics, Faculty of Engineering Arad
Aurel Vlaicu University of Arad Romania
Arad
Romania

Lakhmi C. Jain
University of Canberra
Canberra, SA
Australia

ISSN 2194-5357 ISSN 2194-5365 (electronic)

Advances in Intelligent Systems and Computing
ISBN 978-3-319-62523-2 ISBN 978-3-319-62524-9 (eBook)
DOI 10.1007/978-3-319-62524-9
Library of Congress Control Number: 2017952376

© Springer International Publishing AG 2018

This work is subject to copyright. All rights are reserved 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, express 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.

Printed on acid-free paper

This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

These two volumes constitute the Proceedings of the 7th International Workshop on
Soft Computing Applications, or SOFA 2016, which will be held on August 24–26,
2016, in Arad, Romania. This edition was organized by Aurel Vlaicu University of
Arad, Romania, University of Belgrade, Serbia, in conjunction with Institute of
Computer Science, Iasi Branch of the Romanian Academy, IEEE Romanian
Section, Romanian Society of Control Engineering and Technical Informatics
(SRAIT)—Arad Section, General Association of Engineers in Romania—Arad
Section and BTM Resources Arad.
Soft computing concept was introduced by Lotfi Zadeh in 1991 and serves to
highlight the emergence of computing methodologies in which the accent is on
exploiting the tolerance for imprecision and uncertainty to achieve tractability,
robustness, and low solution cost. Soft computing facilitates the use of fuzzy logic,
neurocomputing, evolutionary computing and probabilistic computing in combi-
nation, leading to the concept of hybrid intelligent systems.
Combining of such intelligent systems' tools and a large number of applications
can show the great potential of soft computing in all domains.
The volumes cover a broad spectrum of soft computing techniques, theoretical
and practical applications find solutions for industrial world, economic, and medical
problems.
The conference papers included in these proceedings, published post conference,
were grouped into the following area of research:
• Methods and Applications in Electrical Engineering
• Knowledge-Based Technologies for Web Applications, Cloud Computing,
Security Algorithms and Computer Networks
• Biomedical Applications
• Image, text and signal processing
• Machine Learning and Applications
• Business Process Management
• Fuzzy Applications, Theory and Fuzzy and Control
• Computational Intelligence in Education

v
vi Preface

• Soft Computing & Fuzzy Logic in Biometrics (SCFLB)

• Soft Computing Algorithms Applied in Economy, Industry and Communication
Technology
• Modelling and Applications in Textiles
• Methods and Applications in Electrical Engineering
In SOFA 2016, we had six eminent keynote speakers: Professor Michio Sugeno
(Japan), Professor Anna Esposito (Italy), Professor Mika Sato-Ilic (Japan),
Professor Valeriu Beiu (Romania), Professor Salil Bose (USA), Professor Rajeeb
Dey (India) and an interesting roundtable of Professor József Dombi (Hungary).
Their summaries talks are included in this book.
We especially thank the Honorary Chair of SOFA 2016 Prof. Lotfi A. Zadeh
who encouraged and motivated us, like to all the other SOFA editions.
We are thankful to all the authors that have submitted papers for keeping the
quality of the SOFA 2016 conference at high levels. The editors of this book would
like to acknowledge all the authors for their contributions and also the reviewers.
We have received an invaluable help from the members of the International
Program Committee and the chairs responsible for different aspects of the
workshop. We appreciate also the role of special sessions organizers. Thanks to all
of them we had been able to collect many papers of interesting topics, and during
the workshop, we had very interesting presentations and stimulating discussions.
For their help with organizational issues of all SOFA editions we express our
thanks to TRIVENT Company, Mónika Jetzin and Teodora Artimon for having
customized the software Conference Manager, registration of conference partici-
pants and all local arrangements.
Our special thanks go to Janus Kacprzyk (Editor in Chief, Springer, Advances in
Intelligent Systems and Computing Series) for the opportunity to organize this
guest edited volume.
We are grateful to Springer, especially to Dr. Thomas Ditzinger (Senior Editor,
Applied Sciences & Engineering Springer-Verlag) for the excellent collaboration,
patience, and help during the evolvement of this volume.
We hope that the volumes will provide useful information to professors,
researchers, and graduated students in the area of soft computing techniques and
applications, and all will find this collection of papers inspiring, informative, and
useful. We also hope to see you at a future SOFA event.

Valentina Emilia Balas

Lakhmi C. Jain
Marius Mircea Balas
Keynote Presentations
 Introduction to Choquet Calculus

Michio Sugeno

Tokyo Institute of Technology, Japan

Abstract. In this talk, we give a brief introduction to a recent topic

“Choquet calculus” where calculus consists of integrals and derivatives.
Modern integrals, namely Lebesgue integrals initiated by Lebesgue in
1902, are associated with the concept of “measures.” Lebesgue measures are
defined as additive set functions with certain conditions, and hence,
Lebesgue integrals hold additivity by inheriting the property of measures. In
the latter half of the twentieth century, a new concept of “measures without
additivity” named fuzzy measures was proposed by Sugeno in 1974. Fuzzy
measures (non-additive measures in general) are defined as monotone set
functions and considered as a natural extension of Lebesgue measures
leading to the concept of non-additive integrals with respect to non-additive
measures, where we note that monotonicity contains additivity in it.
In 1953, Choquet studied the so-called Choquet functionals based on
capacities, where capacities representing “potential energy” of a set were
also monotone set functions but not captured as “measures” as in the sense
of Lebesgue. Together with the appearance of fuzzy measures, Choquet
functionals were finally re-formulated as non-additive integrals with respect
to fuzzy measures by Höle in 1982. Since then, various non-additive inte-
grals with respect to non-additive measures have been suggested. Among
them, we focus on Choquet integrals which are the most general extension
of Lebesgue integrals
Once we obtain the concept of integrals, we become curious about their
inverse operations. In the case of Lebesgue integrals, Radon and Nikodym
gave Radon-Nikodym derivatives as inverse operations in 1913 and 1930,
respectively. It is well-known that with the aid of Radon-Nikodym derivatives,
Kolmogorov proved the existence of conditional probabilities in 1933 and thus
initiated modern probability theory, where probabilities are nothing but
Lebesgue measures. On the other hand in fuzzy measure theory, conditional
fuzzy measures have been not well defined in the sense of Kolmogorov. Very
recently, inverse operations of Choquet integrals were studied as “derivatives
with respect to fuzzy measures” (Sugeno 2013).
In this talk, we deal with Choquet calculus (Sugeno 2015) based on
Choquet integrals and derivatives. So far, most studies on Choquet integrals
have been devoted to a discrete case. In Choquet calculus, we deal with con-
tinuous Choquet integrals and derivatives as well. First, we show how to
calculate continuous Choquet integrals. To this aim, we consider distorted
Lebesgue measures as a special class of fuzzy measures, and nonnegative and
non-decreasing functions. The Laplace transformation is used as a basic tool
for calculations. Distorted Lebesgue measures are obtained by monotone
transformations of Lebesgue measures according to the idea of distorted
x M. Sugeno

probabilities suggested by Edwards in 1953. We remember that Kahneman

was awarded Nobel Prize in Economics in 2002, where his cumulative pro-
spect theory is based on “Choquet integrals with respect to distorted proba-
bilities.” Next, we define derivatives of functions with respect to distorted
Lebesgue measures, where the derivatives correspond to Radon-Nikodym
derivatives in the case of Lebesgue integrals. We also discuss the identification
of distorted Lebesgue measures which is a problem arising particularly in
Choquet calculus. Then, we show some relations between Choquet calculus
and fractional calculus which is recently getting very popular, for example, in
control theory. Also, we consider differential equations with respect to distorted
Lebesgue measures and give their solutions. Lastly, we present the concept of
conditional distorted Lebesgue measures defined with the aid of
Radon-Nikodym-like derivatives.

Biography
Michio Sugeno was born in Yokohama, Japan, in
1940. After graduating from the Department of
Physics, the University of Tokyo, he worked at
Mitsubishi Atomic Power Industry. Then, he served
the Tokyo Institute of Technology as research asso-
ciate, associate professor, and professor from 1965 to
2000. After retiring from the Tokyo Institute of
Technology, he worked as Laboratory Head at the
Brain Science Institute, RIKEN, from 2000 to 2005,
and Distinguished Visiting Professor at Doshisha
University from 2005 to 2010 and then Emeritus
Researcher at the European Centre for Soft
Computing, Spain, from 2010 to 2015. He is Emeritus Professor at the Tokyo
Institute of Technology.
He was President of the Japan Society for Fuzzy Theory and Systems from
1991 to 1993, and also President of the International Fuzzy Systems Association
from 1997 to 1999. He is the first recipient of the IEEE Pioneer Award in Fuzzy
Systems with Zadeh in 2000. He also received the 2010 IEEE Frank Rosenblatt
Award and Kampét de Feriét Award in 2012.
His research interests are Choquet calculus, fuzzy measure theory, nonlinear
control, and preference theory, applications of systemic functional linguistics and
language functions of the brain.
 Emotional Facial Expressions:
Communicative Displays
or Psychological Universals?

Anna Esposito

Department of Psychology, and IIASS, Seconda Università di Napoli, Italy

[email protected]; [email protected]
http://www.iiassvietri.it/anna.html

Abstract. Emotional feelings permeate our everyday experience, con-

sciously or unconsciously, driving our daily activities and constraining our
perception, actions, and reactions.
During daily interactions, our ability to decode emotional expressions
plays a vital role in creating social linkages, producing cultural exchanges,
influencing relationships, and communicating meanings.
Emotional information is transmitted through verbal (the semantic
content of a message) and nonverbal (facial, vocal, gestural expressions, and
more) communicative tools and relations and exchanges are highly affected
by the way this information is coded/decoded by/from the addresser/
addressee.
The accuracy above the chance in decoding facial emotional expres-
sions suggested they can play the role of psychological universals. However,
this idea is debated by data suggesting that they play the role of social
messages dependent upon context and personal motives.
These open questions are discussed in this talk, at the light of experi-
mental data obtained from several experiments aimed to assess the role of
context on the decoding of emotional facial expressions. The reported data
support more the idea that facial expressions of emotions are learned to
efficiently and effectively express intentions and negotiate relations, even
though particular emotional aspects show similarities across cultural
boundaries.
Research devoted to the understanding of the perceptual and cognitive
processes involved in the decoding of emotional states during interactional
exchanges is particularly relevant in the field of Human-Human,
Human-Computer Interaction and Robotics, for building and hardenind
human relationships, and developing friendly, emotionally, and socially
believable assistive technologies.
xii A. Esposito

Biography
Anna Esposito received her “Laurea Degree” summa
cum laude in Information Technology and Computer
Science from the Università di Salerno in 1989 with a
thesis on: The Behavior and Learning of a
Deterministic Neural Net (published on Complex
System, vol 6(6), 507–517, 992). She received her
PhD Degree in Applied Mathematics and Computer
Science from the Università di Napoli, “Federico II” in
1995. Her PhD thesis was on: Vowel Height and
Consonantal Voicing Effects: Data from Italian (pub-
lished on Phonetica, vol 59(4), 197–231, 2002) and
was developed at Massachusetts Institute of
Technology (MIT), Research Laboratory of Electronics (RLE), under the supervi-
sion of Professor Kenneth N Stevens.
She has been a Postdoc at the International Institute for Advanced Scientific
Studies (IIASS), and Assistant Professor at the Department of Physics at the
Università di Salerno (Italy), where she taught courses on cybernetics, neural net-
works, and speech processing (1996-2000). She had a position as Research
Professor (2000-2002) at the Department of Computer Science and Engineering at
Wright State University (WSU), Dayton, OH, USA. She is currently associated
with WSU as Research Affiliate.
Anna is currently working as an Associate Professor in Computer Science at the
Department of Psychology, Seconda Università di Napoli (SUN). Her teaching
responsibilities include cognitive and algorithmic issues of multimodal communi-
cation, human–machine interaction, cognitive economy, and decision making. She
authored 160+ peer-reviewed publications in international journals, books, and
conference proceedings. She edited/co-edited 21 books and conference proceedings
with Italian, EU, and overseas colleagues.
Anna has been the Italian Management Committee Member of:
• COST 277: Nonlinear Speech Processing, http://www.cost.esf.org/domains_
actions/ict/Actions/277 (2001–2005)
• COST MUMIA: Multilingual and Multifaceted Interactive information Access,
• www.cost.esf.org/domains_actions/ict/Actions/IC1002 (2010–2014)
• COST TIMELY: Time in Mental Activity, www.timely-cost.eu (2010–2014)
• She has been the proposer and chair of COST 2102: Cross Modal Analysis of
Verbal and Nonverbal Communication, http://www.cost.esf.org/domains_
actions/ict/Actions/2102 (2006 –2010).
Emotional Facial Expressions: Communicative Displays or Psychological Universals? xiii

Since 2006, she is a Member of the European Network for the Advancement of
Artificial Cognitive Systems, Interaction and Robotics (www.eucognition.org);
She is currently the Italian Management Committee Member of ISCH COST
Action IS1406:
Enhancing children’s oral language skills across Europe and beyond (http://
www.cost.eu/COST_Actions/isch/Actions/IS1406)
Anna’s research activities are on the following three principal lines of
investigations:
• 1998 to date: Cross-modal analysis of speech, gesture, facial and vocal
expressions of emotions. Timing perception in language tasks.
• 1995 to date: Emotional and social believable Human-Computer Interaction
(HCI).
• 1989 to date: Neural Networks: learning algorithm, models and applications
 Soft Data Analysis Based on Cluster Scaling

Mika Sato-Ilic

University of Tsukuba, Japan

[email protected]

Abstract. There is a growing need to analyze today’s vast and complex

societal data; however, conventional data analysis that is dependent on
statistical methods cannot deal with the frequently complex data types that
make up this data. As early as 2000, the following six challenges were
reported as important future challenges in the core area statistical research in
the twenty-first century. The six challenges pointed out are (1) scales of data,
(2) data reduction and compression, (3) machine learning and neural net-
works, (4) multivariate analysis for large p, small n (high dimension low
sample size data), (5) Bayes and biased estimation, and (6) middle ground
between proof and computational experiment.
Soft data analysis which is soft computing-based multivariate analysis is
the core area in which to combine conventional statistical methods and
machine learning or data mining methods, and has a strong capability to
solve the statistical challenges in the twenty-first century. In soft data
analysis, we have developed cluster-scaled models which use the obtained
cluster as the latent scale for explaining data. While the original scale does
not have the capacity to work as the scale for complex data, a scale that is
extracted from the data itself will have the capability to deal with the vast
and complex data.
This presentation outlines the problems and challenging issues of sta-
tistical data analysis caused by the new vast and complex data, how our
cluster-scaled models are related with these issues, and how the models
solve the problems with some applications.

Biography
Prof. Mika Sato-Ilic currently holds the position of
Professor in the Faculty of Engineering, Information
and Systems, at the University of Tsukuba, Japan. She
is the founding Editor in Chief of the International
Journal of Knowledge Engineering and Soft Data
Paradigms, Associate Editor of Neurocomputing,
Associate Editor of Information Sciences, Regional
Editor of International Journal on Intelligent Decision
Technologies, and Associate Editor of the International
Journal of Innovative Computing, Information and
Control Express Letters, as well asserving on the
Soft Data Analysis Based on Cluster Scaling xv

editorial board of several other journals.In addition, she was a Council of the
International Association for Statistical Computing (a Section of the International
Statistical Institute), a Senior Member of the IEEE, where she held several positions
including the Vice-Chair of the Fuzzy Systems Technical Committee of the IEEE
Computational Intelligence Society. In addition, she has served on several IEEE
committees including the administration committee, program co-chair, and special
sessions co-chair. Her academic output includes four books, 10 chapters, and over
120 journal and conference papers. Her research interests include the development of
methods for data mining, multidimensional data analysis, multimode multiway data
theory, pattern classification, and computational intelligence techniques for which
she has received several academic awards.
Why the Brain Can and the Computer Can’t
Biology Versus Silicon

Valeriu Beiu

Universitatea “Aurel Vlaicu” din Arad, Romania

[email protected]

Abstract. This presentation aims to follow on von Neumann’s prescient

“The Computer and the Brain” (Yale Univ. Press, 1958) and—relying on
the latest discoveries—to explain how the Brain achieves ultra-low power
and outstandingly high reliability (nano-)computing, while our silicon-based
computers cannot. The tale will be reversed as starting from the brain, and in
particular from very fresh experimental results, for information processing
and communication.
We shall proceed from the gated ion channels which are the
nano-switches of the brain. Understanding the ways, ion channels com-
municate will allow analyzing the statistical behaviors of arrays of gated ion
channels. These will be followed by unexpected results showing that highly
reliable communication using arrays of absolutely random devices is pos-
sible at amazingly small redundancy factors (<10). For computations, we
will make the case for interweaving arrays of ion channels acting as dis-
tributed amplifiers, and for adapting the classical Kirchhoff current law to
active (i.e.,, “amplified”) electrodiffusion. Afterward, we shall touch upon
the ultra-low power/energy consumption, where we will stress the crucial
role played by hydrated ions, the very fast ion channels, and the much
slower ion pumps, as well as how these behave in extremely crammed
spaces. Finally, moving only briefly to the next higher level, the computa-
tional power of columnar structures will be explored in the context of large
fan-in cyclic circuits.
The information conveyed will expose the reasons why our current
silicon-based approaches are falling short of doing what the brain is able to
do, and also reveal a stringent need for new and more accurate computational
models. As quite a few issues are still being investigated, we will conclude
with a call-to-arms for both the computing and the VLSI/nano-communities.
Why the Brain Can and the Computer Can’t Biology Versus Silicon xvii

Biography
Valeriu Beiu (S’92–M’95–SM’96) received the MSc
in Computer Engineering from the University
“Politehnica” Bucharest in 1980, and the PhD summa
cum laude in Electrical Engineering from the
Katholieke Universiteit Leuven in 1994.
Since graduating in 1980, he has been with the
Research Institute for Computer Techniques, University
“Politehnica” Bucharest, Katholieke Universiteit
Leuven, King’s College London, Los Alamos National
Laboratory, Rose Research, Washington State
University, United Arab Emirates University, and cur-
rently is with “Aurel Vlaicu” University of Arad. His
research interests have constantly been on biological-/neural-inspired circuits and
brain-inspired architectures (low-power, highly reliable, massively parallel), being
funded at over US$ 40M, and publishing over 250 papers (over 40 invited and more
than 10 patents) as well as giving over 190 invited talks, organizing over 100
conferences and working (unfortunately very slowly) on two books: Emerging
Brain-Inspired Nano-Architectures and VLSI Complexity of Discrete Neural
Networks.
Dr. Beiu has received five fellowships and seven best paper awards, and is a
senior member of the IEEE as well as a member of ACM, INNS, ENNS, and
MCFA. He was a member of the SRC-NNI Working Group on Novel
Nano-architectures, the IEEE CS Task Force on Nano-architectures, and the IEEE
Emerging Technologies Group on Nanoscale Communications, and has been an
Associate Editor of the IEEE Transactions on Neural Networks (2005–2008) and
of the IEEE Transactions for Very Large Scale Integration Systems (2011–2015),
while being an Associate Editor of the Nano Communication Networks (since
2010).
 Online Course: A Modern Trend
of Teaching and Learning

Salil Bose

Fellow of the Royal Society of Biology, UK

[email protected]

Abstract. As academics, we have dual responsibilities—doing research and

teaching. We know how difficult, if not impossible, it is to do the one to our
fullest satisfaction without compromising with the other. My presentation is
focused on teaching—how we can make teaching (more) effective and at the
same time make more time and energy available for research. Thanks to
modern Internet technology, a new online avenue of teaching and learning
shows promising potential to reach these goals. Online pedagogical
methodologies are now globally either complementing or replacing in-class
teaching.
An online course is basically a course that is offered in part or wholly
via the Internet. The most important infrastructural component of online
course is the LMS (Learning Management System) software. The LMS is
the online environment which enables the learning to happen. It is the
platform, where the teacher creates a course and the students can interact
with the course content in a dynamic manner. Online courses include means
for the students to contact their professor and access most course materials,
including online readings, videos, audio files, and other resources. This is
also where students go to participate in discussion boards to exchange views
with fellow classmates, and the instructor can monitor and comment on their
discussion. Depending on the versatility of LMS, students can also e-mail
and instantly message their classmates and instructors.
My objective is to describe how an LMS can be used to create a course.
Currently, many learning management systems are available with varying
capabilities, costs, and conveniences; research is ongoing in developing more
with varying features. In this presentation, as a case study, I will describe one
LMS, more specifically known as LCMS (learning content management
system) with brand name Drona (http://drona.netimaginelearning.com/),
which has several unique user-friendly features.
Animation is an elegant component of online courses especially to
illustrate and explain complex functions and abstract concepts of a phe-
nomenon. However, creating a “real-life” animation with sufficient details of
structure and function is a highly skilled job. In this presentation, I will also
show an animation of a biological molecular machine (DNA replication in
eukaryotic cells) to demonstrate how the potential of online courses can be
harnessed to illustrate highly complex biological phenomena in a dynamic
and three-dimensional matter overcoming the barrier of distance.
Online Course: A Modern Trend of Teaching and Learning xix

Biography

Salil Bose obtained his PhD from the University of

Rochester, New York, in 1975. His thesis was on
energy transduction in photosynthetic processes. After
completing postdoctoral training in Brookhaven
National Laboratories in Long Island and University in
Illinois at Urbana-Campaign, he taught and conducted
research on bioenergetics at two universities including
Jawaharlal Nehru University, where he served as
Professor and Chairman of the Center for
Biotechnology. Prof Bose worked for eleven years as a
Senior Scientist in National Institutes of Health (NIH) in USA doing research on
fast kinetics of proton (H+) transduction in bacteriorhodopsin photocycle, fast
kinetics of electron transfer in cytochrome oxidase, and regulation of oxidative
phosphorylation in mitochondria by monitoring electron transfer kinetics and
changes in transmembrane electrical potential. Prof Bose has published over 100
research papers.
Professor Bose served Nanyang Technological University (NTU) in Singapore
for thirteen years, initially as visiting professor and later as senior education con-
sultant. He was involved in advising PhD students, and developing and teaching
interdisciplinary courses at various levels—undergraduate to PhD. He initiated the
development of online courses at NTU and conducted research in developing
effective online assessment methodology. He also served the NTU School of
Computer Science as an adjunct professor to teach systems biology to the master’s
students in bioinformatics.
For many years, Professor Bose has been working as the Editor in Chief of a
Canadian journal, International Journal of Biology published by the Canadian
Center of Science and Education (www.ccsenet.org) and as a member of the
international advisory board of a British journal, Journal of Biological Education
published by the Society of Biology (www.societyofbiology.org/).
Invited as a visiting scientist/professor, Professor Bose has contributed to
research and teaching at many reputed institutions such as Kyungpook National
University in South Korea, Washington University at St. Louis USA, Imperial
College of Science & Technology London, Carnegie Institution of Washington at
Stanford University, University of California at Berkeley, and Michigan State
University at East Lansing, USA.
Professor Bose is an elected Fellow of the Royal Society of Biology, UK.
 Complex Dynamical System: Analysis
and Control via Linear Matrix Inequality
(LMI) Techniques

Rajeeb Dey

National Institute of Technology Silchar, India

[email protected]

Abstract. To analyze/predict the behavior of system at a future time is

always a topic of interest not only in engineering or physics but also in every
subject (like economics, pschycology, finance), and thus, it is a topic of
research to systems and control community too. For understanding the
behavior of a system, one needs to have a dynamical model, but modeling
should be such that it must predict more realistic nature/behavior without
many complications in the models too. More complex is the model more
complex will be the mathematical framework for carrying out the analysis
and further control synthesis of the system under study. Thus, there is
always a trade-off between the complexity of the model of the system and
framework of analysis.
In this talk, few important features of the systems will be considered to
create a model of the complex dynamical system in an attempt to
obtain/predict more realistic situation of the response. One such feature is
delays in the system dynamics along with consideration of parametric
uncertainties, hard and smooth nonlinearities, thereby making the system
model more realistic but bit more complexes. The modeling is oriented in
such a way that analytical framework remains easier, tractable, and further
control development becomes straightforward. Thus, this talk will cover
comprehensive analysis and control design for such system using estab-
lished LMI (linear matrix inequality) approach.
This talk will further focus on solution of certain interesting complex
systems from electric power networks, biomedical engineering, fuzzy and
neural network dynamical systems as case studies.
Complex Dynamical System: Analysis and Control xxi

Biography

Rajeeb Dey received his B.Tech (Electrical

Engineering) from North Eastern Regional Institute of
Science & Technology (NERIST), Itanagar, India, in
2001, M.Tech in Control System Engineering from
Indian Institute of Technology (I.I.T) Kharagpur,
India, in 2007, and PhD in Control Engineering from
Jadavpur University, Kolkata, India, in 2012. He has
been awarded National Associateship by Department
of Bio-Technology, Govt. of India, to work on control
aspect of Artificial Pancreas in Indian Institute of
Science, Bangalore in 2013. In 2014, he was been
awarded TWAS-CONACYT postdoctorate fellowship
by TWAS (The World Academy of Science), Italy, to work on complex dynamical
systems at CINVESTAV-IPN, Mexico City. He received AICTE (Govt. of India)
funded research project for implementation of output delay feedback control for
gantry crane problem in 2009 and in 2010 students working on this project has been
awarded best undergraduate project in India by Indian National Academy of
Engineers (INAE).
Dr. Dey holds a permanent position as Assistant Professor in Electrical
Engineering in National Institute of Technology, Silchar, India. He has around 13
years experience of teaching undergraduate as well as postgraduate students and has
guided one PhD and several master’s thesis in the area of control system engi-
neering. Dr. Dey has around 30 research publications in reputed international
journals and conferences. He is a regular reviewer of over 15 SCI (E) journals
related to applied mathematics and control engineering and also Editor of
International Journal of Artificial Intelligence Paradigm (Inderscience Publishers).
His research interest includes robust control, control using LMI techniques,
time-delay systems, intelligent control, control applications in biomedical engi-
neering. He is a Senior Member of IEEE Control System Society, Member of
SIAM, and Life Member of System Society of India. He is currently executive
committee member of IEEE CSIM (Control system, Instrumentation and
Measurement) Kolkata Chapter, India, and held the position of publication chair for
First IEEE CMI 2016.
Theory and Practice of Business Intelligence

Mihaela I. Muntean

West University of Timişoara, Romania

[email protected]

Abstract. Based on a selective literature review and some of the author’s

recent papers, a unifying theoretical approach of the most relevant business
intelligence (BI)-specific concepts will be introduced. Business Intelligence
is an umbrella term for various business managing approaches based on
well-informed decisions, which lead to a high-performance level within
organizations (Brohman, D.K., 2000; McKnigts, W., 2004; Melfert, F.,
Winter, R., Klesse, M., 2004; Mukles, Z., 2009; Hatch D., Lock M., 2009;
Borysowich, C., 2010; Jamaludin, I. A., Mansor, Z., 2011; Mircea M. (ed.),
2012). Eleven definitions describing the actual BI phenomena are subject
of the debate. In terms of a value proposition, the From-Data-To-
Performance value chain is designed (Muntean, M., 2012), phases like
business analysis, enterprise reporting, and performance management rep-
resent a series of activities that create and build value. Key forces like cloud,
mobile, social and big data enrich the BI framework and contribute to the
diversity of the BI phenomena. Based on the fact that most of the data (80%
of the data) stored in corporate databases have a spatial component, a
business intelligence approach for spatial enablement is recommended to be
developed (Muntean, M., Cabau, L., 2012). Despite the importance of the
technological dimension of the BI approaches, the business process model is
determinant for any initiative (Muntean, M., Muntean C., Cabău, L., 2013).
Considerations regarding business intelligence governance, based on the
author's expertise, establish a framework for the BI life cycle and ensure
consistent project delivery.
Based on these considerations, some practice examples have been
introduced, closing the gap between theory and practice. Data warehouse
proposals (Muntean, M., 2016), an innovative use of QR codes in BI
reporting (Muntean, M., Mircea, G., Băzăvan, S., 2014), and a feasibility
analysis for BI initiatives (Muntean, M., Muntean, C., 2013) are subject
of the debate.
Theory and Practice of Business Intelligence xxiii

Biography
Currently, Professor Mihaela I. Muntean is the Chair
of the Business Information Systems Department at
the West University of Timisoara and an IT inde-
pendent consultant. With a background in computer
science and a Ph.D. obtained both in Technical
Science and in Economic Science (Economic
Informatics), Professor Mihaela I. Muntean focused
her research activity on topics like information tech-
nology, knowledge management, business intelli-
gence, business information system. Over 70 papers in
indexed reviews and conference proceedings and the
involvement with success in eight multiannual
national research grants/projects are sustaining her contributions in the research
fields mentioned above.
 The Theory of Fuzzy Sets and Its
Real-World Applications
(Round Table)

József Dombi

University of Szeged, Hungary

[email protected]

Abstract. The SOFA workshop focuses on theoretical and applied com-

putational intelligence and informatics. Its aim is to provide a platform for
scientists, researchers, and students on soft computing to publish to discuss
and publish the results of their research.
The connection of the theoretical results and the practical applications is
one of the most important tasks of the workshop. In the roundtable meeting,
we will discuss how different operators could be applied in practice, first of
all by using the fuzzy control systems. In this area, several problems arise:
– describing the logical expression (using different operators)
– representing the consequent
– aggregating the consequents
– defuzzification
– interpolative solutions, etc.
The fuzzy control can also be represented by neural network. A very
interesting question is how to handle this approach. Finally, real practical
application shows the effectiveness of the different configurations of fuzzy
control. We can also mention that the input of the fuzzy control can be
obtained from picture processing procedure or from other resources (e.g.,
using signal processing).
In summary, the topic of the roundtable discussion covers the theory of
fuzzy sets and its real-world applications.

Biography
József Dombi’s degrees earned at University of Szeged. Academic degrees:
University doctor’s degree (1977, Summa cum laude), Candidate mathematical
sciences (1994, CSc, Title of dissertation: Fuzzy sets’ structure from the aspect of
multicriteria decision aid.). Visiting positions: Leningrad (1971, 6 months), Leipzig
(1974, 6 months), Bukarest (1978, 1 month), DAAD Scholarship, Aachen (1978,
12 months), Alexander von Humboldt Scholarship, Aachen (1986, 12 months),
European Scholarship, Bristol (1987, 3 months), Academic research exchange
The Theory of Fuzzy Sets and Its Real-World Applications xxv

(Paris, Helsinki, Turku, Tampere), CEEPUS guest

professor, Linz (2000 and 2009, 1 month), and
Klagenfurt (2008, 1 month). Awards: 1991 Minister of
Education award, 1997 Pro Sciencia award for student,
1998 László Kalmár award, in 1997, DataScope won
Information Technology award in Brussels, won the
Best Software of the Year award in 2000 at
COMDEX, Las Vegas. Editorial membership: edito-
rial board’s member of the Foundations of Computing
and Decision Sciences magazine until 1997, editorial
board’s member of the International Journal of
Systems Sciences, and editorial board’s member of the
International Journal of Advanced Intelligence Paradigms. Membership in inter-
national organizations: IFSA (International Fuzzy System Association), Member of
European COST Action on Multicriteria Decision Making, ESIGMA (European
Special Interest Group on Multicriteria Analysis), European Working Group on
Multiple Criteria Decision Aid, MTA Public Body (Operational Research), MOT
(Hungarian Association of Operational Research), Hungarian Humboldt
Association. Business: 1993 founder and head of Cygron Research Ltd., 2000
scientific consultant of Mindmaker Ltd., 2002 founder and head of Dopti research
and development Ltd. Research interest: computational intelligence, theory of fuzzy
sets, multicriteria decision making, genetic and evolutional algorithms, operation
research and visualization. Teaching: information visualization which has an
enormously growing interest in Hungary and abroad, the theory of artificial intel-
ligence, fuzzy systems, multicriteria decision making, and intelligent systems.
Contents

Machine Learning and Applications

Model Evaluation as Approach to Predict a Diagnosis . . . . . . . . . . . . . . 3
Adriana Mihaela Coroiu
The Basics of Implementing Intelligent Social Networks . . . . . . . . . . . . . 17
Stefan Cosmin Popa, Marius Constantin Popescu, and Antoanela Naaji
The Generation of Tests of Knowledge Check
Using Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Doru Anastasiu Popescu, Nicolae Bold, and Alexandru Ion Popescu
Matrix-Valued Bidirectional Associative Memories . . . . . . . . . . . . . . . . . 36
Călin-Adrian Popa
Enhanced Gradient Descent Algorithms
for Quaternion-Valued Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . 45
Călin-Adrian Popa
A Reinforcement Learning Based Approach to Multiple
Sequence Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Ioan-Gabriel Mircea, Iuliana Bocicor, and Gabriela Czibula
Supervised Learning Techniques for Body Mass Estimation
in Bioarchaeology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Vlad-Sebastian Ionescu, Gabriela Czibula, and Mihai Teletin
A Fuzzy Approach of Sensitivity for Multiple Colonies
on Ant Colony Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Camelia-M. Pintea, Oliviu Matei, Rabie A. Ramadan, Mario Pavone,
Muaz Niazi, and Ahmad Taher Azar

xxvii
xxviii Contents

Nature Inspired Partitioning Clustering Algorithms: A Review

and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Behzad Saemi, Ali Asghar Rahmani Hosseinabadi, Maryam Kardgar,
Valentina Emilia Balas, and Hamed Ebadi
Parametric Prediction Model Using Expert System and Fuzzy
Harmonic System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
W. Bel, D. López De Luise, A. Asmis, and D. Mansilla
Predicting Ozone Layer Concentration Using Multivariate Adaptive
Regression Splines, Random Forest and Classification
and Regression Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Sanjiban Sekhar Roy, Chitransh Pratyush, and Cornel Barna
Prediction of Customer Satisfaction Using Naive Bayes,
MultiClass Classifier, K-Star and IBK . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Sanjiban Sekhar Roy, Deeksha Kaul, Reetika Roy, Cornel Barna,
Suhasini Mehta, and Anusha Misra
Spam Email Detection Using Deep Support Vector Machine,
Support Vector Machine and Artificial Neural Network . . . . . . . . . . . . . 162
Sanjiban Sekhar Roy, Abhishek Sinha, Reetika Roy, Cornel Barna,
and Pijush Samui

Business Process Management

Collaborative Business Process Solution Considering
an Ontological Dimension of Process Models . . . . . . . . . . . . . . . . . . . . . . 177
Maria Laura Sebu and Horia Ciocarlie
Logistics Scenario for Wind Turbine Assembly Based on ERP . . . . . . . . 194
Gabriela Prostean, Cristian Vasar, and Andra Badea
Relieving Supply Chain Through Collaboration for Wind Turbine
Assembly Based on AHP and DBR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Andra Badea, Gabriela Prostean, and Cristian Vasar
New Learning and New Teaching in Higher Education. Formal
vs. Informal Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
Camelia Ciuclea, Gabriela Prostean, and Diana Andone
Integrating E-learning Platforms in Teaching/Learning.
A Critical Review Upon the Virtual Campus Usage in UPT . . . . . . . . . . 225
Camelia Ciuclea, Gabriela Prostean, Diana Andone, and Anca Draghici
Application of the Fuzzy-Pay-Off Method in the Valuation
of a Financial Instrument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Mihaela Crăciun and Luiela Magdalena Csorba
Contents xxix

Fuzzy Applications, Theory and Fuzzy and Control

Fuzzy Logic Control Strategy for a Three-Phase Electric
Arc Furnace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
Loredana Ghiormez, Octavian Prostean, Ioan Filip, and Manuela Panoiu
On Fuzzy Estimation of Channel Noise Characteristics
in MIMO Communication Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
V. Nicolau and M. Andrei
Fuzzy-Expert System for Investments in Winemaking . . . . . . . . . . . . . . . 278
Marius M. Balas, Emanuela V. Muller, and Ramona Lile
Closed Loop Blood Glucose Regulation of Type 1 Diabetic
Patient Using Takagi-Sugeno Fuzzy Logic Control . . . . . . . . . . . . . . . . . 286
Anirudh Nath, Rajeeb Dey, and Valentina E. Balas
Experimental Implementation of Fuzzy Vision-Based Tracking
Control of Quad-Rotor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
Hemjyoti Das, Aniket Samrat Mazumdar, Rajeeb Dey, and Lintu Roy
Fuzzy Fractional Derivative: A New Definition. . . . . . . . . . . . . . . . . . . . . 314
Marzieh Najariyan, Mehran Mazandarani, and Valentina Emilia Balas
Solving First Order Linear Fuzzy Differential Equations System . . . . . . 330
Marzieh Najariyan, Mehran Mazandarani, and Valentina Emilia Balas
An LMI Based Integral SMC for Tracking Control of a New
4-D Conservative Chaotic System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
Jay Prakash Singh, Rajeeb Dey, and Binoy Krishna Roy
Tracking Control with Vibration Suppression of a Two-link
Flexible Manipulator using Singular Perturbation with Composite
Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Kshetrimayum Lochan, Rajeeb Dey, Binoy Krishna Roy,
and Bidyadhar Subudhi

Computational Intelligence in Education

Distributive Fairness in Educational Assessment . . . . . . . . . . . . . . . . . . . 381
Paul Hubert Vossen
E-Learning Web Based Grid by Utilising P2P Technology
in Presentation Broadcasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
Rafid Al-Khannak
xxx Contents

Soft Computing and Fuzzy Logic in Biometrics (SCFLB)

A Circular Eccentric Iris Segmentation Procedure for LG2200
Eye Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
Cristina-Madalina Noaica
Iris Segmentation in the Last Decade – A Survey . . . . . . . . . . . . . . . . . . 426
C.M. Noaica, P.S. Penariu, and V.C. Stroescu
Circular Fuzzy Iris Segmentation Using Isolines, Sine Law
and Lookup Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
N. Popescu Bodorin, V.E. Balas, and P.S. Penariu
A Parallel Approach on Airport Runways Detection
Using MPI and CImg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
P.S. Penariu, N. Popescu Bodorin, and V.C. Stroescu
Identifying Down Syndrome Cases by Combined Use of Face
Recognition Methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
Vlad Dima, Anca Ignat, and Cristina Rusu
Iris Recognition Using Color and Texture Features . . . . . . . . . . . . . . . . . 483
Ioan Păvăloi and Anca Ignat
Retinal Vessel Classification Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . 498
Florin Rotaru, Silviu-Ioan Bejinariu, Cristina Diana Niţă, Ramona Luca,
Mihaela Luca, and Adrian Ciobanu

Soft Computing Algorithms Applied in Economy, Industry

and Communication Technology
Reliable Metrics for Image LSB Steganography
on Mobile Platforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
Dominic Bucerzan and Crina Raţiu
Information System Security in the Context of Administrative
Cooperation Through Internal Market Information
System in Romania . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
Camelia Daciana Stoian and Dominic Bucerzan
An Application of the Multi Period Decision Trees in the Sustainable
Medical Waste Investments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 540
Luiela Magdalena Csorba and Mihaela Crăciun
Containerizing Solution for a Highly Distributed Systems . . . . . . . . . . . . 557
Vlad Iovanov, Mihai Buzgau, and Dominic Bucerzan
Contents xxxi

Data Acquisition and Processing with Fusion Sensors,

Used in Algorithms for Increasing Accuracy of Position Measurement
of Mobile Robots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
Sandru Florin, Nanu Sorin, Negrea Romeo, Popescu Anca,
and Fericel Sergiu

Modelling and Applications in Textiles

Specialized Software for Tracking and Management Machinery
Used in Textile Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
Pavel Farcaș, Ciprian Pop-Moldovan, and Marius Constantin Popescu
Treatment Optimization for Bioscouring in Ultrasound
of Cotton/Cottonised Hemp Blended Fabrics . . . . . . . . . . . . . . . . . . . . . . 595
Pustianu Monica, Dochia Mihaela, Sîrghie Cecilia, Gavrilaș Simona,
and Tomescu Daniel
Fuzzy Method Applied on the Control of Textiles Biotechnologies . . . . . 605
Bălaș Marius, Sîrghie Cecilia, Pustianu Monica, Dochia Mihaela,
Gavrilaș Simona, and Tomescu Daniel
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613
Machine Learning and Applications
 Model Evaluation as Approach
to Predict a Diagnosis

Adriana Mihaela Coroiu(B)

Babeş-Bolyai University,
1, M. Kogălniceanu Street, 400084 Cluj-Napoca, Romania
[email protected]

Abstract. The paper presents an approach to a relevant issue of the

supervised learning: classification. Creating the models that are able to
generalize to new data, tuning the model so that the performance can
be increased and models evaluation are relevant task of this sub-field of
machine learning. In this paper, we have chosen to treat the evaluation
of a model. The paper consists in two experiments, each of them with
a particular purpose. First experiment outlines different model evalua-
tion methods using specific performance metrics. This paper analyzes
two models, logistic regression and the k-nearest neighbor with three
methods of evaluation: train-test approach, test-set approach and the
k-cross validation approach. On our analyzed data set, we achieved rea-
sonable results using logistic regression as model with k-cross validation
approach as evaluating method of the model. The second experiment
determines the rank of the observations by predicting the probabilities
depending on the response vector. Based on the predicted probability
we try to improve the metric performance. For our particular task, one
metric that allows us to extract relevant information from data and can
be improved is the sensitivity metric.

1 Introduction
The amount of data from the present increases in an exponential way. Being
able to make sense of all this data will represent an essential skill for people
working in data science industry. Beside the expanse of the data quantity there
is and another issues represented of the type of these data. Considering these
aspects, the analysis of the data, in order to extract relevant knowledge is a
continuous challenge. Therefore the extraction of relevant groups, the discovery
of the relations within these groups, the optimization and the ability to generalize
the new data are important topics [15].
Making prediction process, classification, is the theme that we have treated
in our paper.
Classification, as a component of supervised learning, is one the popular and
essential task of machine learning and it is used to place an unknown observation
into a known group, in other words making prediction for a new observation.


c Springer International Publishing AG 2018
V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9 1
4 A.M. Coroiu

We can use different model to make predictions. Besides, it is important to

check if our prediction is correct and accurate. Here appears the role of eval-
uation for a model. We have several evaluation model methods. In our paper,
are presented three of them: first method is train and test the entire data set,
second method is test set approach of the data and the third is k-cross validation
method of evaluation for a model. As models, we have chosen to handle with the
logistic regression model and the k-nearest neighbor model.
In addition, the analysis of different type of data represent another essential
point, needed considered when we analyze data. Therefore, in this paper, we
evaluated the behavior of some models from supervised learning to a different
data sets.
We will explain all these aspects in the next parts of the paper. Hence, the
structure of our paper is: Theoretical background for a brief presentation of the
notions that will be used in the paper. This section is followed by the Motivation
section where we explain the argument of selecting these working idea. Then
we have the Computational experiments section here we will present our work
flow and the results achieved for our experiments. Finally, we have the section
Conclusions and future work directions where we highlight the relevant ideas
resulted from the paper and point out our next work directions.

2 Motivation
One of the goal of the supervised learning is to build a model that generalizes to
new observations. That is what we really need a method that allows to estimate
how well a given model perform on the new observations. In this case we need
some model evaluation methods. If we can quantify the performance of a certain
model, we can choose a model that is most appropriate for our purpose.
The motivation in choosing this work is related on the following aspects:
Create models that are able to make good predictions is an important issue due
to the fact that nowadays, machine learning with all its sub-fields are used in
different domains in which exist collected data. Extract relevant knowledge form
large data sets, or extract knowledge from data sets with various type of data
represent a continue challenge. Improve these models or improve the computed
metrics is in the same time important because of the same reason.

3 Theoretical Background
One of the purpose for data set it is to make predictions, after making this pre-
diction is necessary to quantify if our prediction was a good or a bad prediction.
This means that we have to measure the performance of a model. To realize
this, we used a test set consisting of data with known labels. We train the model
on the training set, apply the model to the test set, and then to measure the
performance by comparing the predicted labels with the true labels.
As methods applied to determine the performance of a model, from the mul-
titude existed in literature, we have choose three methods: train and test the
 Model to predict a diagnosis 5

entire data set [2], test set approach (also known as train-test split the data
method or validation set approach) [4] and the k-cross validation data set [14].
For the test set approach, a common splitting is to use 70% of the data set
for training and 30% for testing. The distribution of the observation within the
two training and test subsets is usually done randomly, in order to guarantee
that both subsets are random samples from the same distribution.
For k-cross validation it is important to use stratified sampling, that means
to ensure the each output class is present in the exact same proportion in the
training and test subsets. It is important to measure the performance of a clas-
sifier on an independent test set. Every training algorithm looks for patterns
in the training data, because of the correlations between the features and the
response vector, and for the new observation, the model tries to determine how
similar it is to a particular class related to what model already have learned.
The main purpose in constructing a model is to build models that are able to
generalize to a new observation, and making predictions of a high performance,
according to Michalski [11]. Generally, working with some models and after that
verifying their accuracy in predicting values for a new observations, exist few
steps [17,19] that have to be chased:

• select the model that will be used;

• create an instance of the model selected in previous phase;
• train the model;
• predict the response for a new observation.

It is essential to know how a model will perform on a new observation because

there are situations in which we have to analyze different data sets and we do
not have enough information about a new observation for which we will predict
a response using our model. As a definition of a model evaluation it can be says
that estimate how well a model are likely to perform on a new observation (or
a set of observations) using existing data. These evaluations methods help us
to choose a model that represents a good choice for our particular task. In the
same time, model evaluation methods can be used in the following assignments:

• choose between two (or more) different models;

• select the optimal tuning parameters of a model;
• decide between different features set;

In our paper, we will use the evaluation methods as a method for selecting
between two models that generalize to our new data.
To quantify the performance of a model, there are some evaluation metrics
[5]. The metrics that we will use in our experiments are: accuracy score, classi-
fication error, sensitivity, precision, specificity, and false positive rate. All these
metrics can be computed from confusion matrix. A confusion matrix define the
performance of a classification model on a data set used for test for which the
true values are known. Splitting the confusion matrix, we have the following
components and equations [16].
6 A.M. Coroiu

• a is the number of correct predictions that an instance is negative;

• b is the number of incorrect predictions that an instance is positive;
• c is the number of incorrect of predictions that an instance negative;
• d is the number of correct predictions that an instance is positive.

The accuracy is the proportion of the total number of predictions that were
correct. The equation for this metric is:
a+d
accuracy =
a+b+c+d
The sensitivity, the recall or true positive rate is the proportion of positive
cases that were correctly identified, as calculated using the equation:
d
sensitivity =
c+d
The sensitivity metric tries to answer the question, when the actual value
is positive, how often is the prediction correct, or how sensitive is the model in
order to determine the positive instances.
The precision is the proportion of the predicted positive cases that were
correct, as calculated using the equation:
d
precision =
b+d
The precision metric tries to answer the question when a positive value is
predicted, how often is the prediction correct, or how precise our model is when
predicting positive instances. The classification error, or the misclassification
rate is the proportion of the total number of predictions that were incorrect.
The equation is:
c+b
error =
a+b+c+d
The specificity metric is the proportion of negatives cases that were classified
correctly, can be computed using the equation:
a
specif icity =
a+b
The false positive rate (FPR) is the proportion of negatives cases that were
incorrectly classified as positive, can be computed using the equation:
b
FPR =
a+b
Other relevant notions used in our experiments are the receiver operating
characteristics and the area under the ROC curve. The receiver operating char-
acteristics (ROC) can be see as a technique for visualizing, organizing and select-
ing models based on their performance [3]. The ROC measures the performance
 Model to predict a diagnosis 7

of a model at all possible thresholds. The thresholds make sense just in the con-
text of predicted probabilities. The ROC is the plot of true positive rate versus
false positive rate. The threshold allows us to visualize the performance of a
model, across of all possible classification thresholds and because of this we have
the opportunity to choose a threshold that determine an equilibrating sensitivity
value. Furthermore, the ROC performs well for unbalanced data sets.
To compare models we may want to reduce ROC performance to a single
scalar value representing the expected performance. A common method is to
compute the area under the ROC curve (AUC) [1]. The value is between 0
and 1.0. The AUC has an important statistical property: the AUC of a model is
equivalent to the probability that the model will rank a randomly chosen positive
instance higher than a randomly chosen negative instance [9].
In paper, as models, we analyzed the logistic regression and the k-nearest
neighbor models.
Logistic regression is a generalization form of the linear regression model to
classification problems. In logistic regression we use a set of input data to predict
the probabilities of class membership. These probabilities can then be mapped
to class labels, thus predicting the class for each observation. To perform logistic
regression we are using the formula:

eβ0 +β1 x
π = P rob(y = 1|x) =
1 + eβ0 +β1 x
In logistic regression we have two important aspects: the probability and the
odds. The odds of an event represent the ration of the probability of the event
by its complement.
π
Odds =
1−π
For logistic regression it is important to notice the relationship between the
odds ration and the parameter β.
π
log( ) = log(eβ0 +β1 x ) = β0 + β1 x
1−π
In logistic regression, β1 represents the change in the log-odds for a unit
change in x. Therefore, the eβ1 represents the change in the odds for a unit
change in x. If we extend the logistic regression formula to include multiple
inputs, the result is as follows:
π
log( ) = β0 + β1 x1 + β2 x2 + ... + βp xp
1−π
This formula can be rearranged into logistic regression formula:

eβ0 +β1 x1 +β2 x2 +...+βp xp

π=
1 + eβ0 +β1 x1 +β2 x2 +...+βp xp
The logistic regression takes the features and multiply each one by a weight
and then add them up. This result will be put into the sigmoid, with a number
8 A.M. Coroiu

between 0 and 1. Anything above 0.5 will classify as a 1, and anything below
0.5 will be classify as a 0. Therefore, the logistic regression can be considered
as a probability estimate. Using k-Nearest Neighbors (kN N ), an observation
is classified by a majority of its neighbors. K is always a positive integer. The
neighbors are selected from a set of observation for which the correct classifica-
tion is known. kNN is a model that works like this: for a training set that has
labeled data, when a new data without a label is given, it is compared that new
piece of data to the existing data, every piece of existing data. Then is taken
the most similar pieces of data (the nearest neighbors) and looks at their labels.
The greater value of k is the value what is important to us [6].

4 Computational Experiments
In this section we will describe the experiments that we have performed. We have
two experiments: the purpose of the first experiment was to evaluate the per-
formance of a model. We have computed the performance of two models, the
logistic regression and the kNN, we have used all the three method for evaluate
a model enumerated in the Theoretical background section.
The first data set used in our experiment is the breast cancer wisconsin data
set (BCWds). As a short description of the data set, it contains 699 observations,
each observation has 10 features (all of them numerical values) and the vector
response contains the diagnosis of the patients and has values between 2 and 4,
value 2 is interpreted as benign, and value 4 is interpreted malignant [18]. For
this data set, we started with the question related to which model produced the
correct predictions for a new observation, in other words, to predict the type of
cancer (malign or benign) for a patient given his health measurements (a new
set of features collected for a new patients), which is the best model? In this
case, because the vector response has two values will have a binary classification
problem.
The second data set used in our experiment is Dermatology data set
(Dermads). This data set contains 366 observations, each observation has 34
features (from these features, 33 features are ordinal as type of data and one
feature is numerical) and the vector response contains 6 values representing der-
matology disease [8]. Because the vector response can has six possible value, we
have a multi-class problem. For logistic regression model, we will have a clas-
sification task with more than two classes, therefore we will have to perform a
multi-class logistic regression. As a preprocess of this data set we had the fol-
lowing aspects, first, in the numerical feature, there are some missing values. In
order to deal with this missing values, we have compute the mean of the existing
values for this feature and after that we just replaced the missing values with
the computed mean, and the second, the ordinal features were encoded such as
these can be used in our models, the encoding methods was to use a one-of-K,
that is the transformation of each ordinal feature with m possible values into m
binary features, with only one active, once.
Using these two data sets, our goal is to build models that generalize to new
observations. A model evaluation method allow us to use our existing labeled
 Model to predict a diagnosis 9

data to estimate how well our models are likely to perform on our new obser-
vation. These methods allow to choose between different types of models. The
model with a good performance can be used for a new observation.
In order to determine model that is much appropriate to be used to predict
the new observations behavior, we used the work flow presented in the section
Theoretical Background, with all three methods of evaluation on our two data.
Next, we will present our experiment and the achieved results: First method of
evaluation that we choose to analyze is the method called train and test the
entire data set (TTEds). This method has the following phase:

• train the model in the entire data set;

• test the model by checking how it performs on that same data;
• evaluate the result.

In this case, the problem is to make some predictions but without check if these
predictions are correct. By testing our model on the same data set for which we
actually know the true value response, we can check how well our model is doing
through comparing predicted response value with the true response values. For
this evaluation method, the computed value of accuracy is training accuracy.
Using this method of evaluation for a model, the problem is represented by over-
fitting the model because the method creates an unnecessary complex model and
because of this will not work in a very good way when we will make prediction
on the new observations, this is, partially due to the fact that the model learns
and from noisy data. Consequently, training accuracy is not a good estimate
accuracy for new observations.
The second method for evaluation the performance of a model is test set
approach (TSds). This method has the following phases:
• split the data into two pieces, training set and the testing set;
• test the model on the training set;
• test the model on the testing set;
• evaluate the result.
Because we evaluate the model on the data that was not used to train the model,
in this method, we get a more accurately simulating on how well our model will
perform to a new observation. Related to the dimension on the two sets, training
and testing sets, we have used the proportion 40% for the testing set and 60%
for the training set. And for the process of assigning an observation to a training
or a testing set, it is a random process and every time we run the experiment,
we will have different structure of the observations in training and in the testing
set. Using this method of evaluation, when we make some prediction on the
testing set, because we know the response values for the testing set, we can
compare the predicted value with the actual values stored in the test vector. In
this case we have the testing accuracy. Although, the results achieved performing
the experiment using this method of evaluation are significant to the maximum,
there are some drawback and with this evaluation method. The method provides
a high variance estimate of new observations accuracy, which means that the
10 A.M. Coroiu

answer for a new observation can has a different value depending upon which
observation happened to be in the training set comparing with the set.
There is an alternative model evaluation model that largely overcomes this
limitation, called k-partitions cross validation (k-CVds). This is the third method
for evaluate a model that we have used. This method repeats the test set app-
roach multiple times and averaging the results. The working phases of this eval-
uation model method are:

• choose a value for k;

• split the data set into k equal partitions. Partition 1 will be used as a testing
set and the union of the other partitions as the training set;
• train the model on the training set;
• test the model on the testing set;
• compute the testing accuracy;
• repeat the previous three phases k-times, using a different partition as the
testing set each time;
• compute the average of the testing accuracy, the cross-validated accuracy.

The results achieved in evaluation of these two models on the first data set,
breast cancer wisconsin data sets are highlight in Table 1.

Table 1. The measured value for BCWds of the LR and kNN using TTEds, TSds and
TSds

Logistic regression k-nearest neigbour

TTEds TSds k-CVds TTEds TSds k-CVds
Accuracy 0.77 0.70 0.69 0.80 0.72 0.70
Class. error 0.23 0.30 0.31 0.20 0.28 0.30
Sensitivity 0.54 0.24 0.33 0.65 0.48 0.48
Precision 0.73 0.55 0.59 0.75 0.60 0.57

The results achieved in evaluation of these two models on the first data set,
dermatology data sets are highlight in Table 2.
Comparing the train test method with the test set approach and with the k
cross-validation to evaluate a model, we can say that the main reason we prefer
cross validation to the other two evaluation methods is that cross validation
generates some more accurate estimate of new observation accuracy, and this is
what we really need when we have to choose a model to perform to a new data
set. In the same time, cross validation uses the data more efficiently than other
two models because every observations is used for both training and testing the
model. Indeed, we have to admit that method of train-test split data set is more
efficient in term of time complexity since k-partition cross validation repeats the
k-times the test set approach, and this can be a downside for larger data sets.
Concluding, because each partition of the data set access the testing set for one
 Model to predict a diagnosis 11

Table 2. The measured value for Dermds of the LR and kNN using TTEds, TSds and
TSds

logistic regresion k-nearest neighbors

TTEds TSds k-CVds TTEds TSds k-CVds
Accuracy 0.89 0.84 0.81 0.90 0.80 0.78
Class. Error 0.11 0.16 0.20 0.10 0.20 0.22
Sensitivity 0.61 0.64 0.67 0.63 0.63 0.60
Precision 0.69 0.70 0.70 0.79 0.69 0.68

iteration and it is part of the training set for the other k − 1 iterations, the
cross-validated accuracy is the most appropriate solution to achieve our goal.
Comparing the two models used, the k-nearest neighbors and the logistic
regression, there are some advantages and disadvantages for both of them. As the
advantages of k-nearest neighbors we enumerate: it is simple and can be used for
classification, the training model is fast, some disadvantages of k-nearest neigh-
bors, are overcome by logistic regression: all the training data must be stored,
prediction phase can be slow when the number of observations is large, it is sen-
sitive to irrelevant features and to the scale of the data and accuracy. For logistic
regression the main advantages are: model training and the prediction are fast,
no tuning is required, features do not need to be scaled, can perform well with
a small number of observations, outputs well-calibrated predicted probabilities.
The main disadvantages of logistic regression are that presumes a linear relation-
ship between the features and the log-odds of the response can not automatically
learn feature interactions.
Our second part of the experiment is a step forward and consist in the fol-
lowing: After in previous part we have been work to establish what is the model
that is indicated to be used when we have to make some predictions to new data,
now, in this experiment we have analyze the important of a new observation, in
other words we rank the observations by predicting the probabilities dependent
on the response vector, therefore we predicted the probabilities instead of the
class predictions.
For this experiment we have used the data set Diabetes data set from the
UCI Repositories [8]. The vector response in this data set is used to distinguish
patients liable to have diabetes the patients that are not liable to have diabetes.
Because in the previous experiment we already found that we achieve a reason-
able value of accuracy metric, if we have used the logistic regression model using
the k-cross validation method to evaluate a model, and in this experiment, we
will use the logistic regression model with the k-cross validation. Next, the work
flow for this experiment consist in the following:

• determine the feature to be included to determine the predictors of the model,

we have choose the features glucose, insulin, bmi and age;
• train the model using these predictors;
12 A.M. Coroiu

We achieve the following values for the logistic regression coefficients, presented
in Table 3. We converted the log-odds to odds and we gave values of a new
patients and for those patients, we compute the predicted probability. All these
results are highlighted in Table 3. We can see that the probability respects the
trend of the data included in our data set, therefore we can prioritize the patients.

Table 3. The log-odds, odds and accuracy for new patients

Features Log-odds Odds Newp1 Newp2 Newp3

Glucose 0.033 1.035 20 120 70
Insulin −0.001 0.998 93 93 45
bmi 0.083 1.087 32 50 60
Age 0.028 1.029 45 65 30
Accuracy 0,018 0,821 0,429

Heretofore, we have computed the vector of the predicted probabilities for

our predictor X. For these predicted probabilities we have created the histogram
of these predicted probabilities (ypred1 ). The histogram is presented in Fig. 1.

Fig. 1. The histogram for y pred1

Therefore we computed the values of AUC metric. We created the ROC curve
for our actual y and our predicted yp red. We increased the values of the pre-
dicted probabilities through taking the square root of the predicted probabilities
(ypred2 ) and we created again the histogram based on these modified value of
predicted probabilities and it is obvious that the histogram is changed, as we
can see in Fig. 2.
If we computed again the AUC metric and created the ROC curve, these two
did not changed the value and the shape. Another procedure of modify the value
of the predicted probabilities was to decrease the value of the small probabilities
 Model to predict a diagnosis 13

Fig. 2. The histogram for y pred2

and to increase the value of the probabilities with grater values (ypred3 ). After
this, we re-create the histogram of these modified probabilities, as we can see
in Fig. 3.

Fig. 3. The histogram for y pred3

This histogram changed his shape, but the AUC had the same value and
the ROC curve has the same shape. All collected information are presented in
Table 4.
Concluding, the AUC metric conserve the same value, without regard to
the modified values of the predicted probabilities. This is an important aspect,
because a whole of relevant information can be extracted from visualize diagrams
of the date and in some cases to have readable diagrams is necessary to add some
modification to these probabilities.
Constructing the ROC curve, we had some other relevant metrics that can
be used to extract relevant information or to increase or to decrease the val-
ues of some important metric for our purpose. We had the false positive rate
(FPR) and the true positive rate (TPR). Both of them presented in the section
Theoretical Background. And we have the threshold. Because for this data set,
14 A.M. Coroiu

Table 4. The AUC for the predicted probabilities

Predicted prob set Predicted values AUC

y pred1 0,65 0,10 0,81 0,09 0,44 0,26 0,07 0,25 0,90 0,47 0,80
y pred2 0,80 0,30 0,90 0,31 0,66 0,51 0,27 0,50 0,95 0,68 0,80
y pred3 0,80 0,01 0,9 0,009 0,19 0,07 0,005 0,06 0,95 0,22 0,80

we have as value in the vector response, we have a binary problem, therefore,

the default value of the threshold is 0.51 and diagnosis 1 is predicted only if
the value of the threshold is exceeded, otherwise diagnosis 1 is predicted. The
value of threshold can be modify in order to modify some important metrics. We
have varied the value of threshold for modify the value of sensitivity and of the
specificity. These two metrics was also presented in the Theoretical Background
section. The specificity metric relieves how selective is our model in predicting
positive observations. We can increase the sensitivity of a model if we modify the
value of a threshold, therefore, we modify the value of our threshold to achieve a
greater value of the sensitivity. The sensitivity metric and the specificity metric
are inverse metrics, therefore, if we modify the threshold to increase the value
of the sensitivity, we achieved a lower value of the specificity. The values of the
sensitivity and the specificity computed for various value of the threshold are
presented in Table 5.

Table 5. The shifted value of threshold and the modification of the metrics values

Threshold Sensitivity Specificity

0.3 0.977 0.236
0.4 0.944 0.436
0.5 0.8507 0.628
0.6 0.682 0.766
0.7 0.567 0.870

It is obvious that if we decreased the value of the threshold, the value of the
sensitivity metric increased because our model is more sensitive to the positive
observations. The specificity, being an inverse metric to sensitivity, decreased
when we decreased the value of the threshold and increase when we increased
the value of the threshold.
Concluding, second part of our experiment is a step forward after the model
evaluation method, and it can easy became recurrent because the variety of the
feature which can be included in our predictor we can achieve different values of
the metrics that we can compute. In the same time, we saw that an important
component in the sensitivity metric is the threshold and different value can help
us to have a sensitivity oriented model or a specificity one.
 Model to predict a diagnosis 15

All the experiments have been performed using the Scientific Computing
Tools for Python [7,10,12,13].

5 Conclusions and Future Work

The goal of this paper was to determine if a model can deal with different type
of data, furthermore quantify the performance of a model to a new observa-
tion. Quantifying the performance of a model is not the final step because it is
necessary to evaluate this performance.
In order to evaluate the performance of a model, we have used one of the
most common metric for classification problem, classification accuracy, partic-
ularly called training accuracy for the first method of evaluation of a model,
testing accuracy for the second method and cross validated accuracy for the
third method of evaluation of a model.
We have test all these three methods to evaluate a model, and as a conclusion,
considering all the advantages and the disadvantages of them, we choose that for
our purpose and for our type of data, the k-cross validation evaluation method
is the best solution and as a model, the logistic regression.
As future work directions, we intend to use the evaluation model method
to determine what are the features that have to be included in the training
data set such that the model provides prediction with a high value of accuracy.
Moreover, we want to improve the accuracy of a model, in our case, logistic
regression, through tuning the parameters of the model.

References
1. Bradley, A.P.: The use of the area under the ROC curve in the evaluation of
machine learning algorithms. Pattern Recogn. 30(7), 1145–1159 (1997)
2. Dursun, D., Walker, G., Kadam, A.: Predicting breast cancer survivability: a com-
parison of three data mining methods. Artif. Intell. Med. 34(2), 113–127 (2005)
3. Fawcett, T.: An introduction to ROC analysis. Pattern Recogn. Lett. 27(8), 861–
874 (2006)
4. Gama, J., Rodrigues, P.P., Raquel, S.: Evaluating algorithms that learn from data
streams. In: Proceedings of the 2009 ACM symposium on Applied Computing.
ACM (2009)
5. Gunawardana, A., Shani, G.: A survey of accuracy evaluation metrics of recom-
mendation tasks. J. Mach. Learn. Res. 10, 2935–2962 (2009)
6. Harrington, P.: Machine Learning in Action. Manning, Greenwich (2012)
7. Hunter, J.D.: Matplotlib: a 2D graphics environment. Comput. Sci. Eng. 9, 90–95
(2007)
8. Lichman, M.: UCI Machine Learning Repository. University of California, School
of Information and Computer Science, Irvine (2013)
9. Lobo, J.M., Valverde, A.J., Real, R.: AUC: a misleading measure of the perfor-
mance of predictive distribution models. Glob. Ecol. Biogeogr. 17(2), 145–151
(2008)
10. McKinney, W.: Data structures for statistical computing in python. In: Proceedings
of the 9th Python in Science Conference, pp. 51–56 (2010)
16 A.M. Coroiu

11. Michalski, R.S., Jaime, C.G., Mitchell, T.M. (eds.): Machine Learning: An Artificial
Intelligence Approach. Springer Science & Business Media, Heidelberg (2013)
12. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, É.: Scikit-learn: machine
learning in python. J. Mach. Learn. Res. 12, 2825–2830 (2011)
13. Perez, F., Granger, B.E.: IPython: a system for interactive scientific computing.
Comput. Sci. Eng. 9, 21–29 (2007)
14. Seewald, A.K., Johannes, F.: An evaluation of grading classifiers. In: Advances in
Intelligent Data Analysis, pp. 115–124. Springer, Heidelberg (2001)
15. Senthil Kumar, S., Hannah Inbarani, H., Azar, A.T., Own, H.S., Balas, V.E.,
Olariu, T.: Optimistic multi-granulation rough set based classification for neonatal
jaundice. In: Kacprzyk, J. (ed.) Proceedings of the 6th International Workshop
on Soft Computing Applications SOFA 2014. Advances in Intelligent Systems and
Computing, Timisoara, Romania, July 22–24 (2014)
16. Weng, C.G., Josiah, P.: A new evaluation measure for imbalanced datasets. In:
Proceedings of the 7th Australasian Data Mining Conference, vol. 87. Australian
Computer Society, Inc. (2008)
17. Witten, I.H., Eibe, F.: Data Mining : Practical Machine Learning Tools and Tech-
niques. Morgan Kaufmann, Burlington (2005)
18. Wolberg, W.H., Mangasarian, O.L.: Multisurface method of pattern separation
for medical diagnosis applied to breast cytology. Proc. Nat. Acad. Sci. U.S.A. 87,
9193–9196 (1990)
19. Zhu, X.: Semi-supervised learning literature survey (2005)
 The Basics of Implementing Intelligent
Social Networks

Stefan Cosmin Popa(&), Marius Constantin Popescu,

and Antoanela Naaji

Vasile Goldis Western University Arad, 94-96 Revolutiei Blvd., Arad, Romania
[email protected],
{mpopescu,anaaji}@uvvg.ro

Abstract. The paper studies the implementation of a web software application

in node.js language, with a view to analyzing neural networks and artificial
intelligence in a virtual environment, with application in social networks. The
software allows experimenting on the capacity of virtual entities to learn to act
as a group or individually, avoid the margins of the environment where they are
located, or obstacles, and learn to feed in order to live. The application is a
novelty in the sense that, until recently, i.e. the emergence of node.js, imple-
menting the concept of neural networks in a scripting language was not possible.
This concept can be developed in practical applications such as social networks,
search engines with artificial intelligence, personal assistants, e.g. Siri etc. The
final aim is to obtain an efficient software solution, with application in social
networks, which can evolve over time.

Keywords: Neural networks
Scripting
Node.js
Web application
Social

networks

1 Introduction

The application described in the paper is based on an original design with an interface
where artificial intelligence activity can be observed. There are no similar works in the
literature, with a descriptive approach to artificial intelligence, but only mathematical
related ones, less accessible to inexperienced programmers. The paper seeks to facil-
itate the process of developing applications based on artificial intelligence using node.
js, which is a popular language, as it uses JavaScript on both the server and the client
part; this means that developers need only program in one language, across all layers.
The application uses the following front-end technologies: HTML5 (Hyper Text
Markup Language 5) [1], CSS3 (Cascading Style Sheets 3) [2], JavaScript [3], HTML5
Canvas [4]. The node.js language is a runtime environment with libraries running
JavaScript applications outside the browser. Node.js and the Synaptic library [5] are
used in the back-end. The application is hosted and was tested on a Debian 8 personal
server and runs in a modern web browser (recommended, Google Chrome v49,
Microsoft Edge v25, Mozilla Firefox v45).
Usually, node.js is employed to run server applications in real time and excels in
the performance of using non-blocking I/O and asynchronous events. A real advantage

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_2
18 S.C. Popa et al.

is in the fact the same language can be used to construct the backend of the frontend,
already constructed in JavaScript. Although still in its beta phase, the language is used
by many large companies (Microsoft, Yahoo!, LinkedIn, eBay, The New York Times,
etc.), on account of being stable and efficient. Few programming languages have
known such rapid development as node.js. It can be easily seen that most of the
resources found on GitHub are in JavaScript or node.js. As such, it is a good oppor-
tunity for developing applications based on artificial intelligence, which for the time
being are not numerous. The node.js programming language is fairly reliable for cre-
ating services e.g. API (Application Programming Interface), which would allow
multithreading – the application’s capacity to provide simultaneous responses. The
beneficial part of the node.js language is the immense amount of code, already pre-
sented on GitHub. A developer only has to combine the resources, in order to save time
when creating a new application. The following chapters will present a few basic terms
and their implementation using the developed software application in intelligent social
networks.

2 Neural Networks

A neuron is the fundamental unit of a neural network with inputs, processing and
outputs, whose operation depends on a serious of synaptic weights, which, applied to
input data, generate outputs through the activation function [6]. Neural networks
consist of multiple nodes, imitating the biological neurons of the human brain, con-
nected by links, and which interact with each other (Fig. 1). Each node can assume
input data and perform simple operations with these data, the outcome being trans-
mitted to other neurons. The output node is called value node or activation node. Each
link is assigned to a synaptic weight, neurons being thus capable of learning, changing
the value of the link’s synaptic weight in a similar way to biological synapses.

Fig. 1. Example of a simple neural network

Each line is a link between two neurons, and links can have a certain direction
depending on the type of the diagram, indicating the flow of information (Fig. 1). Each
link has a synaptic weight, an integer (variable) controlling the signal between two
neurons. If the network generates a “good” or “desired” output, then there is no need to
adjust the synaptic weight. If, however, it should generate an “unsatisfactory”, “un-
desired” output or an error, then the system changes the synaptic weight, in order to
 The Basics of Implementing Intelligent Social Networks 19

improve future outcomes. The most widely used algorithm for neural networks is the
back propagation algorithm [7]. This is a learning algorithm in which the network’s
neurons learn from examples. Thus, if the algorithm is assigned a task to perform, it
will change the synaptic weight of the network at the end of learning, to produce a
specific output from input data. The algorithm adjusts the synaptic weight using as

Fig. 2. Operation of a perceptron

calculus the minimization of the Gradient Descent function [8]. Usually, the activation
function is a sigmoid function (Fig. 2).
In the case of FeedForward network, the propagation of information is unidirec-
tional and, for this reason, it is one of the simplest neural networks, namely, a neuron
sends information to another neuron from which it does not receive information. Such
networks do not have feedback loops and are used to generate, recognize or classify
forms having fixed inputs and outputs.

3 Implementing the Application

The Synaptic.js library [9] was used to implement the application. Like the other
options, namely Brain.js and Mind.js, it is possible to create a complete neural network
using only a few lines of code. Inside the library there are several preconfigured
architectures of neural networks, which underlay the software described as follows. The
application simulates a virtual set defined by the margins of the HTML5canvas, in
which the entities created will explore a set/map, will have obstacles to avoid, will feed
and will die. As follows we will detail the application’s options and mode of operation.
Entities are divided into two classes, one of entities learning together and behaving as a
group, and the other one in which entities work individually. As such, it can be
observed that, when initializing the application, all entities appear in completely ran-
dom places on the map (Figs. 3 and 4). As they converge, red ones tend to form a
group and move in a group (Fig. 4).
20 S.C. Popa et al.

Fig. 3. Generating entities

Fig. 4. Red entities use the cohesion function

This is possible due to using the cohesion function, which defines the cohabi-
tation between the members of a group, defined as follows:
cohesion:function(neighboors)
{ var sum=new Vector(0,0); var count=0;
for (var i in neighboors)
{if (neighboors[i]!=this)//&&!neighboors[i].special)
{sum.add(neighboors[i].location);count++;
}
}
sum.div(count);return sum;
}

The function will find all the neighbours within its range that pertain to the same
species and add them to the group. This way the neural network, which defines the
entity, learns what its neighbours and what its obstacles are. The method is known as
neural network training. In this case, the entities will be defined as follows:
 The Basics of Implementing Intelligent Social Networks 21

function Creature(world,x,y)
{ this.network=new Architect.Perceptron(40,25,3);
this.world=world;this.mass=.4;
this.maxspeed=3;this.maxforce=.2;
this.lookRange=this.mass*200;
this.length=this.mass*10;
this.base=this.length*.5;
this.HALF_PI=Math.PI*.5;
this.TWO_PI=Math.PI*2;
this.location=new Vector(x,y);
this.velocity=new Vector(0,0);
this.acceleration=new Vector(0,0);
this.color="#C70E4C";
}
Thus, entities are assigned a neural network (a perceptron), which has 40 inputs, 25
hidden layers and 3 outputs. Likewise, each entity must be assigned a set/world. As the
laws of mechanical physics will apply, each entity will have a mass, so that when a
collision occurs with other objects in the map, it will move at maximum speed,
therefore having maximum collision force. For an entity to be able to see, the
lookRange variable will be assigned to it, which will be equal to its mass multiplied
by 200; therefore, the entity will have a visual range of 80 pixels around it. The location
of the entity will be given by the set, and color is a variable that can be set as needed.
Going back to location, it is noted that it instantiates a vector function. This function is
useful for calculating the position of the entity relative to coordinates, angle of impact,
mass and weight. Below are a few lines of code from the function:

angle:function()
{ return Math.atan2(this.y,this.x);
},
setMag:function(m)
{ var angle=this.angle();
this.x=m*Math.cos(angle);
this.y=m*Math.sin(angle);
return this;
},
setAngle:function(a)
{ var mag=this.mag();
this.x=mag*Math.cos(a);
this.y=mag*Math.sin(a);
return this;
},
rotate:function(a)
{ this.setAngle(this.angle()+a);
return this;
},
limit:function(l)
{ var mag=this.mag();
if(mag>l)this.setMag(l);
return this;
},
22 S.C. Popa et al.

The entity is displayed in the set using the moveTo function, which assumes the
coordinates in the memory locations and, depending on the dangers on the map, will
apply the laws of mechanical physics for each element, and will establish the angular
displacement, alignment collision or coexistence. For example, var separation
will use the separate function to separate the entities in the set from those of a
different color, a variable which will be assigned to a force variable representing a
vector, and then this variable will be applied to the entity. The same also happens to the
other variables, defined in the previously described function.
moveTo:function(networkOutput){
var force=new Vector(0,0);
var target=new Vector(networkOutput[0]*this.world.width,
networkOutput[1]*this.world.height);
var angle=(networkOutput[2]*this.TWO_PI)-Math.PI;
var separation=this.separate(this.world.creatures);
var separation2=this.separate(this.world.obstacles);
var
alignment=this.align(this.world.creatures).setAngle(angle);
var cohesion=this.seek(target);
force.add(separation);
force.add(separation2);
force.add(alignment);
force.add(cohesion);
this.applyForce(force);
},

The draw function will access another function of the entity, namely update. The
function calculates the location where the entity will be drawn in the canvas,
depending on speed and current location:
draw:function()
{ this.update();
var ctx=this.world.context;ctx.lineWidth=1;
var angle=this.velocity.angle();
x1=this.location.x+Math.cos(angle)*this.base;
y1=this.location.y+Math.sin(angle)*this.base;
x2=this.location.x+Math.cos(angle+this.HALF_PI)*this.base;
y2=this.location.y+Math.sin(angle+this.HALF_PI)*this.base;
x3=this.location.x+Math.cos(angle-this.HALF_PI)*this.base;
y3=this.location.y+Math.sin(angle-this.HALF_PI)*this.base;
ctx.lineWidth=2;ctx.fillStyle=this.color;
ctx.strokeStyle=this.color;ctx.beginPath();
ctx.moveTo(x1,y1);ctx.lineTo(x2,y2);
ctx.lineTo(x3,y3);ctx.stroke();ctx.fill();
},
 The Basics of Implementing Intelligent Social Networks 23

The update function, as defined in the code, will access the boundaries
function of the entity, which will define the limits and add extra acceleration to the
speed of this entity [9]:
update:function()
{ this.boundaries();
this.velocity.add(this.acceleration);
this.velocity.limit(this.maxspeed);
if(this.velocity.mag()<1.5)
this.velocity.setMag(1.5);
this.location.add(this.velocity);
this.acceleration.mul(0);
},

The seek function will establish the speed of the group, depending on the location
of neighbours in the group and is defined as follows:
seek:function(target)
{ var seek=target.copy().sub(this.location)
seek.normalize();seek.mul(this.maxspeed);
seek.sub(this.velocity).limit(0.3);
return seek;
},

Two functions, blastoff and loop, are used to implement the set, or define the
map. The blastoff function contains: the canvas variable, which will find the div
with the canvas id in the html file and transform it into this set; the ctx variable,
which defines the context of the application, i.e. 2-dimensional; the obs variable,
which represents the number of obstacles; the num variable, which represents the
number of entities in each group; the fps variable, which represents the application’s
display rate; the world variable, which will have width, height and an array of
entities, an array of obstacles, an array of food and the option of context. Cycling will
be conducted in order to populate the map with entities, obstacles and food, meaning
one for to each variable of values, assigning to the map the entity, obstacle or food,
with the key from the for at random values x and y. Likewise, this function is used
to calculate the new location and the angle of incidence for each entity. The loop
function contains settings for the context (genre, color, form, size), the update
function for entities (obstacles or food), foreach for obstacles, entities and food (in
which the input for the neural network will be set, where the learning rate and target
coordinates are), the function for drawing obstacles, entities and food. The function will
be recalled in order to make it recursive. Each entity, when encountering food, will
assume the mass/volume of food, while it disappears from that location and reappears
in a different spot on the map. To avoid an excessive increase in the number of entities
the life function will be utilized, which will decrease the mass of the entity by 0.1
per minute, until it reaches 0; in this situation the die function will be called, which
destroys the entity and transforms it into food. An interesting thing to observe in the
entities’ feeding process is that the entities which move in a group and learn together
24 S.C. Popa et al.

have a higher survival rate (as they learn from the others that mass is essential for
survival), whereas individual entities – if they do not find food – will die, and those that
do will learn how to survive.

4 Future Work

Once the previously described application has been developed, further work might
include creating a platform matching people according to certain criteria, as well as a
search engine, in which the input prediction will be based on the same concepts as
those presented in the paper. It is worth mentioning that there is already a similar
concept of using artificial intelligence to improve search, developed by Google, capable
of learning and improving results, utilizing data derived from people [10]. Short-term
development possibilities refer to implementing predictive search, while attempting to
implement finding gender- or age-related similarities in the search engine. The med-
ium- and long-term development possibilities refer to optimizing the source code, to
developing a support service or an automated registration function, using existing
accounts from the social network Facebook and Google APIs, to increase the number
of users. Likewise, development refers to training neural networks and collecting
comparative data from other conventional matching systems, aiming to analyze per-
formance, or to add a satisfaction form, which will help improve the application.
Another perspective consists of improving application security and preventing server
overload due to robot-generated ports, or optimizing the application for mobile devices
(smartphones, tablets). The node.js application consists of an http server and a restful
API. So far, the following node.js modules/libraries [11] have been used to create this
application:
– express is a module for node.js allowing to create a web application and post it
on one of the available http ports;
– body-parser is a module transmitting post requests to the application;
– ejs is a template manager for node.js, which allows adding a special code into an
html sheet, which can display the variables in the template;
– mongoose is a module linking the application to the mongodb database [12, 13];
– passport are modules needed to create easier and more secure login;
– bcrypt-nodejs is a module used to secure the application and passwords;
– cookie-parser is a module that helps in using cookies for the session;
– Synaptic is a library for the neural network [9].
The application has the following structure [11, 14]: app, views, config,
node-modules, routes and auth, models, user, index, login, reg-
ister, dashboard, profile, search, database.js. The user will be able
to edit their profile, look for friends using their desired qualities for subsequent sorting,
or may find already connected friends by name, using a conventional search function.
By this stage we developed the search engine, based on Hopfield Architecture,
which serves as content-addressable memory. Networks are trained to “remember”
patterns, and when new patterns are inputted, it returns the most similar pattern for
which it was trained.
 The Basics of Implementing Intelligent Social Networks 25

var hopfield = new Architect.Hopfield(10) // creation of a

network for 10-bit patterns

// training the network with two different patterns

hopfield.learn([
[0, 1, 0, 1, 0, 1, 0, 1, 0, 1],
[1, 1, 1, 1, 1, 0, 0, 0, 0, 0]
])
// when we input new patterns, the network will return the
most result similar with the one for which it was trained

hopfield.feed([0,1,0,1,0,1,0,1,1,1])
// the network will return [0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
hopfield.feed([1,1,1,1,1,0,0,1,0,0])
// reteaua va returna [1, 1, 1, 1, 1, 0, 0, 0, 0, 0]

Using this example we created the search engine architecture by extending the
Network class of Synaptic library.
We created an 80-bit Hopfield neural network. The doTrain function is set at
1,000 iterations and an error ratio of 0.0005. If a new username is entered into the
database then the network will be automatically trained for. After converting the result
back into ASCII from the binary string, the network response will be displayed in ai.ejs
where the user will be able to view the search result (Fig. 5) and that person’s data in
order to contact them (Fig. 6).

Fig. 5. Example for searching

26 S.C. Popa et al.

Fig. 6. Searched personal information

Each user will be able to find their friends using the search form, this information
being then stored into the database and submitted to the neural network training
function, according to predefined criteria (sex, age, common friends). Neural networks
have yielded good results in text classification. Although instruction for complex tasks
is cumbersome, we attempted to use them in this application. In theory, the application
will operate by obeying a certain concept. Thus, there will be a neural network trainer,
which will train the neural network according to certain formulas (e.g. age or sex).
Thus, a female person will be assigned a male person, or another female person, all
depending on the preferences set in the profile. Likewise, calculations will be made to
determine and match optimal age. Once the neural network has been trained, when
searching for people it will return as result the person it deems appropriate. This may be
the beginning of a new era in social networking or dating, the innovation consisting of
this very matching, which will be given by an artificial intelligence trained for this
purpose. Since the search will be performed using graph technology, optimizing results
with artificial intelligence will lead to a much shorter calculation time. At the same
time, it is interesting to note that the selection of a partner will be made “intelligently”,
and not according to a set of if-else testing criteria.

5 Conclusions

This application illustrates the implementation of an artificial life form in a virtual

environment with technologies known by most programmers, with unlimited appli-
cation in the field of artificial intelligence, which may lead to faster searches, indexing
and perhaps even to an artificial intelligence that could learn to communicate. By
learning one programming language, namely JavaScript, with the help of Node.js and
using the Synaptic library, an easy opportunity is opened to developing this kind of
applications, offering JavaScript developers the possibility to do whatever they want
with the same programming language. The purpose of the software is to serve an
introduction to understanding neural networks, avoiding the fairly cumbersome
mathematical part, with immediate application to the implementation of intelligent
social networks.
 The Basics of Implementing Intelligent Social Networks 27

References
1. HTML5 - Web developer guide, MDN (2015). https://developer.mozilla.org/en-US/docs/
Web/Guide/HTML/HTML5
2. CSS3-CSS, MDN. https://developer.mozilla.org/en/docs/Web/CSS/CSS3
3. Azat, M.: JavaScript and Node Fundamentals, Washington, DC (2014)
4. Kaufman, C., Perlman, R., Speciner, M. (eds.): Mastering Node.js. Packt Publishing, Livery
Place (2013)
5. Node.js v5.9.0, Documentation. https://nodejs.org/docs/latest/api/
6. Dzitac, I. (ed.): Inteligentă artificiala. Universitatii Aurel Vlaicu, Arad (2008)
7. Popescu, M.C., Balas, V., Olaru, O., Mastorakis, N.: The backpropagation algorithm
functions for the multilayer perceptron. In: Proceedings of the 11th WSEAS International
Conference on Sustainability in Science Engineering, Timisoara, pp. 28–31 (2009)
8. Popescu M.C., Olaru O.: Conducerea optimala a proceselor - Proiectare asistata de calculator
in Matlab si Simulink. Academiei Tehnice Militare, Bucureşti (2009)
9. Synaptic.js. https://github.com/cazala/synaptic
10. Patent Google. http://www.google.com/patents/US20050004905
11. Easy Note Authentication. https://scotch.io/tutorials/easy-node-authentication-setup-and-
local
12. A Basic Introduction to Mongo DB. https://mongodb.github.io/node-mongodb-native/api-
articles/nodekoarticle1.html
13. The Mongo Shell. https://docs.mongodb.org/manual/mongo/#working-with-the-mongo-s
14. An overview of Bootstrap, How to Download and Use, Basic Templates and Examples, and
more. http://getbootstrap.com/getting-started/
 The Generation of Tests of Knowledge Check Using
Genetic Algorithms

Doru Anastasiu Popescu1 ✉ , Nicolae Bold2, and Alexandru Ion Popescu3

( )

1
Faculty of Mathematics and Computer Science, University of Pitesti, Pitesti, Romania
[email protected]
2
Faculty of Management, Economic Engineering in Agriculture and Rural Development,
University of Agronomic Sciences and Veterinary Medicine Bucharest, Slatina Branch, Romania
[email protected]
3
National College “Radu Greceanu”, Slatina, Romania
[email protected]

Abstract. In this paper, we will present a modality for generating knowledge-

check tests using questions characterized by keywords. The obtained tests have
certain restrictions regarding the keywords labeling the questions. These restric‐
tions will be defined in the paper. The test generation will be made using genetic
algorithms, whose chromosomes will be considered the tests and genes will be
considered questions. At the end of the paper some results obtained with the
algorithm implemented in Java programming language will be presented.

Keywords: Learning · Assessment · Algorithm · Genetic · Chromosomes

1 Introduction

The problem of assessment of knowledge gained after a learning period is always up to

date. The way of determining the degree of notion assimilation gained through learning,
the connection between them and the way in which they are used in practice is very
important, because these actions can lead to changes in contents, examples and modal‐
ities of presentation in order to achieve a greater efficiency. Developing these types of
methods is based on a well-projected assessment system or strategy, as shown in paper
[5]. After this step, the teachers must combine this strategy with the suitable methods.
There are shown several ways of optimizing the assessment methods using e-learning
platforms in papers [3, 4]. These platforms are strongly connected with algorithmic
approaches of assessment.
The intrusion of education has effects on education. This is not a perspective which
we will study in this paper, but it is important to be studied. In this matter, papers [1,
2] show some of the effects of technology-based education methods on the education.
Next, we will present a modality of creating knowledge assessment tests using a set
of questions, for which we know a set of keywords. In practice, these keywords are
mostly terms and notions that characterize the question and are used in the learning
process, which have a special significance for the teacher and the student.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_3
 The Generation of Tests of Knowledge Check 29

In practice, a useful thing was the fixation of a set of keywords, the generation of
tests which would use as many keywords as possible from this set and the fact that every
question must contain at least one keyword from this set.
For finding the solution of this problem, in the paper we used a genetic algorithm.
A chromosome is represented by a test and a gene is represented by a question. The
genetic algorithms have a great applicability in the case of NP-complete problems. This
type of problems has the specificity of obtaining several results which have the same
restrictions. Another solving variant would have needed the usage of generation algo‐
rithms based on backtracking method, but this is useless for a large number of questions
using the actual technology.
The genetic algorithms have a great applicability in domains such as web applica‐
tions (as those presented in papers [6–8]), the creation of plans (planning) for different
activities (as those presented in [9, 12]), graph theory [10, 13] and other practical
domains.
Section 2 will present the problem of test generation with restrictions and the nota‐
tions which will appear in the other sections and an example.
In Sect. 3 we will present the general form of the genetic algorithm used in the next
section and in Sect. 4 we will present the algorithm which solves the problem of the
generation of tests with restrictions. The results obtained with this algorithm are
presented in Sect. 5.

2 The Problem of Test Generation with Restrictions

As we presented in the introduction, we want to present a modality of test generation

having at hand a set of n questions given by:
– The statement;
– The keywords which define the question.
These keywords have a special significance in the process of learning and assimila‐
tion of knowledge.
An example of 4 questions of this kind for a course of operating systems (OS) is
presented next:

Question Q1
Statement: Which is the role of the Android operating system?
Keywords: operating, Android

Question Q2
Statement: Which are the functions of the Windows 10 operating system?
Keywords: operating, Windows

Question Q3
Statement: How can we create a folder in Windows 10?
Keywords: creation, folder, Windows
30 D.A. Popescu et al.

Question Q4
Statement: How can we create a file in Windows 10?
Keywords: create, file, Windows

The teacher which holds these questions (statements and keywords) wishes to make
several tests formed from m questions which will use all the keywords from a set of
keywords established by him, denoted by KW. Another request is that every question
from the test must be characterized by at least one keyword from KW.
For example, for m = 3 and KW = {Windows, Android, folder} we can build the
next tests:

Test 1: Q1, Q2, Q3

Test 2: Q1, Q3, Q4

The problem of the generation of these tests is reduced to the generation of subsets
with k elements, subsets of the question set. These subsets have the properties:
– Every question from the subset contains at least one keyword from KW;
– Every keyword from KW is a keyword of at least one question from the subset.
Practically, every question is characterized by a certain number of keywords that
summarize best the subject that the question treats. Given a set of initial keywords given
by the user, the generator forms tests from questions whose at least one of the keywords
matches with one of those given by the user. The choice of these questions is made using
a generator that uses genetic methods and structures.
There can exist the particular case of not receiving any solution to a particular set of
input data. In this case, the algorithm can output the maximum number of questions that
can be generated and used, a number which is lower than m.
In order to obtain a larger number of tests with a polynomial algorithm, we will use
genetic algorithms, which offer a sufficient number of tests in order for our problem to
be solved.

3 Description of the Used Algorithm

In this paper, we will use a variant of the algorithm which allows us to generate tests
formed from questions codified by numbers 1, 2, …, n. Each question has its set of
keywords KW1, KW2, …, KWn specified and for each set of keywords K we want to
generate tests with k questions so that every question i from the test (1 ≤ i ≤ n) has the
property that KWi and K have at least one common keyword and K is a subset in the
union of tests set.
The genetic algorithm used in Sect. 4 will generate chromosomes with k genes, every
gene having values from the set {1, 2, …, n}.
For the generation of the chromosomes we will use the genetic algorithm presented
in Fig. 1.
 The Generation of Tests of Knowledge Check 31

Fig. 1. Scheme of a genetic algorithm for the problem of test generation

The sequence of Java instructions from the genetic algorithm which make the gener‐
ation of chromosomes after NoGeneration is the next one:

The method generate() makes the generation of the initial population. The method
generation() builds the new generation resulted after the chromosomes are crossed-over
using various methods. The method mutation() is used for the modification of some
genes from certain chromosomes chosen randomly. The method sort() is used for sorting
the chromosomes by the value of the fitness function, so that the most performant chro‐
mosomes can be found at the upper part of the population.

4 Description of the Algorithm Used for the Problem

of Test Generation

Next, we will use the notations from the previous sections. In order to represent better
the fact that a question Q is labeled or not with the keywords from the set
KW = {KW1, KW2, …, KWnrKW} we will use a string of numbers v = (v1, v2, …, vnrQ),
with the next significance:
{
1, if the keyword from the position i from Q is i KW
vi =
0, otherwise
32 D.A. Popescu et al.

nrQ represents the number of keywords from the question Q. The structures used
within the algorithm are:
• n (integer value, representing the total number of tests);
• m (integer value, representing the number of questions within the sequence of optimal
tests, m ≤ n);
• nrKW (integer value, the number of keywords given by the user);
• nrcuv (integer value, the number of keywords for each test);
• TKW[nrKW] (String array, containing the keywords given by the user);
• T[n][nrKW] (bi-dimensional array, which contains 0 and 1 values, with the next
significance of T[i][j]: 1 if the keyword j from question i is found in the array KW
and 0 otherwise), adica sirul v pentru intrebarea i;
• NoGeneration (integer value, representing the number of desired generations for the
genetic algorithm);
• NrPop (integer value, which represents the total number of solutions);
• pop[NrPop][m] (integer bi-dimensional array, which contains the solutions).
The chromosomes (stored in the lines of the bi-dimensional array pop) are made
from genes with values from the set of questions {1, 2, …, n}. The number of genes
from the chromosomes is m.
The fitness function will be defined, for a chromosome i, as the number of keywords
from the questions which are also in the set KW and will be denoted by f(i). In order to
calculate this function, we will use the data processing made in the array T.
Input data is represented by n, m, nrKW and KW. Output data is represented by the
first k lines of the bi-dimensional array pop.
The algorithm described in this paper is a sequence of steps which have their specific
operations. These steps are presented below.
Step 1. Input data presented above is read.
Step 2. The arrays T is built, using the next idea (i ≤ n, j ≤ nrcuv):
T[i][j] = 1, if the keyword j from test i is in KW and T[i][j] = 0, otherwise;
Step 3. The sequences are randomly generated, from the former test to the latter, test
by test, after the next form:
pop[i][j] = z; z = 1..n; i ≤ NrPop; j ≤ m
The values of the fitness are calculated for every sequence.
Steps 4, 5 and 6 are repeated for a number of generation times.
Step 4. The sequences obtained at Step 3 are used for generating other sequences
using crossover operation. This means that for two sequences chosen randomly pop[N]
and pop[M] are switched depending on the position P, also chosen randomly (N,
M ≤ NrPop, P ≤ m):

pop[N] = popN, 1, popN, 2, … popN, P, popN, P + 1 … , popN, m

pop[M] = popM, 1, popM, 2, … , popM, P, popM, P + 1 … , popM, m
↓
pop[N] = popN, 1, popN, 2, … , popN, P, popM, P + 1 … , popM, m
pop[M] = popM, 1, popM, 2, … , popM, P, popN, P + 1 … , popN, m
 The Generation of Tests of Knowledge Check 33

Step 5. The sequences obtained are used for generating other sequences using muta‐
tion operation. The mutation for a sequence pop[N] and two positions M and P is made
in this way (N ≤ NrPop, M, P ≤ m):
pop[N] = popN,1, …, popN,M, …, popN,P, …, popN,m
Step 6. Using any type of ordering method, the sequences are ordered depending on
the fitness value of the sequence.
Step 7. The first Nopop ordered solutions are output, because they have the highest
fitness values.

5 Observation

The presented algorithm complexity is O(NoGeneration∙(n2 + Nrpop − log(Nopop))).

6 Implementation and Results

One of the key characteristic that the generator must have is the quickness of the solu‐
tions. This is why we will study the algorithm regarding the runtime of the algorithm
for variations of n, m and NoGeneration, variables that are involved and most important
within the generator. The algorithm presented in Sect. 4 was implemented in Java
programming language. The calculations for the runtime were run in a Microsoft
Windows 8.1 environment, on a machine equipped with a 1.80 GHz processor which
uses a 4 GB RAM memory (Tables 1, 2, 3).

Table 1. Results of runtime for different values of n (m = 10, NoGeneration = 200)

n Runtime (s)
100 1.65
250 1.89
500 3.02
750 2.94
1000 2.86
1000 2.91
(all questions having the keywords from TKW)

Table 2. Results of runtime for different values of m (n = 1000, NoGeneration = 200)

m Runtime (s)
100 5.42
250 9.62
500 22.08
34 D.A. Popescu et al.

Table 3. Results of runtime for different values of NoGeneration (n = 1000, m = 10)

NoGeneration Runtime (s)
100 2.89
250 6.68
500 13.05
750 20.37
1000 25.71

Following the algorithm, we obtained the next results:

These results are showing us that the algorithm runtime depends on m and NoGen‐
eration and less on n (the number of questions). Thus, variability of results (given by a
larger number of generations) is made in a progressively larger period (Fig. 2).

Fig. 2. Scheme of a genetic algorithm for generated test problem

The output shows in this way (n = 100, m = 10, NoGeneration = 200, k = 10):
– 53 17 47 12 68 33 98 18 10 49 no of keywords from TKW: 9
– 12 17 33 10 18 47 98 53 68 49 no of keywords from TKW: 9
– 18 68 98 49 47 17 10 12 53 33 no of keywords from TKW: 9
– 17 12 98 10 47 33 53 49 68 18 no of keywords from TKW: 9
– 12 33 53 10 47 17 98 18 68 49 no of keywords from TKW: 9
– 10 17 53 12 18 33 98 47 68 49 no of keywords from TKW: 9
– 98 17 33 18 47 68 53 12 10 49 no of keywords from TKW: 9
– 10 17 49 12 33 47 98 18 68 53 no of keywords from TKW: 9
– 49 17 53 33 47 12 10 18 98 68 no of keywords from TKW: 9
– 10 17 53 98 47 33 68 18 12 49 no of keywords from TKW: 9
The series of numbers presented above are, actually, the sequences of questions (that
is, the tests). A number represents the order number of a question in the database of
questions where questions are kept and updated over periods of time. For example, 53
is the question with the order number 53, which can be accessed using this number. The
questions are validated, the keywords of the questions in the sequences matching with
those given by the user. Also, these sequences contain the highest number of keywords
which match with those given by the user.
The results obtained with the Java implementation for the algorithm presented in
Sect. 4 were good, confirming its practical utility.
 The Generation of Tests of Knowledge Check 35

7 Conclusions

The model presented in the paper was used in the learning and assessment processes for
different courses and the results were good. In the next period, we would like to develop
an online application which would offer the possibility of the easy usage of this model.
We consider that other types of restrictions related to the questions used at test generation
can be established. The fitness function from the test generation algorithm can be modi‐
fied in order for other conditions to be verified. The questions can be related to one
another. For example, one of them to be part of the same test, this relation being modeled
with an un-oriented graph, as presented in [11]. We also want to make, as a future work,
a comparative study of the methods used for generating tests which exist at this point
and are used in the learning process.

References

1. Cox, M.J., Marshall, G.: Effects of ICT: do we know what we should know? Artic. Educ. Inf.
Technol. 12(2), 59–70 (2007)
2. Catalano, H.: Pedagogical implications of using net support school application in teaching
educational sciences. In: International Conference on Virtual Learning, Timișoara (2015)
3. Baron, C., Şerb, A., Iacob, N.M., Defta, C.L.: IT infrastructure model used for implementing
an E-learning platform based on distributed databases. Qual. Access. Success. J. 15(140),
195–201 (2014)
4. Defta, C.L., Serb, A., Iacob, N.M., Baron, C.: Threats analysis for E-learning platforms.
Knowledge Horizons. Eco. 6(1), 132 (2014)
5. DeLuca, C., Klinger, D.A., Searle, M., Shulha, L.M.: Developing a curriculum for assessment
education. Assess. Matt. 2, 20–42 (2010)
6. Popescu, D.A., Radulescu, D.: Approximately similarity measurement of web sites, ICONIP.
In: Neural Information Processing, Proceedings LNCS, Springer, Heidelberg, 9–12 November
2015
7. Popescu, D.A.: Reducing the navigation graph associated to a web application. Buletin
Stiintific - Universitatea din Pitesti, Seria Matematica si Informatica, Nr. 16, 125–130 (2010)
8. Popescu, D.A., Popescu, I.A.: Model of determination of coverings with web pages for a
website. In: International Conference on Virtual Learning. pp. 279–283 (2015)
9. Popescu, D.A., Nicolae D.: Generating a class schedule with reduced number of constraints.
In: Proceedings the 9th International Scientific Conference eLearning and software for
Education, Bucharest, April 25–26, pp. 297–300 (2013)
10. Popescu, D.A., Ionita, A.E.: Combinatorica si Teoria Grafurilor, Ed. Rhabon. Tg. Jiu (2005)
11. Nijloveanu, D., Bold, N., Bold, A.C.: A hierarchical model of test generation within a battery
of tests. In: International Conference on Virtual Learning (2015)
12. Larrañaga, P., Kuijpers, C.M.H., Murga, R.H., Inza, I., Dizdarevic, S.: Genetic algorithms for
the travelling salesman problem: a review of representations and operators. Artif. Intell. Rev.
13(2), 129–170 (1999)
13. José Fernando Gonçalves, Mauricio, G.C.: Resende, biased random-key genetic algorithms
for combinatorial optimization. J. Heuristics 17(5), 487–525 (2011)
 Matrix-Valued Bidirectional Associative
Memories

Călin-Adrian Popa(B)

Department of Computer and Software Engineering,

Polytechnic University Timişoara, Blvd. V. Pârvan,
No. 2, 300223 Timişoara, Romania
[email protected]

Abstract. Complex-, quaternion-, and Clifford-valued neural networks

have been intensively studied over the last few years. Matrix-valued bidi-
rectional associative memories, a further generalization of the real-valued
neural networks, for which the states, outputs, and thresholds are all
square matrices, are introduced. After giving the definition of these net-
works, an expression for an energy function is provided. A proof is given
that this function is indeed an energy function for the proposed network.

Keywords: Clifford-valued neural networks · Bidirectional associative

memories · Energy function · Matrix-valued neural networks

1 Introduction

There has been an increasing interest in the study of neural networks with values
in multidimensional domains in recent years. Although first introduced in the
1970’s (see, for example, [17]), complex-valued neural networks, the first type of
multidimensional neural networks that was proposed, have received more atten-
tion in the 1990’s and in the past decade. Their wide range of applications include
those in complex-valued signal processing and in telecommunications and image
processing (see, for example, [5,10]).
Another type of neural networks, defined on the 2-dimensional algebra of
hyperbolic numbers, are hyperbolic-valued neural networks, see [3,8,12]. It has
been shown that their decision boundary, consisting of two hypersurfaces, can
have any angle, unlike the complex-valued neural networks, which always have
orthogonal decision boundaries.
Quaternion-valued neural networks, which are defined on the 4-dimensional
algebra of quaternion numbers, were introduced in the 1990’s also, first as a gen-
eralization of the complex-valued neural networks, see [1,2,11]. Popular appli-
cations of these networks include chaotic time series prediction, the 4-bit par-
ity problem, and quaternion-valued signal processing. A promising area which
remains to be explored is that of 3-dimensional and color image processing, in
which data can be more naturally expressed in quaternion form.


c Springer International Publishing AG 2018
V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9 4
 Matrix-Valued Bidirectional Associative Memories 37

Clifford-valued neural networks were defined in [13,14], and later discussed,

for example, in [4,7,8]. The complex, hyperbolic, and quaternion algebras are
special types of Clifford algebras, which have dimension 2n , n ≥ 1. Clifford
algebras or geometric algebras have many applications in physics and engineer-
ing, making them appealing as novel types of data representation for the neural
network domain, also.
Complex, hyperbolic, quaternion, and Clifford numbers√can all be written in
matrix form. For example, a complex number a + ib, i = −1, can be written
in the form  
a −b
,
b a
a hyperbolic number a + ub, u2 = 1, u = ±1, in the form
 
ab
,
ba

and a quaternion a + ib + jc + kd, i2 = j 2 = k 2 = ijk = −1, in the form

⎛ ⎞
a b c d
⎜ −b a −d c ⎟
⎜ ⎟
⎝ −c d a −b⎠ .
−d −c b a

Thus, each of these algebras can be seen as a subalgebra of the algebra of square
matrices, with the natural addition and multiplication of the matrices. This led
to the natural idea of defining a generalization of all the above neural networks
with matrix inputs, outputs, weights and biases, first in the form of feedforward
networks, see [15], and now in the form of bidirectional associative memories.
Because of their degree of generality, these neural networks are bound to have
many applications in the future at solving problems at which traditional neural
networks have failed or performed poorly.
First introduced by Kosko in [6], bidirectional associative memories, an exten-
sion of the unidirectional Hopfield neural networks, were intensely studied, and
have many applications in pattern recognition and automatic control. Because
of the fact that complex-valued bidirectional associative memories were intro-
duced in [9], quaternion-valued bidirectional associative memories in [7], and
Clifford-valued bidirectional associative memories in [16], we considered an inter-
esting idea to introduce matrix-valued bidirectional associative memories. These
networks can be applied to store matrix patterns and to solve difficult matrix
optimization problems.
The rest of this paper is organized as follows: Sect. 2 gives the definition
of matrix-valued bidirectional associative memories, and the expression for the
energy function, together with the proof that the given function is indeed an
energy function for the defined network. Section 3 is dedicated to presenting the
conclusions of the study.
38 C.-A. Popa

2 Matrix-Valued Bidirectional Associative Memories

Consider the algebra Mn of square matrices of order n with real entries.
In the following, we will define bidirectional associative memories for which
the states, outputs, and thresholds are all from Mn , which means that they are
square matrices. The network is described by the set of differential equations
⎧
⎪
⎪ dXi (t) P
⎪
⎪ τ = −X (t) + Wij f (Yj (t)) + Ai , ∀i ∈ {1, . . . , N },
⎨ i dt i
j=1
(1)
⎪
⎪ dYj (t) N
⎪
⎪
⎩υj dt = −Yj (t) + Wji f (Xi (t)) + Bj , ∀j ∈ {1, . . . , P },
i=1

where τi ∈ R, τi > 0 is the time constant of neuron Xi , υj ∈ R, υj > 0

is the time constant of neuron Yj , Xi (t) ∈ Mn is the state of neuron Xi at
time t, Yj (t) ∈ Mn is the state of neuron Yj at time t, Wij ∈ Mn is the
weight connecting neuron Xi to neuron Yj , f : Mn → Mn is the nonlinear
matrix-valued activation function, Ai is the threshold of neuron Xi , and Bj is
the threshold of neuron Yj , ∀i ∈ {1, . . . , N }, ∀j ∈ {1, . . . , P }. The derivative is
considered to be the matrix formed by the derivatives of each element [Xi (t)]ab
of the matrix Xi (t) with respect to t:
 
dXi (t) d([Xi ]ab )
:= .
dt dt 1≤a,b≤n

By denoting Ui (t) := f (Xi (t)) the output of neuron Xi and Vj (t) := f (Yj (t))
the output of neuron Yj , the above set of differential equations becomes:
⎧
⎪
⎪ dXi (t) 
P
⎪
⎪τ = −X (t) + Wij Vj (t) + Ai , ∀i ∈ {1, . . . , N },
⎨ i
dt
i
j=1
⎪
⎪ dYj (t) 
N
⎪
⎪ −Y Wji Ui (t) + Bj , ∀j ∈ {1, . . . , P }.
⎩ j dt
υ = j (t) +
i=1

The activation function is formed of n2 functions f ab : Mn → R, 1 ≤ a, b ≤ n:

 
f (X) = f ab (X) 1≤a,b≤n .

Studying the stability of the above defined network requires a series of

assumptions about the activation function, which we will detail below.
The first assumption is that the functions f ab are continuously differentiable
with respect to each [X]cd , ∀1 ≤ c, d ≤ n, ∀1 ≤ a, b ≤ n, and the function f is
bounded: ∃M > 0, ||f (X)|| ≤ M,∀X ∈ Mn , where ||Y || is the Frobenius norm
of matrix Y , defined by ||Y || = Tr(Y Y T ), and Tr(Y ) represents the trace of
matrix Y . In this case, the n2 × n2 Jacobian matrix of the function f can be
defined as
 Matrix-Valued Bidirectional Associative Memories 39

 
∂f ab (X)
Jacf (X) = .
∂[X]cd 1≤a,b≤n
1≤c,d≤n

The second assumption is that f is injective and Jacf (X) is symmetric and
positive definite, ∀X ∈ Mn . This, together with the above assumption, assures
the existence of the inverse function of f , g : Mn → Mn , g = f −1 . We can thus
write g(Ui (t)) = Xi (t), ∀i ∈ {1, . . . , N }, and g(Vj (t)) = Yi (t), ∀j ∈ {1, . . . , P }.
In these conditions, there exists a function G : Mn → R,
n 
 [X]ab
G(X) = g ab (Y ab )dy,
a,b=1 0

where g ab : Mn → R are the component functions of g and the matrices Y ab

have the following form
⎧
⎪
⎨[X]cd , (c, d) < (a, b)
ab
[Y ]cd = y, (c, d) = (a, b) , ∀1 ≤ a, b ≤ n.
⎪
⎩
0 (c, d) > (a, b)

For example, for 2 × 2 matrices, we have that

 [X]11    [X]12  
y0 [X]11 y
G(X) = g 11 dy + g 12 dy
0 00 0 0 0
 [X]21    [X]22  
21 [X]11 [X]12 22 [X]11 [X]12
+ g dy + g dy.
0 y 0 0 [X]21 y

It can be easily seen that this function satisfies

∂G(X)
= g ab (X), ∀1 ≤ a, b ≤ n,
∂[X]ab

or, in matrix form:

∂G(X)
= g(X). (2)
∂X
Lastly, the third assumption concerns the weights of the network, which must
satisfy:
Wji = WijT , ∀i ∈ {1, . . . , N }, ∀j ∈ {1, . . . , P }.
40 C.-A. Popa

Taking all the above assumptions into account, we can define the energy
function E : MNn
+P
→ R of the bidirectional associative memory (1) as:


N 
P
E(U(t), V(t)) = − Tr(Ui (t)T Wij Vj (t))
i=1 j=1


N 
N
+ G(Ui (t)) − Tr(ATi Ui (t))
i=1 i=1

P 
P
+ G(Vj (t)) − Tr(BjT Vj (t)). (3)
j=1 j=1

A function E is an energy function for the network (1) if the derivative of E

along the trajectories of network, denoted by dE(U(t),V(t))
dt , satisfies the condition
dE(U(t),V(t)) dE(U(t),V(t)) dUi (t) dVj (t)
dt ≤ 0 and dt = 0 ⇔ dt = dt = 0, ∀i ∈ {1, . . . , N },
∀j ∈ {1, . . . , P }. In the following, we will show that the function E defined in
(3) is indeed an energy function for the network (1).
For this, we start by applying the chain rule:

dE(U(t), V(t)) N n
∂E(U(t), V(t)) d[Ui (t)]ab
=
dt i=1
∂[Ui (t)]ab dt
a,b=1


P n
∂E(U(t), V(t)) d[Vj (t)]ab
+
j=1 a,b=1
∂[Vj (t)]ab dt
  T 
N
∂E(U(t), V(t)) dUi (t)
= Tr
i=1
∂Ui (t) dt
  T 
P
∂E(U(t), V(t)) dVj (t)
+ Tr , (4)
j=1
∂Vj (t) dt

where by ∂E(U(t),V(t))
∂[Ui (t)]ab we denoted the partial derivative of the function E with
respect to each element [Ui (t)]ab of the matrices Ui (t), ∀1 ≤ a, b ≤ n, ∀i ∈
{1, . . . , N }, and analogously for ∂E(U(t),V(t))
∂[Vj (t)]ab , ∀j ∈ {1, . . . , P }.
For the partial derivatives
 
∂E(U(t), V(t)) ∂E(U(t), V(t))
=
∂Ui (t) ∂[Ui (t)]ab 1≤a,b≤n

and  
∂E(U(t), V(t)) ∂E(U(t), V(t))
= ,
∂Vj (t) ∂[Vj (t)]ab 1≤a,b≤n
 Matrix-Valued Bidirectional Associative Memories 41

we have from (3) that

∂E(U(t), V(t)) P
=− Wij Vj (t) + g(Ui (t)) − Ai
∂Ui (t) j=1
⎛ ⎞
P
= −⎝ Wij Vj (t) − Xi (t) + Ai ⎠
j=1

dXi (t)
= −τi , ∀i ∈ {1, . . . , N },
dt

∂E(U(t), V(t)) N
=− Wji Ui (t) + g(Vj (t)) − Bj
∂Vj (t) i=1
N 

=− Wji Uj (t) − Yj (t) + Bj
i=1
dYj (t)
= −υj , ∀j ∈ {1, . . . , P },
dt
where we used the fact that
dTr(X T A) dTr(AT X)
= = A,
dX dX
relation (2), the assumption Wji = WijT , and also the set of equations given by
(1). Now, Eq. (4) becomes:

 T 
dE(U(t), V(t)) N
dXi (t) dUi (t)
= Tr −τi
dt i=1
dt dt
   
P
dYj (t)
T
dVj (t)
+ Tr −υj
j=1
dt dt


N   T  
dXi (t) dUi (t)
=− τi vec vec
i=1
dt dt
P    T  
dYj (t) dVj (t)
− υj vec vec
j=1
dt dt
    

N
dUi (t)
T
T dUi (t)
=− τi vec [Jacg (Ui (t))] vec
i=1
dt dt
    

P
dVj (t)
T
T dVj (t)
− υj vec [Jacg (Vj (t))] vec
j=1
dt dt
≤ 0, (5)
42 C.-A. Popa

where we denoted by vec (X) the vectorization of matrix X. We also used the
identity
Tr(AT B) = vec(A)T vec(B), ∀A, B ∈ Mn ,
and, from g(Ui (t)) = Xi (t) and g(Vj (t)) = Yj (t), we obtained that
   
dg(Ui (t)) dUi (t)
vec = Jacg (Ui (t))vec ,
dt dt

   
dg(Vj (t)) dVj (t)
vec = Jacg (Vj (t))vec ,
dt dt
∀i ∈ {1, . . . , N }, ∀j ∈ {1, . . . , P }. Because Jacf (X) is symmetric and positive
definite, we deduce that Jacg (U ) is also symmetric and positive definite, and
thus   T  
dUi (t) dUi (t)
vec T
[Jacg (Ui (t))] vec ≥ 0,
dt dt
  T  
dVj (t) dVj (t)
vec T
[Jacg (Vj (t))] vec ≥ 0,
dt dt
∀i ∈ {1, . . . , N }, ∀j ∈ {1, . . . , P }, which allowed us to write the lastinequality 
in relation (5). Equality is attained when dE(U(t),V(t)) = 0 ⇔ vec dUdt i (t)
=
  dt
dVj (t) dVj (t)
vec dt = 0 ⇔ dUdti (t)
= dt = 0, ∀i ∈ {1, . . . , N }, ∀j ∈ {1, . . . , P }, thus
ending the proof that E is indeed an energy function for the network (1).
Two possible examples of activation functions, inspired by the ones used in
real-valued and complex-valued neural networks, are:
V
f (V ) = , ∀V ∈ Mn ,
1 + ||V ||
 
f ([V ]ab )1≤a,b≤n = (tanh[V ]ab )1≤a,b≤n , ∀V ∈ Mn .

The first one corresponds to the fully complex activation functions, while the sec-
ond one corresponds to the split complex activation functions from the complex-
valued domain.

3 Conclusions

The definition of the matrix-valued bidirectional associative memories was given.

Because complex, hyperbolic, quaternion and Clifford algebras can be seen as
matrix subalgebras of the square matrix algebra with the natural operations
of addition and multiplication of the matrices, matrix-valued neural networks
represent a generalization of the complex-, hyperbolic-, quaternion-, and Clifford-
valued neural networks.
 Matrix-Valued Bidirectional Associative Memories 43

A series of assumptions assuring the existence of the energy function for the
proposed network were made. Then, based on these assumptions, it was shown
that the proposed function is indeed an energy function for the above defined
matrix-valued bidirectional associative memory.
Complex- and quaternion-valued neural networks, and their direct gener-
alization, the Clifford-valued neural networks will most likely have even more
applications in the future. But Clifford algebras have dimension 2n , n ≥ 1, and
it is possible that the applications do not need such a large dimension for the
values of the input data. This is where matrix-valued neural networks might
come into play, because their dimension is only n2 , and thus it allows for more
efficient memory use than in the case of Clifford-valued neural networks.

References
1. Arena, P., Fortuna, L., Muscato, G., Xibilia, M.: Multilayer perceptrons to approx-
imate quaternion valued functions. Neural Networks 10(2), 335–342 (1997)
2. Arena, P., Fortuna, L., Occhipinti, L., Xibilia, M.: Neural networks for quaternion-
valued function approximation. In: International Symposium on Circuits and Sys-
tems (ISCAS). vol. 6, pp. 307–310. IEEE (1994)
3. Buchholz, S., Sommer, G.: A hyperbolic multilayer perceptron. In: International
Joint Conference on Neural Networks (IJCNN). vol. 2, pp. 129–133. IEEE, July
2000
4. Buchholz, S., Sommer, G.: On clifford neurons and clifford multi-layer perceptrons.
Neural Networks 21(7), 925–935 (2008)
5. Hirose, A.: Complex-Valued Neural Networks, Studies in Computational Intelli-
gence, vol. 400. Springer, Heidelberg (2012)
6. Kosko, B.: Bidirectional associative memories. IEEE Trans. Syst. Man. Cybern.
18(1), 49–60 (1988)
7. Kuroe, Y.: Models of clifford recurrent neural networks and their dynamics. In:
International Joint Conference on Neural Networks (IJCNN). pp. 1035–1041. IEEE
(2011)
8. Kuroe, Y., Tanigawa, S., Iima, H.: Models of hopfield-type clifford neural networks
and their energy functions - hyperbolic and dual valued networks. In: Interna-
tional Conference on Neural Information Processing, vol. 7062, pp. 560–569, Lec-
ture Notes in Computer Science (2011)
9. Lee, D., Wang, W.: A multivalued bidirectional associative memory operating on
a complex domain. Neural Networks 11(9), 1623–1635 (1998)
10. Mandic, D.P., Goh, V.S.L.: Complex Valued Nonlinear Adaptive Filters: Noncir-
cularity, Widely Linear and Neural Models. Wiley, Blackwell (2009)
11. Nitta, T.: A quaternary version of the back-propagation algorithm. In: Interna-
tional Conference on Neural Networks, No. 5, pp. 2753–2756. IEEE (1995)
12. Nitta, T., Buchholz, S.: On the decision boundaries of hyperbolic neurons. In:
International Joint Conference on Neural Networks (IJCNN), pp. 2974–2980. IEEE
(2008)
13. Pearson, J., Bisset, D.: Back propagation in a clifford algebra. In: International
Conference on Artificial Neural Networks, vol. 2, pp. 413–416 (1992)
14. Pearson, J., Bisset, D.: Neural networks in the clifford domain. In: International
Conference on Neural Networks, vol. 3, pp. 1465–1469. IEEE (1994)
44 C.-A. Popa

15. Popa, C.A.: Matrix-valued neural networks. In: Matoušek, R. (ed.) International
Conference on Soft Computing (MENDEL). Advances in Intelligent Systems and
Computing, vol. 378, pp. 245–255. Springer International Publishing, June 2015
16. Vallejo, J., Bayro-Corrochano, E.: Clifford hopfield neural networks. In: Interna-
tional Joint Conference on Neural Networks (IJCNN), pp. 3609–3612. IEEE, June
2008
17. Widrow, B., McCool, J., Ball, M.: The complex LMS algorithm. Proc. IEEE 63(4),
719–720 (1975)
 Enhanced Gradient Descent Algorithms
for Quaternion-Valued Neural Networks

Călin-Adrian Popa(B)

Department of Computer and Software Engineering,

Polytechnic University Timişoara,
Blvd. V. Pârvan, No. 2, 300223 Timişoara, Romania
[email protected]

Abstract. This paper proposes enhanced gradient descent learning

algorithms for quaternion-valued feedforward neural networks. The
quickprop, resilient backpropagation, delta-bar-delta, and SuperSAB
algorithms are the most known such enhanced algorithms for the real-
and complex-valued neural networks. They gave superior performances
than the gradient descent algorithm, so it is natural to extend these learn-
ing methods to quaternion-valued neural networks, also. The quaternion
variants of these four algorithms are presented, which are then used to
learn various time series prediction applications. Experimental results
show an important improvement in performance over the quaternion
gradient descent.

Keywords: Quaternion-valued neural networks · Quickprop · Resilient

backpropagation · Delta-bar-delta · SuperSAB · Time series prediction

1 Introduction

Quaternion-valued neural networks have received an increasing interest over the

last few years. These networks were successfully applied in chaotic time-series
prediction [2], color image compression [8], color night vision [11], polarized signal
classification [4], and 3D wind forecasting [10,20,21].
In the signal processing domain, in which some signals are expressed in 3D or
4D form, these networks appear as a natural choice for solving problems such as
time series prediction. Because of these practical applications, there is a natural
interest to increase the performance and speed of the training algorithms for this
type of networks. One obvious method for doing this is to extend the algorithms
that have proven their efficiency in the real- and complex-valued neural networks
domains to the quaternion-valued domain.
One of the simplest classes of algorithms that perform better than the clas-
sical gradient descent in the real and complex [13] domains is the so-called
enhanced gradient descent algorithms. By doing a series of ad hoc modifications
to the gradient descent algorithm, these methods have managed to increase its
performance. Out of the vast literature presenting such methods, we chose some

c Springer International Publishing AG 2018
V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9 5
46 C.-A. Popa

of the most known, robust and theoretically well-founded optimization methods

to extend to the quaternion domain.
The quickprop algorithm approximates the error function by a quadratic
polynomial and finds its minimum. The resilient backpropagation algorithm
replaces the partial derivatives that appear in the expression of the update rule
of the gradient descent with other weight update quantities, which are computed
taking into account only the signs of the partial derivatives. Lastly, delta-bar-
delta and SuperSAB, which differ only slightly, make the learning rate parameter
specific to each weight of the network, and to each training epoch, which means
that they also provide a method to update these specific learning rates.
In this paper, our purpose is to describe all these four training methods in
the case of quaternion-valued feedforward neural networks. Our presentation of
these enhanced gradient descent algorithms will follow that of [14], and of the
classical book [3], with the obvious adaptation to the quaternion domain.
The remainder of the paper is organized as follows: Sect. 2 presents the four
enhanced gradient descent algorithms for quaternion-valued neural networks:
quickprop, resilient backpropagation, delta-bar-delta and SuperSAB. The exper-
imental results of three time series prediction applications of the proposed algo-
rithms are shown in Sect. 3, together with an explanation of the nature of each
problem. Section 4 is dedicated to presenting the conclusions of the present study.

2 Enhanced Gradient Descent Algorithms

2.1 Quickprop
Let H = {q = [q]0 + i[q]1 + j[q]2 + k[q]3 |[q]0 , [q]1 , [q]2 , [q]3 ∈ R} be the algebra of
quaternions, where i, j, k are the imaginary units which satisfy i2 = j 2 = k 2 =
ijk = −1.
Assume we have a quaternion-valued feedforward neural network with L
layers and error function E. If we denote by wjk l
the weight connecting neuron
j from layer l to neuron k from layer l − 1, for all l ∈ {2, . . . , L}, we can define
the weight step at epoch t by
Δwjk
l
(t) = wjk
l
(t + 1) − wjk
l
(t).

Then, for the gradient descent algorithm, the update rule for the weight wjk l
is
given by
 
∂E ∂E ∂E ∂E
Δwjk
l
(t) = −ε (t) + i (t) + j (t) + k (t) ,
∂[wjk
l ]
0 ∂[wjk
l ]
1 ∂[wjk
l ]
2 ∂[wjk
l ]
3

where ε ∈ R is the learning rate.

The updating expression for the quickprop algorithm, which was first pro-
posed by Fahlman in [7], is
∂E
l ] (t)
∂[wjk a
l
[Δwjk (t)]a = l
[Δwjk (t − 1)]a , ∀0 ≤ a ≤ 3. (1)
∂E
l ] (t
∂[wjk a
− 1) − ∂E
l ] (t)
∂[wjk a
 Enhanced Gradient Descent Algorithms 47

The heuristic behind the algorithm is that it approximates the error function
with respect to each weight by a quadratic polynomial. This assumes that each
weight is independent of all the others and that the quadratic polynomial has a
minimum, and not a maximum.

2.2 Resilient Backpropagation

Riedmiller and Braun [15] devised the resilient backpropagation algorithm, or
RPROP for short, in order to eliminate the potential harmful influence of the size
of the partial derivative on the weight step. The algorithm replaces the partial
derivative from the weight update formula with another quantity denoted Δljk ,
which will then be used for updating the weight wjk l
. The update formulas for
computing the quantities Δjk are:
l

⎧
⎪
⎪η [Δjk (t − 1)]a ,
+ l
if ∂E
l ] (t − 1) × ∂E
l ] (t) >0
⎨ ∂[wjk a ∂[wjk a

[Δljk (t)]a = η − [Δljk (t − 1)]a , if ∂E

l ] (t − 1) × ∂E
l ] (t) < 0 , ∀0 ≤ a ≤ 3,
⎪
⎪
∂[wjk a ∂[wjk a
⎩ l
[Δjk (t − 1)]a else

which are calculated taking into account only the signs of the partial derivatives:
each time the partial derivatives ∂[w∂E
l ]
a
change their sign, which indicates that
jk
the last update was too big and the algorithm has jumped over a local minimum,
the quantities [Δljk ]a are decreased by the factor η − , and each time those partial
derivatives have the same sign, the quantities [Δljk ]a are increased by a factor of
η + , in order to accelerate the convergence. The real parameters η + and η − must
satisfy 0 < η − < 1 < η + , and are usually taken to be η − = 0.5 and η + = 1.2.
We can now write the formulas for updating the weight wjk l
based on the
quantities Δjk :
l

⎧
⎪
⎪−[Δljk (t)]a , if ∂E
l ] (t) >0
⎨ ∂[wjk a
l
[Δwjk (t)]a = [Δljk (t)]a , if ∂E
(t) < 0 , ∀0 ≤ a ≤ 3.
⎪
⎪
l
∂[wjk ]a
⎩
0, else

This updating rule is simple: if the derivative is positive, which means an

increasing error, the weight is decreased, and if the derivative is negative, the
weight is increased.

2.3 Delta-bar-delta

The delta-bar-delta algorithm was proposed by Jacobs in [9], who had the idea to
introduce a different learning rate for each weight in the network, and to devise a
procedure for updating these learning rates while learning. The gradient descent
rule becomes in this case
48 C.-A. Popa

∂E ∂E
Δwjk
l
(t) = − [εljk (t)]0 (t) + i[εljk (t)]1 (t)
∂[wjk
l ]
0 ∂[wjk
l ]
1

∂E ∂E
+ j[εljk (t)]2 (t) + k[εljk (t)]3 (t) .
∂[wjk
l ]
2 ∂[wjk
l ]
3

We can see that each learning rate εljk (t) depends on the weight wjk l
and on
the epoch t, and it is a quaternion, unlike in the quaternion gradient descent,
where the learning rate ε is a real number. The update rules for the learning
rates εljk (t) are:
⎧
⎪
⎪ κ + [εljk (t − 1)]a , if ∂[w∂ Ē (t − 1) × ∂[w∂E (t) > 0
⎨ l
jk ]a
l
jk ]a

[εljk (t)]a = η − [εljk (t − 1)]a , if ∂[w∂ Ē (t − 1) × ∂[w∂E (t) < 0 , ∀0 ≤ a ≤ 3,

⎪
⎪
l
jk ]a
l
jk ]a
⎩ l
[εjk (t − 1)]a else

where we denoted
∂ Ē ∂E ∂ Ē
(t) := (1 − θ) (t) + θ (t − 1), ∀0 ≤ a ≤ 3.
∂[wjk
l ]
a ∂[w l ]
jk a ∂[w l
jk ]a

The parameter η − must satisfy 0 < η − < 1, and is usually taken to be, like
in the RPROP algorithm, equal to 0.5. We also took κ = 10−6 and θ = 0.7 in
our experiments, but these values are usually determined empirically.
In this learning algorithm, if two consecutive partial derivatives ∂[w∂E
l ]
a
have
jk
the same sign, then the learning rate is increased by a small constant κ to accel-
erate learning. If the two partial derivatives have different signs, this indicates
that the local minimum was missed, which means that the weight step was too
big. Thus, the learning rate is decreased by multiplying it with a positive factor
η − . To ensure convergence, the previous partial derivative ∂[w∂El ] (t − 1) was
a jk
replaced by an exponentially weighted average of the current and past partial
derivatives with θ as the base and the epoch as the exponent. This weighted
average was denoted by ∂[w∂ Ē
l ] (t), and was defined above.
ajk

2.4 SuperSAB
The last algorithm that we present is called SuperSAB and was proposed by
Tollenaere in [16]. It is very similar to the delta-bar-delta algorithm, in that a
different learning rate is used for each weight in the network. The update rule is
the same as in the previous case:

∂E ∂E
Δwjk (t) = − [εljk (t)]0
l
(t) + i[εljk (t)]1 (t)
∂[wjk
l ]
0 ∂[w l
jk ]1

∂E ∂E
+ j[εjk (t)]2
l
(t) + k[εjk (t)]3
l
(t) ,
∂[wjk
l ]
2 ∂[wjk
l ]
3
 Enhanced Gradient Descent Algorithms 49

with the exception that here the learning rate εljk (t) corresponding to weight
wjk
l
and epoch t has a different update formula:
⎧
⎪
⎪η + [εljk (t − 1)]a , if ∂[w∂E (t − 1) × ∂E
l ] (t) >0
⎨ l
jk ]a ∂[wjk a

[εjk (t)]a = η − [εljk (t − 1)]a , if ∂[wl ] (t − 1) ×

l ∂E ∂E
l ] (t) < 0 , ∀0 ≤ a ≤ 3.
⎪
⎪ jk a ∂[wjk a
⎩ l
[εjk (t − 1)]a else

Thus, SuperSAB can be seen as a combination of the RPROP and delta-bar-

delta algorithms. Exactly like in the resilient backpropagation algorithm, the
parameters also have to satisfy 0 < η − < 1 < η + . Because of this resemblance,
we took in our experiments the same values for these parameters as above,
namely η − = 0.5 and η + = 1.2.
Instead of adding a term to the learning rate if the partial derivatives have
the same sign, like in the delta-bar-delta algorithm, we multiply it by the factor
η + , like in the RPROP algorithm. Also, the last two partial derivatives were
taken into account with their original values, unlike the exponentially weighted
average from the delta-bar-delta algorithm.

3 Experimental Results
3.1 Linear Autoregressive Process with Circular Noise

An important application of quaternion-valued neural networks is at signal pre-

diction. A known benchmark proposed in [12], and used for testing quaternion-
valued neural networks in [6,19,22], is the prediction of the quaternion-valued
circular white noise n(k) = na (k) + inb (k) + jnc (k) + knd (k), where na , nb ,
nc , nd ∼ N (0, 1), passed through the stable autoregressive filter given by
y(k) = 1.79y(k − 1) − 1.85y(k − 2) + 1.27y(k − 3) − 0.41y(k − 4) + n(k).
In our experiments, we trained quaternion-valued feedforward neural net-
works using the classical gradient descent algorithm (abbreviated GD), the
quickprop algorithm (QCP), the resilient backpropagation algorithm (RPR),
the delta-bar-delta algorithm (DBD), and the SuperSAB algorithm (SAB).
The tap input of the filter was 4, so the networks had 4 inputs, 4 hidden
neurons on a single hidden layer, and one output. The activation function for
the hidden layer was the fully quaternion hyperbolic tangent function, given by
q −q
G2 (q) = tanh q = eeq −e
+e−q , and the activation function for the output layer was
the identity G3 (q) = q. Training was done for 5000 epochs with 5000 training
samples.
To asses the effectiveness of the algorithms, we used a measure of performance
σ2
called prediction gain, and defined by Rp = 10 log10 σx2 , where σx2 represents the
e
variance of the input signal and σe2 represents the variance of the prediction
error. The prediction gain is given in dB, and, because of the way it is defined,
a bigger prediction gain means better performance.
50 C.-A. Popa

Table 1. Experimental results for linear autoregressive process with circular noise

Algorithm Prediction gain

GD 4.51
QCP 6.37
RPR 6.41
DBD 5.46
SAB 6.40

After running each algorithm 50 times, the prediction gains are given in
Table 1. We can see that QCP, SAB and RPR performed approximately the
same, followed by DBD, but all of them performed better than the gradient
descent algorithm.

3.2 3D Lorenz System

The 3D Lorenz system is given by the following ordinary differential equations:
dx
= α(y − x)
dt
dy
= −xz + ρx − y
dt
dz
= xy − βz,
dt
where α = 10, ρ = 28, and β = 2/3. This represents a chaotic time series
prediction problem, and was used to test quaternion-valued neural networks in
[2,17,22].
The tap input of the filter was 4, and so the networks had 4 inputs, 4 hidden
neurons, and one output neuron. The networks were trained for 5000 epochs
with 1337 training samples.
The averaged results after 50 runs of each algorithm are given in Table 2. The
measure of performance was the prediction gain, like in the above experiment.
In this case, QCP and RPR performed best, SAB followed, and DBD was again
last of the four, but still better than GD.

Table 2. Experimental results for the 3D Lorenz system

Algorithm Prediction gain

GD 7.56
QCP 10.59
RPR 11.07
DBD 9.35
SAB 10.33
 Enhanced Gradient Descent Algorithms 51

3.3 4D Saito Chaotic Circuit

The last experiment involves the 4D Saito chaotic circuit given by

 dx1  
dt −1 1 x1 − ηρ1 h(z)
=
−α1 α1 β1 y1 − η β11 h(z)
dy1 ρ
dt
 dx2  
−1 1 x2 − ηρ2 h(z)
dt = ,
−α2 α2 β2 y2 − η β2 h(z)
dy2 ρ2
dt

1, z ≥ −1
where h(z) = is the normalized hysteresis value, and z = x1 +x2 ,
−1, z≤1
ρ1 = 1−β β1
1
, ρ2 = 1−β
β2
2
. The values of the parameters are (α1 , β1 , α2 , β2 , η) =
(7.5, 0.16, 15, 0.097, 1.3). This is also a chaotic time series prediction problem,
and was used as a benchmark for quaternion-valued neural networks in [1,2,5,
6,18,19].
Like in the above experiments, the tap input of the filter was 4, and so the
networks had the same architectures as the ones described earlier, and were
trained for 5000 epochs with 5249 training samples.
The average prediction gains after 50 runs of each algorithm are given in
Table 3. In this last experiment, the performances were similar to those in the
previous experiments: QCP, SAB, and RPR had approximately the same per-
formance, followed by DBD, and finally by GD.

Table 3. Experimental results for the 4D Saito chaotic circuit

Algorithm Prediction gain

GD 5.76
QCP 11.49
RPR 11.58
DBD 6.28
SAB 11.55

4 Conclusions
The description of the most known enhancements of the classical gradient descent
algorithm for training quaternion-valued feedforward neural networks was given.
The four variations of the gradient descent method, which had different update
rules and different methods for computing the learning rates, were applied for
training networks used to solve three well-known time series applications.
The experimental results shown that all the enhanced gradient descent algo-
rithms performed better on the proposed problems than the classical gradient
52 C.-A. Popa

descent algorithm, in some cases as much as 6 dB better in terms of prediction

gain.
We can conclude that the proposed enhanced gradient descent algorithms
represent a more efficient method of training feedforward quaternion-valued
neural networks than the classical gradient descent.

References
1. Arena, P., Fortuna, L., Muscato, G., Xibilia, M.: Multilayer perceptrons to approx-
imate quaternion valued functions. Neural Netw. 10(2), 335–342 (1997)
2. Arena, P., Fortuna, L., Muscato, G., Xibilia, M.: Neural Networks in Multidimen-
sional Domains Fundamentals and New Trends in Modelling and Control. Lecture
Notes in Control and Information Sciences, vol. 234. Springer, London (1998)
3. Bishop, C.: Neural Networks for Pattern Recognition. Oxford University Press,
Inc., New York (1995)
4. Buchholz, S., Le Bihan, N.: Polarized signal classification by complex and quater-
nionic multi-layer perceptrons. Int. J. Neural Syst. 18(2), 75–85 (2008)
5. Che Ujang, B., Took, C., Mandic, D.: Split quaternion nonlinear adaptive filtering.
Neural Netw. 23(3), 426–434 (2010)
6. Che Ujang, B., Took, C., Mandic, D.: Quaternion-valued nonlinear adaptive filter-
ing. IEEE Trans. Neural Netw. 22(8), 1193–1206 (2011)
7. Fahlman, S.: An empirical study of learning speed in backpropagation net-
works. Technical report 1800, Carnegie Mellon University, January 1988. http://
repository.cmu.edu/compsci/1800
8. Isokawa, T., Kusakabe, T., Matsui, N., Peper, F.: Quaternion neural network and
its application. In: Palade, V., Howlett, R., Jai, L. (eds.) Knowledge-Based Intel-
ligent Information and Engineering Systems. Lecture Notes in Computer Science,
vol. 2774, pp. 318–324. Springer, Heidelberg (2003)
9. Jacobs, R.: Increased rates of convergence through learning rate adaptation. Neural
Netw. 1(4), 295–307 (1988)
10. Jahanchahi, C., Took, C., Mandic, D.: On HR calculus, quaternion valued stochas-
tic gradient, and adaptive three dimensional wind forecasting. In: International
Joint Conference on Neural Networks (IJCNN), pp. 1–5. IEEE, July 2010
11. Kusamichi, H., Isokawa, T., Matsui, N., Ogawa, Y., Maeda, K.: A new scheme for
color night vision by quaternion neural network. In: International Conference on
Autonomous Robots and Agents, pp. 101–106, December 2004
12. Mandic, D., Chambers, J.: Recurrent Neural Networks for Prediction: Learning
Algorithms, Architectures and Stability. Wiley, New York (2001)
13. Popa, C.A.: Enhanced gradient descent algorithms for complex-valued neural net-
works. In: International Symposium on Symbolic and Numeric Algorithms for Sci-
entific Computing (SYNASC), pp. 272–279. IEEE, September 2014
14. Riedmiller, M.: Advanced supervised learning in multi-layer perceptrons - from
backpropagation to adaptive learning algorithms. Comput. Stand. Interfaces 16(3),
265–278 (1994)
15. Riedmiller, M., Braun, H.: A direct adaptive method for faster backpropagation
learning: the RPROP algorithm. In: IEEE International Conference on Neural
Networks, vol. 1, pp. 586–591. IEEE, March 1993
16. Tollenaere, T.: Supersab: fast adaptive back propagation with good scaling prop-
erties. Neural Netw. 3(5), 561–573 (1990)
 Enhanced Gradient Descent Algorithms 53

17. Took, C., Mandic, D.: The quaternion LMS algorithm for adaptive filtering of
hypercomplex processes. IEEE Trans. Signal Process. 57(4), 1316–1327 (2009)
18. Took, C., Mandic, D.: Quaternion-valued stochastic gradient-based adaptive IIR
filtering. IEEE Trans. Signal Process. 58(7), 3895–3901 (2010)
19. Took, C., Mandic, D.: A quaternion widely linear adaptive filter. IEEE Trans.
Signal Process. 58(8), 4427–4431 (2010)
20. Took, C., Mandic, D., Aihara, K.: Quaternion-valued short term forecasting of wind
profile. In: International Joint Conference on Neural Networks (IJCNN), pp. 1–6.
IEEE, July 2010
21. Took, C., Strbac, G., Aihara, K., Mandic, D.: Quaternion-valued short-term joint
forecasting of three-dimensional wind and atmospheric parameters. Renew. Energy
36(6), 1754–1760 (2011)
22. Xia, Y., Jahanchahi, C., Mandic, D.: Quaternion-valued echo state networks. IEEE
Trans. Neural Netw. Learn. Syst. 26(4), 663–673 (2015)
 A Reinforcement Learning Based Approach
to Multiple Sequence Alignment

Ioan-Gabriel Mircea(B) , Iuliana Bocicor(B) , and Gabriela Czibula(B)

Faculty of Mathematics and Computer Science, Babeş-Bolyai University,

1, M. Kogălniceanu Street, 400084 Cluj-Napoca, Romania
{mircea,iuliana,gabis}@cs.ubbcluj.ro

Abstract. Multiple sequence alignment plays an important role in com-

parative genomic sequence analysis, being one of the most challenging
problems in bioinformatics. This problem refers to the process of arrang-
ing the primary sequences of DNA, RNA or protein to identify regions
of similarity that may be a consequence of functional, structural or evo-
lutionary relationships between the sequences. In this paper we tackle
multiple sequence alignment from a computational perspective and we
introduce a novel approach, based on reinforcement learning, for address-
ing it. The experimental evaluation is performed on several DNA data
sets, two of which contain human DNA sequences. The efficiency of our
algorithm is shown by the obtained results, which prove that our tech-
nique outperforms other methods existing in the literature and which
also indicate the potential of our proposal.

Keywords: Bioinformatics · Multiple Sequence Alignment · Machine

Learning · Reinforcement Learning

1 Introduction
Sequence alignment [9] plays an important role in molecular sequence analy-
sis, being a key element in understanding the evolution and functioning of liv-
ing organisms. As the genomes of an increasing number of organisms are being
sequenced nowadays, there is a constant need to study and investigate the evo-
lutionary, structural and functional relationships between these genomes. The
aim is to find common patterns between sequences of DNA, RNA or proteins
so as to construct an alignment of two or more sequences. When two sequences
are involved, it is called pairwise alignment, when three or more sequences are
involved, it is called multiple sequence alignment. Since aligning each pair of
sequences from a set of several does not reveal the common information of all the
sequences, multiple sequence alignment is often necessary to find the sequences
from the set that share an evolutionary lineage, are descendant from a common
ancestor or that maintain certain domains or structure.
In this paper we approach, from a computational perspective, the problem
of multiple sequence alignment (MSA). We introduce a novel approach for solv-
ing the MSA problem, based on reinforcement learning (RL) - an approach to

c Springer International Publishing AG 2018
V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9 6
A Reinforcement Learning Based Approach to Multiple Sequence Alignment 55

machine intelligence in which an agent can learn an optimal sequence of actions

which lead to a goal, by receiving rewards or reinforcements on its chosen actions
[23]. We have previously conducted in [16] some preliminary experiments using
the RL model for MSA, with the purpose of determining the best setting for
our method. The best alignment algorithm and the best action selection policy
which were decided in [16] will be used in this paper for introducing the RL based
method for MSA. To our knowledge, the MSA problem for DNA sequences has
not been addressed in the literature using reinforcement learning, so far.
The rest of the paper is organized as follows. Section 2 presents the main
aspects related to the topics approached in this paper, the MSA problem and
reinforcement learning. The RL model that we propose is introduced in Sect. 3.
Section 4 provides several experiments, using several real life data sets containing
DNA sequences. Comparisons between our approach and related work from the
literature are given in Sect. 5. Section 6 outlines some conclusions of the paper
and indicates future research directions.

2 Background
We begin by presenting the main aspects of the MSA problem and its relevance
in bioinformatics. Then, we continue to briefly describe some of the existing
techniques in the literature that approach this problem.

2.1 Problem Relevance

Multiple sequence alignment plays a very important role in comparative genomic

sequence analysis, being one of the most challenging problems in bioinformatics.
Sequence alignment refers to the process of arranging the primary sequences of
DNA, RNA or protein to identify regions of similarity that may be a consequence
of functional, structural or evolutionary relationships between the sequences.
The aim is to find common patterns, to construct an alignment of two or more
sequences so as to maximize the similarities between them. Pairwise alignment
only involves two sequences. Still, in practice, situations in which three or more
sequences need to be aligned often occur. In such cases, aligning only pairs of
sequences is not sufficient, as this does not reveal common information for all
the sequences in the data set. Therefore, MSA is necessary in order to find
various significant information about the evolutionary, functional or structural
relationships that might exist between several sequences.
MSA is most often needed for sequences that are assumed to be connected
by evolutionary relationships, that may have descended from common ances-
tors. From the result of such an alignment, multiple pieces of information can
be inferred: the sequences’ evolutionary origins, mutations that happened in the
DNA or in proteins, identification of conserved sequence motifs or secondary
structures. In addition, MSA is one of the first necessary steps required in con-
structing phylogenetic trees or determining sequences’ homologies.
56 I.-G. Mircea et al.

A direct method for producing an MSA uses dynamic programming to iden-

tify the globally optimal alignment solution, one who has a maximum alignment
score. One of the most popular cost measures to evaluate the quality of an
alignment is the sum of pairs SP score method [15]. This score sums the sub-
stitution scores of all possible pairwise combinations of sequence characters over
all columns of a multiple sequence alignment:
N n−1
   n
SP = s(cji , cki ) (1)
i=1 j=1 k=j+1

where:

– n is the number of sequences in the alignment;

– N is the number of columns in the alignment;
– ci , i ∈ {1, . . . , N } represents column i from the alignment;
– cji , j ∈ {1, . . . , n} is the j th character from the ith column;
– s(cji , cki ) is a measure indicating the comparing score between characters cji
and cki .

2.2 Literature Review

A direct method for producing an MSA uses the dynamic programming tech-
nique [19] to identify the globally optimal alignment solution. Still, this method
is computationally very expensive, the running times are growing exponentially
with the size of the problem (the MSA problem is NP-complete [25]). For this rea-
son various heuristics have been proposed in the literature for solving this prob-
lem and most of the practical alignment algorithms are based on such heuristics
that produce quasi-optimal alignments.
Chen and Lin [4] tackle the DNA MSA problem using genetic algorithms in
conjunction with divide-and-conquer techniques to determine the best cut points
for DNA sequences. They tested their approach using several DNA sequence sets
and showed that it outperforms other methods existing in the literature. The
same authors offered a different approach to the same problem, based on genetic
simulated annealing techniques to choose the best cutting point set from a DNA
sequences data set [5].
Nguyen et al. [20] introduce a parallel hybrid genetic algorithm for the protein
MSA problem. They propose a new chromosome representation and correspond-
ing new genetic operators and they use two kinds of local search heuristics. In
addition to these, their method is extended to parallel processing.
Aside from the GA approaches, other solutions to the MSA problem have
been proposed, which use methods from the field of machine learning. A parti-
tioning approach combined with an ant colony optimization (ACO) method was
used by Chen et al. [6] to solve the classical MSA problem. The experimental
results of this study show that the algorithm can significantly reduce the running
time and improve the solution quality on large-scale multiple sequence alignment
benchmarks.
A Reinforcement Learning Based Approach to Multiple Sequence Alignment 57

Rasmussen and Krink [22] propose to solve the MSA problem using hidden
Markov models (HMMs), with a new particle swarm optimization (PSO) based
training method. Nasser et al. [17] used fuzzy logic for approximate matching
of sequences, because fuzzy logic allows tolerance of inexactness or errors in
subsequence matching. They develop a prototype for a fuzzy assembler, that
is designed to work with low quality data, an advantage that most existing
techniques do not possess. The assembler was tested on sequences from two
genome projects and the obtained results demonstrate that it can generate opti-
mal assemblies.
A genetic algorithm based approach to solving the multiple sequence align-
ment problem was proposed by Agarwal in [1]. It uses custom genetic algorithm
operators such as the Gap shuffle mutation in order to produce offsprings of
improved fitness that may describe a better alignment with respect to a “sum-of-
pair” based scoring frame using PAM and BLOSUM scoring matrices. Further-
more, the genetic algorithm approach has been used to enhance the performance
of an ant colony model presented in [26], trained to find a better alignment by
searching the solution inside a planar graph designed to model possible multiple
sequence alignment configurations.
Adapting the crossover and mutation probabilities according to the specific
characteristics of the population, the immune adaptive genetic algorithm pro-
posed by Chao and Shuai in [3] manages to improve the performance of a classical
genetic algorithm approach by enhancing individual diversity when population
variety decreases and favoring the convergence towards a global optimum.
A great amount of research has also been conducted on finding which of the
existing state of the art multiple sequence approaches performs best in practice
[24] and also which protein or DNA benchmarks provide the most reliable sources
of genetic material to be used in research. Caroll et al. propose in [2] a benchmark
for DNA sequences generated on the tertiary structure of the proteins they
encode. The proposed benchmark is successfully used in DNA alignment in order
to test a genetic algorithm enhanced with self-organizing crossover and selection
operations [21].

3 A RL Model for Multiple Sequence Alignment

In this section we introduce the reinforcement learning model proposal for solving
the MSA problem. First, we discuss about sequence to profile alignment, since
it will be further used in our RL based approach.

3.1 Sequence to Profile Alignment

Even thought the pairwise alignment of two sequences is a rather straightforward

task that may be easily achieved by use of a dynamic programming algorithm,
the situation tends to become complicated if three or more sequences need to be
aligned globally with one another.
58 I.-G. Mircea et al.

The main difficulty consists in aligning several pre-aligned sequences with a

given sequence. When a sequence is to be aligned with a set of already aligned
sequences, a symbol from the sequence is matched against a whole column of
corresponding symbols from all the pre-aligned sequences.
In order to surmount this impasse, the alignment is no longer conducted
between a sequence and a set of pre-aligned sequences but between the sequence
to be aligned and a profile characterizing the set.
The profile is computed in the following manner: for each position in the
set of pre-aligned sequences a list is maintained holding information about the
frequency of the appearance of each symbol from the alphabet {A, C, G, T, },
where A, C, G, T represent the 4 nucleotides and represents a gap.
The algorithm we propose for finding the optimal alignment and its corre-
sponding score extends the Needleman-Wunsch algorithm [18] for detecting the
optimal alignment of two sequences.

3.2 The Proposed RL Approach for the MSA Problem

Let us consider, in the following, that Seq is the input data set, consisting of n
(n > 1) multi-dimensional DNA sequences: Seq = (Seq1 , Seq2 , . . . , Seqn ), each
sequence Seqi being a sequence of nucleotides, Seqi = (Seqi1 , Seqi2 , . . . , Seqili ),
Seqij ∈ {A, T, G, C} is a nucleotide and li represents the length of sequence Seqi ,
i.e. the number of nucleotides in the sequence.
Let us denote by AlignSeq the matrix representing the optimal alignment of
sequences from the list Seq (considered in the order Seq1 , Seq2 , . . . , Seqn ). Such
an alignment means that the sequences are progressively aligned as follows: Seq1
is aligned with Seq2 , then the resulted optimal alignment is aligned with Seq3 .
The next step is to align the optimal alignment obtained so far with Seq4 and
the iterative process is continued until all the sequences Seq1 , Seq2 , . . . , Seqn are
optimally aligned.
As indicated above, the MSA problem consists of determining the alignment
of the sequences from Seq which has the maximum associated score. From a
computational point of view, the M SA problem can be viewed as the problem
of generating a permutation σ = (σ1 , σ2 , . . . σn ) of {1, 2, . . . , n} that maximizes
the score Score(Seqσ ) of the alignment of sequences considered in the order σ:
Seqσ = (Seqσ1 , Seqσ2 , . . . , Seqσn ). The score of the alignment Seqσ represents
the score of the corresponding matrix AlignSeqσ and it has to be maximized.
We define the RL task associated to the M SA problem as follows.
n+1
The state space S (the M SA agent’s environment) will consist of n n−1−1
states, i.e. S = {s1 , s2 , ..., s nn+1 −1 }. The initial state of the agent in the envi-
n−1
n+1
ronment is s1 . A state sik ∈ S (ik ∈ {1, . . . , n n−1−1 }) reached by the agent
at a given moment after it has visited states si1 , si2 , si3 , ...sik−1 (here si1 is the
initial state s1 ) and has selected actions ai1 , ai2 , ...aik−1 is a terminal (final or
goal) state if the number of states visited by the agent in the current sequence
is n + 1 (i.e. k = n + 1) and all the selected actions (ai1 , ai2 , ...ain represent a
 A Reinforcement Learning Based Approach to Multiple Sequence Alignment 59

permutation of {1, 2, . . . , n} (the selected actions are distinct), i.e. aij = aik ,
∀ j, k ∈ {1, . . . , n} j = k.
The action space A consists of n actions available to the problem solving
agent and corresponding to the n possible values 1, 2, . . . , n used to represent a
solution (permutation of {1, 2, . . . , n}), i.e. A = {a1 , a2 , . . . , an }, where ai = i,
∀ i ∈ {1, . . . , n}. From a state s ∈ S the agent can move in n successor states,
by executing one of the n possible actions.
A graphical representation of the states space and the transitions between
the states are given in Fig. 1. The circles represent states and the transitions
between states are indicated by arrows labeled with the action that leads the
agent from a state to another.

Fig. 1. Environment for the MSA agent

The reward function will be defined in the following (Formula (2)).

Let us consider a path π in the above defined environment from the initial
to a final state, π = (π0 π1 π2 · · · πn ), where π0 = s1 and ∀ k ∈ {0, . . . , n − 1} the
state πk+1 is a neighbor of state πk , i.e. πk+1 ∈ δ(πk ).
The sequence of actions obtained following the transitions between the suc-
cessive states from the path π will be denoted by aπ = (aπ0 aπ1 aπ2 · · · aπn−1 ),
where πk+1 = Δ(πk , aπk ), ∀ k ∈ {0, . . . , n − 1}. The sequence aπ will be referred
to as the list of actions associated to the path π.
The list of actions associated to a path π gives us a list of sequences
Seqπ = (Seqaπ0 , Seqaπ1 , . . . , Seqaπn−1 ). If all the actions from the list of actions
aπ associated to a path π are distinct (i.e. aπi = aπj ∀0 ≤ i < j ≤ n−1), then aπ
is a permutation of {1, 2 . . . n} and can be viewed as a possible order for obtain-
ing an alignment of the input sequences Seq (i.e. the alignment considered in the
order Seqaπ0 , Seqaπ1 , . . . , Seqaπn−1 ). Consequently, the value Score(Seqaπ ) cor-
responding to a path π whose associated list of actions contains distinct actions
would represent the optimal score of aligning the list of input sequences in the
60 I.-G. Mircea et al.

order given by aπ , i.e. in the order Seqaπ0 , Seqaπ1 , . . . , Seqaπn−1 . Let us denote
by M (k,r)=(mi,j )1≤i≤k,1≤j≤r (1 < k ≤ n) the matrix representing the optimal
alignment of the list of sequences (Seqaπ0 , Seqaπ1 , . . . , Seqaπk−1 ) (k denotes the
number of rows and r the number of columns of the alignment).
The M SA problem formulated as a RL task will consist of training the
agent to find a path π from the initial to a final state having the maximum score
Score(Seqaπ ) of the corresponding alignment. After the reinforcement learning
process, the agent will learn to execute those transitions (actions) that maximize
the sum of rewards received on the path from the initial to a final state.
Since we aim at obtaining a path π having the maximum score computed
by progressively aligning the sequences in the order corresponding to the asso-
ciated list of actions aπ , we define the reinforcement function as indicated in
Formula (2).
⎧
⎪
⎪ 0 if k = 1
⎪
⎨ −∞ aπk−1 ∈ {aπ0 , aπ1 , . . . , aπk−2 }
r(πk |πk−1 , ...π0 ) =  r k−1

⎪
⎪
⎪
⎩ s(mk,j , ml,j ) otherwise
j=1 l=1
(2)
In Formula (2) r(πk |πk−1 , ...π0 ) denotes the reward received by the agent in
state πk , after its history in the environment is π = (π0 = s1 , π1 , π2 , . . . , πk−1 )
and s(x, y) represents the match/mismatch score between two characters x and y.
The agent receives a negative reward on paths that are not valid (the associ-
ated list of actions contains actions that are not distinct), therefore it will learn
to explore only valid paths. Considering the reward defined in Formula (2), as
the learning goal is to maximize the total amount of rewards received on a path
from the initial to a final state, the agent is trained to find a valid path π that
maximizes the score of the alignment that is indicated by its corresponding list
of actions aπ .
Considering the way the reward was defined, it is very likely that after the
training of the M SA agent was completed, it will learn the optimal alignment of
the sequences Seq1 , Seq2 , . . . , Seqn . Once the alignment is obtained, the optimal
score of the alignment is computed.
During the training step of the learning process, the agent will determine its
optimal policy in the environment, i.e. the mapping from states to actions that
maximizes the sum of the received rewards.
For training the MSA agent, we propose a Q-learning approach, in which the
agent learns an action value function, denoted by Q, function giving the expected
utility of taking a given action in a given state [23]. The idea of the training
process is the following. After the Q values are initialized, during some training
episodes, the agent will experiment (using an action selection mechanism) some
(possible optimal) valid paths from the initial to a final state, updating the Q-
values estimations according to the Q-learning algorithm [8]. At the end of the
training process, the Q-values estimations will be in the vicinity of the exact
values.
A Reinforcement Learning Based Approach to Multiple Sequence Alignment 61

The action selection mechanism we have used is the -Greedy mechanism

extended with a look-ahead step in order to guide the exploration of the search
space. The look-ahead step consists in selecting, with a certain probability, the
action a that maximizes the score of aligning the sequence Seqa with the profile
corresponding to the alignment associated to the list of actions of the current
path.

4 Computational Experiments
This section presents experimental evaluations of our RL based model for the
MSA problem. For the computational experiments developed in order to test our
approach, we used several data sets, consisting of real DNA sequences. These
data sets have previously been used in the literature to test other approaches
to the MSA problem, which gives us the opportunity to compare our algorithm
with related work. The information about all the public data sets is synthesized
in Table 1. All the experiments presented in this section were carried out on a
PC with an Intel Core i3 Processor at 2.4 GHz, with 3 GB of RAM.

Table 1. Detailed information of the data sets used for our experiments.

Data set No. of Average Max. seq. Min. seq.

sequences length length length
Hepatitis C virus 10 211.9 212 211
Papio anubis 5 1093 1098 1088
469 from oxbench mdsa all [2] 3 332 339 321
429 from oxbench mdsa all [2] 12 171 183 153
Data set 1 from [26] 3 39.33 40 39
Lemur, gorilla, mouse data set 3 93 94 92
from [26]
Rat, lemur, opossum data set 3 129 129 129
from [26]
Mitochondrial human DNA - 8 365.75 372 363
first data set
Mitochondrial human DNA - sec- 6 362.83 366 360
ond data set

4.1 Parameters Setting and Implementation Details

In our experiments, for applying the RL based approach introduced in Sect. 3, we
have used a software framework that we have previously introduced for solving
combinatorial optimization problems using reinforcement learning techniques [7].
The MSA agent was trained using the following parameter setting: the dis-
count factor for the future rewards is γ = 0.9; the learning rate is α = 0.8;
62 I.-G. Mircea et al.

the number of training episodes is 104 ; the action selection mechanism was used
with  = 0.8. We are considering the following values for the scoring matrix: −2
for the gap penalty, −1 for the mismatch penalty and +2 for the match score.
We have chosen these values mainly because they have been used in other works
from the literature and this allows us to compare the results of our algorithm
with other results reported so far. The resulted alignments are evaluated by three
measures: the sum of pairs score [15], the number of perfectly matched columns,
or exact match (EM) and the column score (CS), computed as the ratio between
the number of columns that are exact matches and the total alignment length
(AL). All these evaluation measures have to be maximized in order to obtain
better alignments.

4.2 Hepatitis C Virus DNA Sequences

The first real life DNA sequence data set that we used contains 10 sequences
belonging to the organism Hepatitis C virus (subtype 1b). These sequences were
taken from the EMBL database [12], they have lengths of either 211 or 212
nucleotides and they can be retrieved by the following European Nucleotide
Archive (ENA) accession numbers: Z75841, Z75850, Z75854, Z75852, Z75858,
Z75851, Z75853, Z75859, Z75860, Z75861.
We ran our RL algorithm for this data set, using the parameter setting
described in Sect. 4.1. The algorithm converges towards the solution rapidly,
in less than 2 s in this case. The obtained alignment has a score of 18627 and
the order of sequences is the one in which they were specified in the previ-
ous paragraph. The number of matched columns is 198 and considering that
the sequences’ length is approximately 212, we may state that this number is
quite high. Table 2 shows the score, number of matched columns, the number
of epochs and the time needed for convergence. We mention that the last two
metrics (number of epochs and time) are averaged over 5 runs of the algorithm.

Table 2. Results for the Hepatitis C virus (subtype 1b), the Papio anubis data sets
[12], for data sets 469 and 429 from oxbench mdsa all [2] and for three different species
data sets [26] and for the two mitochondrial human DNA data sets, obtained by our
RL based algorithm. In this table the following abbreviations are used: AL - alignment
length, EM - exact matches (number of perfectly matched columns) and CS - column
score (CS = EM AL
).

Data set Score AL EM CS No. of epochs Time (sec.)

Hepatitis C virus 18627 212 198 0.93 <100 <2
Papio anubis 18719 1110 918 0.82 160 190
469 from oxbench mdsa all 565 378 171 0.45 <100 <1
429 from oxbench mdsa all 8668 205 43 0.21 2950 186
Data set 1 from [26] 167 42 31 0.73 <100 <1
Lemur, gorilla, mouse data set from [26] 345 99 66 0.66 <100 <1
Rat, lemur, opossum data set from [26] 486 133 87 0.65 <200 <1
A Reinforcement Learning Based Approach to Multiple Sequence Alignment 63

4.3 Papio Anubis (Olive Baboon) DNA Sequences

We tested our algorithm on a second publicly available data set, containing 5
DNA sequences belonging to the organism Papio anubis (olive baboon). Similar
to the prior data set, the sequences were taken from the EMBL database [12],
using the following ENA accession numbers: U35624, U35625, U35626, U35627,
U35628. The average length of these sequences is 1093. We observe that in this
case the number of sequences is smaller, but their lengths are approximately
5 times greater than in the previous case.
We used the parameter setting described in Sect. 4.1 to run our algorithm.
Our RL algorithm obtained an alignment with a score of 18719 and the permu-
tation of sequences is 1, 3, 2, 5, 4, if the sequences are considered to be numbered
in the order they were specified above. As in the previous case, the number of
matched columns, 918 is high, considering the average length of the sequences
(1093). The number of epochs and the computational time until convergence
to the optimum solution are averaged over 5 runs of the algorithm and they
are shown in Table 2, together with the sum of pairs score and the number of
matched columns.

4.4 Data Set 469 from oxbench mdsa all

For the third experiment, we used the data set 469 from oxbench mdsa all [2],
which contains three DNA sequences. The average length of the sequences is 332.
Using the same parameter setting that we described in Sect. 4.1, our algorithm
obtained an alignment with a score of 565 and the permutation of sequences
0, 2, 1, considering the sequences in the order they are presented in the file 469.afa
from the above mentioned data set. The resulted alignment contains 378 columns
and the number of perfectly matched columns is 171. Just as for the other data
sets, the results are also shown in Table 2, averaged over 5 runs of the algorithm.

4.5 Data Set 429 from oxbench mdsa all

For the fourth experiment, we used another data set from the oxbench mdsa all
[2] database: data set 429, which contains twelve DNA sequences. The aver-
age length of the sequences is 171. Using the same parameter setting that we
described in Sect. 4.1, our algorithm obtained an alignment with a score of
8668. The resulted alignment contains 205 columns and the number of perfectly
matched columns is 43. The results are shown in Table 2.

4.6 DNA Belonging to Three Different Species

The fifth experiment focuses on three data sets that were used for evaluating an
ant colony with genetic algorithm based on planar graph [26]. Each one contains
three sequences: the first data set contains shorter sequences, while the last two
have sequence segments belonging to three kind of species. To be able to compare
the results obtained by our RL based algorithm to the ones obtained by the ant
64 I.-G. Mircea et al.

colony genetic algorithm [26], we used as values for the scoring matrix the same
values used in the previously mentioned work. These values are identical to the
ones presented in Sect. 4.1, with only one observation: the alignment score of two
gaps is 0.

First Data Set. The first data set was used in Experiment 1 of [26]. For
the exact DNA sequences, we invite the reader to consult the paper [26]. The
average length of the sequences is 39. The score of the alignment obtained by
our algorithm is 167, the alignment length is 42 and the number of matched
columns is 31. The optimum permutation is 0, 1, 2, considering the sequences in
the order they are given in [26].

Second Data Set: Lemur, Gorilla, Mouse DNA. The second data set was
used in Experiment 2 of [26] and it contains DNA sequences from three species:
lemur, gorilla and mouse. For the exact DNA sequences, we invite the reader
to consult the paper [26]. The average length of the sequences is 93. The score
of the alignment obtained by our algorithm is 345, the alignment length is 99
and the number of matched columns is 66. The optimum permutation is 0, 1, 2,
considering the sequences in the order they are given in [26].

Third Data Set: Rat, Lemur, Opossum DNA. The third data set was used
in Experiment 3 of [26] and it contains the third exon sequences of the beta-
globin gene from three species: rat, lemur and opossum. For the exact DNA
sequences, we invite the reader to consult the paper [26]. The average length of
the sequences is 129. The score of the alignment obtained by our algorithm is
486, the alignment length is 133 and the number of matched columns is 87. The
optimum permutation is 0, 1, 2, considering the sequences in the order they are
given in [26].

5 Comparison to Related Work

Most works in the literature that approach the MSA problem use protein data
sets for evaluation. This is most probably due to the fact that there are consid-
erably more multiple protein sequence alignment benchmarks (i.e. BAliBASE,
OXBench, PREFAB and SMART) than DNA benchmarks, making the evalua-
tion of MSA programs difficult for DNA sequences [2].
For our comparisons, we selected four papers presenting MSA approaches for
the DNA sequence alignment [4,5,21,26]. The experimental evaluation of these
methods are made on six of the data sets we used in Sect. 4: DNA sequences
from the organisms hepatitis C virus and papio anubis, the data set 469 from
oxbench mdsa all [2] and three other data sets with DNA belonging to different
species.
In the following, we will discuss the results obtained by our RL based algo-
rithm in comparison with other results reported in the literature, for the same
data sets.
A Reinforcement Learning Based Approach to Multiple Sequence Alignment 65

Table 3. Comparative results for the Hepatitis C virus (subtype 1b) and the Papio
anubis data sets [12].

Data set Our RL approach GA + Genetic

Divide and Simulated
Conquer [4] Annealing [5]
Score Exact match Score Exact match Score Exact match
Hepatitis C virus 18627 198 17768 193 18287 194
Papio anubis 18719 918 17912 908 18030 914

Table 4. Comparative results for the data sets containing DNA belonging to three
different species. In this table the following abbreviations are used: AL - alignment
length, EM - exact matches (number of perfectly matched columns) and CS - column
score (CS = EM
AL
).

Data set Our RL approach Ant Colony with GA [26]

Score AL EM CS Score AL EM CS
Data set 1 from [26] 167 42 31 0.73 167 42 31 0.73
Lemur, gorilla, mouse data set from [26] 345 99 66 0.66 348 96 62 0.64
Rat, lemur, opossum data set from [26] 486 133 87 0.65 471 131 84 0.64

First the alignments obtained by our method were compared to the align-
ments obtained by two other methods in the literature: [4,5], for the two data
sets containing DNA sequences from the organisms hepatitis C virus and papio
anubis. We note that our RL based algorithm proved to outperform the other
approaches for the considered data sets. Table 3 shows the results obtained by our
RL algorithm, in comparison with other results obtained by Chen et al. [4,5]. We
notice that the RL approach proves to be more efficient, as the sum of pairs score
of the alignment, as well as the number of matched columns have greater val-
ues that the ones reported by the approaches based on genetic algorithms (GA):
the alignment method based on genetic algorithms and divide-and-conquer tech-
niques [4] and the method based on genetic simulated annealing techniques [5].
We remark that the results we use for comparison from [5] are the best results
reported in this paper, for the specified data set.
Another approach for the MSA problem, using an ant colony with genetic
algorithm based on planar graph is presented in [26]. The authors experiment

Table 5. Comparative results for the data set 469 from oxbench mdsa all [2]. In this
table the following abbreviations are used: AL - alignment length, EM - exact matches
(number of perfectly matched columns) and CS - column score (CS = EM AL
).

Data set Our RL approach Self-Organizing GA [21]

AL EM CS AL EM CS
Data set 469 from oxbench mdsa all [2] 378 171 0.45 406 45 0.11
66 I.-G. Mircea et al.

Table 6. Comparative results for all the data sets: our RL algorithm is compared to
ClustalW [14] and to MAFFT [13]. In this table the following abbreviations are used:
AL - alignment length, EM - exact matches (number of perfectly matched columns)
and CS - column score (CS = EM
AL
).

Data set Our RL approach ClustalW [14] MAFFT [13]

Score AL EM CS Score AL EM CS Score AL EM CS
Hepatitis C virus 18627 212 198 0.93 18627 212 198 0.93 18627 212 198 0.93
Papio anubis 18719 1110 918 0.82 18827 1098 905 0.82 18860 1098 907 0.82
469 from oxbench mdsa all 565 378 171 0.45 464 359 134 0.37 549 342 122 0.35
429 from oxbench mdsa all 8668 205 43 0.20 9575 186 35 0.18 10218 183 35 0.19
Data set 1 from [26] 167 42 31 0.73 149 42 28 0.66 134 41 25 0.60
Lemur, gorilla, mouse data set [26] 345 99 66 0.66 345 94 59 0.62 345 94 59 0.62
Rat, lemur, opossum data set [26] 486 133 87 0.65 480 130 83 0.63 471 129 81 0.62

on three DNA data sets and we used these data sets in our experiments as well,
to allow the comparison of the results (Sect. 4.6). Table 4 illustrates the results
of our RL based algorithm in comparison with the algorithm described in the
previously mentioned work. We observe that for the first data set the score
obtained by our algorithm, as well as the alignment length and the number of
columns that match perfectly are the same as those previously obtained in the
literature. For the second data set, however, we notice that even if the score is
lower and the length of the alignment is greater, the number of perfectly matched
columns, as well as the column score are slightly better. As for the third data set,
the results obtained by our algorithm outperform the ones reported in [26]: all
considered evaluation measures (score, exact match and column score) indicate
that our RL based approach leads to a better alignment.
We also compared our algorithm to the approach based on a self organizing
genetic algorithm presented in [21]. The authors of this paper use the DNA
data set 469 from oxbench mdsa all to evaluate their technique, therefore we
also made experiments on this data set (Sect. 4.4). The comparative results are
presented in Table 5. In this table we only show the alignment length, exact
match and column score, as only the values of these evaluations measures are
presented in the paper [21]. We remark that our algorithm clearly outperforms
the one presented in [21], as the alignment length is shorter (meaning less gaps
introduced), the number of columns that match perfectly is almost four times
greater, these two measures inferring that the column score is also much better.
All the alignments that we have obtained with our algorithm have also been
compared to alignments obtained with 2 different tools often used by biologists,
both tools available via web-services on the EMBL-EBI website [10]. The tools
in question are ClustalW [14] (which aligns the sequences by constructing a phy-
logenetic tree) and MAFFT [13] (which employs fast Fourier transform). Table 6
illustrates the comparative results for all the data sets that we experimented on.
The comparative results considering the CS and EM measures are depicted in
Figs. 2 and 3.
A Reinforcement Learning Based Approach to Multiple Sequence Alignment 67

Fig. 2. Comparative results for all the data sets: our RL algorithm is compared to
ClustalW [14] and to MAFFT [13] using the CS (column score) measure

Fig. 3. Comparative results for all the data sets: our RL algorithm is compared to
ClustalW [14] and to MAFFT [13] using the EM (exact match) measure

The CS measure is considered as highly relevant in the evaluation process of

an MSA algorithm since it favors alignments of less length and greater column
matching scores. The comparisons of our Q-learning approach for MSA against
other approaches in the literature (see Tables 4, 5 and 6) with respect to the CS
measure reveal the following: out of 24 comparisons, our approach is better in
17 cases, equal in 6 cases and slightly worse in 1 case (with a difference of 0.01).
These results indicate the effectiveness of our proposal.
The running times of our algorithm have been emphasized in Table 2. The
algorithm converges quite fast, in most cases in less than 2 s. However, for
larger data sets (containing more sequences or having very long sequences),
the algorithm needs more time to converge: for the considered data sets between
68 I.-G. Mircea et al.

17 s and 7 min, depending on the data set. MAFFT and ClustalW are rapid algo-
rithms as well. When measuring the computational time taken by these tools we
notice that their running time does not change very much for larger data sets:
we measured the running time for MAFFT for two of the evaluated data sets,
one with 8 and one with 3 sequences, where the sequences have approximately
equal lengths (first mitochondrial human DNA data set and data set 469 from
oxbench mdsa all). The running times are very similar, being between 3 and 5 s,
for both these data sets. However, we remark that our algorithm is somehow
sensitive to the size of the input data: greater input data does not always mean
more time, but there are cases in which this statement is true. For one of the
data sets our algorithm needed 17 s to converge to the optimal alignment, but
for the other it needed less than a second. We could therefore conclude that for
smaller data sets our algorithm might be a better choice, from a computational
time perspective and also from the point of view of the accuracy of its results.
The main approaches existing in the MSA literature (see Sect. 2.2) usually
follow GA-based algorithm variations to solve the MSA problem. As opposed to
the genetic algorithms solutions in the literature which do not alter the order of
the sequences in the sequence set but randomly insert gaps in any of them by
use of genetic operators in order to eventually improve the score of the candi-
date alignments, the proposed reinforcement learning approach tries to preserve
the incremental pairwise nature of the major currently employed MSA tools
(such as ClustalW, MAFFT) while exploring the search space in an effective
reward-targeted manner. The results therefore obtained can be easily used for
the construction of phylogenetic trees due to their progressive nature.

6 Conclusions and Further Work

In this paper we have approached, from a computational perspective, the multi-
ple sequence alignment problem, using a reinforcement learning based approach
and focusing on DNA sequences. To our knowledge, reinforcement learning has
not been used in the literature to address this problem, so far.
The experimental evaluations have been conducted using real DNA data
sets. The obtained results indicate the efficiency of our approach and prove that
the RL based algorithm outperforms, in most cases, other techniques from the
literature that approach the same problem. Moreover, another advantage of our
algorithm is the low computational time, as the convergence is achieved in a
matter of seconds or at most minutes, for the considered data sets.
We will investigate possible improvements of the model by using different
reinforcement functions, adding new local search mechanisms to increase the
model’s performance, using function approximation methods for the Q-values,
in case the state space becomes very large. In addition, we plan to investi-
gate a second RL based model, which uses different definitions of the state and
action spaces and of the reinforcement function. Regarding the biological per-
spective, we will consider a modification of the alignment algorithm so as to
favor both global alignment and the finding of motifs that are locally similar in
the sequences [11].
 A Reinforcement Learning Based Approach to Multiple Sequence Alignment 69

Acknowledgment. This work was supported by a grant of the Romanian National

Authority for Scientific Research, CNCS-UEFISCDI, project number PN-II-RU-TE-
2014-4-0082.

References
1. Agarwal, P.: Alignment of multiple sequences using GA method. Int. J. Emerg.
Technol. Comput. Appl. Sci. (IJETCAS) 13–177, 412–421 (2013)
2. Carroll, H., Beckstead, W., O’Connor, T., Ebbert, M., Clement, M., Snell, Q.,
McClellan, D.: Dna reference alignment benchmarks based on teritary structure of
encoded proteins. Bioinformatics 23(19), 2648–2649 (2007)
3. Chao, L., Shuai, L.: The research on DNA multiple sequence alignment based on
adaptive immune genetic algorithm. In: International Conference on Electronics
and Optoelectronics (ICEOE), vol. 3, pp. V3–75–V3–78, July 2011
4. Chen, S.M., Lin, C.H.: Multiple DNA sequence aalignment based on genetic algo-
rithms and divide-and-conquer techniques. Int. J. Appl. Sci. Eng. 3, 89–100 (2005)
5. Chen, S.M., Lin, C.H.: Multiple DNA sequence alignment based on genetic simu-
lated annealing techniques. Inf. Manag. Sci. 18, 97–111 (2007)
6. Chen, Y., Pan, Y., Chen, L., Chen, J.: Partitioned optimization algorithms for
multiple sequence alignment. In: Proceedings of the 20th International Conference
on Advanced Information Networking and Applications, pp. 618–622 (2006)
7. Czibula, I., Bocicor, M., Czibula, G.: A software framework for solving combina-
torial optimization tasks. Studia Universitatis “Babes-Bolyai”, Informatica, LVI,
3–8 (2011). Proceedings of KEPT 2011, Special Issue
8. Dayan, P., Sejnowski, T.: TD(λ) converges with probability 1. Mach. Learn. 14,
295–301 (1994)
9. Eger, S.: Sequence alignment with arbitrary steps and further generalizations, with
applications to alignments in linguistics. Inf. Sci. 237, 287–304 (2013)
10. EMBL-EBI, The european bioinformatics institute. http://www.ebi.ac.uk/about
11. Gotoh, O.: An improved algorithm for matching biological sequences. J. Mol. Biol.
162, 705–708 (1982)
12. Kanz, C., Aldebert, P., Althorpe, N., et al.: The EMBL nucleotide sequence data-
base. Nucleic Acids Res. 36, D29–D33 (2005)
13. Katoh, S.: MAFFT multiple sequence alignment software version 7: improvements
in performance and usability. Mol. Biol. Evol. 30, 772–780 (2013)
14. Larkin, M., Blackshields, G., Brown, N., Chenna, R., McGettigan, P., McWilliam,
H., Valentin, F., Wallace, I., Wilm, A., Lopez, R., Thompson, J., Gibson, T.,
Higgins, D.: ClustalW and clustalX version 2.0. Bioinformatics 23(21), 2947–2948
(2007)
15. Lipman, D., Altschul, S., Kececioglu, J.: A tool for multiple sequence alignment.
Proc. Natl. Acad. Sci. U.S.A. 86, 4412–4415 (1989)
16. Mircea, I., Bocicor, M., Dı̂ncu, A.: On reinforcement learning based multiple
sequence alignment. Studia Universitatis “Babes-Bolyai”, Informatica LIX, 50–
65 (2014)
17. Nasser, S., Vert, G., Nicolescu, M., Murray, A.: Multiple sequence alignment using
fuzzy logic. In: Proceedings of the IEEE Symposium on Computational Intelligence
and Bioinformatics and Computational Biology, pp. 304–311 (2007)
18. Needleman, S., Wunsch, C.: A general method applicable to the search for simi-
larities in the amino acid sequence of two proteins. J. Mol. Biol. 48(3), 443–453
(1970)
70 I.-G. Mircea et al.

19. Nelwamondo, F.V., Golding, D., Marwala, T.: A dynamic programming approach
to missing data estimation using neural networks. Inf. Sci. 237, 49–58 (2013)
20. Nguyen, H., Yoshihara, I., Yamamori, K., Yasunaga, M.: Neural networks, adaptive
optimization, and RNA secondary structure prediction. In: Proceedings of the 2002
Congress on Evolutionary Computation, CEC 2002, pp. 309–314 (2002)
21. Nizam, A., Shanmugham, B., Subburaya, K.: Self-organizing genetic algorithm for
multiple sequence alignment. Glob. J. Comput. Sci. Technol. 11(7) (2011)
22. Rasmussen, T., Krink, T.: Improved hidden Markov model training for multi-
ple sequence alignment by a particle swarm optimization-evolutionary algorithm
hybrid. BioSystems 72, 5–17 (2003)
23. Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. MIT Press,
Cambridge (1998)
24. Thompson, J.D., Linard, B., Lecompte, O., Poch, O.: A comprehensive bench-
mark study of multiple sequence alignment methods: current challenges and future
perspectives. PLoS ONE 6(3), e18093+ (2011)
25. Wang, L., Jiang, T.: On the complexity of multiple sequence alignment. Comput.
Biol. 4, 337–348 (1994)
26. Xiang, X., Zhang, D., Qin, J., Yuanyuan, F.: Ant colony with genetic algorithm
based on planar graph for multiple sequence alignment. Inf. Technol. J. 9(2), 274–
281 (2010)
 Supervised Learning Techniques for Body Mass
Estimation in Bioarchaeology

Vlad-Sebastian Ionescu(B) , Gabriela Czibula, and Mihai Teletin

Faculty of Mathematics and Computer Science, Babeş-Bolyai University,

1, M. Kogălniceanu Street, 400084 Cluj-Napoca, Romania
{ivlad,gabis}@cs.ubbcluj.ro, [email protected]

Abstract. The task of estimating the body mass from human skeletal
remains based on bone measurements is an important one in bioarchaeol-
ogy and forensic anthropology. Most of the current literature deals with
this problem through mathematical linear regression formulas applied to
various bones. In order to improve the existing results, two supervised
learning-based regression models are proposed, using artificial neural
networks and support vector machines, which are useful for express-
ing good (usually nonlinear) mappings between skeletal measurements
and body mass. Several experiments performed on an open source data
set show that the proposed applications of machine learning-based algo-
rithms lead to better results than the current state of the art. Thus, the
proposed methods are useful for producing good body mass estimations
from skeletal measurements.

Keywords: Bioarchaeology · Body mass estimation · Machine learn-

ing · Artificial neural networks · Support vector machines

1 Introduction
Body mass estimation from human skeletal remains represents a problem of
major importance in paleontological and archaeological research, since it can
provide useful information about past populations [21], such as their healthi-
ness, different social aspects, the influence of environmental factors and others.
Obtaining correct estimations for body mass from skeletons is also important in
forensic death investigations regarding unknown skeletal remains [18].
In the bioarchaeological literature it is agreed that postcranial measurements
have a direct relationship to the body size and produce the most accurate
estimates [3]. Estimation of the body mass from human skeletons represents
a challenge in forensic death investigations concerning unidentified remains [18].
A major problem in research related to body mass estimation is caused by a
lack of research collections. The current methods for estimating body mass from
the skeletons are: mechanical, morphometric [2] and a combination of biome-
chanical and morphometric methods [18]. The morphometric methods consist
of directly reconstructing the body size from the skeletal elements, while the
mechanical methods provide functional associations between skeletal elements
and body mass [2].

c Springer International Publishing AG 2018
V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9 7
72 V.-S. Ionescu et al.

In the machine learning literature, artificial neural networks [6,26] and sup-
port vector machines [12] are well known supervised learning models, with a
wide range of applications. They are adaptive systems that are able to learn
an input/output mapping (target function) from data. After the systems were
trained by receiving some examples for the target function, they will be able to
provide an approximation of the target output for new inputs.
In this paper, two machine learning-based regression models are used for
estimating the body mass from human skeletons, using bone measurements:
artificial neural networks and support vector machines. The experiments are
carried out on a publicly available database. The obtained results are compared
with the results of several body mass estimation regression formulas existing
in the bioarchaeological literature. The results obtained by the application of
machine learning models outperform the previous state of the art. Performed
experiments indicate a superior performance of the proposed models relative to
other similar approaches existing in the literature and highlight the efficiency of
using machine learning-based regression techniques for the problem of human
body mass estimation. To the best of the authors’ knowledge, the proposed
approaches are novel, since there are no existing machine learning approaches
to learn, in a supervised manner, to estimate body mass from human skeletons.
The remainder of the paper is structured as follows. Section 2 presents the
problem of archaeological body mass estimation and its relevance in bioarchae-
ology, as well as the fundamentals of the machine learning-based methods which
will be used in this paper. Section 3 introduces the machine learning-based meth-
ods for estimating the body mass from human skeletons that were applied.
Experiments on a publicly available data set are given in Sects. 4 and 5 con-
tains an analysis of the obtained results and comparisons with related work
from the literature. Section 6 provides the conclusions of the paper and indicates
several future research directions.

2 Background

This section starts by briefly presenting the body mass estimation problem,
emphasizing its relevance and importance within the archaeological research.
Then, the fundamentals of artificial neural networks and support vector machines
are presented.

2.1 Body Mass Estimation

Body mass estimation is a very important problem for modern archeology. It

can provide certain knowledge about past populations, such as indicators for
[21]: the past population’s health, the effects of different environmental factors
on past populations (e.g. subsistence strategy, climatic factors), social aspects
etc. The ability to obtain accurate body mass estimations from skeletons is also
essential in forensic death investigations concerning unidentified skeletal remains
 Supervised Learning Techniques for Body Mass Estimation 73

[18]. Consequently, it is essential for biaorchaeologists to develop and use body

mass estimation methods that are as accurate as possible.
However, designing an accurate method for solving this problem remains
a great challenge, because there are many factors which should be taken into
account [2]. Some decisions that are to be taken in this process [21]: which
are the most relevant skeletal measurements to use, which is the appropriate
statistical approach to apply, which should be the skeletal sample to use etc.
Most of the existing statistical methods for body mass estimation use lin-
ear regression formulas that usually consider one or a few bones measurements.
These formulas are usually developed on particular data sets, and it is question-
able whether or not they would perform well on previously unknown data.
Supervised machine learning-based models are therefore likely to be a good
alternative to existing methods for body mass estimation, since they can be
retrained on new data sets easily. Moreover, particular techniques to avoid the
overfitting problem can be used in order to develop models that show good
potential to generalize well on unseen data.

2.2 Artificial Neural Networks and Support Vector Machines

Artificial neural networks [15,25] are machine learning models with a wide appli-
cation area in domains like pattern recognition, speech recognition [27], predic-
tion [14], system identification and control. In structural similarity with biologi-
cal neural systems, artificial neural networks [17] consist of a set of interconnected
computational units, also referred to as neurons. One motivation for the distrib-
uted representation is to capture the parallelism expressed by natural neural
systems [17].
An artificial neural network (ANN) [1,20] is an adaptive system that learns
a function (an input/output map) from data. Adaptive means that the system
parameters are repeatedly adjusted during the training phase. After the training
phase, the Artificial Neural Network parameters are fixed and the system is used
to solve the problem at hand (the testing phase). The Artificial Neural Network
is built with a systematic step-by-step procedure to optimize a performance cri-
terion or to follow some implicit internal constraint, which is commonly referred
to as the learning rule.
In a supervised learning scenario, an input instance is presented to the neural
network together with the corresponding target response [13]. These input-
output pairs are often provided by an external supervisor. An error is repre-
sented by the difference between the desired response and the system’s output.
This error information is fed back to the network and the system parameters
are adjusted in a systematic manner (the learning rule). The process is repeated
until an acceptable performance is achieved.
Support Vector Machines were developed for classification by Cortes and
Vapnik [7], but they can be adapted for regression quite easily, resulting in the
so-called ε-support vector regression algorithm (SVR). The hyperparameter ε
controls the level to which the algorithm is allowed to make mistakes.
74 V.-S. Ionescu et al.

Formula (1) describes the underlying SVR numerical optimization problem.

The following notations are used: xi is a training instance, yi is its target output,
b is a bias term, w is the weights vector being searched and C is a regularization
hyperparameter. The variables ξi− and ξi+ are used to make the learning feasible
by allowing for some degree of error [28].
m

1
minimize w2 + C (ξi− + ξi+ )
2 i=1
⎧
−
⎪
⎨ iy − (w · xi + b) ≤ ε + ξi
(1)
subject to (w · xi + b) − yi ≤ ε + ξi+
⎪
⎩ − +
ξi , ξi ≥ 0

3 Machine Learning-Based Body Mass Estimation

This section introduces the proposed machine learning-based approaches (artifi-

cial neural networks and support vector machines) for estimating the body mass
of human skeletal remains, based on bone measurements.
Consider a data set of human skeletal remains denoted by H =
{h1 , h2 , . . . , hn } in which each instance hi represents a human skeleton. Each
skeleton is characterized by m features representing different bone measurements
which are relevant for the body mass estimation problem. The measurements are
numerical values and correspond to several significant bones in the body. Thus,
an instance hi may be viewed as an m-dimensional vector hi = (hi1 , hi2 , . . . , him )
where hij ∀1 ≤ i ≤ n represents the value of the j-th measurement applied to
the i-th skeleton. For each human remain hi , the body mass of the individual is
known and is denoted by bmi . No feature extraction or selection was performed.
The features were used as they appear in the data set from the archaeological
literature.
The ANN and SVM models are used as regressors for the body mass estima-
tion problem - they provide an estimation ei for the target body mass for each
individual hi from the data set H.
Before applying the machine learning-based methods, the data set is pre-
processed. First, the data is normalized using the Min-Max normalization
method. ANN and SVM models are known to be sensitive to data normalization,
requiring the values inputted to them to be of identical orders of magnitude in
order to be able to perform well.
After normalization, a statistically-based feature selection is applied in order
to determine how well the measurements are correlated with the target body
mass output. The dependencies between the features and the target body mass
are determined using the Pearson correlation coefficient [30]. A high Pearson
correlation implies that the two variables are linearly related, which makes linear
and non-linear learning models likely to perform well. A low Pearson correlation
means that a linear model will likely not perform well, but does not give any
information about non-linear models.
 Supervised Learning Techniques for Body Mass Estimation 75

Also as a data preprocessing step, a self organizing feature map (SOM) [29]
is used in order to obtain a two dimensional view of the data that will learned.
The trained SOM is visualized using the U-Matrix method [11]. The U-Matrix
values can be viewed as heights giving a U-Matrix landscape, with high places
encoding instances that are dissimilar while the instances falling in the same
valleys represent data that are similar. On the U-matrix, it is possible to observe
outliers, i.e. input instances that are isolated on the map. The visually observed
outliers are eliminated from the training data. This process is detailed in the
experimental part of the paper.
After the data preprocessing step, as in a supervised learning scenario, the
regression ANN and SVM models are trained during training and then tested
in order to evaluate the performance of the obtained models. These steps are
further detailed in their respective sections.

3.1 Training

The following sections present details about how the ANN and SVM models are
built during the training step.

The ANN Model. The ANN model’s capability to represent non-linear func-
tions is used in order to solve the weight estimation problem. The requirement
for the problem at hand is to design a model which uses the known measure-
ments as input features, processes them in its hidden layers and gives the results
corresponding to the estimation.
For the experiments, a feedforward neural network is built. It will be trained
using the backpropagation-momentum learning technique [24].
The ANN is composed of three layers: input, hidden and output. The input
layer is composed of ni (equal to the dimensionality of the input space) neurons.
 is used, the number nh of hidden neurons being computed
A single hidden layer
√
as nh = ni · no . Bias neurons are also used in the input and hidden layers.
On the output layer, there is a single neuron corresponding to the weight output.
All hidden and output units use the sigmoid activation function.
During training phase, the squared error between the network output value
and the target value for the current training instance is computed. The error is
propagated backwards through the network, followed by the gradient computa-
tion [17]. All the weights are updated in order to reduce the observed error for
the current example.
In order to avoid the convergence to a local minimum, stochastic gradient
descent is used during training, for building the network. The idea behind this
variation [17] is to approximate the true gradient descent search by updating the
weights incrementally, following the calculation of the error for each individual
example.
The momentum technique is also added to the ANN model, in order to speed
up the convergence and to avoid local minimums [23].
76 V.-S. Ionescu et al.

A validation subset consisting of [n/10] randomly chosen instances (n denotes

the number instances from the input data set) is extracted from the training data
set and used for optimizing the number of training iterations. Ten randomly
initialized ANNs are trained in parallel and at the end of the training, the ANN
that has the best performance on the validation set is kept. The obtained ANN
is then used for testing.

The SVM Model. The SVM model consists of ε-SVR with multiple choices
of kernels and a randomized hyperparameter search. A random search has been
shown to outperform classical grid searches for this problem [4]. The random
search consists of specifying either a probabilistic distribution for each hyperpa-
rameter or a discrete set of possible values. Then, for a given number of itera-
tions, random values are selected for each hyperparameter, according to the given
probability distributions or discrete sets of values. A model is constructed each
iteration, and evaluated using 10-fold cross validation. The best model returned
by the random search is then tested using Leave One Out cross validation, as
described below.

3.2 Testing
For evaluating the performance of both the ANN and SVM models, a leave-one
out cross-validation is used [31]. In the leave-one out (LOO) cross-validation on
a data set with n instances, the learning model is trained on n-1 instances and
then the obtained model is tested on the instance which was left out. This is
repeated n times, i.e. for each instance from the data set.
During the cross-validation process, the mean of absolute errors (MAE) is
computed using the Formula (2)
n
1
M AE = |bmi − ei | (2)
n i=1
In the above formula ei is the estimated body mass for the i-th test instance
(as provided by the ANN/SVM regressor), and bmi represents the real body
mass for the i-th test instance (known from the data set).
For the ANN, 20 LOO cross-validations will be performed, since the ran-
domness of several steps (weights initialization, the selection of the validation
set). As a preliminary step within a leave-one out cross-validation process, the
validation subset (see Sect. 3.1) will be extracted. The remaining subset will be
used for evaluating the model performance using LOO, as described above. The
MAE values reported for the 20 runs are averaged and a statistical analysis on
the obtained results is performed.
The SVM is tested using a single LOO cross-validation, since there is no
randomness in building the SVM regressor, thus the obtained results are always
the same. There is randomness in the initial shuffling of the data, but experiments
have shown that this did not influence the results in any significant manner. The
random search for the hyperparameters optimization is run for 200 iterations.
 Supervised Learning Techniques for Body Mass Estimation 77

4 Experimental Evaluation
In this section, an experimental evaluation of the proposed machine learning
models (described in Sect. 3) is provided on several case studies derived from an
open source archaeological data set. Original implementations were used for the
ANN and the SOM employed in the data preprocessing step. The scikit-learn
machine learning library was used for the SVM implementation [19].

4.1 Data Set and Case Studies

The data set considered in the experiments is an archaeological data set publicly
available at [10]. The database at [10] was developed through a research project
[9] and is a skeletal database composed of forensic cases to represent the ethnic
diversity and demographic structure of the United States population [10]. The
database contains 1009 human skeletons and 86 bones measurements represent-
ing: postcranial measurements (length, diameter, breadth, circumference, and
left and right, where appropriate) of the clavicle, scapula, humerus, radius, ulna,
sacrum, innominate, femur, tibia, fibula, and calcaneus [10]. Only the instances
for which the (forensic and/or cadaver ) body mass was available were extracted
from this database.
An analysis of the bioarchaeological literature revealed 10 measurements
which are used for human body mass estimation and have a good correlation
with the body size (stature and/or body mass) [2]: (i) femoral head diameter (ii)
iliac breadth, femur bicondrial length (iii) clavicle, humerus, radius, ulna, femur,
tibia and fibula. From these measurements, the femural measurements seem to
produce the most accurate body mass estimations [2].
In case of two measurements for any of these bones (left and right), their
mean was used. If only one measurement existed in the database, it was used as
is. Instances containing missing bone measurements were not considered in any
of the case studies.
The following case studies are considered, with the aim of analyzing the
relevance of the previously mentioned measurements for the problem of body
mass estimation.

– The first case study consists of 200 instances characterized by 3 measure-

ments - (i) and (ii).
– The second case study consists of 146 instances characterized by 8 measure-
ments - (i) and (iii).
– The third case study consists of 135 instances characterized by all 10 mea-
surements - (i), (ii) and (iii).

4.2 Data Preprocessing

As mentioned in Sect. 3, before building the ANN and SVM models, the
training data is preprocessed. After normalizing the data using the Min-Max
normalization method, the Pearson correlation coefficients between the features
78 V.-S. Ionescu et al.

Fig. 1. Pearson correlation and U-Matrix visualization.

(measurements) and the target weights are computed. Figure 1(a) illustrates the
correlations between the 10 measurements and the target output on the third data
set (case study).
Figure 1(a) shows that the first three features have the highest correlation
with the body mass. The femoral head diameter has the maximum correlation
with the body mass (0.5086), while the length of the tibia has a correlation of
only 0.3858 with the body mass. Analysing the correlations from Fig. 1(a) it is
expected that the best performance of the ANN and SVM models is on the first
case study (using only the first three measurements).
In order to determine possible outliers within the training data, a self orga-
nizing map is trained on the first data set (consisting of 200 instances). The
U-Matrix visualization of the trained SOM is depicted in Fig. 1(b). On the map
one can see four small regions (with boundaries outlined in white). The instances
from these regions may be viewed as possible outliers in the data set, since they
are isolated from the other instances. This way, 8 instances can be visually
identified as possible outliers. These instances are removed from the training
data sets.

4.3 Results
In this section, an experimental evaluation of the proposed application of ANN
and SVM machine learning regressors (described in Sect. 3) is provided on the
case studies described in Sect. 4.1.
For the ANN parameters, a learning rate of 0.3 momentum values of 0.2 were
used. These were chosen by using a classic grid search over several values lower
than 1, and choosing the values which consistently provided the best results.
For the random search of SVM hyperparameters, the uniform probability
distribution was used. The intervals or sets from which each hyperparameter is
drawn are as follows:
 Supervised Learning Techniques for Body Mass Estimation 79

– kernel: from the set {rbf, linear, sigmoid, poly} (i.e. RBF, linear, sigmoid,
polynomial kernels).
– ε: from [0, 1).
– γ: from [0, 1).
– b: from [0, 1).
– Polynomial degree for the polynomial kernel: from the set {1, 2, 3, 4, 5}.
– C: from [0, 1).
Note that not all hyperparameters apply to every kernel choice.
For the case studies considered for evaluation (Sect. 4.1), after preprocessing
the training data set as indicated in Sect. 4.2, the ANN and SVM regressors are
built through training (see Sect. 3.1).
For evaluating the performance of the trained ANN and SVM, 20 runs of
Leave One Out (LOO) cross-validation were used for the ANN model and a
single run for the SVM model, because the latter does not use any random
numbers, so its results will always be the same.

ANN Results. The MAE values along with the minimum, maximum, mean
value and standard deviation, obtained using the ANN, for each case study
performed, are given in Table 1. In this table, the MAE values are given in
kilograms.

Table 1. Results obtained using the ANN, considering all 20 LOO cross-validations.

Case study Mean (kg) Min (kg) Max (kg) Stdev (kg)
First case study 8.766 7.751 10.195 0.682
Second case study 9.593 8.301 11.259 0.77
Third case study 8.945 7.63 10.338 0.649

Table 1 indicates, as expected, that the best results were obtained for the
first case study, when only the first 3 measurements are used for characterizing
the skeletal remains. Figure 2 depicts the values for the MAE measure obtained
during the 20 runs of the LOO cross-validation process applied on the first case
study. The average of these values, as well as their standard deviations are also
indicated. The small values for the standard deviation reveal a good performance
of the proposed ANN model application.

SVM Results. The results obtained using the SVM are presented in Table 2.
The best values used for the hyperparameters (including the used kernel func-
tion) are depicted in the last column of the table.
As shown in Table 2, the SVM obtained a performance similar to the ANN:
the best performance on the first case study and the worst performance on the
second case study.
80 V.-S. Ionescu et al.

Fig. 2. MAEs for the ANN on the first case study.

Table 2. Results obtained using the SVM.

Case study MAE (kg) Example best hyperparameters

First case study 8.4897 kernel: linear
ε: 0.027578797176389447
C: 0.641027338498186
Second case study 9.4918 kernel: linear
ε: 0.09576656837738184
C: 0.24862712124238084
Third case study 8.6843 kernel: polynomial
ε: 0.04798752786733451
C: 0.1465703832031966
degree: 2
γ: 0.7681948068876795
b: 0.42975272058542213

5 Discussion and Comparison to Related Work

In this section, an analysis is provided for the approaches introduced in Sect. 3

for body mass estimation from bone measurements. Then, a comparison with
similar approaches from the literature is conducted.
As shown in the experimental part of the paper, both the ANN and SVM
models have provided about the same performances for the body mass estimation
problem. The SVM slightly outperformed the ANN, with at most 0.28 MAE.
 Supervised Learning Techniques for Body Mass Estimation 81

The experimental values obtained for the (average) MAE considering all case
studies are summarized in Table 3. The last column from Table 3 contains the
MAE value (averaged over all LOO cross-validations). The best values, for each
case study, are highlighted.

Table 3. MAE values obtained using the ANN and the SVM models on the considered
case studies

Case study Model MAE (kg)

First case study ANN 8.766
SVM 8.4897
Second case study ANN 9.593
SVM 9.4918
Third case study ANN 8.945
SVM 8.6843

Table 3 shows that the best machine learning-based regressor for estimating
the human body mass from skeletal remains is the SVM, when only 3 measure-
ments (femoral head diameter, iliac breadth and femur bicondrial length) are used
for the skeletal elements. This is to be expected, since these three measurements
showed the highest correlation with the target body mass (Fig. 1(a)).
Analysing the results from Table 3 it can also be seen that the worst results
were obtained, for both ANN and SVM, on the second case study. Therefore, the
iliac breadth and femur bicondrial length are also important in estimating the
body mass and the measurements for the clavicle, humerus, radius, ulna, femur,
tibia and fibula do not improve the body mass estimation results.
It is worth mentioning that the outliers removal step performed during the
data preprocessing step has increased the performance of the ML regressor.
Table 4 illustrates the effect of removing the outliers from the training data
of the ANN. A significant reduction for the MAE value was obtained for the
third case study.

5.1 Comparison to Related Work

In the following, a brief review of the recent human body mass estimation liter-
ature is given, with the aim to compare the used ML regressors to the existing
related work. As far as the authors are aware, there are no existing machine
learning-based models (like the ANN and SVM models applied in this paper)
for the problem of body mass estimation from skeletal remains.
A comparison between several body mass estimation methods was conducted
by Auerbach and Ruff in [2] (2004). The authors proposed to test some exist-
ing methods on a great variety of subjects. They used skeletal remains of 1173
adult skeletons of different origins and body sizes, both males and females. Three
82 V.-S. Ionescu et al.

Table 4. Comparative MAE values - with and without outliers removal using the
ANN.

Case study MAE reduction (%) Outliers removal MAE (kg)

First case study 19.39 Yes 8.766
No 10.874
Second case study 19.22 Yes 9.593
No 11.875
Third case study 25.68 Yes 8.945
No 12.036

femural head-based regression formulas were tested and compared on the con-
sidered skeletal sample: Ruff et al. [22] (1991), McHenry [16] (1991) and Grine
et al. [8] (1995). The authors concluded that for a very small body size range
(Pygmoids), the formula of McHenry (1992) can provide a good body mass esti-
mation. For very large body sizes, the formula of Grine et al. (1995) should be
used, whereas for the other samples the formula of Ruff (1991), or the average
of the three techniques would be the best approach.
Ruff et al. provided in [21] (2012) new body mass estimation equations that
are generally applicable to European Holocene adult skeletal samples. Body mass
estimation equations were based on femoral head breadth. 1145 skeletal speci-
mens were obtained from European museum collections, from time periods rang-
ing from Mesolithic to the 20th century [21]. On these data sets, the regression
formulas introduced in [21] provided better results than the previous formulas
from Ruff et al. [22] (1991), McHenry [16] (1991) and Grine et al. [8] (1995).
The data sets used in the previously mentioned papers are not publicly avail-
able, that is why an exact comparison of the approaches introduced in this
paper to the previously mentioned approaches cannot be performed. Since the
authors have not been able to find experiments in the literature related to body
mass estimation using the same data set as in this paper, the following com-
parison to related work was conducted. The regression formulas introduced in
Ruff et al. [8,16,21,22] were applied on the data sets used in this paper (all
three case studies) and the obtained MAE values were compared with the ones
provided by the ANN and SVM regression models. The results of the compar-
ison are given in Table 5. In this table, 95% confidence intervals [5] were used
for the obtained results. The comparison is graphically depicted, for all three
case studies, in Fig. 3. In this figure, the first two dashed bars correspond to the
ANN and SVM models used in this paper, while the other bars correspond to
the above mentioned four approaches from the literature. For the ANN, the 95%
confidence intervals of the average MAE are also illustrated.
In Table 5 and Fig. 3 it can be observed that the MAE values obtained by
the ANN and SVM methods are smaller than those obtained using regression
formulas from the literature. One can notice that for the ANN even the upper
limit of the 95% confidence interval of the mean MAE is below the MAE from the
 Supervised Learning Techniques for Body Mass Estimation 83

Fig. 3. Comparison to related work

literature. This is somehow predictable because the previously stated regression

formulas only use one measurement, the femoral head anterior-posterior breadth,
while the proposed approaches are using multiple measurements. This is the
best measure of the performance of the machine learning approach, since the
experiments are performed on the same data sets. Note that the ANN and SVM
models provided better results, even if the evaluation of the machine learning
based models was obtained using multiple cross-validation runs (in order to
avoid overfitting), whereas the formulas from the literature were obtained using
the entire data set. Another major advantage of the proposed approaches with
respect to the literature is that estimations are made without knowing the sex
characteristics, which is mainly used in the existing literature.
Table 5 shows that, considering both the approaches from the literature as
well as the ANN and SVN, the best performance was obtained on the first case
study.
It has to be mentioned that most of the researchers from the bioarchaeological
fields develop regression formulas for body mass estimation based on a data set
which is also used for testing the developed formulas, without using a testing
set independent from the training data and without using any type of cross-
validation. This may lead to overfitting, as for the regression formulae from the
literature which provided good performances on the data they were trained on
(about 4–5 MAE), but when applied on an unseen test data (our case studies)
they provide large MAE values (see Sect. 4.3). It is likely that the methods from
the body mass estimation literature would provide larger errors under the testing
methodology used for the proposed machine learning applications.
84 V.-S. Ionescu et al.

Table 5. Comparison between the machine learning approaches and similar related
work. 95% confidence intervals are used for the ANN results.

Approach First case study Second case Third case study

MAE (kg) study MAE (kg) MAE (kg)
ANN 8.766 ± 0.299 9.593 ± 0.338 8.9452 ± 0.284
Implementation
SVM 8.4897 9.4918 8.6843
Implementation
Ruff et al. (2012) [21] 10.115 11.43 11.202
Ruff et al. (1991) [22] 10.476 11.556 11.359
Mc Henry (1992) [16] 10.176 11.514 11.324
Grine et al. (1995) [8] 10.609 11.656 11.431

The main advantage of the machine learning approaches proposed in this

paper over the simple mathematical ones is that the machine learning models can
be retrained on new data, and they are able to generalize well if specific methods
for avoiding overfitting are used, like the ones described in Sect. 3.1. Moreover,
it can be computationally costly to develop mathematical formulas on new data
sets, which will likely not generalize well anyway. Instead, a machine learning
regressor (like ANN or SVM) can easily learn from new data sets, resulting in
learning models that are able to perform well on unseen data having the same
features as the training one.

6 Conclusions and Future Work

Two supervised learning regression models were applied for estimating the body
mass from human skeletal remains based on bone measurements. The idea to
apply machine learning for the body size estimation problem is a novel one.
Several case studies were conducted on a publicly available data set from the
bioarchaeological literature. The obtained results outperformed classical statis-
tical methods for body size estimation, which makes the proposed approaches
useful for the field of bioarchaeology.
Further work will be done in order to improve the data preprocessing step of
learning, by automatically detecting and removing the outliers from the train-
ing data set. Since the archaeological data sets usually contain missing mea-
surements, methods for dealing with these missing values will be investigated.
Applying other machine learning-based regression models for the body mass esti-
mation problem, like radial basis function networks and k-nearest neighbors, is
also a direction of interest for the authors.

Acknowledgment. This work was supported by a grant of the Romanian National

Authority for Scientific Research, CNCS–UEFISCDI, project number PN-II-RU-TE-
2014-4-0082.
 Supervised Learning Techniques for Body Mass Estimation 85

References
1. Arora, P., Srivastava, S., Singhal, S.: Analysis of gait flow image and gait Gaussian
image using extension neural network for gait recognition. Int. J. Rough Sets Data
Anal. 3(2), 45–64 (2016)
2. Auerbach, B., Ruff, C.: Human body mass estimation: a comparison of “mor-
phometric” and “mechanical” methods. Am. J. Phys. Anthropol. 125(4), 331–342
(2004)
3. Ruff, R.B.: Body mass prediction from skeletal frame size in elite athletes. Am. J.
Phys. Anthropol. 113(4), 507–517 (2000)
4. Bergstra, J., Bengio, Y.: Random search for hyper-parameter optimization. J.
Mach. Learn. Res. 13, 281–305 (2012)
5. Brown, L.D., Cai, T.T., Dasgupta, A.: Interval estimation for a binomial propor-
tion. Statistical Science 16, 101–133 (2001)
6. Chatterjee, S., Sarkar, S., Hore, S., Dey, N., Ashour, A.S., Balas, V.E.: Particle
swarm optimization trained neural network for structural failure prediction of mul-
tistoried RC buildings. Neural Comput. Appl. 1–12 (2016)
7. Cortes, C., Vapnik, V.: Support-vector networks. Mach. Learn. 20(3), 273–297
(1995)
8. Grine, F., Jungers, W., Tobias, P., Pearson, O.: Fossil homo femur from berg aukas,
northern namibia. Am. J. Phys. Anthropol. 26, 67–78 (1995)
9. Jantz, R.J., Moore-Jansen, P.H.: Database for Forensic Anthropology in the
United States, 1962–1991 (ICPSRversion) (1998). University of Tennessee, Dept.
of Anthropology, Knoxville, TN [producer] (1998). Inter-university Consortium for
Political and Social Research, Ann Arbor, MI (2000)
10. Jantz, R.J., Moore-Jansen, P.H.: Database for Forensic Anthropology in the United
States, 1962-1991 (ICPSR). University of Tennessee, Department of Anthropology
(2000). http://www.icpsr.umich.edu/icpsrweb/NACJD/studies/2581
11. Kaski, S., Kohonen, T.: Exploratory data analysis by the self-organizing map:
structures of welfare and poverty in the world. In: Neural Networks in Financial
Engineering. Proceedings of the Third International Conference on Neural Net-
works in the Capital Markets, pp. 498–507. World Scientific (1996)
12. Kriti Virmani, J., Dey, N., Kumar, V.: PCA-PNN and PCA-SVM based CAD
systems for breast density classification. In: Hassanien, A.E., Grosan, C., Fahmy
Tolba, M. (eds.) Applications of Intelligent Optimization in Biology and Medicine,
pp. 159–180. Springer International Publishing, Cham (2016)
13. Kröse, B., van der Smagt, P.: An introduction to neural networks. The University
of Amsterdam (1996)
14. Linstrom, K.R., Boye, A.J.: A neural network prediction model for a psychiatric
application. In: International Conference on Computational Intelligence and Mul-
timedia Applications, pp. 36–40 (2005)
15. Maji, P., Chatterjee, S., Chakraborty, S., Kausar, N., Samanta, S., Dey, N.: Effect
of Euler number as a feature in gender recognition system from offline hand-
written signature using neural networks. In: 2015 2nd International Conference
on Computing for Sustainable Global Development (INDIACom), pp. 1869–1873,
March 2015
16. McHenry, H.: Body size and proportions in early hominids. Am. J. Phys. Anthro-
pol. 87, 407–431 (1992)
17. Mitchell, T.M.: Machine Learning. McGraw-Hill Inc., New York (1997)
86 V.-S. Ionescu et al.

18. Moore, M.K.: Body mass estimation from the human skeleton. Ph.D. thesis,
The University of Tennessee, Knoxville (2008)
19. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: machine
learning in python. J. Mach. Learn. Res. 12, 2825–2830 (2011)
20. Rojas, R.: Neural Networks: A Systematic Introduction. Springer, Heidelberg
(1996)
21. Ruff, C., Holt, B., Niskanen, M., Sladk, V., Berner, M., Garofalo, E., Garvin,
H., Hora, M., Maijanen, H., Niinimki, S., Salo, K., Schuplerov, E., Tompkins, D.:
Stature and body mass estimation from skeletal remains in the European holocene.
Am. J. Phys. Anthropol. 182(4), 601–617 (2012)
22. Ruff, C., Scott, W., Liu, A.C.: Articular and diaphyseal remodeling of the proximal
femur with changes in body mass in adults. Am. J. Phys. Anthropol. 86, 397–413
(1991)
23. Rumelhart, D.E., Hinton, G.E., Williams, R.: Learning internal representations
by error propagation. In: Rumelhart, D.E., McClelland, J.L. (eds.) Parallel Data
Processing, vol. 1, pp. 318–362. The MIT Press, Cambridge (1986)
24. Russell, S., Norvig, P.: Artificial Intelligence - A Modern Approach. Prentice Hall
International Series in Artificial Intelligence. Prentice Hall, Upper Saddle River
(2003)
25. Samanta, S., Ahmed, S.S., Salem, M.A.M.M., Nath, S.S., Dey, N., Chowdhury,
S.S.: Haralick features based automated glaucoma classification using back propa-
gation neural network. In: Satapathy, C.S., Biswal, N.B., Udgata, K.S., Mandal, J.
(eds.) Proceedings of the 3rd International Conference on Frontiers of Intelligent
Computing: Theory and Applications (FICTA) 2014, vol. 1, pp. 351–358. Springer
International Publishing, Cham (2015)
26. Sirshendu, H., Tanmay, B., Nilanjan, D., Hassanien, A.E.: Encryption using
speeded up robust features extraction technique and artificial neural network. In:
Innovations, and Applications, Image Feature Detectors Foundations (2015)
27. Skowronski, M.D., Harris, J.G.: Automatic speech recognition using a predictive
echo state network classifier. Neural Netw. 20(3), 414–423 (2007)
28. Smola, A., Schlkopf, B.: A tutorial on support vector regression. Stat. Comput.
14(3), 199–222 (2004)
29. Somervuo, P., Kohonen, T.: Self-organizing maps and learning vector quantization
for feature sequences. Neural Process. Lett. 10, 151–159 (1999)
30. Tufféry, S.: Data Mining and Statistics for Decision Making. Wiley, New York
(2011)
31. Wahba, G., Lin, Y., Zhang, H.: GACV for support vector machines, or, another
way to look at margin-like quantities. In: Advances in Large Margin Classifiers,
pp. 297–309 (2000)
 A Fuzzy Approach of Sensitivity for Multiple Colonies
on Ant Colony Optimization

Camelia-M. Pintea1 ✉ , Oliviu Matei1, Rabie A. Ramadan2, Mario Pavone3,

( )

Muaz Niazi4, and Ahmad Taher Azar5,6

1
TU Cluj-Napoca, Cluj-Napoca, Romania
[email protected], [email protected]
2
Cairo University Egypt, Giza, Egypt
[email protected]
3
University Catania, Catania, Italy
[email protected]
4
COMSATS Pakistan, Islamabad, Pakistan
[email protected]
5
Faculty of Computers and Information, Benha University, Banha, Egypt
6
Nile University, 6th of October City, Egypt
[email protected], [email protected]

Abstract. In order to solve combinatorial optimization problem are used mainly

hybrid heuristics. Inspired from nature, both genetic and ant colony algorithms
could be used in a hybrid model by using their benefits. The paper introduces a
new model of Ant Colony Optimization using multiple colonies with different
level of sensitivity to the ant’s pheromone. The colonies react different to the
changing environment, based on their level of sensitivity and thus the exploration
of the solution space is extended. Several discussion follows about the fuzziness
degree of sensitivity and its influence on the solution of a complex problem.

Keywords: Ant colony optimization · Stigmergy · Sensitivity · Uncertainty ·

Fuzzy data · Traveling salesman problem · Combinatorial optimization problems

1 Introduction

Complex problems from different real-life domains, as for example transportation and
medicine, use heuristics to efficiently find high-quality near optimal solutions in a reason‐
able running time [1, 10]. Ant Colony Optimization (ACO) it is a class of metaheuristic
optimization methods using the concepts of distributed optimization for solving Combi‐
natorial Optimization Problems [7]. In ACO metaheuristic [7], the ants simulate real ant
behavior. The ants produce pheromone trails used as indirect communication between
ants. The stigmergic behavior of ants [3, 9] it is used to obtain a problem solution.
In real-life incompleteness and uncertainty are usual for nowadays information. This
is a very challenging task when it is the time to take decisions. A fuzzy set is a class of
objects with a continuum of grades of membership [30]. Fuzzy logic handles the concept
of “partial truth”, where the value of “truth” it is in a range from “completely true” to
“completely false”. In the literature there are several directions to handle uncertain

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_8
88 C.-M. Pintea et al.

events. For example processing uncertain databases [1] or using resilient algorithms to
tolerate degrees of errors in data without losing performance, storage capability and
correctness [8, 22–24, 29].
The current paper investigates the Ant Colony Optimization when the artificial ants
have to take decisions influenced by fuzzy degrees of sensitivity. The identification of
solutions for complex problems it is based on the ants’ colonies self-organization and
by the indirect interactions between ants. The indirect communication occur when the
environment is modified, this process is stigmergy and influences the artificial ants
decisions.
The paper is organized as follows: Sect. 2 is about the Ant Colony Optimization;
Sect. 3 introduces the fuzzy approach of Sensitive Multiple Colonies followed in
Sect. 4 by the fuzzy approach for ACO with Multiple Sensitive Colonies. The last part
of the paper includes several discussions on the proposed technique and concludes with
future work and final remarks.

2 Prerequisites

The concept of stigmergy it is behind the collective behavior of social individuals as for
example the ants or why not, the humans. Grassé [9] studied the stigmergic behavior of
ants and in [13] it is described the indirect communication between bees within a colony.
“Stigmergy occurs when an insect’s actions are determined or influenced by the conse‐
quences of another insect’s previous action” [2].
Nowadays are made artificial environments [28] called smart-dust network with tiny
micro-electromechanical systems as sensors, robots or other devices installed with wire‐
less communication able to detect vibrations, light, smells or others.
The environment mediates the indirect communication between individuals, and ants
in particular in Ant Colony Optimization (ACO) [7]. The ACO meta-heuristic includes
several ant algorithms.
In ant algorithms are used artificial ants; the artificial ants mimic the behavior of real
ants, mainly the using the indirect communication based on the amount of pheromone
deposited on a path of a network (graph), the probabilistic preference for a network paths
with a larger amount of pheromone and in time, the shorter paths will have larger amount
of pheromone.
The natural behavior of ants includes cooperation and adaptation. There are several
variants of ant mechanisms [11, 19, 26].
A combinatorial optimization problem represented as a network /graph could be
solved with an ant algorithm. In general, all ant algorithms are based on the following
items, regarding the problems’ candidate solution.
• In the problem related network, each path followed by an ant has a solution associ‐
ated.
• The amount of the pheromone deposited by an ant on a network’s path it is propor‐
tional to the quality of the corresponding solution.
• When an ant has the possibility to choose between several paths, it will chose, more
probably, the path with the larger amount of pheromone.
 A Fuzzy Approach of Sensitivity for Multiple Colonies 89

So, finally the majority/all the ants will follow the paths with the larger pheromone
trail that will be, hopefully, the shortest tour and also the better solution for the optimi‐
zation problem.
ACO it is successfully used to solve complex real world problems in different
domains [7]: transport, routing, communication, scheduling, semantic web, bioinfor‐
matics etc.

3 A New Fuzzy Approach of Multiple Sensitivity Colonies for ACO

An ant colony it is a system endowed with stigmergy where not all the individuals have
the same reaction to the pheromone trails. An ant has a certain pheromone sensitivity
level (psl) expressed as a real number in the [0;1] range. When psl it is null the ants
ignores the indirect stigmergic information, so it is considered pheromone blind.
The ants are environment explorers when their sensitivity is rather small, so they are
more independent then the others considered environment exploiters. The independent
ants have the goal to sustain search diversification. The ants with high pheromone sensi‐
tivity will exploit further intensively the potential search areas [4, 15–17, 19].
In the newly Multiple Sensitive Colonies, each colony it is endowed with a new
parameter called sensitivity level range (rsl). An ant from a given colony has its pher‐
omone level included in the specific range sensitivity level, psl it is included in slr. As
for example an ant j with pslj = 0,15 it is included into a colony with slr = [0,10; 0,25].
Could be considered two variants for the psl of an ant:
– The psl could be considered fixed, a static value during the search ant’s activity.
– The psl could be considered variable, a dynamic value, but with the constraint to be
limited by slr.
In the current approach it will be used the second variant with a particular fuzzy
approach on the slr range. Other related fuzzy approaches are in [5, 6]. A sensitivity
degree is based on the uncertain environment influence, so it is called the fuzziness
degree of sensitivity, fds; in particular are used the dimension regularity and scale regu‐
larity. Each psl is relocated randomly in an interval I, with the bounds: psl and the nearest
psl bound of slr.
For example, a particular case for the new sensitive psl value could be considered
the average between the initial sensitivity level, psl, and the nearest bound range,
Average(Min(l, psl); Min(psl, r)), where l and r are the extremes left (l) and right (r) of
the slr range. So, for example, the ant j with the initial pslj = 0,15, included into the
colony with slr = [0,10; 0,25] will have the potential to change its psl value into another
psl value, based on a fuzzy approach. For the considered example, the newly pheromone
sensitivity level of the ant j could be: pslj = 0,125, but when using the fuzziness degree
of sensibility could be any real value between [0,10; 0,15].
The global parameters, considered as integers values: a and b on the (0,100) are
expressing the uncertainty, as in [18]. The parameter a, the dimension regularity, speci‐
fies here how many ant’s sensitivity values are modified, with the same impact irre‐
spective of the number of ants.
90 C.-M. Pintea et al.

The parameter b, the scale regularity, specifies how far the psl value moves, but
keeping the same impact irrespective of the distance [18]. Each psl value it is relocated
randomly within a range. In Fig. 1 are shown the modified values of ant j sensitivity
pslj, randomly chosen in the interval I, with the bounds: psl and the nearest psl bound
of slr.

Fig. 1. A symbolic representation of the initial map (left) and the fuzzified map (right)

In time the ant colonies, with different levels of sensitivity obtain beneficent result
for complex problems. Engaging multiple ant colonies the proposed model aims to
achieve robust solutions in an evolutionary manner as further is illustrated by the numer‐
ical experiments.
The introduced Ant Colony Optimization with Multiple Colonies endowed with
Fuzzy Degrees of Sensitivity for solving combinatorial optimization problems can be
described as follows.
ACO with Multiple Colonies endowed with Fuzzy Degrees of Sensitivity

4 Numerical Experiments and Discussions

The main software used is the public software of T. Stützle ACOTSPV1.0 [25] consid‐
ering also the 3-opt optional heuristic [12] for all algorithms. Based on [15] with
 A Fuzzy Approach of Sensitivity for Multiple Colonies 91

Sensitive Ant Model (SAM), there are used colonies with different pheromone sensitiv‐
ities non-overlapping interval distribution. The newest tests are made on the fuzzyified
data-sets.
The fuzzy data are obtain similar with [18]: several nodes i are modified, randomly
chosen using C(i; radius), the circle with the center in the current node i; the radius is
y = x · b = 100, where x is the distance from node i to the nearest node. For four of the
initial instances included in [15] were obtain eight new fuzzified instances. From each
original instance were obtain new instances using the parameters a and b: (10,25) and
(10,50).
The following cases and notations are considered as in [15].
• SAM: the Sensitive Ant Model use a random pheromone sensitivity psl in [0,1] for
the colonies of ants;
• h-SAM: SAM with half of the colonies of ants use a random psl in [0,1] and the other
colonies of ants, psl in [0,0.25];
• q-SAM: SAM with a quarter of ants use a random psl in [0,1] and the others with PSL
in [0,0.25].
There are considered different parameter sets: a = 1, b = 2, τ0 = 0.01; a number of
ten ants are considered, r = 0, 5, q0 = 0, 7 and the termination criteria is 60 s./trial for
ten trials and maximum 500 tours.
As in [18], is included a statistical analysis based on the Percentage Change (PC)
of the solution related to the optimal solution, where, in our case, the solution is the best
solution found and optimum is the optimum found in the TSPLib [27] for each consid‐
ered instance: PC = (solution-optimum)/optimum× 100%,
Tables 1, 2, 3 and 4 and Figs. 2, 3 and 4 show that multiple-colonies with different
levels of pheromone sensitivity have good results when solving TSP. Practically, the
results shows that implementation of ACO is fairly stable when some uncertain data are
involved. So, this could show the ability of multiple colonies, with different pheromone
sensitivity, to be adaptable to many difficult situations.

Table 1. Average of best results: ACS on TSP instances; fuzzy instances: (a; b)∈{(10;50);
(10;25)}; initial value of ACS [15]; percentage change (PC) is computed.
Instance Optimal ACS PC (10,50) (10,25)
ACS PC ACS PC
lin318 42029.00 42050.16 5.03% 42314.6 67.95% 42086.6 13.70%
pcb442 50778.00 50842.86 12.77% 50735.9 −8.29% 50761.4 −3.27%
att532 27686.00 27702.08 5.81% 27679.1 −2.49% 27720.2 12.35%
rat783 8806.00 8817.44 12. 99% 8804.8 −1.36% 8817.5 13.06%
92 C.-M. Pintea et al.

Table 2. Average of best results: Sensitive Ant Model (SAM) on TSP instances; fuzzy instances:
(a;b)∈{(10;50); (10;25)}; initial value of SAM [15]; percentage change (PC) is computed.
Instance Optimal SAM PC (10,50) (10,25)
SAM PC SAM PC
lin318 42029.00 42038.92 2.36% 42298.4 64.10% 42054.6 6.09%
pcb442 50778.00 50831.42 10.52% 50735.5 −8.37% 50751.8 −5.16%
att532 27686.00 27701.12 5.46% 27683.1 −1.05% 27724.4 13.87%
rat783 8806.00 815.02 10.24% 8803.6 −2.73% 8811.8 6.59%

Table 3. Average of best results: q-SAM (quarter of ants use random psl ∈[0;1], others: psl∈
[0;0:25]) on TSP instances; fuzzy instances: (a;b) ∈{(10;50); (10;25)}; initial value of q-SAM [15];
percentage change (PC) is computed.
Instance Optimal q-SAM PC (10,50) (10,25)
q-SAM PC q-SAM PC
lin318 42029.00 42031.28 0.54% 42304.5 65.55% 42040 2.62%
pcb442 50778.00 50801.58 4.64% 50704.7 −14.44% 50759.2 −3.70%
att532 27686.00 27705.50 7.04% 27679 −2.53% 27726 14.45%
rat783 8806.00 8825.74 22.42% 8803.4 −2.95% 8827.7 24.64%

Table 4. Average of best results: h-SAM (half of ants use random psl ∈[0;1], others: psl∈[0;0:25])
on TSP instances; fuzzy instances: (a;b)∈{(10;50); (10;25)}; initial value of h-SAM [15];
percentage change (PC) is computed.
Instance Optimal h-SAM PC (10,50) (10,25)
h-SAM PC h-SAM PC
lin318 42029.00 42031.28 0.54% 42302.4 65.05% 42073 10.47%
pcb442 50778.00 50804.14 5.15% 50706.2 −14.14% 50769.5 −1.67%
att532 27686.00 27702.94 6.12% 27679.7 −2.28% 27723.7 13.62%
rat783 8806.00 8823.42 19.78% 8805.2 −0.91% 8826.3 23.05%

Fig. 2. Comparison of Percentage Change (PC) for the initial data-sets: lin318 (left) and pcb442,
att532, rat783 instances (right).
 A Fuzzy Approach of Sensitivity for Multiple Colonies 93

Fig. 3. Comparison of Percentage Change (PC) for pcb442, att532, rat783 fuzzy data-sets:
(10,25) (left) (10,50) (right).

Fig. 4. Comparison of Percentage Change (PC) for the lin318 data set: (10,25) (left) and (10,50)
(right) fuzzy data.

Future work will include testing the sensitivity degree of fuzziness on real-life data
of complex problems as generalized problems [14, 16, 19–21].

5 Conclusions

Ant Colony Optimization is today a powerful bio-inspired tool for solving difficult opti‐
mization problems. Inter-operation among ants is based on indirect communication
mediated by pheromone trails. The article introduces fuzzy degrees of sensitivity when
using multiple colonies of ants with different levels of pheromone sensitivity on ACO.
The exploration and exploitation are enhanced and lead to potentially better ACO solu‐
tion of optimization problems. Ongoing research focuses on several numerical experi‐
ments on several optimization problems, as Traveling Salesman Problem and its gener‐
alized version, to test the effects of the proposed fuzzy approach.

Acknowledgement. The study was conducted under the auspices of the IEEE-CIS
Interdisciplinary Emergent Technologies task force.
94 C.-M. Pintea et al.

References

1. Aggarwal, C.C., Yu, P.S.: A survey of uncertain data algorithms and applications. IEEE Trans.
Knowl. Data Eng. 21(5), 609–623 (2009)
2. Bonabeau, E., et al.: Swarm intelligence from natural to artificial systems. Oxford University
Press, Oxford (1999)
3. Camazine, S., et al.: Self Organization in Biological Systems. Princeton University Press,
Princeton (2001)
4. Chira, C., Dumitrescu, D., Pintea, C.-M.: Learning sensitive stigmergic agents for solving
complex problems. Comput. Inform. 29(3), 337–356 (2010)
5. Crisan, G.C., Nechita, E., Palade, V.: Ant-based system analysis on the traveling salesman
problem under real-world settings. In: CIMA, pp. 39–59 (2014)
6. Crisan, G.C., Pintea, C.-M., Pop, P.C.: On the resilience of an ant-based system in fuzzy
environments. An empirical study. In: FUZZ-IEEE, pp. 2588–2593 (2014)
7. Dorigo, M., Di Caro, G., Gambardella, L.M.: Ant algorithms for discrete optimization. Artif.
Life 5(2), 137–172 (1999)
8. Finocchi, F., et al: Designing reliable algorithms in unreliable memories algorithms, LNCS,
vol. 3669, pp. 1–8 (2005)
9. Grasse, P.-P: La reconstruction du nid et les coordinations interindividuelles chez
bellicositermes natalensis et cubitermes sp. La theorie de la stigmergie: essai d’interpretation
du comportement des termites constructeurs. Insect Soc. 6, 41–80 (1959)
10. Holzinger, A.: Interactive machine learning for health informatics: when do we need the
human-in-the-loop? Brain Inf. 3(2), 119–131 (2016)
11. Holzinger, A., et al.: Towards interactive machine learning: applying ant colony algorithms
to solve the traveling salesman problem with the human-in-the-loop approach. In: CD-ARES
(2016)
12. Helsgaun, K.: An effective implementation of the Lin Kernighan TSP heuristic. Eur. J. Oper.
Res. 126, 106–130 (2000)
13. Michener, C.D.: The social behavior of bees: a comparative study. Harvard University Press,
Massachusetts (1974)
14. Lahrichi, N., et al.: An integrative cooperative search framework for multi-decision-attribute
combinatorial optimization: application to the MDPVRP. Eur. J. Oper. Res. 246(2), 400–412
(2015)
15. Pintea, C.-M., Chira, C., Dumitrescu, D.: Sensitive ants: inducing diversity in the colony.
Stud. Comput. Intell. 236, 15–24 (2008)
16. Pintea, C.-M., Chira, C., Dumitrescu, D., Pop, P.C.: Sensitive ants in solving the generalized
vehicle routing problem. Int. J. Comput. Commun. 6(4), 734–741 (2011)
17. Pintea, C.-M., Pop, P.C.: Sensor networks security based on sensitive robots agents. A
conceptual model. In: Conference CISIS, Czech Republic, vol. 89, pp. 47–56 (2012)
18. Pintea, C.-M., Ludwig, S.A., Crisan, G.-C.: Adaptability of a discrete PSO algorithm applied
to the traveling salesman problem with fuzzy data. In: FUZZ-IEEE, pp. 1–6 (2015)
19. Pintea, C.-M.: A unifying survey of agent-based approaches for equality-generalized traveling
salesman problem. Informatica. 26(3), 509–522 (2015)
20. Pop, P.C., Matei, O., Sabo, C.: A new approach for solving the generalized traveling salesman
problem. In: International Workshop on Hybrid Metaheuristics, pp. 62–72 (2010)
21. Pop, P.C., Matei, O., Sitar, C.P.: An improved hybrid algorithm for solving the generalized
vehicle routing problem. Neurocomputing 109, 76–83 (2013)
22. Popescu-Bodorin, N., Balas, V.E.: Fuzzy membership, possibility, probability and negation
in biometrics. Acta Polytechnica Hung. 11(4), 79–100 (2014)
 A Fuzzy Approach of Sensitivity for Multiple Colonies 95

23. von Neumann, J.: Probabilistic logics and the synthesis of reliable organisms from unreliable
components. In: Automata Studies, pp. 43–98 (1956)
24. Balas-Timar, D.V., Balas, V.E.: Ability estimation in CAT with fuzzy logic. In: IEEE ISCIII
2009, pp. 55–62 (2009)
25. Software ACO. http://iridia.ulb.ac.be/mdorigo/ACO/aco-code/public-software.html
26. Stutzle, T., Hoos, H.H.: MAX-MIN ant system. Future Gen. Comput. Syst. 16, 889–914
(2000)
27. TSPLibrary. http://comopt.ifi.uni-heidelberg.de/software/TSPLIB95/
28. Warneke, B., et al.: Smart dust: communicating with a cubic-millimeter. Computer 34, 44–
51 (2001)
29. Yordanova, S., Merazchiev, D., Jain, L.: A two-variable fuzzy control design with application
to an air-conditioning system. IEEE T. Fuzzy Sys. 23(2), 474–481 (2015)
30. Zadeh, L.A.: Fuzzy Sets. Inf. Control 8, 338–353 (1965)
 Nature Inspired Partitioning Clustering
Algorithms: A Review and Analysis

Behzad Saemi1, Ali Asghar Rahmani Hosseinabadi2(&),

Maryam Kardgar2, Valentina Emilia Balas3, and Hamed Ebadi4
1
Computer Department, Kavosh Institute of Higher Education,
Mahmood Abad, Mazandaran, Iran
[email protected]
2
Young Researchers and Elite Club, Ayatollah Amoli Branch,
Islamic Azad University, Amol, Iran
[email protected],
[email protected]
3
Aurel Vlaicu’ University, Bd. Revolutiei 77, 310130 Arad, Romania
[email protected]
4
Department of Electrical and Electronics,
Knowledge-Based Institute of Technology, Asre- Enghelab Complex,
5179895848 Tabriz, Iran
[email protected]

Abstract. Clustering algorithms are developed as a powerful tool to analyze

the massive amount of data which are produced by modern applications. The
main goal of these algorithms is to classify the data in clusters of objects, so that
data in each cluster is similar based on specific criteria and data from two
different clusters be different as much as possible. One of the most commonly
used clustering methods is partitioning clustering method. So far various par-
titioning clustering algorithms are provided by researchers, among them
inspiring the nature algorithms are the most popular used algorithms. In this
paper some partitioning clustering algorithms inspiring by nature are described,
and then these algorithms are compared and evaluated based on several stan-
dards such as time complexity, stability and also in terms of clustering accuracy
on real and synthetic data sets. Simulation results have shown that combina-
tional methods have good influence to increase the efficiency of algorithms and
also the use of different operators can maintain population diversity and cause to
reach a good answer in a reasonable time.

Keywords: Partitioning clustering
Inspiring by the nature algorithms

Stability
Hybrid

1 Introduction

Clustering algorithms are developed as a powerful tool to analyze the massive amount
of data which are produced by modern applications. The main goal of these algorithms
is to classify the data in clusters of objects, so that data in each cluster is similar based
on specific criteria and data from two different clusters be different as much as possible.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_9
 Nature Inspired Partitioning Clustering Algorithms 97

Clustering algorithms can be divided into several categories, including partitioning

Clustering algorithms, hierarchical algorithms and algorithms based on density. In
partitioning algorithms clustering is based on the objective function, this means that a
set of data is divided to k categories and then by using a repetitive motion technique,
objects move from on cluster to another cluster to improve the clustering. Many
clustering methods have been introduced so far. K-Means algorithm is one of the most
important and widely used algorithms in clustering issue, and is known due to the
simplicity and rapid implementation for large test data set [1].
In clustering problems it is not possible using full search method, due to challenges
such as define a proper objective function, and ambiguity in data clustering definition
(so that clusters have a good quality) and also because of the large search space. That’s
why heuristic techniques are designed by the aim of achieving optimal or near-optimal
solutions [2].
The structure of this paper is as follow, in Sect. 2 data clustering problem is
presented. Section 3 groups the inspired by the nature algorithms and shows the fea-
tures of heuristic algorithms used by researchers and also investigates the history of
methods based on them used for data clustering. In last section these algorithms are
compared in terms of time complexity, stability and reliability, as well as clustering
accuracy on actual data sets from UCI repository. With the hope that specialists in this
area have better view to each type of these algorithms for data clustering. In Sect. 4
these algorithms are evaluated in terms of time complexity, stability and reliability and
also data clustering accuracy on real data test of UCI repository to give better view to
experts about theses algorithms. We show results of simulations in Sect. 5, and con-
clude the work in Sect. 6.

2 Data Clustering Problem Statement

Data clustering is one of the important tools in data mining, machine learning and
pattern classification [3]. The purpose of clustering is to divide existing data into
several groups so that data from different groups are different as much as possible and
data within one group are similar. In partitioning clustering which is called objective
function based clustering, N record of data is divided into k class such a way that each
class represents a cluster and k <= n. partitioning clustering algorithm starts a primary
classification according to k (number of clusters) and try to move objects from on
cluster to another cluster by using a repetitive motion technique to improve categories.
This improvement is possible by minimize an objective function (cost function, error
function). In clustering problem Euclidean distance is used as a measurement of
similarity to classify n points in k categories in search space Rn. Equation (1) shows the
amount of similarity of difference between to data x and y. in this equation m represents
the feature or dimensions of target data.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xm ffi
2
d ðx; yÞ ¼ i¼1
ðx i  y i Þ ð1Þ
98 B. Saemi et al.

3 Meta-Heuristic Algorithms for Data Clustering

Nature is a great source of inspiration with powerful and complex phenomena.

Meta-heuristic algorithms inspired by nature given into various branches of science and
industry because of importance of optimization. These algorithms are smart tech-
nologies which mimics the nature in order to solve optimization problems and create a
new method for computer calculation Fig. 1 shows the classification of inspired by
nature algorithms provided by researchers.

Fig. 1. Classification of inspired by nature algorithms provided by researchers

Figure 1, shows the classification of the most commonly used algorithms in data
clustering as standard algorithms. Each algorithm has its own pattern to solve clustering
problem. The most important steps in designing a heuristic algorithm which always
increases the complexity of the algorithm and also speed of reaching to optimal
solutions, is to find a suitable method to represent candidate solutions. Candidate
solution represents a solution for optimization problem. Performance and complexity of
evolutionary search algorithms is highly dependent on display candidate solutions.
Various heuristic methods use different way to display solutions and most algorithms,
display solutions as a vector of a specific data type. Three of the most common
methods to display solutions in inspired by the nature algorithms for data clustering
problem is as follow:
• Cluster center based display
In this way, each solution represents the cluster center and that is used when it is
needed to search for the best cluster center. The procedure is either represent each
solution as a vector of size n*k (n is dimensions or number of features and K is
number of clusters and the contents is center of a cluster) [9] or a two dimensional
array in which rows include data as centers. For example Fig. 2 shows a sample of
this way.
 Nature Inspired Partitioning Clustering Algorithms 99

Fig. 2. Cluster center based display

There are 3 clusters in Fig. 2, and data dimensions are as 3 features. And array size
is equal 9. For example center of the first cluster represents data with features 2.3,
3.4, 4.5. A difficulty of this method is array size. To solve that we can numbering all
input data sets and then use a one dimension array with the length of the number of
centers to represent solutions. Content of each cell is centers of clusters [10].
Figure 3 shows an example of this method by 3 cluster centers.

Fig. 3. Cluster center based display based on data numbering

• Display based on two statuses 0, 1

In this method, each solution is displayed as a one-dimensional with the size of data
length, index represents data number. Contents of each array are zero or one. One
means that given data is cluster center. This way doesn’t need to predetermine the
number of clusters and it can be varied. To create a solution in this way, first 0 and 1
are putted in each cell randomly, then data that have value 1 are chosen as cluster
center and remain data transferred to a suitable cluster based on a proper objective
function [11]. Figure 4 shows a sample of this method. One of the outstanding
characteristics of this method is its simplicity and the way of updating.

Fig. 4. Display based on two statuses 0, 1

• Display as two parts

This method uses a two-part one-dimensional array, first part is for members and
second part is for groups. In member parts, array index represents data number and
its content which shows that target data corresponds to which group. Contents of
group part is group number and its size can be dynamic [12]. Figure 5 shows this
method.
100 B. Saemi et al.

Fig. 5. Display as two part

A good feature of this method is its simplicity and low complexity because it is
easy to determine that each data correspond to which group, so it is good for problems
that need dynamic grouping.
Each of these methods can be used any heuristic algorithms inspired by nature. It is
essential to choose the best display method based on n the characteristics of a heuristic
algorithm and the way of updating in order to reduce the complexity and increase the
capabilities of the search algorithm. The characteristics of each inspired by nature
algorithm are investigated in the following sub-sections.

3.1 Evolutionary Meta-Heuristic Algorithms

Evolutionary meta-heuristic algorithm, as the most intelligent search methods, uses
natural revelation in its exploration in an irregular search space [13]. This methods use
evolutionary theory of Charles Darwin. Genetic algorithm is one of the most widely
used meta-heuristic algorithms for data clustering.

3.1.1 The Genetic Algorithm (GA)

The first algorithm to simulate gene-based system was Genetics algorithm (GA), which
was proposed by Fraser [14, 15] at first, and then by Bremermann [16] and Reed [17].
Holland valuable work [18] had led to popular GA. GA models gene evolution, such a
way that uses Genotype to show the attributes of the organisms. Main operations of GA
are: crossover and mutation operations to produce new generation. Figure 6 shows the
pseudo code of genetic algorithm for data clustering. Many studies using GA have been
conducted for data clustering so far. The following will introduce some of these studies.

Fig. 6. Pseudo code of Genetic algorithm for data clustering

 Nature Inspired Partitioning Clustering Algorithms 101

In [19] GA combined with K-means was used for data clustering. The aim was to
choose the best center using GA and put data into proper cluster using K-means.
Chromosome represented as a two-dimensional array and content of each array was
cluster center. In [9] each solution is represented as a one-dimensional array of size
n  k. In which n reflects number of dimensions or features and k reflects number of
clusters. The contents of each array were a string represents the cluster center. To
evaluate the performance of proposed algorithm, first, data is transferred to a cluster
center that has less distance from that. Then in second stage, new cluster centers are
calculated and cluster centers replaced with an average of each cluster. In [12] A
Grouping genetic algorithm is proposed for data clustering. In fact it is a modified
algorithm based on Genetic Algorithm for problems need grouping and classification.
Solutions are coded as a one-dimensional array which contains two parts, first part
defines the members of each group and the second specifies group number. In this
coding, the length of each part can be varied and it is possible to determine the best
number of categories by using an appropriate objective function. The main charac-
teristic of this algorithm was to use local search and island model for chromosome
migration. In order to migrate, the best person is selected in each island at first, and this
person is replaced on a chromosome in another island randomly. The main objective of
this procedure is to escape from the local optimum and diversity.
In [20] GA combined with K-means was used for large scale data clustering. Each
solution is represented as a one-dimensional array. The contents of each array were
cluster centers. After generate a child from crossover and mutation operations,
K-means algorithm is executed to do the best clustering. In [21] a set of centers is
created by using hyper-quad tree algorithm to fix the problem of k-means algorithm.
One feature of this method is its suitable coding for problems that have dynamic and
varied number of clusters.

3.2 Collective Intelligence Meta-Heuristic Algorithm

These types of algorithms are inspired by social creatures’ behavior. And they try to
understand Social creatures’ search mechanism by examining the collective intelli-
gence and communication methods that they use to reach their goals. After finding this
mechanism, will be tried to provide a model to solve the mentioned problem. Here are
some algorithms used for data clustering.

3.2.1 The Ant Colony Optimization (ACO)

Ant colony or more generally insect societies are distributed systems that despite the
simplicity of the colonies create a complex social organization. This organization is able
to perform complex tasks which are beyond the ability of a single member. Ant Colony
Optimization Model is formed based on the behavior of real ants. This model is as a
source of inspiration to design a new algorithm to solve optimization problems. Ant
colony algorithm was proposed by Dorigo et al. [22] as a multi-agent solution for opti-
mization problems such as traveling salesman problem and then a variety of algorithms
have been developed by studying various behaviors like food seeking, task division, build
nests, and so on. In all examples, ants coordinate their activities through stigmergy.
102 B. Saemi et al.

Stigmergy is an indirect communication which is done by change on the environment

[22]. Figure 7 shows the pseudo code of ant colony algorithm for data clustering.

Fig. 7. Pseudo code of ant colony algorithm for data clustering

In [23] Ant swarm chaos algorithm Have been used for data clustering. Proposed
algorithm could find a global optimum solution based on the objective function and is
not sensitive to the size and density of clusters and also can be used for
multi-dimensional data. Each ant represents the gravity of cluster center. Each ant
chooses one data as cluster center based on a probability. After finding cluster centers,
data are transferred to a proper cluster based on an objective function.

3.2.2 The Artificial Bee Colony Algorithm

Many studies have been done on Artificial Bee Colony for optimization problems and
different versions of it were proposed. Honey bee performs nectar finding as a social
task to produce honey and use a complex communication system for nectar finding. In
the communication system, bee broadcasts the information about distance and quality
of food resources to inform other insects intended to seek the nectar of flowers. Life of
Bees shows that they have a special communication language named language of dance
to communicate with each other [24]. Artificial Bee Colony optimization was proposed
by Abadeh for the first time in 2008 [25]. Figure 8 shows the pseudo code of Artificial
Bee Colony algorithm for data clustering.
In [26] Artificial Bee Colony was used for data clustering, and one-dimensional
array was used for solution representation. Each row of array indicates a cluster and its
content is cluster center. Row length is equal to the number of dimensions or features.
And after finding cluster centers, other data will be transferred to the best cluster based
on an objective function. [27] Tries to escape from local optimum in Artificial Bee
Colony optimization by using a mutation operation of Genetic algorithm. And new
solutions generated using genetic operations. Artificial Bee Colony was used for data
clustering in [28] with that aim of finding the best cluster centers.
 Nature Inspired Partitioning Clustering Algorithms 103

Fig. 8. Pseudo code of Artificial Bee Colony algorithm for data clustering

3.2.3 The Particle Swarm Optimization (PSO)

Particle Swarm Optimization Algorithm is a family of collective intelligence methods
and also one of the most successful algorithms in continuous and discrete optimization.
This method was introduced for the first time in 1995 by inspiring form collective
behavior of birds and fish, and evolutionary algorithms concepts [29, 30]. Similar to
evolutionary algorithms, particle swarm optimization algorithm is a population based
algorithm in which particles that are candidate solutions for a function or a problem,
constitute a swarm (population).
These particles move in problem space and try to find an optimal solution in search
space based on their personal experiences and collective experiences. This method is
not affected much by the size and nonlinear feature of the problem and had shown good
results in static, noisy environments, and also continuous environments. These features,
plus ease of implementation, no need to be continuous for objective function and the
ability to adapt to a dynamic environment, make the algorithm applicable for different
areas. Originally there are two PSO algorithms named gbest and ibest, which are
different in neighbor sizes [31, 32]. In gbest whole swarm is considered as neighbor for
each particle. The social component of velocity updating is obtained according to the
information obtained from Particle swarm. In ibest smaller neighbor is defined for each
particle, and social information Component, includes local knowledge of environment.
Figure 9 shows the pseudo-code of particle swarm optimization algorithm.
104 B. Saemi et al.

Fig. 9. Pseudo code of Particle Swarm Optimization algorithm for data clustering

Much researches based on particle swarm optimization have been done for data
clustering so far. In [33] data clustering was based on modified particle optimization
algorithm, and proposed algorithm includes two features: 1. Set parameters for rapid
convergence, and 2. Reconstruction or modification of particles to escape from local
optimum. Particle modification is done such a way that particles that are stuck in local
optimum are chosen based on a certain rate and their position is changed randomly. In
fact their method is based on to move toward to gbest or mutation operation in Genetic
algorithm. This method works as follow, distance from particle to gbest is calculated at
first, and then d% of particle have smallest distance to best particles are chosen and
then f% of them are chosen for reconstruction. In next step the position of each particle
such as how to make each particle will be changed far from particle modeling. Since
every particle in this step may lose some good features, so to solve this problem, in next
step, target particle take some features from gbest and then its position changed again
and in last step particle velocity will be updated. In [34] a combination of particle
swarm algorithm, k-means algorithm, and Nelder-Mead algorithm was used for data
clustering. The aim was to take advantages of global search of particle swarm algo-
rithm, and optimized search ability of Nelder-Mead optimization algorithm. In [35]
particle swarm optimization had been used for data clustering. Their objective function
was based on distance of each member within a cluster. The less distance would be
better. in proposed algorithm, particle swarm optimization algorithm selects the best
centers at first, and then k-means algorithm executes to select member for each cluster.

3.2.4 The Imperialist Competitive Algorithm (ICA)

Imperialist Competitive Algorithm provided by Atashpaz in 2007 as a meta-heuristic
search method which is inspired of Human social phenomenon [36]. Particularly, this
algorithm sees the imperial process as a human social and political evolution and uses it
as a source of inspiration in optimization algorithms by mathematical modeling of this
 Nature Inspired Partitioning Clustering Algorithms 105

historical phenomenon. Like other evolutionary optimization methods, Imperialist

Competitive Algorithm starts with initial population. Each element of population is a
country. Countries are divided into two categories: imperialism and colony. Each
imperialism controls a number of colonies depending on his strength. Absorption policy,
Imperialist Competitive, and Revolution are the main elements of this algorithm. Fig-
ure 10 shows the Pseudo code of Imperialist Competitive Algorithm for data clustering.

Fig. 10. Pseudo code of Imperialist Competitive Algorithm for data clustering

In recent years, Imperialist Competitive Algorithms have been used as a new

algorithm for data clustering by researchers. A combination of harmonic k-means
algorithm and Imperialist Competitive Algorithm was used in [37] for data clustering.
In fact harmonic k-means algorithm is used to solve the problem of being sensitive to
the initial centers in k-means algorithm. As well as mutation and crossover operations
of Genetic algorithm is used for better global search. In mutation operation, a colony of
an imperialist is moved to imperialism. In crossover operation two colonies of two
imperialisms are switched. Another feature of this method is to perform harmonic
k-means algorithm in each imperialism. The objective function was based on the
maximum similarity between internal elements and minimum similarity between
external elements. In [38] an effective combination algorithm based on Imperialist
Competitive Algorithm and k-means algorithm was proposed for data clustering. The
aim was to find the best cluster center by using Imperialist Competitive Algorithm.
Each country represented as a one-dimensional array and includes cluster centers. In
this method, after generation an initial population including countries and imperialisms,
K-means algorithm will be performed on each imperialism and colonies for each
imperialism are selected. This procedure is executed for each colony. And then
Imperialist Competitive Algorithm will be executed by applying Absorption and rev-
olution operations with the aim of finding better members for each cluster based on the
objective function. In [39] an extension of Imperialist Competitive Algorithm is used
for data clustering. This study showed that one problem of Imperialist Competitive
Algorithm is angle in absorption operator, which is visible in two-dimensional
106 B. Saemi et al.

problems, but n-1 angels are needed for multi-dimensional problems while its execu-
tion is difficult, and Imperialist Competitive optimization is introduced to solve the
mentioned problem. This concept includes colonies trends to their imperialisms about
public dissatisfaction about the existing policy. In other words, this concept is
Equivalent to the initial velocity in particle swarm optimization. This concept, which
includes two states: public dissatisfaction and trend to imperialism, caused to form an
equation to update a country in each imperialism.
In [40] a combination of Imperialist Competitive Algorithm and k-means algorithm
was proposed for data clustering too. In order to avoid rapid convergence in Imperialist
Competitive algorithm a mutation operator was used. After executing Imperialist
Competitive Algorithm in a certain repetition, the best solution of Imperialist Com-
petitive Algorithm is transferred to simulated annealing algorithm and this algorithm
tries to improve the solution. To perform mutation operation in K-means algorithm,
initial imperialism generation would be performed for all imperialisms and cluster
centers for colonies selected and the output would be the beginning of Imperialist
Competitive Algorithm.

3.3 Abiotic Meta-Heuristic Algorithms

Generally these kinds of algorithms are modeled from physics. These algorithms can be
with memory or without memory and a set of rules may be used during search. In the
following Tabu Search method and Gravitational Search Algorithm will be investigated
as Abiotic Meta-heuristic algorithms for data clustering.

3.3.1 The Tabu Search Algorithm (TS)

Tabu Search is meta-heuristic global search algorithm and based on neighborhood
search which is proposed by Glover in 1989 [41]. In this method the function of human
memory is simulated. Human memory stores what it had been seen by using an
effective structure while simple of information. And also sets a list of forbidden actions.
This list is formed according to the latest searches. This method prevents to perform
any operations repeatedly. The aim of Tabu Search method is to consider not men-
tioned solutions. So move toward recently searched solutions would be forbidden.
Figure 11 shows the pseudo code of Tabu Search Algorithm for data clustering.

Fig. 11. Pseudo code of Tabu Search Algorithm for data clustering
 Nature Inspired Partitioning Clustering Algorithms 107

In [42] Tabu Search Algorithm is used for data clustering. The main characteristic of
this method is the variable number for k (number of clusters). And three indexes have
been used for cluster validation: 1. When data collection is separable, it is expected that
the distance between clusters be large and diameter be small. 2. Davies Bouldin Index: is
defined based on cluster dispersion and non-similarity between two clusters. 3. PMA
index: the higher index, members of cluster is compressed and distance between clusters
is larger. In this method an one-dimensional array is used for solution representation.
Array index indicates data number, the contents of each array is zero or 1. Value 1
indicates that this data is cluster center; the number of clusters can be vary.
Researchers in [43] used K-harmonic algorithm and tabu search algorithm for data
clustering. The aim was to discover K-harmonic drawbacks by using tabu search
method. Two random exchange strategies were used to select cluster centers and
change the center based on validation function.

3.3.2 The Gravitational Search Algorithm (GSA)

Gravitational Search Algorithm is proposed by Rashedi in 2007 [44], this algorithm is
based on Newton’s law of gravity [45, 46]. In gravity law, each particle is absorbed by
other particles with a force that have direct proportional to the product of masses and
reverse proportional to the square of the distance. Gravitational Search Algorithm is
introduced by simulating gravity law and Newton’s laws in an environment with a
discrete time in search space [47–49]. In this method optimization is done by using a
design of gravity laws and move in an artificial system with discrete time. System
environment is similar to a problem space. According to gravity law, each agent (mass)
recognizes location and status of other masses through gravitational attraction. So this
force can be used as a means to exchange information [50–52]. Figure 12 shows the
pseudo code of Gravitational Search Algorithm for data clustering.

Fig. 12. Pseudo code of Gravitational search algorithm for data clustering

A combination of k-means algorithm and gravitational search algorithm was used

for data clustering [53]. Each mass represented as a one-dimensional array with the size
of the number of clusters. The aim was to generate the target population of masses.
108 B. Saemi et al.

Initial population is created based on the nature of input data and minimum, maximum
and average values of them and also output of K-means algorithm. In [54] a grouping
algorithm based on gravitational search algorithm is proposed in which mass repre-
sentation is based on a two part one-dimensional array. The first part includes members
and the second part indicates groups. To create solutions, group is determined at first
and then members are allocated to the groups randomly.

4 Select Appropriate Algorithms from Each Group

for Evaluation

In this section to evaluate and compare inspired by the nature algorithms, we select the
most widely used algorithms to solve optimization problems, especially for data
clustering from each group. Genetic algorithm (GA) from meta-heuristic algorithms
group, imperialism Competitive Algorithm (ICA) and particle swarm optimization
(PSO) from collective intelligence Meta-heuristic algorithms group, Tabu search
(TS) from Abiotic Meta-heuristic algorithms and also combine algorithm based on GA
and PSO (GA-PSO) are selected for evaluation.

4.1 Evaluation Parameters

Various criteria are considered to provide a more accurate evaluation on each algo-
rithm. These criteria will be introduced in the following.
• Clustering accuracy criteria: As mentioned in the previous sections, a clustering
algorithm must divide data into categories that Members within each category have
the maximum similarity and external members have maximum difference. Clus-
tering accuracy is obtained as Eq. 2:
Xk X q
EP ¼ DistanceðX; lðci ÞÞ þ Pk1 Pk   
i¼1 Xci
i¼1 j¼i þ 1 Distance lðci Þ; l cj
ð2Þ

As it is clear in Eq. (2), clustering accuracy is calculated based on both internal and
external distance. ci is ith cluster, lðci Þ indicates center of ith cluster.
DistanceðX; lðci ÞÞ is the distance between each object from its center and
Dstance lðci Þ; l cj is the distance between center of ith cluster and center of jth
cluster. q is an input array which shows the importance of considering external
distance. The less EP, is the more accuracy and the solution will be more efficient.
• Stability criteria: One of the features that each inspired by the nature algorithm
must have, is independent reliability for each execution. To evaluate stability, we
perform algorithms on data sets and then compare the results with the best, worse
and average solutions. And we evaluate the standard deviation of each algorithm for
each data sets o show the solution dispersion. The less standard deviation, the more
reliability.
 Nature Inspired Partitioning Clustering Algorithms 109

• Time complexity: in this case, algorithms are compared in terms of the best exe-
cution time to reach an optimal clustering. The time complexity of an algorithm
would be less if it has lees execution time.
Figure 13 shows the pseud code to evaluate the selected algorithms for data
clustering. As it seen, the objective function is based on to minimize the internal
distance and maximize external distance. Termination condition is based on conver-
gence condition. If algorithms couldn’t improve the best solution in a certain iterations,
after reaching to a converge iterations end.

Fig. 13. Pseudo code to evaluate data clustering algorithms.

5 Simulation Results

Selected algorithms were implemented using C++ programming language and ran on a
personal computer with 3 GHz Pentium-IV processor and 2 GB RAM. And 5 data set
are provided to test all features of algorithms: one Synthetic data set and four real data
sets are provided from UCI repository. Table 1 shows the features of each data set.
110 B. Saemi et al.

Table 1. Features of each data sets for data clustering algorithms

Data Number of Number of Number of data Repetitive
test clusters features convergence
Soybean 4 35 47 (10,10,10,17) 100
Iris 3 4 150 (50,50,50) 100
Wine 3 13 178 (59, 71,48) 200
Glass 6 9 214 (70, 76,17, 13, 9,29) 300
Art 4 2 300 (75,75,75,75) 300

As mentioned above, for a more detailed evaluation on each data test, each algo-
rithm was run 10 times. Table 2 shows the results of each algorithm based on internal
distance within clusters.

Table 2. Results of each algorithm based on internal distance within clusters

Data Test Evaluation Internal distances in clusters
GA PSO ICA TS GA-PSO
Soybean best 99.992 99.912 103.432 108.322 96.486
Worst 101.832 102.932 105.527 111.435 101.686
Average 100.1 100.3 104.4 109.6 99.1
Standard deviation 0.83 1.18 0.66 1.2 1.57
Iris best 97.222 98.351 98.36 102.363 97.222
Worst 98.342 104.324 103.324 106.435 99.651
Average 101.5 100.7 100.7 104.2 97.9
Standard deviation 1.62 2.05 1.67 1.66 0.7
Wine best 16392.354 16453.362 16530.573 16931.178 16424.352
Worst 16472.435 16551.475 16721.439 17123.981 16458.932
Average 16432.1 16515.4 16630.7 17071.3 16442.9
Standard deviation 24.05 30.11 52.96 27.65 9.73
Glass best 217.302 223.763 218.794 222.244 214.69
Worst 224.353 232.849 225.235 229.543 219.643
Average 220.6 227.2 220.1 224.2 215.4
Standard deviation 2.37 2.13 2.38 1.41 1.28
Art best 15721.946 15721.946 15721.946 15721.946 15721.946
Worst 15723.235 15725.947 15725.584 15724.679 15722.436
Average 15721.9 15722.8 15722.7 15722.6 15721.5
Standard deviation 0.83 1.31 1.34 1.10 0.50

Table 2 shows that the hybrid algorithm GA-PSO always do better than other
compared algorithms. And also GA is better than PSO, ICA and TS. As well as the
hybrid algorithm GA-PSO has better average than compared algorithms. TS have better
standard deviation and reliability than ICA, PSO and GA. But in most cases have lower
 Nature Inspired Partitioning Clustering Algorithms 111

performance than other algorithms. Generally simulation results show that

population-based algorithms such as GA, ICA, and PSO have better performance due
to have various solutions and various operations.
The hybrid algorithm GA-PSO, preserved population diversity and discover the
advantages of global search in genetic algorithm and also search ability of PSO and
cause to obtain better solutions. In the following we evaluate the results in terms of the
speed to reach the optimal solution. Figure 14 shows the best execution time for all
algorithms on data sets of Table 1.
Figure 14 shows that PSO needs more execution time in most results. It means that
PSO needs more time to reach its optimal solution. The hybrid algorithm GA-PSO has
better execution time compared to other algorithms and could obtain better solution in a
reasonable time. So one solution based algorithms such as TS are better than population
based algorithms, but they failed to reach an acceptable solution. The hybrid algorithm
GA-PSO leads to investigate more solutions due to operation diversity and obtain
optimal solution in a reasonable time.

Fig. 14. Best execution time on data sets of Table 1

In the following we evaluate algorithms in terms of convergence rate to reach

optimal solution. Figure 15 shows the merit chart in each iteration of each algorithm on
data set Iris.
Figure 15 shows that convergence rate of genetic algorithm is very low and TS has
more convergence rate. GA-PSO performs better in initial iterations than other algo-
rithms. And obtained better solutions in each iteration. Figure 16 shows the merit chart
in each iteration of each algorithm on data set Wine.
112 B. Saemi et al.

Fig. 15. Merit chart in each iteration of each algorithm on data set Iris

Figure 16 shows that PSO algorithm has good convergence rate in the early, but in
the next iterations couldn’t improve the best solutions. Also convergence rate of
imperialist competitive algorithm is lower than compared algorithms. Figure 17 shows
the merit chart in each iteration of each algorithm on data set Glass.

Fig. 16. Merit chart in each iteration of each algorithm on data set Wine
 Nature Inspired Partitioning Clustering Algorithms 113

Fig. 17. Merit chart in each iteration of each algorithm on data set Glass

Figure 17 shows that ICA algorithm has a low convergence rate and TS algorithm
has a good convergence rate. But GA-PSO algorithm always obtained better solutions
in earlier iterations
After investigating merit charts, we conclude that using various operations and
increase diversity in the population cause to obtain optimal solutions in earlier itera-
tions. And TS due to investigate more neighbors of a solution, has a good convergence
rate in earlier iterations but in next iterations couldn’t improve the best solution.

6 Conclusion and Future Work

Clustering algorithms are developed as a powerful tool to analyze the massive amount
of data which are produced by modern applications. The main goal of these algorithms
is to classify the data in clusters of objects, so that data in each cluster is similar based
on specific criteria and data from two different clusters be different as much as possible.
Various meta-heuristic searching methods with the aim to achieve optimal or
near-optimal solutions are proposed for data clustering problem due to challenges such
as define a suitable objective function, ambiguity in data clustering definition. One of
the most efficient algorithms are inspired by nature algorithms. These algorithms
explore a wide search space of a complex optimization problem by inspiring from a
natural phenomenon. The purpose of this paper is to compare and evaluate inspired by
the nature algorithms proposed by researchers for partitioning data clustering. At first
we described the main meta-heuristic algorithms for data clustering and then choose the
best ones for comparison and evaluation in terms of criteria such as clustering accuracy,
stability, time complexity, and convergence rate. The result shows that population-
based algorithms are always better than one population algorithms for data clustering.
114 B. Saemi et al.

The hybrid heuristics increase the speed to reach an optimal solution in a reasonable
time due to suitable search in problem space. Despite existence of different approaches
to solve data clustering problem, it is needed to optimize clusters and increase accuracy
of data clustering, and also it is important to use high rate algorithms and different
features for dynamic environments with a lot of data.
In this field some challenges should be overcome. In the following some directions
to improve data clustering algorithms are provided as future work.
1. One of the most important problems in heuristic algorithm is solution diversity. In
simulation results, it is concluded that the use of various populations cause to
optimum data clustering. It is needed to provide appropriate solutions in order to
increase solution diversity in a population, and also providing a proper stricter to
represent solutions and use various operations to update solutions can increase the
population diversity in heuristic algorithms.
2. One of the results in simulation section is that the use of searching ability of two
algorithms can cause to reach an acceptable solution in a reasonable time and
increase stability. Our future work will consider hybrid algorithms.
3. the most important problem in heuristic algorithms is facing a large amount of data.
Our proposal is to use parallel evolutionary algorithms to increase the algorithms
speed when facing a large amount of data.

References
1. Jajuga, K., Sokolowski, A., Bock, H.H. (eds.): Classification, Clustering, and Data Analysis:
Recent Advances and Applications. Springer Science & Business Media, Heidelberg (2012)
2. Das, S., Abraham, A., Konar, A.: Metaheuristic Clustering. Studies in Computational
Intelligence, pp. 1–266. Springer (2009). ISBN 978-3-540-92172-1
3. Garey, M.R., Johnson, D., Witsenhausen, H.: The complexity of the generalized Lloyd-max
problem (corresp.). IEEE Trans. Inf. Theor. 28(2), 255–256 (1982)
4. Xu, R., Wunsch, D.: Survey of clustering algorithms. IEEE Trans. Neural Netw. 16(3), 645–
678 (2005)
5. MacQueen, J.: Some methods for classification and analysis of multivariate observations. In:
Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability,
vol. 1, No. 14, pp. 281–297, June 1967
6. Selim, S.Z., Ismail, M.A.: K-means-type algorithms: a generalized convergence theorem and
characterization of local optimality. IEEE Trans. Pattern Anal. Mach. Intell. 1, 81–87 (1984)
7. Li, H.: Study on subspace clustering algorithm (2014)
8. Huang, Z.: A fast clustering algorithm to cluster very large categorical data sets in data
mining. In: DMKD (1997)
9. Maulik, U., Bandyopadhyay, S.: Genetic algorithm-based clustering technique. Pattern
Recogn. 33(9), 1455–1465 (2000)
10. Chatterjee, S., Mukhopadhyay, A.: Clustering ensemble: a multiobjective genetic algorithm
based approach. Procedia Technol. 10, 443–449 (2013)
11. Liu, Y., Wu, X., Shen, Y.: Automatic clustering using genetic algorithms. Appl. Math.
Comput. 218(4), 1267–1279 (2011)
 Nature Inspired Partitioning Clustering Algorithms 115

12. Agustı, L.E., et al.: A new grouping genetic algorithm for clustering problems. Expert Syst.
Appl. 39(10), 9695–9703 (2012)
13. Bezdek, J.C., Keller, J., Krishnampuram, R., Pal, N.R.: Fuzzy Models and Algorithms for
Pattern Recognition and Image Processing. Kluwer Academic Publishers, Dordercht (1999)
14. Fraser, A.S.: Simulation of genetic systems by automatic digital computers VI. Epistasis.
Austr. J. Biol. Sci. 13(2), 150–162 (1960)
15. Barker, J.S.F.: Simulation of genetic systems by automatic digital computers. III. Selection
between alleles at an autosomal locus. Austr. J. Biol. Sci. 11, 603–612 (1958)
16. Bremermann, H.J.: Optimization through evolution and recombination. Self-organ. Syst. 93,
106 (1962)
17. Reed, J., Toombs, R., Barricelli, N.A.: Simulation of biological evolution and machine
learning: I. Selection of self-reproducing numeric patterns by data processing machines,
effects of hereditary control, mutation type and crossing. J. Theor. Biol. 17(3), 319–342
(1967)
18. Holland, J.H.: Adaptation in natural and artificial system: an introduction with application to
biology, control and artificial intelligence. University of Michigan Press, Ann Arbor (1975)
19. Rahman, M.A., Islam, M.Z.: A hybrid clustering technique combining a novel genetic
algorithm with K-Means. Knowl. Based Syst. 71, 345–365 (2014)
20. Laszlo, M., Mukherjee, S.: A genetic algorithm that exchanges neighboring centers for
k-means clustering. Pattern Recogn. 28(16), 2359–2366 (2007)
21. Laszlo, M., Mukherjee, S.: A genetic algorithm using hyper-quadtrees for low-dimensional
k-means clustering. IEEE Trans. Pattern Anal. Mach. Intell. 28(4), 533–543 (2006)
22. Dorigo, M., Birattari, M., Stützle, T.: Ant colony optimization. IEEE Comput. Intell. Mag.
1(4), 28–39 (2006)
23. Wan, M., Wang, C., Li, L., Yang, Y.: Chaotic ant swarm approach for data clustering. Appl.
Soft Comput. 12(8), 2387–2393 (2012)
24. von Frisch, K.: The Dance Language and Orientation of Bees. Harvard University Press,
Cambridge (1967)
25. Abadeh, M.S., Habibi, J., Lucas, C.: Intrusion detection using a fuzzy genetics-based
learning algorithm. J. Netw. Comput. Appl. 30(1), 414–428 (2007)
26. Zhang, C., Ouyang, D., Ning, J.: Intrusion detection using a fuzzy genetics-based learning
algorithm. Expert Syst. Appl. 37(7), 4761–4767 (2010)
27. Yan, X., Zhu, Y., Zou, W., Wang, L.: A new approach for data clustering using hybrid
artificial bee colony algorithm. Neurocomputing 97, 241–250 (2012)
28. Karaboga, D., Ozturk, C.: A novel clustering approach: Artificial Bee Colony (ABC) algo-
rithm. Appl. Soft Comput. 11(1), 652–657 (2011)
29. Poli, R., Kennedy, J.: Particle swarm optimization. Swarm Intell. 1(1), 33–57 (2007)
30. Kennedy, J., Eberhart, R.C.: Particle swarm optimization. In: Proceedings of the IEEE
International Conference on Neural Networks, Piscataway, pp. 1942–1948 (1995)
31. Eberhart, R.C., Shi, Y.: Particle swarm optimization: developments, applications and
resources. In: 2001. Proceedings of the 2001 Congress on Evolutionary Computation, vol. 1,
pp. 81–86. IEEE (2001)
32. Kennedy, J., Eberhart, R.C.: Swarm Intelligence. Morgan Kaufmann, San Francisco (2001)
33. Tsai, C.Y., Kao, I.W.: Particle swarm optimization with selective particle regeneration for
data clustering. Expert Syst. Appl. 38(6), 6565–6576 (2011)
34. Kao, Y.T., Zahara, E., Kao, I.W.: A hybridized approach to data clustering. Expert Syst.
Appl. 34(3), 1754–1762 (2008)
35. Zhao, M., Tang, H., Guo, J., Sun, Y.: Data clustering using particle swarm optimization. In:
Future Information Technology, pp. 607–612. Springer, Heidelberg (2014)
116 B. Saemi et al.

36. Atashpaz-Gargari, E., Lucas, C.: Imperialist competitive algorithm: an algorithm for
optimization inspired by imperialistic competition. In: IEEE Congress on Evolutionary
computation, CEC 2007, pp. 4661–4667. IEEE, September 2007
37. Abdeyazdan, M.: Data clustering based on hybrid K-harmonic means and modifier
imperialist competitive algorithm. J. Supercomput. 68(2), 574–598 (2014)
38. Niknam, T., Fard, E.T., Pourjafarian, N., Rousta, A.: An efficient hybrid algorithm based on
modified imperialist competitive algorithm and K-means for data clustering. Eng. Appl.
Artif. Intell. 24(2), 306–317 (2011)
39. Zadeh, M.R.D., Fathian, M., Gholamian, M.R.: A new method for clustering based on
development of imperialist competitive algorithm. China Commun. 11(12), 54–61 (2014)
40. Niknam, T., Fard, E.T., Ehrampoosh, S., Rousta, A.: A new hybrid imperialist competitive
algorithm on data clustering. Sadhana 36(3), 293–315 (2011)
41. Glover, F.: Tabu search-part I. ORSA J. Comput. 1(3), 190–206 (1989)
42. Pan, S.M., Cheng, K.S.: Evolution-based tabu search approach to automatic clustering. IEEE
Trans. Syst. Man Cybern. Part C Appl. Rev. 37(5), 827–838 (2007)
43. Güngör, Z., Ünler, A.: K-harmonic means data clustering with tabu-search method. Appl.
Math. Model. 32(6), 1115–1125 (2008)
44. Rashedi, E., Nezamabadi-Pour, H., Saryazdi, S.: GSA: a gravitational search algorithm. Inf.
Sci. 179(13), 2232–2248 (2009)
45. Hosseinabadi, A.R., Yazdanpanah, M., Rostami, A.S.: A new search algorithm for solving
symmetric traveling salesman problem based on gravity. World Appl. Sci. J. 16(10), 1387–
1392 (2012)
46. Hosseinabadi, A.R., Ghaleh, M.R., Hashemi, S.E.: Application of modified gravitational
search algorithm to solve the problem of teaching Hidden Markov Model. Int. J. Comput.
Sci. 10(3), 1–8 (2013)
47. Tavakkolai, H., Hosseinabadi, A.R., Yadollahi, M., Mohammadpour, T.: Using gravitational
search algorithm for in advance reservation of resources in solving the scheduling problem
of works in workflow workshop environment. Indian J. Sci. Technol. 8(11), 1–16 (2015)
48. Tavakkolai, H., Yadollahi, N., Yadollahi, M., Hosseinabadi, A.R., Rezaei, P., Kardgar, M.:
Sensor selection wireless multimedia sensor network using gravitational search algorithm.
Indian J. Sci. Technol. 8(14), 1–6 (2015)
49. Hosseinabadi, A.R., Kardgar, M., Shojafar, M., Shamshirband, S., Abraham, A.: Gravita-
tional search algorithm to solve open vehicle routing problem. In: 6th International
Conference on Innovations in Bio-Inspired Computing and Applications (IBICA 2015).
Advances in Intelligent Systems and Computing, kochi, India, pp. 93–103. Springer (2016)
50. Rostami, A.S., Bernety, H.M., Hosseinabadi, A.R.: A novel and optimized algorithm to
select monitoring sensors by GSA. In: International Conference on Control, Instrumentation
and Automation (ICCIA), pp. 829–834 (2011)
51. Hosseinabadi, A.R., Siar, H., Shamshirband, S., Shojafar, M., Nizam, M., Nasir, M.H.:
Using the gravitational emulation local search algorithm to solve the multi-objective flexible
dynamic job shop scheduling problem in Small and Medium Enterprises. In: Annals of
Operations Research, vol. 229, Issue 1, pp. 451–474 (2015)
52. Hosseinabadi, A.R., Farahabadi, A.B., Rostami, M.S., Lateran, A.F.: Presentation of a new
and beneficial method through problem solving timing of open shop by random algorithm
gravitational emulation local search. Int. J. Comput. Sci. 10(1), 745–752 (2013)
53. Hatamlou, A., Abdullah, S., Nezamabadi-pour, H.: A combined approach for clustering
based on K-means and gravitational search algorithms. Swarm Evol. Comput. 6, 47–52
(2012)
54. Dowlatshahi, M.B., Nezamabadi-pour, H.: GGSA: A grouping gravitational search
algorithm for data clustering. Eng. Appl. Artif. Intell. 36, 114–121 (2014)
 Parametric Prediction Model Using Expert
System and Fuzzy Harmonic System

W. Bel(&), D. López De Luise, A. Asmis, and D. Mansilla

CI2S Lab, Pringles 10 2nd FL, C1183ADB Buenos Aires, Argentina

{wbel,mdldl}@ci2s.com.ar, [email protected],
[email protected]

Abstract. Timing of events is a compelling subject under research. This paper

presents a model that aims to refine critical parameters in systems focused on
detection of peculiar pattern of time in parameter’s variations for predefined
events. that works in real-time. The relevant parameters are presented and the
fuzzification process is depicted. Furthermore the global architecture of the
KRONOS prototype (the one that implements the proposed model) is also
described as well as a case of study. Since a good test set is very difficult to
obtain, it remains as a work in progress task and not included as part of this
paper.

Keywords: Expert systems
Harmonic system
Fuzzy logic
Risk prediction

1 Introduction

Process Mining is generally used when there is no formal description of processes, or

when existing information has poor quality. Event logs can also be used in contrast of
probabilistic models to compare real world with the proposed model.
But mining processes is only one of the areas that are affected by the behavior in
time of sub-events. Web mining is also Important to study navigation users, predict
loads and balances of resources at the nodes of a network, etc.
Other systems typically employ temporary extensively are the DSS (Decision
Support Systems) which are often used to evaluate alternative decisions or predict the
impact of them (Tables 1, 2, 3, 4 and 5).
There are many strategies in each case to evaluate the events and their character-
istics with time. But the strategy based on “harmonic systems” [21, 22], focus of the
study described as temporary behavior patterns of events, extracting from there the
main features that identify the causes of this pattern [23–25].
Fuzzy logic is based on the extension of traditional logic to a set of values (con-
tinuous or discrete) associated with certain semantic problem. This kind of logic was
formulated in 1965 by engineer and mathematician Lotfi Zadeh [18].
Artificial Intelligence, fuzzy logic is used to solve a variety of problems, mainly
related to control of complex industrial processes and decision systems in general,
resolution and high data compression no explicit semantic content. The fuzzy logic

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_10
118 W. Bel et al.

Table 1. Weather entity

Attribute Type Unit of measurement Fuzzy pattern
temperature real °C 1
pressure real hPa 2
wind real Km/h 3
humidity real % 4
visibility real Km 5
dewpoint real °C 6
sunrise time without time zone HH:MM:SS 7
sunset time without time zone HH:MM:SS 7
weather character varying (15) – –

Table 2. Vehicle_Type entity

Attribute Type Unit of measurement Fuzzy pattern
name character varying (25) – –
width integer cm –
height integer cm –
length integer cm –

Table 3. Vehicle entity

Attribute Type Unit of measurement Fuzzy pattern
patent character varying (10) – –
color character varying (20) – –
type integer – –

Table 4. Risk_Type entity

Attribute Type Unit of measurement Fuzzy pattern
level character varying (15) – –
description character varying (100) – –

Table 5. Risk entity

Attribute Type Unit of measurement Fuzzy pattern
type integer – –
date date dd/mm/yyyy –
“time” time without time zone HH:MM:SS 7
idweather integer – –
idzone integer – –
pedestrian boolean – –
currentspeed integer Km/h 8
idperson character varying (20) – –
belt_helmet boolean – –
alcohol real g/l 9
 Parametric Prediction Model Using Expert System 119

based systems mimic the way humans make decisions, with the advantage of being
much faster. These systems are generally robust and tolerant of inaccuracies and noise
in the input data [19, 20].
It consists of the transposition of principles of human logic (imitating the imprecise
reasoning) in computer programming. With classical logic, computers can manipulate
strictly dual such values as true/false, yes/no, connected/disconnected, etc. [26–28] In
fuzzy logic, mathematical models are used to represent subjective notions (for example
hot, tall, old, etc.) for specific values that can be manipulated by computers. In this
paradigm, it can also give special value to the variable time, case control systems; they
may need feedback in a specific space of time interpreting previous data to make an
assessment of the situation in an earlier period. Using fuzzy logic allows the system is
put on alert limit situations, giving additional plasticity in cases with momentary
changes by several factors [19, 20, 26–28].
The rest of this paper presents the proposal of an harmonic system improved with
fuzzification of the parameters (Sect. 2), an outline of the prototype that implements the
model (Sect. 3), a case study (Sect. 4) and finally conclusions and future work (Sect. 5)

2 Proposal

The data model currently used by the KRONOS prototype, uses many variables and
complex. The following diagram shows the variables and relationships so you can
understand what type of problem is under the scope of the modeling strategy presented
here (Tables 6, 7, 8, 9 and 10).

Table 6. Pedestrian entity

Attribute Type Unit of measurement Fuzzy pattern
dni integer – –
gender character varying (10) – –
dbirth date años 10
name character varying (25) – –

Table 7. Driver entity

Attribute Type Unit of measurement Fuzzy pattern
dni integer – –
gender character varying (10) – –
dbirth date años 10
patent character varying (10) – –
name character varying (25) – –

Table 8. Cause entity

Attribute Type Unit of measurement Fuzzy pattern
type character varying (50) (*) – –
120 W. Bel et al.

Table 9. Zone entity

Attribute Type Unit of measurement Fuzzy pattern
idname character varying (50) – –
risklevel character varying (20) – –
idmap integer – –
p1x (latitud) real ° –
p1y (longitud) real ° –
p3x (latitud) real ° –
p3y (longitud) real ° –

Table 10. Map entity

Attribute Type Unit of measurement Fuzzy pattern
scale integer – –
zoom integer – –
row integer – –
column integer – –
lat real ° –
lon real ° –
place character varying (256) – –

As shown in Fig. 1, this model is divided into two main activities:

Fig. 1. Class diagram

 Parametric Prediction Model Using Expert System 121

Fig. 2. Graphic temperature fuzzification

ACTIVITY 1: Module that handles the registration of accidents on which different

data mining techniques are applied to obtain patterns and parameters to be able to
predict many types of situations (In the case of KRONOS accidents related to transit),
and thus derive information through a module that feeds back the system (Fig. 2).
ACTIVITY 2: Module that handles the registration of risk, such as real-time pro-
cessing and prediction of risk situations for pedestrians and vehicular traffic. It is
similar to the previous module since calculations are being recorded for performing
Data Mining and obtaining relevant patterns. At Real Time an Expert System and
Harmonic System are applied in parallel, in order to predict risky situations two dif-
ferent strategies that complement each other. This is similar to simulating the opinion
of two human experts with different evaluation criteria on the same situation (Fig. 3).
The system works with the following main entities: pedestrian, driver, weather
conditions and areas, which are composed of a set of variables, from which patterns are
extracted for both Expert System (ES) and Harmonic System (HS) risk predictions
(Fig. 4).

Fig. 3. Graphic reduced temperature fuzzification

Fig. 4. Graphic pressure fuzzification

122 W. Bel et al.

2.1 Variables Description

The next step to describe the model is to make the fuzzification of the variables to be
included in the data model, and then to the harmonic model. It is first important to note
the variables and data types used by the model, since the deeply affect the results. Units
of measurement and pattern of fuzzification used which are described below (Fig. 5).

Fig. 5. Graphic wind fuzzification I

2.2 Variables Fuzzyfication

This section describes the preliminary steps to apply fuzzy logic, covering the fuzzi-
fication process on each variable presented in the tables of the previous sections. To do
this authors simply take the most significant parameters (that is the descriptive attri-
butes). Although it is possible fuzzify any parameter, there are some of them that are
not relevant for the model and makes no sense to include them in the final model. For
example (name, patent, DNI, ID foreign keys, etc.
Temperature: The temperature can be measured with a thermometer sampling the
current environment. If the value is taken at various locations within the same area, at
the same time, it can vary. For scientific calculations, ambient temperature usually
ranges from 20 to 25 degrees Celsius (293 or 298 K, 68 or 77 degrees Fahrenheit)
[16, 17] (Fig. 6).

Fig. 6. Graphic wind fuzzification II

Heat stress is defined as the feeling of discomfort that is experienced when staying
in a particular environment, and requires the body performs an unreasonable effort. The
stress of mechanisms available in the body to maintain the internal temperature is really
high, in order to preserve the exchange of water and other substances that are being
carried out [16, 17] (Tables 11, 12, 13, 14 and 15).
 Parametric Prediction Model Using Expert System 123

Table 11. Accident_Type entity

Attribute Type Unit of measurement Fuzzy pattern
name character varying (50) – –

Table 12. Accident entity

Attribute Type Unit of measurement Fuzzy pattern
date date dd/mm/yyyy –
“time” time without time zone HH:MM:SS 7
type integer – –
die boolean – –
injured boolean – –
alcohol real g/l 9
belt_helmet boolean – –
accompanist boolean – –
idweather integer – –
idcause integer – –
idzone integer – –
pedestrian boolean – –
idperson character varying(20) – –

Table 13. Fuzzification function of temperature

Low Tempered Normal Elevated High
−10 1 3 0 18 −10 1 3 0 18
3 1 5 1 20 3 1 5 1 20
5 0 18 1 25 5 0 18 1 25
20 0 27 20 0 27

Table 14. Reduced fuzzification function of temperature

Low Normal High
−5 1 10 −5 1 10
10 1 12 10 1 12
12 0 27 12 0 27
29 29

Table 15. Fuzzification function of pressure

Low Normal High
885 1 980 885 1 980
980 1 990 980 1 990
990 0 1016 990 0 1016
1026 1026
124 W. Bel et al.

Fig. 7. Graphic humidity fuzzification

Such stress is the cause of the various pathological effects that occur upon an
excessive heat (heat stress) or sudden heat removal (cold stress). By heat stress is
understood as the pressure exerted on a person being exposed to extreme temperatures.
At constant values of temperature, humidity and wind, it exhibits a different reaction
depending on individual susceptibility and adaptation [16, 17] (Fig. 7).
The evaluation of heat stress indexes is complex and under the field of other
technical areas [16, 17]. Among others may derive in:
• Risk of hyperthermia
• Heat discomfort
• Thermal comfort
• Cold discomfort
• Risk of hypothermia
For this case study is discarded in the process of fuzzification scale of 5 categories
and a new 3 categories is drawn up, since in areas where developed prototype tests are
performed is not an area of extreme heat or cold ends, so that the new scale is more
than enough (Fig. 8).

Fig. 8. Graphic visibility fuzzification I

The scale of values proposed for this attribute is presented in the following table,
the measurement unit is degrees Celsius.
Pressure: The atmospheric pressure at a point coincides numerically with the weight
of a static column of air of unitary straight section extending from that point to the
upper limit of the atmosphere. As the air density decreases as altitude increases, it can
not calculate the weight unless we are able to express the variation in air density
depending on the altitude or pressure, so it is not easy to make an exact calculation of
 Parametric Prediction Model Using Expert System 125

atmospheric pressure on a place of the earth’s surface. In addition both the temperature
and air pressure are continuously varying, in a temporal and spatial scales, making it
difficult calculation. You can obtain a measure of atmospheric pressure in a particular
place but it can not draw many conclusions; however, the variation of said pressure
over time allows to obtain useful information which, together with other meteorological
data (air temperature, humidity and winds), can give a fairly accurate picture of the
weather in this place and even a forecast [2] (Tables 16, 17, 18, 19 and 20).

Table 16. Beaufort wind scale

Beaufort wind scale
Force Wind speed in km/h Denomination Visual features
0 <2 Calm Vertical smoke
Anemometer not rotate
1 2 to 5 Light air Smoke deviates
Weather vanes move
2 6 to 11 Light Breeze It feels in the face
Flags are raised
3 12 to 19 Gentle Breeze Move leaves
Flags flutter
4 20 to 28 Moderate Breeze Dust rises
Waving flags
5 29 to 39 Regular or Fresh Wind Sand fly
Branches jiggle
6 39 to 49 Strong Wind Whistling rigging
Thick branches move
7 50 to 60 Wind Force Large trees move
Annoyed walk against
8 61 to 74 Gale Hinds trees
Very difficult to walk and breathe
9 75 to 88 Strong Gale Drag items
Impossible to walk
10 89 to 102 Storm Damage and varied consequences
11 103 to 117 Violent storm Damage and varied consequences
12 118 to + Hurricane Damage and varied consequences

Table 17. Fuzzification function of wind

Quiet Breeze Moderate Strong
0 1 1.8 0 11 0 38 0
1.8 1 2.3 1 12 1 39 1
2.3 0 11 1 38 1 60 1
12 0 39 0
126 W. Bel et al.

Table 18. Fuzzification function of humidity

Low Normal High
0 1 50 0 80 0
50 1 60 1 90 1
60 0 80 1 100 1
90 0

Table 19. International air visibility scale

International air visibility scale
Key Visibility limits Designation
00 From 0 to 25 m Super thick fog, no visibility
0 From 25 to 50 m Very dense fog, no visibility
1 From 50 to 100 m Thick fog, very poor visibility
2 From 100 to 500 m Fog, very poor visibility
3 From 500 to 1000 m Fog, low visibility
4 From 1000 to 2000 m Mist or haze, poor visibility
5 From 2000 to 4000 m Mist or haze, poor visibility
6 From 4000 to 10000 m Diaphanous atmosphere, moderate visibility
7 From 10000 to 20000 m Diaphanous atmosphere, good visibility
8 From 20000 to 50000 m Diaphanous atmosphere, very good visibility
9 Over 50000 m Diaphanous atmosphere, excellent visibility

Table 20. Fuzzification function of visibility

Nothing Minimum Little Limited Limited
0 1 0.025 0 0 0.1 1 0 4 0
0.025 1 0.05 1 1 0.125 1.025 1 6 1
0.05 0 0.1 1 1 1 4 1 50 1
0.125 0 0 1.025 6 0

According CONAGUA (National Water Commission of Mexico) and other

national weather services in different countries. The atmospheric pressure is not stable
and varies between 885 hPa among the deepest cyclones (low pressure) and 1,077 hPa
on the most powerful Siberian anticyclone (high pressure). 1013 hPa pressure being
considered normal or average [2] (Fig. 9).

Fig. 9. Graphic visibility fuzzification II

 Parametric Prediction Model Using Expert System 127

The high pressures also called centers of high pressure or anticyclones. The high
pressures also called centers of high pressure or anticyclones. And low pressures are
called low pressure centers, also called cyclones, depressions or squalls [2].
The scale of values proposed for this attribute is presented in the following table,
the unit of measurement is hPa (Tables 21, 22, 23, 24 and 25).

Table 21. Picture of feeling of comfort

Picture of feeling of comfort
Dewpoint Denomination General comment
−5 to −1 °C Very dry air There is some feeling of irritability and mild
dehydration (especially with strong wind)
Oº to 4 °C Dry air Comfortable environment in winter
5º to 7 °C Dry welfare Wellness with temperatures 20 to 26 C. Nice heat
with temperatures of 27–30 °C
8º to 13 °C Maximum Maximum feeling of comfort and convenience, with
wellness temperatures 20 to 26 °C (no wind and in the shade).
Temperatures are well tolerated 27 to 30 °C
14º to 16 °C Wellness wet Pleasant climate inside temperature of 20–26 °C.
A higher thermal brands, heat is becoming
uncomfortable
17º to 19 °C Moist heat Climate “Heavy” with temperatures of 20–26 °C.
Climate A higher thermal brands, heat is becoming
uncomfortable.
20º to 24 °C Sweltering humid Intense discomfort and discomfort, especially with
heat temperatures of 30 °C or more. There is danger of
“heat stroke”. Sun, danger of insolation
25 °C or more Very humid and Unhealthy and dangerous climate, especially with
unbearable heat thermal marks 30 °C or higher

Table 22. Fuzzification function of dewpoint

Dry Welfare Warm Insufferable
−5 1 4 0 16 0 24 0
4 1 5 1 17 1 25 1
5 0 16 1 24 1 30 1
17 0 25 0

Table 23. Fuzzification function of sunrise-sunset

Dawn Morning Midday Afternoon Sunset Night
3 0 6 0 11 0 13 0 17 0 20 0
4 1 7 1 12 1 14 1 18 1 21 1
6 1 11 1 13 1 17 1 20 1 23 1
7 0 12 0 14 0 18 0 21 0 0 1
3 1
4 0
128 W. Bel et al.

Table 24. Fuzzification function of current speed

Low Middle High Excess
0 1 35 0 60 0 100 0
35 1 40 1 65 1 105 1
40 0 60 1 100 1 240 1
65 0 105 0

Table 25. Table blood alcohol level

Breathalyzer Effects Multiplied risk
0.15 Decreased reflexes 1.2
0.2 False appreciation of distances, underestimation of speed 1.3
0.3 Motor disorders-euphoria 2
0.5 Increased response time 3
0.8 General behavioral disorder 4.5
1.2 Tiredness, fatigue, loss of visual acuity 9

Wind: The wind is the gas flow scale. On Earth, wind is the mass movement of air in
the atmosphere in horizontal movement. Günter D. Roth [3] defines it as compensation
for atmospheric pressure differences between two points. For this an international scale
is used called Beaufort scale, the which were used by Hamblyn [5] and Deblieu [6] for
measurement of force and categorization of the wind, and used for weather forecasting
(Fig. 10).

Fig. 10. Graphic dewpoint fuzzification

For the present case study is discarded fuzzification process forces 8 to 12 to reduce
the number of categories or sets and increase the speed of calculation result, since in
areas where tests developed prototype is made is not an area of strong storms or
hurricanes (Table 26).
The scale of values proposed for this attribute is presented in the following table,
the unit of measurement is Km/h.
The chart is divided into two parts by a difference of scales, magnifying the first
part and minimize the second (Fig. 11).
Humidity: It is called humidity permeating water vapor or a body present in the
atmosphere. Water is present in all living bodies, whether animal or plant, and that
presence is of great importance for life. It is the amount of water vapor present in the
 Parametric Prediction Model Using Expert System 129

Table 26. Table legal rates blood-alcohol

Country Limit allowed Country Limit allowed
Argentina 0.5 g/l Switzerland 0.5 g/l
Spain 0.5 g/l Serbia 0.5 g/l
France 0.5 g/l Denmark 0.5 g/l
Canada 0.8 g/l Belgium 0.5 g/l
United States 0.8 g/l Portugal 0.5 g/l
United Kingdom 0.8 g/l Germany 0.5 g/l
Ukraine 0.0 g/l Turkey 0.5 g/l
Albania 0.0 g/l Russia 0.3 g/l
Czech Republic 0.0 g/l Norway 0.2 g/l
Estonia 0.0 g/l Poland 0.2 g/l
Hungary 0.0 g/l Sweden 0.2 g/l
Romania 0.0 g/l

Fig. 11. graphic sunrise-sunset fuzzification

air, it can be expressed in absolute terms by the absolute humidity, or a relative by

relative humidity or degree humidity. Relative humidity is the percentage ratio of actual
water vapor content of the air and would need to contain saturated at the same tem-
perature [7] (Table 27).
The scale of values proposed for this attribute (relative humidity) is presented in the
following table, the unit of measurement is percentage (%). The values used were
averages of various meteorological institutes (Fig. 12).
Visibility: The visibility is perceivable quality that allows viewing objects at a dis-
tance. A lower visibility, objects will be worse off into the distance, while greater
visibility will be better distant objects. Weather visibility, which is the ability you have
to see in any cardinal point the obstacles that surround it, in a previously known
distance. This form of measurement is distinguished with the abbreviation MOR

Table 27. Fuzzification function of alcohol

Little Low High Excess
0 1 0.2 0 0.5 0 0.8 0
0.2 1 0.25 1 0.55 1 0.85 1
0.25 0 0.5 1 0.8 1 1.5 1
0.55 0 0.85 0
130 W. Bel et al.

Fig. 12. Graphic current speed fuzzification

(Meteorological Optical Range). For the level of visibility with this technique, it is
necessary to set an observation point, where the distance with the nearest obstacles
(houses, trees, hills, etc.), as serving as a reference in a situation of reduced visibility as
National Meteorological Institute (INS) [8] and National Weather Service (SNM) [9].
There is an international scale air visibility, for measuring visibility [10] which use
different organizations (NMI, SNM, Civil Protection, etc.).
The scale of values proposed for this attribute is presented in the following table,
the unit of measurement is Km (Fig. 13).
The chart is divided into two parts by a difference of scales, magnifying the first
part and minimize the second.
Dewpoint: The dewpoint is the value which should lower the air temperature to the
existing water vapor begins to condense. Humans have the physiological ability to
maintain constant 36 °C (37 °C or) to the internal temperature of the human body.
Simultaneously, the temperature of the skin exposed to the outside is usually 32 °C
(provided it is not exposed to sunlight). When the ambient air has a dew point of 20 to
25 °C, perspiration of people becomes copious. Then, the sense of unease is increasing,
and it is said that “time is very heavy”. Discomfort or “heaviness” of the weather
because sweat evaporates with difficulty (or does not evaporate) and therefore the body
does not get cool. A typical value from which it begins to become noticeable envi-
ronmental discomfort is for a value of 20 °C in the dew point. This mentioned value is
obtained under the following weather conditions [11, 12]:

Fig. 13. Relative risk based on levels of BAC (blood alcohol concentration) accident.
 Parametric Prediction Model Using Expert System 131

• An afternoon with temperature 34 °C and 45% humidity.

• One night with 26 °C and 70% humidity.
• Early one morning with 22 °C temperature and 90% humidity.
When our region gets very hot and humid air from the south and central Brazil and
northeastern Argentina, as does winds north, northeast or northwest. Under these
extreme weather conditions, the dew point can reach exceptional value of 24 °C. In this
brand up for example with the following meteorological conditions [11, 12]:
• In an afternoon temperature 36 °C and 50% humidity.
• On a night with 30 °C and 72% humidity.
• In one morning with 25 °C and humidity 95.
These climatic conditions usually appear on some summer days in the area of
Buenos Aires, Rio Plata River and south coast. Once the dew point (obtained by means
of the table) is known you can use the box comfort feeling climate [11, 12].
The scale of values proposed for this attribute is presented in the following table,
the unit of measurement is degrees Celsius (Fig. 14).

Fig. 14. Graphic current alcohol fuzzification

Sunrise-Sunset: Sunrise and sunset are calculated from the average of historical data
obtained from various meteorological institutes, this categorization is divided into the
four seasons spring, summer, autumn and winter. Therefore the system based on the
date that is processing, select the corresponding to the current station fuzzy function for
this paper, only the function corresponding to the winter occurs.
The scale of values proposed for this attribute is presented in the following table,
the unit of measurement is HH:MM:SS (Fig. 15).

Fig. 15. Graphic date birth fuzzification

132 W. Bel et al.

The entity (weather) is responsible to manage attributes of weather aforementioned

and fuzzy. It also has an attribute called “weather” (with values possible rain, snow,
hail, rain, cloudy, etc.) which is also a very important factor in calculating pedestrian or
vehicular risk, which directly influences the visibility attribute. All these data are
obtained from a web service.
In addition to these aspects, since it is a model of a complex process, there are many
variables that are not processed directly by the system but are inferred directly in
calculating the risk of the areas to which it is associated. Among others may be
mentioned dust, smoke, fire, storm or fallen trees, flooded streets and other post
resulting from natural disasters.
Current Speed: The speed at which the vehicles inside and outside the city moving,
constantly changing. In the present context called “current speed” at the speed of a
vehicle at an instant of time. This can range from 0 to 240 km/h or more in some
exceptional cases such as some highways (which means that the speed range would not
be within the maximum speed limit). According Traffic Safety and the Center for
Experimentation and Road Safety (CESVI) speed limits are set for different types of
places [13]. These limits and speed values vary in some countries.
The scale of values proposed for this attribute is presented in the following table,
the unit of measurement is Km/h.
Alcohol: The breathalyser or blood alcohol test measures the blood alcohol concen-
tration. It is obtained through a percentage of body mass being the mass by the volume
or a combination. When measured alcohol detected in exhaled air, the unit used is the
grams per liter of air, which in the usual practice becomes conventionally in grams per
liter of blood.
According to the CESVI alcohol it is introduced in such a way in our culture that
many find it difficult to admit the drawbacks generated in driving. Often there is a great
social tolerance towards consumption and to the problems arising from their use.
Within the context of driving errors, alcohol is almost always associated with speed,
sleep and distractions [14].
In our country, the law that regulates the amount of alcohol allowed to drive a car,
set as alcohol limit 0.5 g of alcohol per liter of blood. It is considered that from this
amount the driver begins to show significant psychomotor reactions that hinder driving.
This becomes more strict limit for driving motorcycles and mopeds where tolerance is
decreased to 0.2 g/l and to 0 for those who drive freight or passengers. This limit is
arbitrary to some extent, it does not contemplate the reality of all drivers in terms of
tolerance that has each of them to the effects of alcohol [14].
It should be noted that alcohol produces a depression of the central nervous system,
impairing psychomotor function, sensory perception (vision and hearing), changing
individual behavior that often translates into a false sense of security and a more risk
taking [14].
The table lists the main effects of alcohol are presented in relation with psy-
chomotor needs necessary for driving, depending on the level of blood alcohol.
Legal rates blood-alcohol levels in different countries.
These rates and blood alcohol levels also vary within the same country under
different circumstances; such as in Argentina the legal blood alcohol limit for driving a
 Parametric Prediction Model Using Expert System 133

car is 0.5 g/l of blood, while it is having less tolerance for motorcycle drivers down to
0.2 g/l. In the case of professional drivers, driving freight or passengers, tolerance is
zero. There is also zero tolerance in the province of Cordoba, of this country. In Spain
the law does not allow driving if the BAC exceeds 0.5 g/l in blood; except in cases of
drivers with less than 2 years of license or professional, where the figure is 0.3 g/l in
the blood. Above these figures, the corresponding sanctions are applied in the so-called
penalty points, which provides for the withdrawal of 4 to 6 points; or even jail.
The scale of values proposed for this attribute is presented in the following table,
the unit of measurement is g/l.
Date birth: Accident statistics indicate that all run risks of an accident; each year in
Spain occur about 140,000 victims (injuries and deaths) due to traffic accidents.
However, the risk of an accident is not the same for all road users. The human factor is
key to the study of these accidents. The age is a factor that is closely related to the
likelihood of one or another type of accident and the consequences thereof [15].
Teenagers often have impulsive behavior driving, poor road training and are guided by
the decisions of the group. They like to show off, leading them to risk maneuvers, and
lead rather unstable vehicles, such as mopeds. Teenagers from 14 to 17 years old
represents 4% of the population. Each year, four of every thousand adolescents are
victims (injuries and deaths) of a traffic accident. In 2004, 8,242 teenagers have been
involved in a traffic accident [15].
The 64% of the victims of this group circulates on a moped when he suffers the
accident. Three out of five adolescents is user rugged moped and two of every five
deaths circulates in this type of vehicle [15].
Older people have accidents mainly during the day and in the central hours of the
morning and afternoon. As drivers, the most common accident is a collision, but as age
increases, the abuse becomes the most dangerous accident [15].
In this group a large decline of the victims who are involved in accidents as drivers
of tourism is appreciated; especially in those over 75 years old. The highest percentage
of deaths recorded as a pedestrian: 36% versus 24% as drivers [15].
According Research Institute of Traffic and Road Safety (INTRAS) of the
University of Valencia in the years 2000–2004 in which they established the following
classification in relation to the data collected and analyzed [15]:
• Children (under 13 years old)
• Adolescents (14–17 years old)
• Youth (18–30 years old)
• Adults (31–65 years old)
• Old people (over 65 years old)
The scale of values proposed for this attribute is presented in the following table,
the unit of measure is years old.
In addition to these aspects, since it is a model of a complex process, there are many
variables that are not processed directly by the system, but inferred directly in calcu-
lating the risk of the areas to which it is associated. Some variables that are inferred
directly on the calculation of risk in the area, among others we can mention the
influence of drugs on the driver, traffic of areas, state of the streets, stray animals and
134 W. Bel et al.

evacuation zones of large amounts of people (out of schools, theaters, events, etc.) at
specific times.

3 Description of Prototype KRONOS

The prototype is implemented with a basic GUI data acquisition, and an expert system
regulated traditional and its alternative modeling harmonic systems, as seen in Figs. 16
and 17 respectively [23–25].

Fig. 16. Expert system model

Fig. 17. Harmonic system model

 Parametric Prediction Model Using Expert System 135

4 Case Study

As mentioned previously, the prototype partially implements the model. The ES core is
functional; it has many rules according to expert recommendations. The interfaces to
user, DBMS, Maps, DBMS, are working and also part of the Harmonics system. The
tests in this section use ES knowledge but not the Harmonics system, since the goal of
this paper is to provide the improvement acquired by considering GPS in the risk
inference.
Database: The test set are risk situations evaluated post-Morten to be able to find the
accuracy of the results. All the dataset belong to real situations of traffic in Concepción
del Uruguay city (Entre Ríos, Argentina). Figure 18 shows the Map of the city.
As can be seen the map has a grid. Each cell represents a zone, described in the
Database as follows: (ID-Zone, Description, Latitude, Longitude, Risk). ID-Zone is an
integer that identify the cell. Description outlines the zone. Table 28 describes the
reference values and the labels in Fig. 18.

Fig. 18. City Map

Table 28. Fuzzification function of date birth

Children Adolescent Youth Adults Old people
3 1 13 0 17 0 30 0 65 0
13 1 15 1 19 1 23 1 67 1
15 0 17 1 30 1 65 1 92 1
19 0 32 0 67 0

The testing resulting in the results summarized in Table 29.

In Table 29, column Desc. (Description field), has the values: (c)enter (r)ound-
about, c(o)untry road, country (z)one. Column S (Subject) has the values: (d)river, (p)
edestrian. To assess the impact of considering GMT (Zone) information, the dataset
was re-evaluated avoiding that variable from the rules. Then, the predictions were
compared against to an expert predictions. The number of hits and errors are in
Table 30.
136 W. Bel et al.

Table 29. Zone labeling

Risk Risk type Label
0 No N
1 High A
2 High A
3 Medium M
4 Low None
5 Low None

Table 30. Testing results

Risk DD/MM/YY Time Weather Desc. S KM/H Belt/Helmet Alc.
P/G (G/L)
0 01/05/15 13:55 sunny C D 40 Si 0.9
0 22/05/15 12:10 sunny R D 90 No 0
0 12/06/15 22:29 sunny O D 40 No 0.3
0 24/11/15 08:32 foggy O D 30 No 0
0 01/05/15 13:56 sunny C P 0.3
0 08/05/15 11:30 cloudy Z P 1
1 12/06/15 22:29 cloudy Z D 122 No 0
1 28/06/15 17:43 cloudy C D 140 No 0.4
1 12/06/15 23:44 cloudy C P 0.2
1 12/06/15 23:45 cloudy C P 0.5
1 23/11/15 23:32 cloudy C D 190 No 0
1 23/11/15 23:50 cloudy C D 230 No 0.9
1 12/06/15 22:20 cloudy Z D 130 No 0.5
3 12/06/15 23:03 cloudy C D 100 No 0
3 12/06/15 23:13 cloudy Z D 100 No 0.2
3 08/09/15 21:45 cloudy C D 150 No 0.1
3 02/10/15 19:21 sunny C D 190 No 0.9
3 02/10/15 19:48 sunny C D 190 No 0.9
3 12/06/15 23:44 cloudy C P 0.51
3 12/06/15 23:44 cloudy C P 1.25
0 25/11/15 11:41 cloudy C P 0
0 25/11/15 12:36 fog Z P 2
0 25/11/15 19:36 sunny C P 1.9
0 25/11/15 02:36 sunny C P 1.9
0 25/11/15 02:36 sunny Z P 1.5
1 25/11/13 13:53 sunny C D 120 1
1 12/08/15 19:30 cloudy Z D 140 0.19
1 17/08/10 19:30 cloudy Z D 140 0.16
1 17/08/10 19:30 cloudy Z D 125 0.18
1 17/08/10 20:30 cloudy Z D 125 0.25
2 25/11/15 20:14 cloudy C D 90 1
2 25/11/15 14:21 cloudy C D 40 2
(continued)
 Parametric Prediction Model Using Expert System 137

Table 30. (continued)

Risk DD/MM/YY Time Weather Desc. S KM/H Belt/Helmet Alc.
P/G (G/L)
2 25/11/15 14:21 cloudy C D 50 2
2 25/11/15 14:21 cloudy C D 50 2
2 25/11/15 14:21 cloudy C D 60 2
3 25/11/15 14:36 sunny Z D 40 0.1
3 25/11/15 20:36 sunny Z D 40 0.1
3 25/11/15 20:36 sunny Z D 60 0.1
3 25/11/15 20:36 sunny Z D 90 0.1
3 25/11/15 14:36 sunny Z D 35 1
5 31/05/11 23:22 cloudy C P 0.13
5 04/05/15 23:22 cloudy C P 0.1
5 25/11/15 15:24 sunny C P 0.1
5 25/11/15 20:00 sunny C P 0.14
5 25/11/15 20:00 sunny C P 0.1
5 09/07/14 23:48 sunny C P 0.11
5 09/07/14 23:48 sunny C P 0.1
5 09/07/14 23:48 sunny C P 0.05
5 09/07/14 23:48 sunny C P 0.03
5 09/07/14 23:48 sunny C P 0.09

The legend (*G) means without GMT information, and (G) means with GMT
information. Results indicate higher accuracy (94% versus 42%). Also the system
trends to underestimate risk. Analyzing these 3 test cases, the deviation occurs for
drivers that do not use helmet/belt and are in a safe zone of the city, during the daylight
hours but when weather has reduced visibility. Figure 19 shows results test by test
(Table 31).

Fig. 19. Risk according Expert (E), ES without GMT (*G) and with GMT (G)
138 W. Bel et al.

Table 31. Results with and without GMT

Error(*G) % Error (G) %
Overestimated 11 22.00 0 00.00
Underestimated 18 36.00 3 06.00
OK 21 42.00 47 94.00

The (E)xpert prediction is the darkest curve. Without GMT information (*G) the
ES prediction results more erratic.

5 Conclusions and Future Work

This paper has presented a series of extracted parameters of the international standard
literature and are based on criteria fuzzy also commonly used. As an initial proposal,
this is a first stage of research where the expert knowledge and consensus on the
relevant parameters is critical to generate inference systems risk.
It has been also presented a case study where it can be appreciated the usefulness of
the approach proposed here but has not yet been possible a more comprehensive
statistical evaluation, which while it is ongoing, still requires a number of cases and the
generation time (due to its characteristics) is quite high. The statistics of public
knowledge will turn to initially validate the quality of the results.

References
1. Union of Pure and Applied Chemistry (IUPAC) - Compendium of Chemical Terminology
Gold Book (2014)
2. Comisión Nacional del Agua (CONAGUA) (2016). www.conagua.gob.mx
3. Roth, G.D.: Meteorología. Formaciones nubosas y otros fenómenos meteorológicos.
Situaciones meteorológicas generales. Pronósticos del tiempo (2003)
4. O’Connor, J.J., Robertson, E.F.: Evangelista Torricelli. MacTutor History of Mathematics
and Science (2002)
5. Hamblyn, R.: The Invention of Clouds: How an Amateur Meteorologist Forged the
Language of the Skies. Farrar, Straus, and Giroux, New York (2001)
6. DeBlieu, J.: Wind: How the Flow of Air Has Shaped Life, Myth, and the Land. Houghton
Mifflin Co., New York (1998)
7. Sendiña Naval, I., Pérez Muñuzuri, V.: Fundamentos de meteorología. Universidad de
Santiago de Compostela (2006)
8. Instituto Nacional de Meteorología (2016). http://www.gob.mx/inm
9. Servicio Meteorológico Nacional (2016). http://www.smn.gov.ar/
10. http://www.loseskakeados.com/joomla1cinco/index.php?option=com_content&view=
article&id=10001:meteorologa-visibilidad-nieblas-brumas-calimas-100-km-10-km&catid=
347:meteorologa&Itemid=298 (2016)
11. Celemín, A.H.: “El tiempo y la salud humana” del libro “Meteorología Práctica”. Chap. 18.
dellic (2012)
12. http://www.paranauticos.com/notas/meteorologia/punto-de-rocio.htm (2016)
13. Gobierno de la Ciudad de Buenos Aires (2016). http://www.cba.gov.ar
 Parametric Prediction Model Using Expert System 139

14. http://www.cesvi.com.ar/seguridadvial/Recomendaciones/SeguridadRecomendaciones3.
aspx (2016)
15. https://www.lineadirecta.com/ZZRecursos/recursos/ES/documentos/GruposdeRiesgo06.pdf
(2004)
16. NTP 322: Valoración del riesgo de estrés térmico: índice WBGT - Pablo Luna Mendaza
(1999)
17. Estrés térmico y sobrecarga térmica: evaluación de los riesgos - Eugenia Monroy Martí
(2001)
18. 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)
19. Zadeh, L.A.: A note on prototype set theory and fuzzy sets. Cognition 12, 291–297 (1982)
20. Zadeh, L.A.: Fuzzy sets and applications. In: Yager, R.R., Ovchinnikov, S., Tong, R.M.,
Nguyen, H.T. (eds.) Selected Papers. Wiley, Nueva York (1987)
21. De Luise, D.L.: Harmonics systems for temporal mining. IEEE CIS Magazine, September
2013
22. De Luise, D.L.: Introducción a los Sistemas armónicos. In: Advanced Research and Trends
in New Technologies, Software, Human-Computer Interaction and Communicability. IGI
Global (2013)
23. Mansilla, D., Sandillú, M., De Luise, D.L., Bel, W.: Un Modelo con Conocimiento Experto
y Sistemas Armónicos para Evaluación de Riesgos de Tráfico. EnIDI Argentna (2015)
24. Acuña, I., García, E., De Luise, D.L., Paredes, C., Celayeta, A., Sandillú, M., Bel, W.:
Traffic & pedestrian risk inference using harmonic systems. In: SOFA, Romania (2014)
25. De Luise, D.L., Bel, W., Mansilla, D.: Risk prediction on time with GPS information. In:
CITCEP (2015)
26. Rakus-Andersson, E.: Fuzzy probabilities of biological continuous fuzzy events. Int.
J. Comput. Intell. Stud. IJCSysE 1(2), 191–204 (2009)
27. Zhao, L.: Application of interval type-2 fuzzy neural networks to predict short-term traffic
flow. Int. J. Comput. Appl. Technol. IJAIP 43(1), 67–75 (2012)
28. Ma, C., Wu, F., Qi, B., Gong, L., Wang, L., An, X.: Study on urban three-lane mixed traffic
flow with buses based on the Nagel-Schreckenberg model. Int. J. Wirel. Mob. Comput.
IJAIP 6(6), 615–620 (2013)
 Predicting Ozone Layer Concentration Using
Multivariate Adaptive Regression Splines,
Random Forest and Classification
and Regression Tree

Sanjiban Sekhar Roy1(&), Chitransh Pratyush1, and Cornel Barna2

1
School of Computer Science and Engineering, VIT University, Vellore, India
[email protected], [email protected]
2
Automation and Applied Informatics, Aurel Vlaicu University of Arad,
Arad, Romania
[email protected]

Abstract. Air pollution is one of the major environmental worries in recent

time. Abrupt increase in the concentration of any gas leads to air pollution. The
cities are mostly affected due to the abundance of population there. One of the
worst gaseous pollutants is OZONE (O3). In this paper, we propose three
predictive models for estimation of concentration of ozone gases in the air which
are Random Forest, Multivariate Adaptive Regression Splines and Classification
and Regression Tree. Evaluation of the prediction models indicates that the
Multivariate Adaptive Regression Splines model describes the dataset better and
has achieved significantly better prediction accuracy as compared to the Ran-
dom Forest and Classification and Regression Tree. A detailed comparative
study has been carried out on the performances of Random Forest, Multivariate
Adaptive Regression Splines and Classification and Regression Tree. MARS
gives the result by considering less variables as compared to other two. More-
over, Random Forest takes a little more time for building the tree as the elapsed
time was calculated to 45 s in this case. In addition, variable importance for each
model has been predicted. Observing all the graphs Multivariate Adaptive
Regression Splines gives the closest curve of both train and test set when
compared. It can be concluded that multivariate adaptive regression splines can
be a valuable tool in predicting ozone for future.

Keywords: Ozone
Multivariable adaptive regression spline
Random

Forest
Classification and Regression Tree

1 Introduction

Urban atmospheric pollutants are increasing day by day. They are considered as one of
the main causes of increased incidence of respiratory illness in citizens. It is now
irrefutable that air pollution is being caused by large amount of Total Suspended
Particulates (TSP) and respiratory particulate of Particulate Matter less than 10 µm in
aerodynamic diameter that has numerous undesired consequences on human health [1].
Air pollutants in an area with good airflow quickly mixes with air and disperses
© Springer International Publishing AG 2018
V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_11
 Predicting Ozone Layer Concentration 141

however when trapped in an area, pollutant concentration can increase rapidly which
ultimately leads to degradation of air quality. In order to measure how polluted the air
is Air Quality Index is examined while for properties of air we see Qualities of Air [2].
All these factors affect the ozone layer which is Earth’s protective layer. It is a belt of
naturally occurring ozone gas that sits 9.3–18.6 miles above Earth and serves as shield
from harmful ultraviolet B radiation emitted by sun. Several steps like Montreal Pro-
tocol which declines emission of ODS (ozone depleting substances) have been taken. It
is expected to result in a near complete recovery of ozone layer near the middle of 21st
century. By 2012, the total combined abundance of anthropogenic ODS in the tropo-
sphere has decreased by nearly 10% from its peak value in 1994 [3, 5].
In the present era, there is a wide spread concern for ozone layer depletion due to
the release of pollution. As particulate matter causes several kind of respiratory and
cardiovascular disease, it also leads to ozone depletion which attracts more and more
attention for air quality information. This shows the need for integration of different
information system in a similar way as done by Birgersson et al. [6] for data integration
using Machine Learning [6]. Prediction of air quality has thus become a necessary need
to save the future. Machine learning has been applied in various fields [14, 15, 17].
Medical and other fields have also been covered by various classification techniques
[18, 19, 22–25]. Just as application of rough set technique was done for data investi-
gation by Roy et al. (2013), in this paper application of Random Forest, Multivariate
Adaptive Regression Splines and Classification and Regression Tree techniques has
been applied for predicting the concentration of ozone [13, 20, 21]. Chuanting Zhang
and Dongfeng Yuan (2015) worked on Grained Air Quality Index Level Prediction
Using Random Forest Algorithm on Cluster Computing of Spark [4]. Previously
existing methods could not meet the demand of real time analysis so a distributed
random forest algorithm is implemented using Spark on the basis of resilient distributed
dataset and shared variable. Parallelized random forest is also used as prediction model.
Estimation of benzene by on field calibration of an electronic nose has been carried by
Vito et al. (2008) in which gas multi-sensor played an important role which helps to
raise the density of the monitoring network. But their concentration estimation capa-
bilities are seriously limited by the known stability and selectivity issues of solid-state
sensors they often rely on. Sensor fusion algorithm used in regression need to be
properly tuned via supervised learning. But this training was revealed to be unsuc-
cessful [7, 12]. Forecasting and prediction of things has become an essential part for
future life. Roy et al. (2015) worked on prediction of Stock Market Forecasting using
Lasso Linear regression model [16]. Vito et al. (2009) worked on CO, NO2 and NOx
urban pollution monitoring. Some authors have used gas multisensor devices as a tool
for densening the urban pollution monitoring mesh due to the significantly low cost per
unit [8]. But the drawback is that these sensors are not reliable for long term and
selectivity issues. In this paper we concentrate on regression technique Multiple
Adaptive Regression Spline (MARS) for Air Quality dataset. Hui et al. (2013) used this
regression model for prediction of emission of CO2 in ASEAN countries. A compar-
ative study of multiple regression (MR) and multiple adaptive regression splines
(MARS) was carried for statistical modelling of CO2 over period of 1980–2007 [9].
MARS model was concluded as more feasible and with better predictive ability. This
paper shows the comparison of regression techniques like Random Forest, Multivariate
142 S.S. Roy et al.

Adaptive Regression Splines and Classification and Regression Tree on Air Quality
data showing the prediction using Salford Predictive Modeller.
This paper is organised as follows. Section 2 overviews proposed techniques of
Random Forest, Multivariate Adaptive Regression Splines and Classification and
Regression Tree. Section 3 gives the experimental setup and the steps involved in
performing the regression techniques on the given dataset. Section 4 displays the
results and discussion. Section 5 concludes the paper.

2 Proposed Techniques

To work with Salford Modeller it is necessary to know the working of the regression
techniques that are going to be used. All are type of machine learning like a computer
program is said to learn from experience ‘E’ with respect to some class of tasks ‘T’ and
performance measure ‘P’ if its performance at tasks in ‘T’ as measured by ‘P’ improves
with experience ‘E’. All these have been used for prediction of ozone concentration by
extracting knowledge from dataset.
A. Random Forest algorithm
It is a tree-based ensemble learning method involving the combination of several
models to solve a single prediction problem. The first algorithm for random decision
forests was created by Tin Kam Ho using the random subspace method. It may also be
said as a collection of many CART trees that are not influenced by each other when it is
constructed [4]. It works as a large collection of decorrelated decision trees. It comes
under bagging technique (average noisy and unbiased models to create a model with
low variance). Tress are based on random selection of data as well as variable. This
develops lots of decision trees based on random selection of data and random selection
of variables. After all the tree are built the data get fed in the tree and proximities are
calculated for each pair of cases. If any two cases occupy the same terminal node, their
proximity is changed and incremented by one. At last, the proximities get normalized
by dividing it by the number of trees. Proximities can be used in replacing missing data,
locating outliers, and producing illuminating low-dimensional views of the data. It
serves as one of the most useful tools in random forests. The proximities originally
form an N  N matrix. After a tree is built, both training and test data are pulled down
the tree. At the end, the proximities get normalized by dividing by the number of trees.
Since the large data set could not fit an N  N matrix into fast memory, a modification
reduced the required memory size to N  T where T stands for the number of trees in
the forest. In order to speed up the computation-intensive scaling and iterative missing
value replacement, the user is provided with the option of retaining only the nrnn
largest proximities for each case. When the dataset is presented, the proximities of each
case in the test set with each case in the training set can also be computed and
compared. The amount of additional computing is moderate. The dataset contains
thousands of data from which concentration of ozone is predicted. Thus Random Forest
is useful in handling thousands of input variables without variable deletion. Hence
Random Forest gives variable importance to each and every variable involved.
 Predicting Ozone Layer Concentration 143

B. Classification and Regression Tree algorithm

Classification Regression Tree was introduced by Breiman et al. (1984) for classifi-
cation or regression predictive modelling problems. It is often referred as ‘Decision
Tree’ but now named as CART in modern software. It provides a foundation for
important algorithms like bagged decision trees, random forest and boosted decision
trees. It is a binary tree that splits a node into two child nodes repeatedly beginning with
the root that contains whole learning sample. Say for x being a nominal categorical
variable of I categories, there are 2I−1-1 possible splits for the predictor. If X is an
ordinal categorical or continuous variable with K different values there are K − 1
different splits on X. At any node say t, the best split s is chosen to maximize a splitting
criterion Diðs,tÞ [11]. There are 3 splitting criteria available.

2.1 Gini Criterion

The impurity measure at a node t is defined as
X
iðtÞ ¼ i;j
cðijjÞpðijtÞpðjjtÞ ð1Þ

It is the decrease of impurity given by

Diðs; tÞ ¼ iðtÞ  pL iðtL Þ  pR iðtR Þ ð2Þ

where pL and pR are probabilities of sending case to left child node and right

pL ¼ pðtL Þ=pðtÞ ð3Þ

child node where

And

pR = pðtR Þ=pðtÞ ð4Þ

X
Diðs; tÞ ¼ pL pR j
½jpðjjtÞ  pðjjtÞj2 ð5Þ

2.2 Twoing Criterion

CART does not require any special data preparation other than a good representation of
the problem.
C. Multivariate Adaptive Regression Splines algorithm
It is a form of regression analysis developed by Friedman in 1991 [10] with the aim to
predict dependent variable from set of independent variable. It is simpler than other
models like random forest and neural network. It is seen as an extension of linear
models that automatically models non linearity interaction between in variables. Mars
is not affected by outliners. It produces model that can be written as an equation. It
144 S.S. Roy et al.

models both classification and regression tasks. It accepts a large number of predictor
and chooses important predictor variable. The extensive use of MARS model can be
done for prediction as it has been done for concentration of ozone in this paper. It is
used for prediction and classification problems (Islam et al. 2015; [4]). The details of
the MARS model can be observed through a website by Salford Systems. Also, this
regression is influenced by the recursive partitioning method for which any criteria can
be chosen for the selection of basis function of multivariate spline. One of the
advantage of the mars model is that MARS can reduce the outliers. The proposed
MARS forms the model with the use of two sided truncated functions of the predictor x
which has the form below.
 
x  t, x[t
ðx  tÞ þ ¼ ð6Þ
0, otherwise

Equation (6) works as a basis function for linear and non-linear functions. Also,
this Eq. (6) works to approximate any function f(x). In MARS, let us assume that the
dependent variable (output) is y and number of terms is M. The output variable can be
represented as following,

X
M
y ¼ f ðxÞ ¼ b0 þ bm H km ðxv ðk; mÞÞ ð7Þ
m1

In the Eq. (7) MARS works over M term. The terms b0 , bm are the parameter.
Hinge function i.e. the H can be written as the following equation,
  Y k
Hkm xvðk;mÞ ¼ 1 hkm ð8Þ
k

In the above Eq. (8) the product of kth of the mth term is given as, xv(k, m)
The value of K = 1 and K = 2 gives additive and pairwise interaction respectively.
For this work, the opted value of K is 2.

3 Experiment

The experiment to predict concentration of ozone is carried out by using software

named ‘Salford Predictor Modeller 7.0’ founded in 1983 which supports the three
techniques Random Forest, Multivariate Adaptive Regression Splines and Classifica-
tion and Regression Tree.
A. DATASET
This dataset is given by Saverio De Vito in 2006 from UCI dataset containing the
response of multisensor device deployed on the field in an Italian City. It has 9358
instances with 15 attributes which was recorded from March (2004) to February
(2005). The attributes are Ground Truth hourly averaged concentrations for CO, Non
Metanic Hydrocarbons, Benzene, Total Nitrogen Oxides (NOx) and Nitrogen Dioxide
 Predicting Ozone Layer Concentration 145

(NO2) and were provided by a co-located reference certified analyser. The attributes
included Date, Time, True hourly averaged concentration CO in mg/m3, PT08.S1
(tin oxide) hourly averaged sensor response, True hourly averaged overall Non Metanic
HydroCarbons concentration in microg/m3, True hourly averaged Benzene concen-
tration in microg/m3, PT08.S2 (titania) hourly averaged sensor response, True hourly
averaged NOx concentration in ppb, PT08.S3 (tungsten oxide) hourly averaged sensor
response, True hourly averaged NO2 concentration in microg/m3, PT08.S4 (tungsten
oxide) hourly averaged sensor response, PT08.S5 (indium oxide) hourly averaged
sensor response, Temperature in °C, Relative Humidity (%), AH Absolute
Humidity [7].
B. Air Quality Prediction Steps
This describes the various steps taken for prediction by Salford Modeller.
Step 1: The database is opened in the software as it supports all type of file.
Step 2: The model is designed by selecting the predictor. A total of 12 predictors
are selected for MARS/Random Forest/CART for this dataset. Date and time are not
chosen as they don’t have any effect. PT08_S5_O3_ is set as the target in all the cases.
Step 3: The analysis method is selected as MARS/Random Forest/CART with
analysis type being ‘regression’ in all three cases.
Step 4: It’s time to separate the dataset as learning set and test set. This is done by
selecting Fraction of cases selected at random for testing by assigning any value.
Remember the values are in terms of percent. Here we put the test set as 0.30.
Step 5: Now the model is started and resulting graph pops up showing the infor-
mation required for future prediction of target variable. It also provides summary for all
other details which is discussed in the next section.

4 Result and Discussion

In this section we compare the results given for the target variable PT08_S5_O3_
through Random Forest, Multivariate Adaptive Regression Splines and Classification
and Regression Tree by Salford Predictive Modeller. Out of 15 attributes, 12 are being
used for used as predictors while 1 is selected as the targeted variable and the targeted
variable being PT08.S5 (O3). 30% of the 9358 instances are selected for test case while
rest go in for the learn case. This paper contains the graphical representation of the
learn and test value, summary of important terms, list of variable importance by the
three models used. On applying Multivariate Adaptive Regression Splines, we get
Fig. 1 which shows the graph shows result of learn and test case where the Y-axis
represents MSE with an interval of 50,000 and X-axis representing Basis functions
which was taken as 15 initially. From the graph conclusion can be made that there are
least error as both the learn and test cases are same. Initially the MSE value starts from
200,000 drops till 5000 and gradually becomes constant.
There are several important parameters that give the model error measure. These
important parameters have been listed in Table 1 showing their value for both learn and
test. The variables include RMSE, MSE, GCV, MAD, MRAD, SSY, SSE, R^2, R^2
Norm, GCV R-Sr. Out of 15 attributes, 12 were set as predictors but after the regression
146 S.S. Roy et al.

Fig. 1. MSE vs. basis functions

model was prepared it was deducted that only 8 variables were important for prediction
of PT08.S5 (O3). The most important variable was found to be PT08_S2_NMHC. The
scores of all the variables are given in decreasing order of their importance in predicting
the predictor in Table 2. The number of basis function was set as 15 initially. The
model assigns special variables to make a new equation to cover all points of non-
linearity. These variables are termed as basis variables. The model is a weighted sum of
basis function. Each basis function takes one of the following forms

Table 1. MARS result for learn and test

Name Learn Test
RMSE 126.46071 131.97553
MSE 15,992.31082 17,417.54040
GCV 16,138.69982 n/a
MAD 97.47557 100.28940
MRAD 0.11444 0.11721
SSY 1,376,184,805.13191 57,713,8537.64731
SSE 105,309,366.83668 48,281,422.13143
R^2 0.92348 0.91634
R^2 Norm 0.92348 0.91634
GCV R-Sq 0.92280 n/a

(1) Constant. Only one term i.e. the intercept

(2) Hinge function
(3) Product of 2 or more Hinge function. Table 3 consists all the basis function and
their combination to give the final equation of Y.
Random Forest was started by setting the number of trees to be built as 500 and
number of predictors for each node as 3. The frequency of the report was set 10 along
with parent minimum case as 2. The elapsed time was nearly 45 s for creating the trees.
 Predicting Ozone Layer Concentration 147

Table 2. Variable importance in MARS

Variable Score
PT08_S2_NMHC_ 100.00
PT08_S1_CO_ 96.61
RH 55.18
NO2_GT_ 46.47
PT08_S3_NOX_ 44.38
PT08_S4_NO2_ 18.28
T 16.76
NOX_GT_ 4.64

Table 3. MARS basis functions

BF6 max(0, 11 - NOX_GT_)
BF7 max(0, PT08_S2_NMHC_ - 609.25)
BF8 max(0, 609.25 - PT08_S2_NMHC_)
BF9 max(0, T - 2.1)
BF11 max(0, RH - 63.15)
BF12 max(0, 63.15 - RH)
BF13 max(0, NO2_GT_ - 98)
BF14 max(0, 98 - NO2_GT_)
BF15 max(0, PT08_S4_NO2_ + 200)
Y 769.987 + 0.584446 * BF1 + 0.860364 * BF2 − 0.0722947 * BF3 + 0.317973 *
BF4 + 0.0305292 * BF5 − 0.42804 * BF6 + 0.747153 * BF7 − 1.56998 * BF8 −
2.74185 * BF9 − 5.13236 * BF11 − 3.58688 * BF12 + 1.15626 * BF13 + 0.359236
* BF14 − 0.102166 * BF15

Separate graphs were obtained showing the comparison of train as in Fig. 2 and test
cases as in Fig. 3 with 500 trees having maximum terminal node of 3293. Observing
the curve in both the cases, a clear difference can be seen in the curve. Both start with
MSE 40,000 but the train set is less steep when compared with test set. Train set shows
a turn at 16th tree (18,502.625) while test set shows a turn at 8th tree (16,911.168).
Both the graph for train and test are given in Figs. 2 and 3 respectively.
There are several important parameters that give the model error measure. These
important parameters have been listed in Table 4 showing their value for both learn and
test. The variables include RMSE, MSE, MAD, MRAD, SSY, SSE, R^2, R^2 Norm.
Unlike MARS in random forest all 12 variables have their own importance and show
their contribution for building trees. Importance of each variable has been shown in
Table 5.
Classification and Regression Tree model also leads in building up of trees. It gives
a graph where the Y-axis shows Relative Error while X-axis shows the number of
nodes. The graph value shows the relative error value as 0.083 at 150th node. So by
examining the graph directly we can get the relative error of test from the train set as
shown in Fig. 4.
148 S.S. Roy et al.

Fig. 2. Train set plot in random forest

Fig. 3. Test set plot in random forest

Table 4. Random forest result for learn and test

Name OOB Test
RMSE 112.45649 117.28208
MSE 12646.46110 13755.08634
MAD 84.08904 86.55029
MRAD 0.10388 0.10746
SSY 1,376,184,805.13192 577,138,537.64731
SSE 83,276,946.34398 38,129,099.33380
R^2 0.93949 0.93393
R^2 Norm 0.94676 0.94078
 Predicting Ozone Layer Concentration 149

Table 5. Variable importance in random forest

Variable Score
PT08_S1_CO_ 100.0000
C6H6_GT_ 59.0625
PT08_S2_NMHC_ 57.9778
PT08_S3_NOX_ 28.8258
PT08_S4_NO2_ 19.8222
NOX_GT_ 17.0169
RH 13.3625
T 12.6999
NO2_GT_ 11.0866
AH 10.2157
CO_GT_ 5.6714
NMHC_GT_ 1.0818

Fig. 4. Relative error through CART

There are several important parameters that give the model error measure. These
important parameters have been listed in Table 6 showing their value for both learn and
test. The variables include RMSE, MSE, MAD, MRAD, SSY, SSE, R^2, R^2 Norm,
AIC, AICc, BIC, Relative Error. CART also requires the use all 12 predictor variables
just as Random Forest does. Table 7 lists the variable according to their importance.
150 S.S. Roy et al.

Table 6. Random forest result for learn and test

Name Learn Test
RMSE 105.96244 131.25001
MSE 11,228.03897 17,226.56562
MAD 80.66569 96.96731
MRAD 0.08853 0.10464
SSY 1,376,184,805.13191 577,138,537.64731
SSE 73,936,636.62739 47,752,039.90055
R^2 0.94627 0.91726
R^2 Norm 0.94627 0.91736
AIC 61,436.82555 27,062.66453
AICc 61,436.87303 27,062.77762
BIC 61,518.33615 27,133.79242
Relative Error 0.05373 0.08274

Table 7. Variable importance in CART

Variable Score
PT08_S1_CO_ 100.0000
PT08_S2_NMHC_ 97.3907
C6H6_GT_ 97.3598
PT08_S3_NOX_ 60.6814
PT08_S4_NO2_ 53.5883
CO_GT_ 44.9134
T 22.8821
NOX_GT_ 7.2299
NO2_GT_ 3.2165
AH 1.9955
RH 1.4960
NMHC_GT_ 0.2262

5 Conclusion

In this paper we have proposed to show the prediction of ozone concentration by using
three regression model. By keeping the train and test in the ratio 7:3 we compare the
result from all three cases. Evaluation of the prediction models indicates that the
Multivariate Adaptive Regression Splines model describes the dataset better and has
achieved significantly better prediction accuracy as compared to the Random Forest
and Classification and Regression Tree. Multivariate Adaptive Regression Splines
gives the result by considering less variables as compared to other two. It evaluates on
basis of 8 variables while other two require all variables. Moreover, Random Forest
takes a little more time for building the tree as the elapsed time was calculated to 45 s
in this case. PT08_S2_NMHC_ is the most important variable as given by Multivariate
Adaptive Regression Splines while PT08_S1_CO_ is most important variable as given
 Predicting Ozone Layer Concentration 151

by Random Forest and Classification and Regression Tree. Observing all the graphs
Multivariate Adaptive Regression Splines gives the closest curve of both train and test
set when compared. It can be concluded that multivariate adaptive regression splines
can be a valuable tool in predicting ozone for future.

References
1. Ozer, P., Laghdaf, M.B.O.M., Lemine, S.O.M., Gassan, J.: Estimation of air quality
degradation due to Saharan dust at Nouakchott, Mauritania, from horizontal visibility data.
Water Air Soil Pollut. 178(1–4), 79 (2007)
2. Zhang, W.Y., Han, T.T., Zhao, Z.B., Zhang, J., Wang, Y.F.: The prediction of surface layer
ozone concentration using an improved AR Model. In: International Conference of
Information Technology, Computer Engineering and Management Sciences, Nanjing,
Jiangsu, pp. 72–75 (2011)
3. Assessment for Decision-Makers: Scientific Assessment of Ozone Depletion: 2014, World
Meteorological Organization, Global Ozone Research and Monitoring Project—Report
No. 56, Geneva, Switzerland (2014)
4. Zhang, C., Yuan, D.: Fast fine-grained air quality index level prediction using random forest
algorithm on cluster computing of spark. In: 2015 IEEE 12th International Conference on
Ubiquitous Intelligence and Computing and 2015 IEEE 12th International Conference on
Autonomic and Trusted Computing and 2015 IEEE 15th International Conference on
Scalable Computing and Communications and Its Associated Workshops
(UIC-ATC-ScalCom), Beijing, pp. 929–934 (2015)
5. Wang, Y., Yue, W.: Target data association in communication constrained environment
using CART: compressed adaptive reference topology. In: 2016 8th IEEE International
Conference on Communication Software and Networks (ICCSN), Beijing, China, pp. 333–
338 (2016)
6. Birgersson, M., Hansson, G., Franke, U.: Data integration using machine learning. In: 2016
IEEE 20th International Enterprise Distributed Object Computing Workshop (EDOCW),
Vienna, Austria, pp. 1–10 (2016)
7. De Vito, S., Massera, E., Piga, M., Martinotto, L., Di Francia, G.: On field calibration of an
electronic nose for benzene estimation in an urban pollution monitoring scenario. Sens.
Actuators B Chem. 129(2), 750–757 (2008)
8. De Vito, S., Piga, M., Martinotto, L., Di Francia, G.: CO, NO2 and NOx urban pollution
monitoring with on-field calibrated electronic nose by automatic bayesian regularization.
Sens. Actuators B Chem. 143(1), 182–191 (2009). ISSN: 0925-4005
9. Hui, T.S., Rahman, S.A., Labadin, J.: Comparison between multiple regression and
multivariate adaptive regression splines for predicting CO2 emissions in ASEAN countries.
In: 2013 8th International Conference on Information Technology in Asia (CITA), Kota
Samarahan, pp. 1–5 (2013)
10. Friedman, J.H.: Multivariate adaptive regression splines. Ann. Stat. 19, 1–67 (1991)
11. Breiman, L., Friedman, J.H., Olshen, R., Stone, C.J.: Classification and Regression Tree.
Wadsworth & Brooks/Cole Advanced Books & Software, Pacific California (1984)
12. Abbasi, R., Moradi, M.H., Molaeezadeh, S.F.: Long-term prediction of blood pressure time
series using multiple fuzzy functions. In: 2014 21th Iranian Conference on Biomedical
Engineering (ICBME), Tehran pp. 124–127 (2014)
152 S.S. Roy et al.

13. Roy, S.S., Viswanatham, V.M., Krishna, P.V., Saraf, N., Gupta, A., Mishra, R.:
Applicability of rough set technique for data investigation and optimization of intrusion
detection system. In: International Conference on Heterogeneous Networking for Quality,
Reliability, Security and Robustness, pp. 479–484. Springer, Berlin Heidelberg, January
2013
14. Roy, S.S., Viswanatham, V.M., Krishna, P.V.: Spam detection using hybrid model of rough
set and decorate ensemble. Int. J. Comput. Syst. Eng. 2(3), 139–147 (2016)
15. Roy, S.S., Viswanatham, V.M.: Classifying spam emails using artificial intelligent
techniques. Int. J. Eng. Res. Afr. 22, 152–161 (2016)
16. Roy, S.S., Mittal, D., Basu, A., Abraham, A.: Stock market forecasting using LASSO linear
regression model. In: Afro-European Conference for Industrial Advancement, pp. 371–381.
Springer International Publishing (2015)
17. Basu, A., Roy, S.S., Abraham, A.: A Novel diagnostic approach based on support vector
machine with linear kernel for classifying the Erythemato-Squamous disease. In: 2015
International Conference on Computing Communication Control and Automation
(ICCUBEA), pp. 343–347. IEEE, February 2015
18. Mittal, D., Gaurav, D., Roy, S.S.: An effective hybridized classifier for breast cancer
diagnosis. In: 2015 IEEE International Conference on Advanced Intelligent Mechatronics
(AIM), pp. 1026–1031. IEEE, July 2015
19. Roy, S.S., Gupta, A., Sinha, A., Ramesh, R.: Cancer data investigation using variable
precision Rough set with flexible classification. In: Proceedings of the Second International
Conference on Computational Science, Engineering and Information Technology, pp. 472–
475. ACM, October 2012
20. Samui, P., Roy, S.S., Kurup, P., Dalkiliç, Y.: Modeling of Seismic Liquefaction Data Using
Extreme Learning Machine, Nova publishers, Natural Disaster Research, Prediction and
Mitigation, pp. 6x9-(NBC-R). ISBN: 978-1-53610-356-4
21. Roy, S.S., Krishna, P.V., Yenduri, S.: Analyzing intrusion detection system: an ensemble
based stacking approach. In: 2014 IEEE International Symposium on Signal Processing and
Information Technology (ISSPIT), pp. 000307–000309 (2014)
22. Popescu-Bodorin, N., Balas, V.E., Motoc, I.M.: Iris codes classification using discriminant
and witness directions. arXiv preprint arXiv:1110.6483 (2011)
23. Azar, A.T., Balas, V.E., Olariu, T.: Artificial neural network for accurate prediction of
post-dialysis urea rebound. In: 2010 4th International Workshop on Soft Computing
Applications (SOFA), pp. 165–170. IEEE, July 2010
24. Roopaei, M., Balas, V.E.: Adaptive gain sliding mode control in uncertain MIMO systems.
In: 3rd International Workshop on Soft Computing Applications, SOFA 2009, pp. 77–82.
IEEE, July 2009
25. Ponalagusamy, R., Senthilkumar, S.: Investigation on time-multiplexing cellular neural
network simulation by RKAHeM (4, 4) technique. Int. J. Adv. Intell. Paradig. 3(1), 43–66
(2011)
 Prediction of Customer Satisfaction
Using Naive Bayes, MultiClass Classifier,
K-Star and IBK

Sanjiban Sekhar Roy1(&), Deeksha Kaul1, Reetika Roy1,

Cornel Barna2, Suhasini Mehta1, and Anusha Misra1
1
School of Computer Science and Engineering, VIT University, Vellore, India
[email protected], [email protected],
[email protected],
{suhasini.mehta2014,anusha.misra2014}@vit.ac.in
2
Automation and Applied Informatics,
Aurel Vlaicu University of Arad, Arad, Romania
[email protected]

Abstract. Customer satisfaction is an important term in business as well as

marketing as it surely indicates how well the customer expectations have been
met with by the product or the service. Thus a good prediction model for
customer satisfaction can help any organization make better decisions with
respect to its services and work in a more informed matter to improvise on the
same. The problem considered in this study is optimization of customer satis-
faction for the customers of San Francisco International Airport. This paper
adopts three classification models Naive Bayes, MultiClass Classifier, K-Star
and IBK as potential classifiers for prediction of customer satisfaction. The
customer satisfaction depends on various factors. The factors which we consider
are the user ratings for artwork and exhibitions, restaurants, variety stores,
concessions, signage, directions inside SFO, information booths near baggage
claim and departure, Wi-Fi, parking facilities, walkways, air train and an overall
rating for the airport services. The ratings are obtained from a detailed customer
survey conducted by the mentioned airport in 2015. The original survey focused
on questions including airlines, destination airport, delays of flights, conveyance
to and from the airport, security/immigration etc. but our study focuses on the
previously mentioned questions. Graphs are plotted for actual and predicted
values and compared to find the least amount of deviation from the actual
values. The model which shows least deviation from actual values is considered
optimal for the above mentioned problem.

Keywords: Customer satisfaction
Prediction
Classification
Clustering

Bayes Theorem
Naïve bayes
Multiclass classifier
K-Star
IBK
Weka

San francisco
International airport
Survey data

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_12
154 S.S. Roy et al.

1 Introduction

Customer’s appraisal of quality of service is crucial information for organizations that

aim to enhance their performance and better their position in the market. Customer
satisfaction is said to be “the number of customers, or percentage of total customers,
whose reported experience with a firm, its products, or its services (ratings) exceeds
specified satisfaction goals [5]. Customer satisfaction is a vital metric in marketing as it
is an indicator of consumer loyalty and their intention of repurchase, retains customer
and consequently reduces churns, increases customer lifetime value, increases positive
word of mouth and reduces cost of acquiring new customer base [1–4]. Customer
loyalty is a driving factor of most established companies in the face of a competitive
environment as the cost of retaining customer is always lower than acquiring new
customers. Customer loyalty can be ensured by surpassing the expectation of the
customer and providing them with products or services that live up to their expectations
or exceed them. This culminates in the customer a tendency to choose the product or
service over other similar options when faced with a choice and also to suggest it to
other people. Another very important metric that comes into the paly is customer
lifetime value (CLV or CLTV) which is a mathematical expression that approximately
predicts the total profit obtained from a customer from the entire relationship with
them. It helps to focus on long term effects of the customer relationship and the
determine the cost that should be put into retaining or acquiring the customer. More-
over, organizations with better quality of service gain higher returns on investments
and have more satisfied customers. It is becoming more and more important for
businesses overcome competition by providing superior service and gaining customer
loyalty. Zeithaml (1988) defines service quality as the customer’s determination of the
overall performance of the product/service. Customer satisfaction can be equated to
influencing loyalty of customers, encouraging them to reuse the organization’s ser-
vices, which, effectively, brings in future revenue and profits. As it can directly affect
profits, service quality and customer satisfaction have become an important part of
marketing strategy for all organization including San Francisco International Airport.
Similar kind of research can be seen by a number of airports using different techniques
[6–8]. Increasing number of organizations are compelled to improve their service
quality and target the satisfaction of their customers in an effort to earn their loyalty and
are willing to invest in various sophisticated computational methods to center their
efforts for optimal results that help them enhance the customer satisfaction. Advanced
machine larning algorithms are being applied in different fields like medicine like
cancer investigation [9, 10] and disease classification [11, 12], stock market prediction
[13], spam detection [14, 15], intrusion detection [16, 17] as well as other engineering
fields like civil and mechanical engineering [17, 18]. The use of machine learning in
business and market analytics is hardly unknown. Classification, also known as doc-
ument classification, is a fundamental technique of data analysis which can be used for
predicting future data trends. The forecasting is based on presently available infor-
mation, i.e. if we are supposed to construct a procedure which is already applied to a
continuous sequence of cases where a new case is assigned to a set of pre-defined
classes based on the features of observed data. This method is termed as supervised
 Prediction of Customer Satisfaction Using Naive Bayes 155

learning or pattern recognition. Classification models entail assigning the data to a

predefined category. The models referred to in this study are Naive Bayes [19],
Multiclass Classifier [20, 21], K-Star and IBK [2] classification models. These models
will be discussed in detail in the following sections. The main aim of the algorithms is
to reduce the classification error. This is done in two stages. First, the classification
algorithm is applied on the training set to obtain a suitable model. Further, the created
model is tested against an existing test set to identify the validity and performance of
the model. In this paper, we have used classification model to measure customer
satisfaction. The two classifiers analyzed here are Bayesian (Bayes net, Naïve Bayes
etc.) and Lazy (IBL, IBK, K Star etc.). For implementation purpose we have used
WEKA (Waikato Environment for Knowledge Analysis) [22] which is a Data Analysis
Tool with algorithms embedded for statistical modeling and data analysis. It provides a
variety of algorithm options and evaluation metrics [23, 24] that were used in this
research. The various applications available in WEKA are Explorer, Experimenter,
Knowledge Flow, Workbench and Simple CLI. Our work was predominantly done in
WEKA Explorer. The data was collected through an elaborate customer survey con-
ducted by the San Francisco International Airport in the year 2015. The opinion poll
involved questions ranging from cleanliness, Wi-Fi, signage, immigration, road, food
and similar such factors. Among a wide range of data the factors which have been
considered for this study include art, food, stores, Wi-Fi, road, parking, etc. The
number of times a customer has flown from the airport is considered as the measure for
customer satisfaction. The purpose of this study is to provide the most appropriate
model which can be used in the future to predict customer satisfaction. From the
various models and algorithms in WEKA we have considered Naive Bayes, Multiclass
Classifier, K-star [25] and IBK classification models. The model with minimum error
percentage is concluded as the most efficient method for prediction of customer sat-
isfaction for the referred data. The predictive values and actual values are plotted in
MS-Excel and compared through graph and their statistical performance chart.

2 Methodology
2.1 Naïve Bayes
According to Bayesian classification, a data item is sorted into a class in order to
predict the values of features for other members of the same class. Previously known
data items are grouped into classes based on the similarity in the values of the features.
The basic idea of Bayesian classification method is that if the algorithm can success-
fully identify the class, it can also predict the values of the other features. If the class
cannot be identified, it is predicted using Bayes’ rule if some of the feature values are
known. In such a classifier, a probabilistic is built by the learning algorithm which
further classifies the new data items. The Naive Bayes classifier is a model of super-
vised learning algorithm which applies Bayes’ theorem with the ‘Naïve’ assumption of
pairwise independence between all the features. Assuming y is a class variable and x1,
x2 ….xn are dependant feature vectors, then according to Bayes’ Theorem, the below
relationship follows.
156 S.S. Roy et al.

Pð yÞPðx1 ; . . .; xn jyÞ
Pðyjx1 ; x2 . . .:xn Þ ¼ ðiÞ
Pðx1 ; . . .; xn Þ

Assuming pairwise independence,

PðyÞpni¼1 Pðxi j yÞ
Pðyjx1 ; x2 . . .xn Þ ¼ ðiiÞ
Pðx1 ; . . .; xn Þ

Since Pðx1 ; . . .; xn Þ is constant for any given input, following classification rule can be
applied.
Yn
Pðyjx1 ; x2 ; . . .; xn Þa PðyÞ i¼1
Pðxi jyÞ ðiiiÞ

+
Yn
^y ¼ arg max PðyÞ i¼1
Pðxi jyÞ ðivÞ
y

MAP estimation can be used to estimate P(y) and P(xi|y).

Although Naive Bayesian classifier produces unsatisfactory results in some cases, it
is very simple and time and cost effective in most cases. Generally, the Naive Bayesian
classifier produces efficient results when the class can be effectively used to predict
other features and the other features are independent of the classes.

2.2 Multiclass Classifier

In MultiClass Classifier, each training point is categorized to a class out of N, such that
N > 2. If N is 1 then the entire set of points becomes a class and if N is equal to the size
of the training set then every individual point becomes a class. It is often important to
realise the optimum value of N to properly create classes, which is done by minimizing
intra-class distance and maximizing inter-class distance (that is, every feature is more
similar to features within its class and less to those in other classes). MultiClass
Classifier constructs a function which, for every new data point, correctly predicts its
class. There can be multiple categories but any point can belong to a single class only.
In this way, the problem is divided into independent problems which can be efficiently
solved through Binary Classification techniques.
If the density Pi(x) is known, prediction can be done using

F ð xÞ ¼ arg max Pið xÞ where i 2 1; . . .; N ðvÞ

But the densities are unknown and hard to estimate. In case of binary classification,
direct estimation of a separating function provides better results than estimating
probability density. Using one-vs-all and one-vs-one approaches this can be applies to
MultiClass classifiers as well.
 Prediction of Customer Satisfaction Using Naive Bayes 157

2.2.1 One-vs-All
This technique requires building N binary classifiers. Assuming, for the ith classifier,
the positive examples are the points belonging to the ith class and negative examples
are the points that do not belong to the ith class. If fi be the ith classifier, then,

f ð xÞ ¼ arg max fið xÞ; for all values of I ðviÞ

2.2.2 One-vs-One
This technique requires building N*(N − 1) classifiers, one classifier to distinguish
every single pair of classes. Let fij be the classifier where the ith and jth class has
positive and negative examples respectively.
Also,
fij ¼ fji ðviiÞ

Then,
!
X
f ðxÞ ¼ arg max fij ðxÞ ðviiiÞ
i
j

2.3 IBK
The IBK classifier (instance bases learning with parameter K) comes under the category of
lazy class folder. Lazy learners maintain a set of training instances and start the real work
during classification. The classifier implements k-nearest neighbor algorithm which is
useful for calculation of the known values near the unknown data point and using the
calculated values to predict the unknown values. The classifier finds the closest training
instance (in Euclidean distance) as well as various attributes such as numeric, date, unary,
binary, missing attributes when classifying a training instance. IBK implementation uses
cross validation to select the best value of K. The output varies with the model used.
Various search algorithms can be used for finding nearest neighbors such as linear search,
KD-trees, ball trees. In this model the number of instances is kept constant therefore as the
new training sets are added the old ones are detached.
K-Nearest neighbor algorithm
Train
Set S of training data is built.
Classify
Let xq be an instance to be classified. Let x1 … xk denote k instances from
S closest to xq
Return
where
(a,b)=1 ; a=b
(a,b)=0 ; otherwise.
158 S.S. Roy et al.

2.4 K-Star
In IB classification problems, “based on a distance metric, every new instance is
compared with all the existing ones, and based on the closest existing instance, a class
is assigned to the new instance”. The basic difference between Kstar and the other IB
classification methods is the concept used for defining the distance metric. Kstar is a
lazy learning algorithm that aims at partitioning n data points into k clusters such that
every data point is classified into the cluster with the closest mean. K* algorithm can be
described as an instance based learner which applies K Nearest Neighbour method with
entropy as a measure of distance. New data points, x, are assigned to the nearest class,
yj, where j = 1…k. Using entropic distance for the retrieval of similar data points
allows handling missing values and real valued attributes. The K* function is calcu-
lated as follows:

K  ðyi; xÞ ¼ ln P ðyi; xÞ ðixÞ

where P* represents the probability that x will reach y through a random path.
Many studies use K* for various classification problems with satisfactory results. In
case of large variables and small k values K* provides faster computations than
hierarchical clustering. However, predicting the value of k can be a challenge and the
method might not give similar results in case the size and density of clusters vary.

3 Result and Discussion

In this paper, the customer satisfaction target variable has been considered as the
overall satisfaction variable in the questionnaire for the survey. Based on the ratings
given on the remaining questions which become our independent variables, the model
tries to predict the customer overall satisfaction level. The customer satisfaction has
classes 0 to 6− with 1 to 5 as unacceptable to outstanding, 0 fro blank and 6 for not
applicable. The dataset was split into train and test data with a 70–30 ratio. The models
were built on the train data and the test data was used to make new predictions to see
model performance. The predicted classes are then compared to the actual class by
WEKA and the number of correct and incorrect classification can be seen.
Table 1 represents the statistical performance for the Naïve Bayes, K-Star, IBK and
Multiclass Classification models based on the classification made on the test data. From
the table it can be observed that Naïve Bayes and Multiclass Classifier have similar

Table 1. Comparative Study of Naïve Bayes, K-Star, IBK, MultiClass

Models Statistical performances
Correctly Incorrectly % correct % incorrect Kappa
classified classified statistic
Naïve Bayes 1235 836 59.63 40.37 0.3933
K-Star 2029 42 97.97 2.02 0.9684
IBK 2039 32 98.45 1.55 0.9759
Multiclass 1280 791 61.81 38.19 0.3052
classifier
 Prediction of Customer Satisfaction Using Naive Bayes 159

performances while the lazy classifiers have similar performance. The lazy classifiers
however, perform better than the others with IBK giving the most accurate results for
the considered data set. Moreover, Figs. 1, 2, 3 and 4 show the area under the ROC
curve for all the classifiers.

Fig. 1. AUC-ROC for Naïve Bayes

Fig. 2. AUC-ROC for multiclass classifier

Fig. 3. AUC-ROC for K-Star

160 S.S. Roy et al.

Fig. 4. AUC-ROC for IBK

4 Conclusion

Considering the statistical performance table and the above graphs, the IBK model
shows maximum accuracy in classifying the customer satisfaction level, hence for this
data set we can conclude that IBK classifier model provides the optimal results.
Artificial customer data can be generated and the satisfaction levels for different kinds
of customer can be studied. We can also do an analysis on which factors affect the
customer satisfaction in what way. This kind of work comes under the umbrella of
sentiment analysis by which customer behavior can be studied further. For further
optimization of the prediction, the data quality can be improved. Moreover, in this
study only a limited number of factors have been considered, but by considering more
factors like delays, facilities like e-checkin, quality of WiFi, conveyance to and from
the airport, etc. The questionnaire can be refined and detailed for better quality
information for more accurate prediction results.

References
1. Ali, S., Smith, K.A.: On learning algorithm selection for classification. Appl. Soft Comput. 6
(2), 119–138 (2006)
2. Vijayarani, S., Muthulakshmi, M.: Comparative analysis of bayes and lazy classification
algorithms. Int. J. Adv. Res. Comput. Commun. Eng. 2(8), 3118–3124 (2013)
3. Deepajothi, S., Selvarajan, S.: A comparative study of classification techniques on adult data
set. Int. J. Eng. Res. Technol. (2012)
4. Funa, L.: Customer satisfaction analysis. (2011)
5. Farris, P.W., Neil, T.B., Phillip, E.P., Reibstein, J.B.: Marketing Metrics: The Definitive
Guide to Measuring Marketing Performance. Pearson Education, Inc., Upper Saddle River
(2010). ISBN 0-13-705829-2
6. Al-Refaie, A., Bata, N., Eteiwi, D., Jalham, I.: Examining factors that affect passenger’s
overall satisfaction and loyalty: evidence from jordan airport. Jordan J. Mech. Ind. Eng. 8(2),
94–101 (2014)
7. Suki, N.M.: Passenger satisfaction with airline service quality in Malaysia: a structural
equation modeling approach. Res. Trans. Bus. Manag. 10, 26–32 (2014)
 Prediction of Customer Satisfaction Using Naive Bayes 161

8. Harvey, G.: Airport choice in a multiple airport region. Trans. Res. A Gen. 21(6), 439–449
(1987)
9. Mittal, D., Gaurav, D., Roy, S.S.: An effective hybridized classifier for breast cancer
diagnosis. In: 2015 IEEE International Conference on Advanced Intelligent Mechatronics
(AIM), pp. 1026–1031 (2015)
10. Roy, S.S., Gupta, A., Sinha, A., Ramesh, R.: Cancer data investigation using variable
precision Rough set with flexible classification. In: Proceedings of the Second International
Conference on Computational Science, Engineering and Information Technology,
New York, pp. 472–475. ACM (2012)
11. Basu, A., Roy, S.S., Abraham, A.: A novel diagnostic approach based on support vector
machine with linear kernel for classifying the erythemato-squamous disease. In: 2015
International Conference on IEEE Computing Communication Control and Automation
(ICCUBEA), pp. 343–347 (2015)
12. Azar, A.T., Balas, V.E., Olariu, T.: Artificial neural network for accurate prediction of
post-dialysis urea rebound. In: 2010 4th International Workshop on IEEE Soft Computing
Applications (SOFA), pp. 165–170 (2010)
13. Roy, S. S., Mittal, D., Basu, A., Abraham, A.: Stock market forecasting using LASSO linear
regression model. In: Afro-European Conference for Industrial Advancement, pp. 371–381.
Springer International Publishing (2015)
14. Roy, S.S., Viswanatham, V.M., Krishna, P.V.: Spam detection using hybrid model of rough
set and decorate ensemble. Int. J. Comput. Syst. Eng. 2(3), 139–147 (2016)
15. Roy, S.S., Viswanatham, V.M.: Classifying spam emails using artificial intelligent
techniques. Int. J. Eng. Res. Afr. 22 (2016)
16. Roy, S.S., Viswanatham, V.M., Krishna, P.V., Saraf, N., Gupta, A., Mishra, R.:
Applicability of rough set technique for data investigation and optimization of intrusion
detection system. In: International Conference on Heterogeneous Networking for Quality,
Reliability, Security And Robustness, pp. 479–484. Springer, Berlin, Heidelberg (2013
17. Roy, S.S., Krishna, P.V., Yenduri, S.: Analyzing intrusion detection system: an ensemble
based stacking approach. In: 2014 IEEE International Symposium on Signal Processing and
Information Technology (ISSPIT), pp. 000307–000309 (2014)
18. Roopaei, M., Balas, V.E.: Adaptive gain sliding mode control in uncertain MIMO systems.
In: 3rd International Workshop on IEEE Soft Computing Applications, SOFA 2009, pp. 77–
82 (2009)
19. McCallum, A., Nigam, K.: A comparison of event models for naive bayes text classification.
In: AAAI 1998 Workshop on Learning for Text Categorization, vol. 752, pp. 41–48 (1998)
20. Huang, G.B., Zhou, H., Ding, X., Zhang, R.: Extreme learning machine for regression and
multiclass classification. IEEE Trans. Syst. Man Cybern. Part B (Cybern.) 42(2), 513–529
(2012)
21. Tax, D.M., Duin, R.P.: Using two-class classifiers for multiclass classification. In IEEE
Proceedings of 16th International Conference on Pattern Recognition, vol. 2, pp. 124–127
(2002)
22. Okhovvat, M., Naderi, H.: Comparing various classification algorithms by WEKA (2015)
23. Sharma, T.C., Jain, M.: WEKA approach for comparative study of classification algorithm.
Int. J. Adv. Res. Comput. Commun. Eng. 2(4), 1925–1931 (2013)
24. Raviya, K.H., Gajjar, B.: Performance evaluation of different data mining classification
algorithm using WEKA. Indian J. Res. 2(1), 19–21 (2013)
25. Cleary, J.G., Trigg, L.E.: K*: an instance-based learner using an entropic distance measure.
In: Proceedings of the 12th International Conference on Machine learning, vol. 5, pp. 108–
114 (1995)
 Spam Email Detection Using Deep Support
Vector Machine, Support Vector Machine
and Artificial Neural Network

Sanjiban Sekhar Roy1(&), Abhishek Sinha1, Reetika Roy1,

Cornel Barna2, and Pijush Samui3
1
School of Computer Science and Engineering, VIT University, Vellore, India
[email protected], [email protected],
[email protected]
2
Automation and Applied Informatics, Aurel Vlaicu University of Arad,
Arad, Romania
[email protected]
3
Department of Civil Engineering, NIT Patna, Patna, India
[email protected]

Abstract. Emails are a very important part of our life today for information
sharing. It is used for both personal communication as well as business pur-
poses. But the internet also opens up the prospect of an enormous amount of
junk and useless information which overwhelms and irritates us. These unnec-
essary and unsolicited emails are what comprise of spam. This study presents
the application of a classification model to classify spam emails from using a
model- Deep Support Vector Machine (Deep SVM). Moreover, other classifier
models like Support Vector Machine (SVM), Artificial Neural Network models
have also been implemented to compare the performance of proposed
Deep SVM model. Furthermore analysis has been done to compare all the
performances using available numerical statistics obtained from these models to
find the best model for the purpose. Spam filtering is a very essential feature in
most email services and thus effective spam classification models are pertinent
to the current digital communication scenario and various work has been done in
this area.

Keywords: Spam
Classification
Deep Support Vector Machine
Support

Vector Machine
Artificial Neural Network

1 Introduction

Most communication today, be it personal or professional is done via emails. This can
be attributed to the ease of using emails for communication. They are lightning fast,
inexpensive, instantaneous and accessible from any part of the globe. Moreover, they
may hold text, audio, video, images, links and documents itself. As people’s habit of
reading emails increased, more and more people started using it to send out informa-
tion. From there came the brilliant idea of advertising over email. Emails provide the
perfect opportunity for sending out the same advertisement to millions of people at one

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_13
 Spam Email Detection Using Deep Support Vector Machine 163

go and low cost. However, these emails can be often pestering and unsolicited and
consequently unwanted by people. This kind of bulk email forms our “junk” or “spam”
[1]. These spam mails not only irritate people but may also use up a lot of bandwidth of
our mail server or cause transmission delays. Hence they pose a great botheration in the
internet community.
There may be a variety of ways to reduce spam including legal methods (laws),
personal methods (not forwarding chain mails, being careful while sharing email id) as
well as technical method like blocking of some IPs but there is no concrete method yet
to avoiding spam emails. Email service providers also come with spam filters that tries
to separate out the junk mails into a separate folder by often mistakes can happen that
can block legitimate correspondence with people. It makes a lot of time to check the
inbox for spam manually and delete it and important emails could be deleted by
mistake. Thus, automatic spam filtering is one of the best methods available today, and
as spam filters get more accurate, the tricks used by spammers also improve ensuing a
constant need to find better spam filter models.
Machine learning has founds its use in every domain be it medical, civil [2],
mechanical, electrical [3], business or stock market. From predicting [4] and classifying
[5] diseases and diagnosing cancer [6, 7], to forecasting the stock [8], weather forecast,
intrusion detection in a network [9, 10], computer vision, risk prediction in banks,
machine learning has found its use in multiple disciplines. Within the umbrella of
machine learning, classification and regression are very popular supervised learning
techniques. Supervised learning forms a class of algorithms where the train data is
present with a set of labelled output and the model tries to predict a value closest to this
desired output. Spam classification is a very common binary classification problem
where the classifier has to decide if the mail is spam or not. One of the most popular
statistical model used for spam filtering is the Naïve Bayes algorithm. This is one of
theoldest algorithms and is used because it adapt with the users email preferences
andgives quite low false positive hits. There has been a lot of research in this field using
different algorithms [11–13] and also in optimizing the existing algorithms for better
results.
Deep learning is a fairly newer branch of machine learning which provides an array
of algorithms that tries to model more complex levels of abstraction in data using
multiple layers with linear or non-linear transformations. It shows promise in being
more efficient in learning features than the simpler algorithms and thus try to make
better representations of data before building a model. In this study, we try to build
different classification models for a spam-emails data set. Three classification models
have been built using Deep SVM, SVM, Neural Network and tested on the test data.
Fundamental criteria like Area under curve (AUC), Accuracy, F1 score, Recall, Pre-
cision were used to compare the models. Curves for ROC, Lift, Precision-Recall are
used for graphical analysis of the performances of these aforementioned models. The
fundamental idea is to analyze performance of above mentioned models with each
other and possibly find the best among them.
164 S.S. Roy et al.

2 Theoretical Background
2.1 Support Vector Machine Model
The Support Vector Machine (SVM) model [14, 15] was first proposed by Vapnik and
has became one of the popularly used model in machine learning. According to the
recent studies SVM [16] generally is known to give better accuracy opposed to other
data classification model. SVM has been applied in vast areas of problems such as text
categorization, handwritten word recognition, tone recognition, image classification
and detection, micro-array gene expression analysis etc. It also is superior in supervised
learning models. In SVM each data point is considered as a n-dimensional vector and
tries to dissect data points using (n − 1) dimensional hyperplane. Out of many
hyperplanes that can divide the data set, it’s best choice is to find the hyperplane which
farthest in distance from its nearest data point SVM [17] is a linear model which tries to
simultaneously minimize the empirical classification error and maximize the geomet-
rical distance between data points and classifying hyperplane.
We can consider data points of the form

fðx1; y1Þ; ðx2; y2Þ; ðx3; y3Þ; ðx4; y4Þ. . .. . .. . .:; ðxn; ynÞg

where,
yn = 1/−1, a constant denoting the class to which that point xn belongs.
n = number of samples
each x = k (a n-dimensional real vector)
The scaling is important to guard against variable (attributes) with larger variance.
We can view this training data, by means of the dividing (or separating) hyper plane
[18], which takes

w
x þ b ¼0 ð1Þ

Here b is a scalar and w is any n-dimensional vector. The vector w is in the

direction that is perpendicular to the direction of the separating hyper-plane. The
parameter is used to regulate distance. If there is no any parameter, the hyper-plane is
forced to go through the origin, resulting in confined solution [19].
Parallel hyper-plane give us the maximum margin and therefore

w:x þ b ¼ 1 ð2Þ

w:x þ b ¼ 1 ð3Þ

2.2 Neural Network Model

Neural networks ape the behavior of the human brain and are a very popular machine
learning algorithm that has been used in different classification and prediction problems
[20–22]. Any system with a set of inputs and outputs and is made of many simple and
 Spam Email Detection Using Deep Support Vector Machine 165

identical processing elements. These processing elements have a number of internal

parameters called weights. The behavior of any element can be altered by updating the
weights of elements that are feeded to desired element and hence it will also alter the
behavior of the whole network. The ultimate goal is to choose the weights of the
network such that it achieves a desired input or output relationship.
ADALINE, the processing element used in the work here has input vector
X = {xi}, which contains n components, a single output y, and a weight vector
W = {wi}, which also has n components. The output y is obtained from the sum of the
nonlinear function f(s) and the one used here is the sigmoid function inputs multiplied
by the weights and then passed through a nonlinear function.
SðXÞ ¼ 1=ð1 þ ex Þ ð4Þ

Y ð X Þ ¼ f ðsð X ÞÞ ð5Þ

F ðsÞ ¼ ½1  expð2sÞ=½1 þ expð2sÞ ð6Þ

F ðsÞ ¼ tanhðsÞ ð7Þ

With this non-linearity, the ADALINE behaves similar to a linear filter when its
output is small, but saturates to +1 or −1 as the output magnitude increases. It should be
noted that one of the ADALINE input is set to +1 which helps in a way of adding constant
bias to the weighted sum. A good measure of the ADALINE performance is the
mean-squared error J, where E(.) denotes an expectation over all available (X, d(X)) pairs.

J ¼ E ð d ð X Þ  yð X Þ Þ 2 ð8Þ

¼ Eðd ð xÞ  f ðxÞÞ2 ð9Þ

2.3 Deep Learning Using Support Vector Machines

Deep support vector machine is a widely used unconventional method for classifica-
tion. Convolutional nets along with linear SVM’s have been proposed in the past as
part of a multistage process. In particular, a deep convolutional net is first trained using
supervised or unsupervised objectives to learn good invariant hidden latent represen-
tations. Then corresponding hidden variables of data samples are treated as input and
fed into linear SVMs. This technique usually improves performance but the drawback
is that lower level features are not been fine-tuned w.r.t. the SVM’s objective. Deep
learning for classification using fully connected layers and convolutional layers have
almost always used soft max layer objective to learn the lower level parameters. There
are exceptions, notably the supervised embedding with nonlinear NCA, and semi
supervised deep embedding. Here we propose using multi class SVM’s objective to
train deep neural nets for classification tasks. The equation used here is given below

Ctn wðpf1 [ wT hn tn gÞ ð10Þ

166 S.S. Roy et al.

Where p{
} is the indicator function. Likewise, for the L2-SVM, we have



2Ctn w max 1 [ wT hn tn; 0 ð12Þ

From this point on, back propagation algorithm is exactly the same as the standard
soft max based deep learning networks. We found L2-SVM to be slightly better than
L1-SVM most of the time and will use the L2-SVM in the experiments section.

3 Data Set Description

Spam-email was created by Mark Hopkins, Erik Reeber, George Forman, Jaap
Suermondt. It was donated by George Forman. This collection is constituted from spam
filed by postmasters and individuals in the time period of June − July 1999. This data
set is a multivariate characteristic data set which are real and continuous in nature. It
consists of 57 attributes with no missing values. These attributes mostly shows whether
a word or character occurrence in the e-mail. The run-length attributes gives the length
of sequence of consecutive capital letters. The last column indicates whether that email
was considered spam or not. This data set consists of 4601 instances, out of which
39.4% (1813 instances) accounts for spam emails. Tool used for this experiment is
Azure machine learning studio. It was used to implement each model and gather
results. The data was first split into train and test sample with a 70–30 probability. The
model was built on the train data and then the model was tested on the unknown test
data to see how it performs when it sees new data.

4 Result and Discussion

Figures 1 and 2 show the comparison area under the ROC curve for artificial neural
network and SVM with respect to the Deep SVM, while the Figs. 3, 4 and 5 show the
area under ROC for the individual models. Figures 6, 7 and 8 show the lift for the
individual models and Figs. 9, 10 and 11 show the precision/recall for the models
(Table 1).
The best classification model for spam-base email classification is observed to be
Deep Support Vector Machine model. The aforementioned model was compared with
models as Neural Network and Support Vector Machine. Figures 1 and 2 compares the
Deep SVM model’s ROC curve with Neural Network’s and SVM’s respectively. The
result of Deep SVM (in red) outperforms both models with an accuracy of 92.8.
Whereas NN and SVM model stands at an accuracy of 91.8 and 89.2 respectively.
ROC curve gives illustration of the performance of a binary classifier system as its
discrimination threshold is varied. The curve is created by plotting the true positive rate
(TPR) against the false positive rate (FPR) at various threshold settings. ROC obtained
 Spam Email Detection Using Deep Support Vector Machine 167

Fig. 1. Comparison of ROC curves of Deep SVM and Artificial Neural Network model.

Fig. 2. Comparison of ROC curves of Deep SVM and SVM model.

from the experiment shows larger AUC (Area under curve) value for Deep SVM model
than for AUC’s for SVM and Neural Network model. This result indicates better
sensitivity or probability of detection for Deep SVM model. AUC value obtained for
Deep SVM model is 0.973 and AUC value obtained for SVM and Neural Network
model is 0.951 and 0.956. Lift curve compares between true positives and positive rate
for a classification. Deep SVM gives smoother curve and higher inclination towards
true positives as positive rate increases. In information retrieval with binary classifi-
cation, precision (also called positive predictive value) is the fraction of retrieved
168 S.S. Roy et al.

Fig. 3. ROC curve for Deep SVM model

Fig. 4. ROC curve for Neural Network model.

instances that are relevant, while recall (also known as sensitivity) is the fraction of
relevant instances that are retrieved [23]. Both precision and recall are therefore based
on an understanding and measure of relevance. Deep SVM model turns out to be the
one with greater balance between two factors. It gives highest precision value at the
point of plunge in the graph. The following values show the comparison between
performance of the three chosen models. Deep SVM model came out as a better model
in terms of Accuracy, AUC, F1 score, Recall and Precision.
Spam Email Detection Using Deep Support Vector Machine 169

Fig. 5. ROC curve for SVM model

Fig. 6. Lift for Deep SVM model

170 S.S. Roy et al.

Fig. 7. Lift for Neural Network model

Fig. 8. Lift for SVM model.

 Spam Email Detection Using Deep Support Vector Machine 171

Fig. 9. Precision/recall curve for Deep SVM model

Fig. 10. Precision/recall curve for NN model

172 S.S. Roy et al.

Fig. 11. Precision/recall curve for SVM model

Table 1. Output result comparison of the three models

True False True False Precision Recall F1 AUC Accuracy
Positive Negative Positive Negative Score
Deep 640 60 1068 72 0.914 0.899 0.907 0.973 0.928
SVM
Neura 1 627 66 1062 85 0.905 0.881 0.893 0.956 0.918
Network
SVM 565 52 1076 147 0.916 0.794 0.850 0.951 0.892

5 Conclusion

In this work, Deep SVM model was found to be effective and more accurate than SVM
and Neural Networks models’ performance in classifying spam emails. Deep SVM
surpassed these models in every evaluating aspect as in accuracy, precision, recall, F1
score, AUC and confusion matrix. The performance of Deep SVM was significantly
greater than that of other models with an accuracy of 92.8% while SVM and Artifi-
cial NN models fared at values of 89.2% and 91.2% respectively. More algorithms in
future can be applied to spam email detection to benchmark the performances.

References
1. Tretyakov, K.: Machine learning techniques in spam filtering. Data mining problem-oriented
Seminar, MTAT.03.177, pp. 60–79 (2004)
2. Samui, P., Roy, S.S., Kurup, P., Dalkiliç, Y.: Modeling of Seismic Liquefaction Data Using
Extreme Learning Machine. Nova publishers, Natural Disaster Research, Prediction and
Mitigation, pp. 6x9 - (NBC-R), ISBN: 978-1-53610-356-4
 Spam Email Detection Using Deep Support Vector Machine 173

3. Roopaei, M., Balas, V.E.: Adaptive gain sliding mode control in uncertain MIMO systems.
In: 3rd International Workshop on Soft Computing Applications, SOFA 2009, pp. 77–82.
IEEE, July 2009
4. Azar, A.T., Balas, V.E., Olariu, T.: Artificial neural network for accurate prediction of
post-dialysis urea rebound. In: 2010 4th International Workshop on Soft Computing
Applications (SOFA), pp. 165–170. IEEE, July 2010
5. Basu, A., Roy, S.S., Abraham, A. A novel diagnostic approach based on support vector
machine with linear kernel for classifying the erythemato-squamous disease. In: 2015
International Conference on Computing Communication Control and Automation
(ICCUBEA), pp. 343–347. IEEE, February 2015
6. Roy, S.S., Gupta, A., Sinha, A., Ramesh, R.: Cancer data investigation using variable
precision rough set with flexible classification. In: Proceedings of the Second International
Conference on Computational Science, Engineering and Information Technology, pp. 472–
475. ACM, October 2012
7. Mittal, D., Gaurav, D., Roy, S.S.: An effective hybridized classifier for breast cancer
diagnosis. In 2015 IEEE International Conference on Advanced Intelligent Mechatronics
(AIM), pp. 1026–1031. IEEE, July 2015
8. Roy, S.S., Mittal, D., Basu, A., Abraham, A.: Stock market forecasting using LASSO linear
regression model. In: Afro-European Conference for Industrial Advancement, pp. 371–381.
Springer International Publishing (2015)
9. Roy, S.S., Viswanatham, V.M., Krishna, P.V., Saraf, N., Gupta, A., Mishra, R.:
Applicability of rough set technique for data investigation and optimization of intrusion
detection system. In: International Conference on Heterogeneous Networking for Quality,
Reliability, Security and Robustness, pp. 479–484. Springer, Berlin (2013)
10. Roy, S.S., Krishna, P.V., Yenduri, S.: Analyzing intrusion detection system: an ensemble
based stacking approach. In 2014 IEEE International Symposium on Signal Processing and
Information Technology (ISSPIT), pp. 000307–000309. IEEE (2014)
11. Sharma, R., Kaur, G.: Spam detection techniques: a review. Int. J. Sci. Res.
12. Roy, S.S., Viswanatham, V.M., Krishna, P.V.: Spam detection using hybrid model of rough
set and decorate ensemble. Int. J. Comput. Syst. Eng. 2(3), 139–147 (2016)
13. Roy, S.S., Viswanatham, V.M.: Classifying spam emails using artificial intelligent
techniques. Int. J. Eng. Res. Afr. 22, 152–161 (2016)
14. Srivastava, D.K., Bhambhu, L.: Data classification using support vector machine. J. Theor.
Appl. Inf. Technol.
15. Mhetre, P., Bapat, M.S.: Classification of teaching evaluation performance using support
vector machine. Int. J. Latest Res. Sci. Technol. 4(6), 37–39
16. Tang, Y.: Deep learning using linear support vector machines. In: International Conference
on Machine Learning 2013: Challenges in Representational Learning Workshop (2013)
17. Bhavsar, H., Panchal, M.H.: A review on support vector machine for data classification. Int.
J. Adv. Res. Comput. Eng. Technol. 1(10), 185 (2012)
18. Raiff, B.R., Karatas, C., McClure, E.A., Pompili, D., Walls, T.A.: Laboratory validation of
inertial body sensors to detect cigarette smoking arm movements
19. Sabeena, S., Priyadharshini, G.: Feature selection and classification techniques in data
mining. Int. J. Eng. Sci. Res. Technol.
20. Nguyen, D.H., Widrow, B.: Neural networks for self-learning control systems. IEEE Control
Syst. Mag.
21. Ponalagusamy, R., Senthilkumar, S.: Investigation on time-multiplexing cellular neural
network simulation by RKAHeM (4, 4) technique. Int. J. Adv. Intell. Paradig. 3(1), 43–66
(2011)
174 S.S. Roy et al.

22. Popescu-Bodorin, N., Balas, V.E., Motoc, I.M.: Iris codes classification using discriminant
and witness directions. arXiv preprint arXiv:1110.6483 (2011)
23. Nareshkumar, S., Vijayarajan, V.: A study on image retrieval by low level features. Int.
J. Comput. Appl. 43(18), 18–21 (2012)
Business Process Management
 Collaborative Business Process Solution
Considering an Ontological Dimension
of Process Models

Maria Laura Sebu(&) and Horia Ciocarlie

Computers and Software Engineering Department,

Politehnica University of Timisoara, Timisoara, Romania
[email protected], [email protected]

Abstract. Considering the potential of collaborative business as a trend, we

start from the assumption that collaborations between companies have a higher
success rate if a good compatibility from process perspective is detected. The
solution is based on the business processes used internally in all participating
organizations to collaborations with common business objectives. For
un-standardized organizations, we used process mining techniques to formalize
processes and to attach them to business scenarios. The process models are
reduced to graph format and graph comparison algorithms are used to compute a
compatibility factor. Syntactic similarity is unfeasible for real business process
comparison in an un-standardized environment. Semantic similarity is resource
consuming and could lead to un-predictable results. We propose in the current
paper a methodology based on domain specific ontology to calculate the simi-
larity between organizations from process perspective and to offer a collabo-
rative solution, business process based.

Keywords: Business Process Management
Process mining
Ontology

domain
Similarity

1 Introduction

This paper aims to identify a collaborative solution, business process based, able to
support multi-organizational collaborations. The start point is represented by the
business processes used internally inside organizations for specific business cases.
Comparison and merge are performed in a semantically ontology oriented space.
Ontologies became very popular nowadays exceeding the scope of artificial intelli-
gence. The reasons are multiple but the most important one is the separation of domain
knowledge from operational knowledge. Domain ontologies allow the specification of
the semantics for the terminology used to represent a system. A domain ontology could
be based on a top level ontology as foundation. The top-level ontology is the base for
linking semantically two or more ontologies. We are considering comparing business
processes each one created in a specific ontology. The comparison is performed by
considering the relation between the elements of the ontologies supporting business
process definition.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_14
178 M.L. Sebu and H. Ciocarlie

Collaborations are kicked off by organizations for several reasons which could be
reduced mainly to the common problem: challenges caused by the high level of
dynamism exceed the scope of single organizations and should be addressed in a multi
organizational environment. The topic of organizational collaboration with shared
business objectives has been studied intensely and the results show that many of the
collaborations initiated are finalized with failures recorded. This result is independent
of the maturity level of the collaborating organizations. For this research, we are
offering technical methods and tools to support organizations in realizing projects in
collaborations and reaching the desired objectives.
Outsourcing is the most common form of collaboration and implies more organi-
zations working on a common project using a common process. Though the benefits of
setting up a successful outsourcing collaboration are easily identifiable, a lot of col-
laborations end up in failures jeopardizing projects. Gartner Group considers that more
than half of outsourcing projects are closed with undesirable results. The major reason
is represented by the resistance to change inside participating organizations when a
project is realized in a collaboration. One of the reasons we’ve considered when
designing the collaborative business process based solution was that a collaborative
relation is easier to accept when it is initiated between companies following similar
working modes and processes. The similarity metric proposed evaluates the compati-
bility between organizations from process perspective. The collaborative solution
composed of input from all participating business processes would contribute to
eliminating the resistance to change. In (Lonsdale 1999) the authors state that most
outsourcing failures are due to the lack of a methodology. The collaborative solution
offers a solution to cope with this reason.
Several other types of collaborations could be targeted between organizations.
Joint-venture is a relationship between organizations sharing common resources and
assets. Such a collaboration could be also modelled with a business process based
solution.
Any collaboration is initiated from a business need identified and captured in a
project. The more compatible the organizations are, the easiest the setup performed is.
The solution proposed is to compute a compatibility factor able to quantify the match
level between organizations and to identify if a collaboration is feasible from business
process perspective. For this purpose the first step is to identify the business case
representing the subject of the collaboration and the business processes followed by
organizations individually. A syntactic representation of process models cannot cover
the business needs of collaborations initiated between organizations operating in dif-
ferent domain areas, using a specific vocabulary and terminology for defining activities.
However this should not be a blocking point in setting up possible collaborations.
Flexibility is a major target in business nowadays. Companies must adapt to
changes and to a high level of dynamism of the market. A condition for survival is the
ability to construct collaborations with different business partners. Due to this con-
straint we considered our solution in a semantic space. Feasibility of the collaboration
is evaluated by analyzing the meaning of the activities and not only their textual
representation. However this search in a semantics space is high resource consuming.
Introducing ontology domain and reducing the search to a limited space fills the
resource gap and ensures the meaning of the activities is interpreted by considering a
 Collaborative Business Process Solution 179

limited amount of possibilities. Introducing ontology space for computing the feasi-
bility and identifying the collaborative solution limits the possible faulty interpretations
of business process activity names.
Once the compatibility level of organizations is computed, compatible organiza-
tions could be the subject of possible collaborations. The collaboration is constructed
by using the merged business processes interpreted in a common ontological space.
A formalized process model could be improved further and synchronized with com-
monly defined business objectives. The collaboration could be monitored at runtime for
ensuring execution support in real time in case of deviations and uncertainties.
In Fig. 1 the collaborative business process solution is graphically illustrated.
Organizations contribute with event log data tracking their actions performed to
achieve business results. Business processes extracted with process mining techniques
are attached to use cases. For each process, an ontology is identified abstracting the
semantics of the area. Ontology is used to permit the analysis of the domain knowledge
and to re-use it. Once the ontologies are formalized, a mapping is performed in between
them to provide a common layer for calculating the business process compatibility
factor. This metric is computed considering the names of the activities and the topology
of the business process. High similarity means compatible organizations, a collabo-
rative business process is constructed by merging the input business processes.

Fig. 1. Collaborative business process solution.

For exemplifying the algorithm presented in this paper, we used the Business
Process Modeling Notation (BPMN) for business process representation. After
studying possible business representations we concluded that BPMN is the closest to
the real business needs and that it is an easy and understandable way to represent
process information.
180 M.L. Sebu and H. Ciocarlie

2 Semantic BPM

The use of semantics in Business Process Management (BPM) has as a major goal
filling the gap between business and IT perspective on how business should be orga-
nized and executed. Semantics opened up new perspectives in BPM and brought new
technical challenges. In practice a BPM System is put in place by a business analyst
identifying a business need. The steps performed for reaching the defined business
objectives are abstracted in a process model which is executed inside a business process
engine. Several BPM implementations are available on the market. The major chal-
lenge is converting business needs to steps executed for achieving business objectives.
Semantics in BPM has as a major goal ensuring the translation is performed accurately
and verifying the lack of gaps in between business and IT.
In (Wetzstein et al. 2007) the requirements for Semantic Business Process Man-
agement (SBPM) are formalized. In the first phase of Semantic Business Process
Modeling, the process models are semantically annotated. Each element in the process
flow has a semantic description associated by referencing an ontology concept. In
(Wetzstein et al. 2007) this step is performed by the business analyst annotating the
business process components with ontology entities. The base for the business process
annotation is represented by the process ontology composed separately for each
industry/domain. With this enhancement, several functionalities become available:
business process discovery, domain search. The major benefit noted is the possibility of
reusing the process models and the discovery of process fragments able to support a
business case.
Semantics in BPM were introduced mostly because of the limitation of the current
business process modeling languages such as BPEL4WS, unable to capture the busi-
ness need of initiating process instances by considering a business objective. Semantic
Business Process Management is defined over Semantic Web technologies: ontology
language representations, knowledge management, query support.
Business Processes and Web Services have been integrated and several standard-
ized approaches aiming to describe a business process from a Web service orchestra-
tion perspective have been proposed. The most relevant example is BPEL4WS focused
on the representation of a limited part of the process space, the patterns of the message
exchange.
In a collaborative scenario each process model could use a separate ontology. As
we start from the premise that collaborations should be targeted even with partners with
a less formal and structured approach, we extract the business process definitions from
the event log data tracking the execution. We construct our solution considering
automatization in calculating the compatibility level and defining the business solution.
This premise is incompatible with the manual approach presented in (Wetzstein et al.
2007) in which the ontology is defined and considered as a reference for the business
analyst manually annotating the business process activities. We are proposing an
extraction of the ontology from business process definition for each business process.
Compared process models have an ontology associated and the search is performed
only in the ontological space. The collaborative solution is composed by considering a
shared ontology identified by merging the ontologies of the source business processes.
 Collaborative Business Process Solution 181

Once a semantic business process is formalized, it can be deployed in a process

engine able to support process execution.
Several standards for Semantics are available on the market. OWL-S is the most
familiar ontology representation language for semantics used mostly for Web Service
description. This has been further developed into Web Service Modeling Ontology
(WSMO). Web Service Modeling Language (WSML) is a group of ontology language
representations and Web Service Execution Framework (WSMX) is the most relevant
implementation of WSMO.
OWL extends RDF (Resource Description Framework) and RDF Schema and
defines a terminology used in RDF documents composed of classes and properties.
A class represents a set of resources. OWL defines two set of properties: object and
data types. Object properties identify an attribute capturing the relation between a
resource and a data type value. Properties could be defined by using sub elements.
OWL supports the relationships between instances.
An ontology is a pair O = (S, A), where S is the vocabulary and A is a set of
axioms operating and describing vocabulary elements. Ontology mapping was intro-
duced mostly because a single ontology was not able to support a wide distributed
environment of Semantic Web. Mapping allows the existence of multiple ontologies
and provides a common layer for access. An ontology mapping process is able to
associate vocabulary elements of two ontologies. The mapping considers the validity of
axioms. In (Kalfoglou and Schorlemmer 2003) a total ontology mapping is theoreti-
cally defined. Having two ontologies O1 = (S1, A1) and O2 = (S2, A2), a total ontology
mapping is a morphism f: S1!S2 of ontological signatures in such a way that all
interpretations that satisfy O2 axioms also satisfy O1 translated axioms. A partial
ontology mapping from O1 = (S1, A1) to O2 = (S2, A2) could be identified if there
exists a sub-ontology O01 ¼ ðS01 ; A01 Þ in which S01 S1 and A01 A1 for which there is a
total mapping from O01 to O2.
Our target is to use the ontology mapping technique for business process com-
patibility computation and for the collaborative solution obtained by merge. The
ontologies for the process models are extracted and mapped for finding common
elements.

3 Organizational Process Compatibility Factor

Collaboration between organizations nowadays became a condition for survival in

business. A successful collaboration opens up new perspectives for any organization in
the area of business opportunities, market position, benefits, and costs. An incompatible
partner identified early in the collaboration setup phase permits the companies to
reorganize and mitigate the side effects of the differences before this incompatibility
jeopardizes the projects, the customers and the companies.
Many collaborations are set up between organizations from financial and strategic
reasons without considering any factor concerning social elements or internal processes
used. Organizations which don’t share common workflows and working methods are
considered less compatible and are less likely to succeed naturally in a collaboration.
182 M.L. Sebu and H. Ciocarlie

We are proposing to use a single metric to quantify the similarity of the organi-
zations from business process perspective. The technique is based on the method
proposed in (Dijkman et al. 2009). The business process similarity considers the
topology of the process graph and the labels of the activities composing the business
process.
For business process similarity computation all process models are transformed in
directed attributed graphs. k is the set of labels and s is the set of types of nodes.
A business process graph is a tuple (N, E, s, k) in which N is the set of nodes, E is the
set of edges, s: N!Ί is a function that maps nodes to types and k: N ! K is a function
that maps nodes to labels. G = (N, E, s, k) is a graph and n € N is a node, n = {m |
(m,n) € E} is the pre-set of n and m = {m | (n,m) € E} is the post set of n.
The algorithm presented in (Dijkman et al. 2009) is extended not to consider a
simple syntactic representation of the activities but to interpret the semantic meaning in
a limited dimension given by the ontologies identified for each business process.
As the semantic space is extremely wide and could lead to incorrect results, we will
limit the semantic match between activities of business processes to an ontological
space. Because we target a fully automated method to compute the similarity, we
perform intermediate steps to achieve the desired results. We are extracting the
Semantics of the Business Vocabulary and Business Rules (SBVR) of both business
processes subject of the comparison. Once the SBVR are available for each business
process, the SBVR business vocabularies are represented in Web Ontology Language
OWL2.
Each business process will have associated an ontology and the comparison is
performed in the ontological space. The ontologies extracted are mapped and for each
activity of the business process the corresponding activity (if exists) is identified in the
partner business process, subject of a comparison.
In Fig. 2 the activity label match is represented by considering an ontological
space. The business processes are composed of activities. An identification of the
activities of the first business process which have a correspondent in the second
business process is performed. The information is used further in the business process

Fig. 2. Similarity technique.

 Collaborative Business Process Solution 183

similarity algorithm proposed in the current paper to quantify the compatibility between
organizations from process perspective. Instead of relying only on the label content of
the activities, or to perform the search in the full semantics space, the search is limited
to a domain constructed by mapping and merging the domains of each individual
process model. The major goal is to ensure optimum reliable results in reasonable
execution time and with a limited amount of resources.
Once the ontologies are available for all participating business processes, the
business process similarity factor is computed. For this purpose the technique presented
in (Dijkman et al. 2009) is extended in the current paper for graph comparisons using
the graph edit distance. The graph edit distance (GED) is the cost of transforming one
graph into another. The transformation is performed in a set of operations: node sub-
stitution, node insertion and deletion, edge insertion and deletion. Node substitution is
the operation in which the node from one graph is replaced with a node of the partner
graph if such a node is found considering the ontology mapping determined. Similarity
of nodes is the similarity of node labels in the ontological space.
Each operation (substitution, insertion and deletion) has a cost function associated
and default values are used for weighting the operation and computing the business
process similarity factor. Default values of 0.5 for edges and 0.2 for nodes are proposed
in (Dijkman et al. 2009) and used for computation. If several sets of operations are
possible to transform one graph into another, GED is represented by the least costly
operation.
For computing the similarity of nodes to identify node substitution operations, the
syntactic string similarity algorithm presented in (Dijkman et al. 2009) is replaced with
a search in a semantic space represented by the ontologies of the source business
processes.

3.1 Extract SBVR Rules from Business Processes

The first step in comparing business processes by using the ontologies extracted and
integrated for each business process is the extraction of business vocabularies from
process models.
According to (Skersys et al. 2013) business rules could be considered as part of the
Business Rules Management (BRM) discipline. As stated in (Skersys et al. 2013) a
business rule is a logical statement that defines or constraints aspects of the business in
the concrete situation. A Business Vocabulary contains a set of terms specialized and
concepts that a given organization uses in performing business.
In (Nemuraite et al. 2010) an approach to integrate a business process with a
business vocabulary is presented by mapping certain elements of the business process
meta model with elements of the business vocabulary meta model. For example, the
BPMN activity could be mapped to SBVR noun concepts, fact type and rules asso-
ciated with that activity. Certain elements of the business process might not be mapped
to business vocabulary elements.
As a simple example, the activity “Submit Change” for the change management
business process could be mapped into a noun concept “change” and a fact “support
184 M.L. Sebu and H. Ciocarlie

users register a change request”. SBVR is able to formalize the definition of a business
vocabulary particular to a business process.
We used for this purpose the Magic Draw UML tool extended with BP!BV and
VeTIS utility (Nemuraite et al. 2010) which could be used either as a standalone or as a
plugin.
For exemplification, the following business processes Figs. 3 and 4 are compared.

Fig. 3. Business process 1.

Fig. 4. Business process 2.

The business processes used for exemplification in Figs. 3 and 4 are also used for
exemplification in (Sebu and Ciocarlie 2016).
The vocabulary is extracted for each business process. A fragment of the vocab-
ulary for Business Process 1:
Collaborative Business Process Solution 185
186 M.L. Sebu and H. Ciocarlie

The business rules for Business Process 1 are identified and a fragment is listed
below:

A fragment of the vocabulary for Business Process 2 presented:

 Collaborative Business Process Solution 187

And a fragment of the rules for Business Process 2 is extracted:

3.2 Transform SBVR Representation into an OWL Representation

of the Ontology
Once the vocabulary and the rules are formalized, a conversion is performed to extract
the OWL2 representation of the business process ontology. For this purpose the tool
s2o converter is used, a transformation utility able to identify an OWL2 ontology
representation of the SBVR vocabulary and rules. The technique is presented in
(Karpovic and Nemuraite 2011). Transformation of SBVR is still an area under
development and improvement. Current research is directed to include a larger subset
of SBVR concepts worth to represent in OWL2. Our goal is to use the available
transformation methodology to extract automatically the OWL2 representation of the
ontologies for the business processes, subject of similarity computation.
The OWL 2 representations of Business Process 1 is graphically summarized in
Fig. 5:

Fig. 5. Ontology Domain Business Process 1

188 M.L. Sebu and H. Ciocarlie

The OWL 2 representations of Business Process 2 is graphically summarized in

Fig. 6:

Fig. 6. Ontology Domain Business Process 2

Mapping of the ontologies is captured in Fig. 7.

Fig. 7. Ontology mapping

3.3 Perform a Mapping of the Ontologies Extracted for the Business

Processes
In this phase the correspondence between the concepts of the ontologies identified for the
business processes is determined. The need for such a mapping is resulted from the need
to perform the semantic match of the activities in a limited semantic space. This space is
the mapping of the corresponding ontologies. For this purpose the OLA– OWL Lite
Alignment tool presented in (Euzenat et al. 2004) is used. This utility has several capa-
bilities which were considered when making the selection. The most important one is the
automation, the production of the alignment does not require any user intervention.
Another aspect considered was the comparison presumption, only entities of the same
 Collaborative Business Process Solution 189

categories are compared. The same similarity function is used for comparing. The values
of the similarity measures are normalized and normalization is enforced during the full
alignment process. Another aspect considered when using the OLA tool was the
implementation in Java and the availability of an OWL Java API for parsing and merging
OWL files. This is a clear advantage as the collaborative business process modeling
solution is designed in Java by using jGraph for representation and visualization.
A glimpse of the mapping performed to limit the search space to the ontology space
is listed below:
190 M.L. Sebu and H. Ciocarlie
 Collaborative Business Process Solution 191

3.4 Compute Business Process Similarity

In (Sebu and Ciocarlie 2015) a metric to compute the compatibility level between
organizations from process perspective is proposed. For operating with business pro-
cesses once the semantic space is constructed, a reduction to graph format is performed.
The business process similarity metric is named for the current research as organiza-
tional compatibility process factor. The comparison considers the names of the activ-
ities and the topology of the graphs.
In (Dijkman et al. 2009) for activity label match, syntactic string similarity algo-
rithms are used. The current paper offers an extension of the search to a semantic space
composed of the common elements for the ontologies extracted for each business
process participating to the collaboration.
For identifying the substituted nodes, instead of comparing syntactically the labels
of the activities, we are executing a search in a semantic space limited to the ontologies
supporting the business processes.

4 Collaborative Process Model

As the compatibility is high, organizations are feasible for collaboration. The setup is
straightforward and could be based on the common business process obtained after
merging the individual business processes of the participating organizations. The
solution is presented in (Sebu and Ciocarlie 2015) and it is not the subject of the current
work presented in this paper. This methodology could be extended to include more
participating organizations. The business processes used for exemplification in Fig. 8
are also used in (Sebu and Ciocarlie 2016).

Fig. 8. Solution supporting collaboration between organization 1 and 2

192 M.L. Sebu and H. Ciocarlie

5 Future Work

The current solution is constructed on an automatic extraction of the ontology domains

for the business processes used inside organizations planning collaborations. The
mapping is automatically executed in a semantical space and the corresponding ele-
ment in the ontology is automatically identified. The advantage of an automatic
solution is clear but the major limitation is represented by specific situations in which
the mapping between ontology elements is custom and should be manually done. The
current proposal could be extended to use a manual ontology mapping able to identify
specific relations between the elements. Search for business process comparison and
merge is performed in an ontological space manually composed.
Another extension possible foreseen for the current approach would be to consider
a single ontology as the base for the business process comparison and merge for
computing the collaborative solution. Assuming that only a set of ontology elements
should be taken into account, this could be accomplished by manually defining the
ontology domain as the base for the collaborative solution and manually annotating
business process activities by referencing the elements of the ontology domain pre-
defined. The search uses the annotations instead of the mapping between two ontology
domains.
The current solution will be tested in a real collaborative organizational environ-
ment represented by organizations in the area of software systems aiming to collaborate
for accomplishing business objectives.

6 Limitation

The current solution is designed considering the ontology domains for the business
processes used in collaborating organizations. The search is performed in the domain
space computed by mapping the corresponding ontologies. If the mapping has
inconsistencies, these will be reflected in the computation of the similarity factor and
the collaborative solution. Such inconsistencies are hard to detect and difficult to
correct.

7 Conclusion

A multi-organizational context for achieving business objectives is a clear target for

many organizations and the number is exponentially growing. For the current research
we are offering technical support to ensure an easier setup of the collaboration and
perspectives for success in realizing business output. For the organizational compati-
bility factor that we are referring in the current paper, we performed a limitation of the
search space to the ontologies of input business processes automatically identified. The
business vocabulary and rules are computed and used for ontology specification.
Business process comparison is performed in a limited search space of the ontology
mapping result. The organizational compatibility factor is calculated in an execution
time of miliseconds.
 Collaborative Business Process Solution 193

The solution presented in this paper is based on an abstraction of business process

definition, input business processes are reduced to graph format, with the purpose of
considering the topology in the compatibility factor computation. The independence of
the maturity level of the participating organizations is achieved; collaborations should
not be limited by the standardization level of the organizations. The compatibility
factor is a single float number easy to interpret and use further.
In case business processes are not formalized and are not available as input, these
could be extracted by analyzing the event log data and applying process mining
algorithms. Like this the compatibility could be measured even if organizations don’t
have clear processes followed internally. The common business process obtained could
represent the base for constructing a collaboration with shared resources and shared
objectives.

References
Lonsdale, C.: Effectively managing vertical supply relationships: a risk management model for
outsourcing. Supply Chain Manag. Int. J. 4(4), 176–183 (1999)
Wetzstein, B., Ma, Z., Filipowska, A., Kaczmarek, M., Bhiri, S., Losada, S.: Semantic business
process management: a lifecycle based requirements analysis. In: Proceedings of the
Workshop on Semantic Business Process and Product Lifecycle Management, pp. 1–12
(2007)
Kalfoglou, Y., Schorlemmer, M.: Ontology mapping: the state of the art. Knowl. Eng. Rev.
18(01), 1–31 (2003)
Dijkman, R., Marlon, D., Garcıa-Banuelos, L.: Graph matching algorithms for business process
model similarity search. In: Proceedings of 7th International Conference of Business Process
Management, pp. 48–63 (2009)
Skersys, T., Butleris, R., Kapocius, K., Vileiniskis, T.: An approach for extracting business
vocabularies from business process models. Inf. Technol. Control 42, 178–190 (2013)
Nemuraite, L., Skersys, T., Sukys, A.: VETIS tool for editing and transforming SBVR business
vocabularies and business rules into UML&OCL models. In: Proceedings of the 16th
International Conference on Information and Software Technologies, pp. 377–384 (2010)
Karpovič, J., Nemuraitė, L.: Transforming SBVR business semantics into Web ontology
language OWL2: main concepts. In: Proceedings of the 17th International Conference on
Information and Software Technologies, pp. 231–238 (2011)
Euzenat, J., Loup, D., Touzani, M., Valtchev, P.: Ontology alignment with OLA. In: Proceedings
of the 3rd EON Workshop, 3rd International Semantic Web Conference, Hiroshima (2004)
Sebu, M., Ciocarlie, H.: Cross-organizational compatibility detection with process mining. In:
Buletinul Institutului Politehnic din Iasi, Universitatea Tehnică Gheorghe Asachi. din Iaşi,
Tomul LVIII (LXII), Fasc. 1, pp. 9–17 (2015)
Sebu, M., Ciocarlie, H.: Business Process Modelling Solution Supporting Cross-Organizational
Collaboration (2016). Submitted for publication
 Logistics Scenario for Wind Turbine Assembly
Based on ERP

Gabriela Prostean1 ✉ , Cristian Vasar2, and Andra Badea1

( )

1
Politehnica University of Timisoara, no. 14 Remus Street, 300191 Timisoara, Romania
[email protected], [email protected]
2
Politehnica University of Timisoara, no. 2 Parvan Blvd., 300223 Timisoara, Romania
[email protected]

Abstract. Integrating scientific concepts in the business and logistics areas have
given a particular attention to process automation from production to distribution
chains. In addition, the rise of business in wind turbines industry in the last three
decades has generated numerous shortcomings in the procurement process for
wind turbines manufacture and assembly. This paper presents a logistics scenario
setup for assembling wind turbines, specifically MRP scenario logic for the
process acquisition, according to necessary weekly demand for the finished
product assembly.

Keywords: Wind turbine · Stock balance · Purchase acquisition

1 Introduction

Coverage of all shorter products life cycle, due to the speed strategies of the new global
economic environment has increasingly integrated much more scientific concepts for
logistics management elements. Starting from the logistics definition [1] - “organizing,
planning, controlling and conducting goods flows from conception to production supply
and distribution to the end customer to satisfy market demands with minimal cost” can
say that logistics is effective and efficient manner through which industrial organizations
ensure their competitive advantage through low cost and high level of customer service.
Council of Logistics Management believes that “logistics” is the process of planning,
implementation and effective control of the flow and storage of actual goods, services
and information relevant to these movements, from point of origin to point of destination
in order to comply with customer requirements [2].
In conclusion logistics efficiently manage all materials flows, service goods and
information, from point of origin to destination, namely to reduce costs and to obtain
for the organization a competitive behavior internally and externally.
In this respect, logistics primarily involves properly planning for the necessary
materials. The main function of material requirements planning is to guarantee material
availability, which is used to procure or produce the required quantities on time both for
internal purposes and for sales and distribution. This process involves stocks monitoring
and, in particular, an automatic creation of procurement proposals for production.
Currently, logistics chain concept as support for production-distribution binder is in

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_15
 Logistics Scenario for Wind Turbine Assembly Based on ERP 195

complete computerization connection with its activities, which is a phenomenon that

was expressed and developed over more than 50 years. Onwards, most of the production-
distribution chain activities were managed manually generating errors in the document
and material flows, due to their inability efficient and effective synchronization. There‐
fore, since the beginning of computer use in business was given special attention to
processes automation for production and distribution chains. The study designed in this
paper shows how SAP system manages the production and process acquisition logic for
manufacturing and assembling a rotor for a wind turbine.

2 Logistics Scenario for Wind Turbine Assembly

The study focuses on presenting calculation algorithms Manufacturing Resource Plan‐

ning (ERP) type, integrated in SAP system and focused on manufacturing rotors for
wind turbines. This logistic scenario system based on Enterprise Resource Planning
(ERP) type has become absolutely necessary as industry in wind turbines experienced
a rapid rise in the last 3 decades and in addition, taking into account that in the last 5
years the size of the turbines has increased dramatically from less than 1 to 2 MW or
more [3]. This aspect has generated numerous shortcomings in the procurement process
for the manufacture and assembly of wind turbines, because raw materials suppliers,
which must satisfy high demand in a reliably manner, have decreased considerably. ERP
systems are designed to connect in an integrated manner the departments of a business
organization, namely to automate the flow of information in the registration orders chain,
acquisition, production, storage, delivery. MRP algorithms allows purchases planning
of raw materials, in connection with production orders and supplies available in ware‐
houses so that customer deliveries to be made as scheduled. In this way it streamlines
the purchasing process, storage and redistribution of raw materials or finished products
of the company. Assuming that was done previously, in the SAP database, called
MasterData, in which they are designed database for finished products (FP) database for
semifinished products (SP) and database for raw materials (RM) in order to generate a
production order in the system must be set up first finished product configuration in the
so-called Bill of Material (BOM) [4]. To finalize manufacturing processes, SAP system,
BOM information from the bottom to top, more specifically it is ensuring that the raw
materials from the last level are purchased, in order to manufacture subassemblies
necessary, then the other acquisitions necessary ensure the final assembly of the finished
product, in this paper is about the “ROTOR” (Figs. 1 and 2).
196 G. Prostean et al.

Fig. 1. BOM for the finish product “ROTOR”

Fig. 2. Lead time for “ROTOR”

To explain certain aspects that are based on the operating algorithms of the MRP
methodology, it is considered Table 1, from which can be notify sales orders for
“ROTOR” finished product recorded in week 2, 3, 5, 6 and 8. Purchase order is made
in the SAP system via transaction code “VA01”.

Table 1. Sales order which generates weekly consumption for the rotor assembly
Name code Weekly consumption
ROTOR (finish product) 1 2 3 4 5 6 7 8
Sales order (VA01) 1 2 3 2 2

Further, in Table 2 is added row containing Stock balance (SB) and row containing
planned orders, resulted from the automatically run of the MRP algorithm. Planned order
is converted into production orders, in case of the company’s internal production, or into
purchase requisitions, in case of the need for external purchases. Considering that in the
first week Stock balance is 2, last row numbers in cells, corresponding weeks 2–8 are
calculated as follows:
 Logistics Scenario for Wind Turbine Assembly Based on ERP 197

Table 2. Sales transactions activated by sales order

Name code Weekly consumption
ROTOR (finish product) 1 2 3 4 5 6 7 8
Sales order (VA01) 1 2 3 2 2
Stock balance (MD04) 2 1 1 0 0 0 0 0
Planned orders 0 0 1 0 3 2 0 2

1. If the stock balance viewed through transaction “MD04” is more than 0, then in
the transaction Purchase Requisition created by the code ‘ME51 N “ it will calculate the
necessary purchase quantity according to the result of the difference between the quantity
contained in the Sales Order (SO) and the quantity of Stock Balance, as follows:

PR = SO − SB (1)
2. If the stock balance is equal to 0, then

PR = SO (2)
In addition, purchase requisitions are calculated to take into account the lead time
necessary for raw materials acquisition process and achieving subassemblies on lower
levels of the BOM (Fig. 1). Given the BOM in Fig. 1, the acquisition process (Fig. 3),
which also consumes time and final time needed to manufacture subassemblies, resulted
in a 2-week lead time, necessary purchase requisitions for ROTOR final product,
according to Table 3. This is why the demand resulting in a planned order for a particular
week is activated in transaction Purchase Requisition “ME51 N” 2 weeks earlier, gap
imposed by lead time (Table 3).

Table 3. Semi-finished product net requirements

Procurement process in SAP ERP is more effective than the traditional one because
of traceability and steps automation improvements. This facilitates acquisition under the
plans needed, complete inventory management and reporting of all activities. In addition
to facilitating access to tools for managing relationships with suppliers, the system
allows selecting and assessing suppliers and negotiating contracts with them. In the
picture below is presented full flow of activities that make the procurement process.
198 G. Prostean et al.

Fig. 3. Process acquisition flow

Within the application has resulted 2 weeks lead time (Fig. 2.) since it is required
one week to perform the complete purchase process for Fiberglass (Figs. 1 and 3), plus
one week to complete the process for blades manufacture (Fig. 1). For Hub is necessary
a lead time for just 1 week since it only goes through the acquisition process (Figs. 1
and 3). Starting from branching BOM and the fact that each branch has resulted each a
different lead time, algorithm triggering the procurement process for each branch of the
BOM is unique, differences arise because different lead time and because different
necessary amount. According to the explanations described above for MRP algorithm,

Table 4. Purchase Requisition for Fiberglass

 Logistics Scenario for Wind Turbine Assembly Based on ERP 199

the table for determining the Purchase Requisition for subassembly Blade should be
configured according to Table 4.
In the SAP procurement system is integrated and facilitated Purchasing Info Record,
with transaction code “ME11”. This facility is useful, offering information about buying
a certain raw materials or subassembly in connection with their suppliers. This database
is important, besides providing these types of material history, offers clear information
on purchase prices for certain quantities or discounts offered by suppliers when
purchasing larger quantities. For example, for the purchase of raw materials “Fiberglass”
in the scheme offered by the facility “Purchasing Info Record” can notify the price
discount measured in monetary units (MU) offered for different quantities ordered,
measured in pieces (PC), (Table 5).

Table 5. Purchasing Info Record

From (PC) Price (MU)
1 59.95
100 39.95
200 29.95

Purchase Requisition concept incorporates customer needs, arguing the need for such
acquisitions, and useful for production process. Purchase Requisition design focuses on
the user needs and supports the user when purchasing requisitions is processing but keep
in mind at the same time the facility of “Purchasing Info Record” for economic efficiency
of the entire logistics chain. Thus, even if the necessary material is clearly determined,
it is preferable to introduce a major amount in the Purchase Requisition, allowing
acquisition of raw materials at a low price, but also takes into account the time period
of storage for additional quantities, imposed by the facility of “Purchasing Info Record”.
The application case presented in this paper was taken into account the “Purchasing Info
Record” and thus gives a reconfigured overview in Table 6 both within certain transac‐
tions ME51 N for Fiberglass as well as in transaction MD04 for the Blades.

Table 6. Purchase requisition for fiberglass taking into account the Purchasing Info Record
200 G. Prostean et al.

3 Conclusions

Material determination of is done either by adding the requirement reports prepared by

each department (PNM) or using automated functionality for planning and material
control MRP (Material Requirements Planning -MD02 and MD03). MRP is a control
and planning system of production and inventory used to manage manufacturing
processes. Most MRP systems are software-based, but it is also possible to perform
manual MRP based on an algorithm.
MRP system is designed to simultaneously meet three objectives:
– Ensuring the flow of necessary production materials at the right time or the flow of
products available to them at the time set by the customer delivery sales order.
– Maintaining the safety stock deposit as reduced both raw materials and finished
products.
– Maintaining and synchronization key activities within a supply chain: production,
sale and purchase.
Logistics scenario presented in this paper highlight very clearly the advantage of
applying MRP algorithm, which notifies the facility “Purchasing Info Record” adjusts
the amount purchased during weeks 3 and 4, cumulating an additional amount of raw
material, which however provides the necessary production process of week 8, since
week 6, without requiring activation of another specific procurement process. There was
thus ensured optimum logistics flow, with materials purchased just in time, at prices
which were offered discounts and the additional amount was consumed in the next week.

Acknowledgement. This paper was developed through the Partnerships in Priority Areas - PN
II, with the support of ANCS, CNDI - UEFISCDI, Project No. 36/2012.

References

1. Council of Supply Chain Management Professionals (CSCMP): Supply chain management:

Terms and glossary (2010). http://cscmp.org/digital/glossary/document.pdf
2. Balan, C.: Logistics, second edition, Ed Uranus Bucharest, pp. 11–34 (2006)
3. Keith Hays of EER, Wind directions, January/February 2007. http://www.altenergy
stocks.com/assets/Wind%20Directions.pdf
4. Weidner, S: Integration Case Study LO, SAP, University Competence Center (2009)
 Relieving Supply Chain Through
Collaboration for Wind Turbine Assembly
Based on AHP and DBR

Andra Badea1(&), Gabriela Prostean1, and Cristian Vasar2

1
Politehnica University of Timisoara,
No. 14 Remus Street, Timisoara 300191, Romania
[email protected], [email protected]
2
Politehnica University of Timisoara,
No. 2 Parvan Blvd., Timisoara 300223, Romania
[email protected]

Abstract. This paper provides an important use of Analytical Hierarchy Pro-

cess (AHP) method that has helped to identify important wind blades criteria
and alternative performance, suitable for isolated region, performance expressed
through composite materials properties. Specifically, AHP algorithm identifies
optimal composite materials combination for blade structure that must strictly
respect unique technical parameters. An impediment in providing composite
materials is given by the restriction of supply chain collaboration, that propa-
gates negative orders for the necessary raw materials, fiber glass and carbon
fiber used in manufacturing wind blades. The use of Drum-Buffer-Rope
(DBR) philosophy, has streamlined to dimension a buffer stock for the move-
ment of fiber glass in the chain, in which was reduced assembly time of the
blades in the wind farm, but also the duration of the implementation project.

Keywords: Supply chain
Collaboration
Wind turbine
AHP and DBR

1 Introduction

Wind energy projects fall among the main promoters of green energy because of its
superior characteristics in abundance, renewability, wide distribution and zero pollu-
tion. However, the complexity of implementing wind energy projects imposed finding
collaborative alliances between suppliers and project developers to support supply
chain activities in such projects. Previous studies conducted by the research team in a
renewable energy projects (RES), specifically wind power projects have shown bot-
tlenecks encountered in supply chain and communication difficulties between research,
development and implementation teams, as well barriers in regulation, investment,
infrastructure, manufacturing and logistics activities procedure. Researches in this field
illustrates methods and techniques that these types of blockages can be easily overcome
through effective management programs and multi-criteria analysis methods. To pre-
vent these barriers, the research team notes that the collaborative activities between
links in the supply chain in RES projects require special attention to what is behind this
evolutionary process, which must be understood and documented to minimize
© Springer International Publishing AG 2018
V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_16
202 A. Badea et al.

unproductive regressions of RES project implementation. The research team identify

important criteria for wind blades assembly in isolated region and proposed a
dimensioning buffer to deliver quality raw materials to ensure that the wind blades
respects unique technical parameters indicated in the first step of the research project.
In this scenario, to optimize the minimum cost of energy was required a multi-
disciplinary method, which involved making an aerodynamic model and a structural
design for blades. [1] The design of the blade structural model involving a reduction in
blade weight corresponding to the wind turbine power, less than 1 MW, by using a
combination of optimal composite material. Thus, the project has been studied and
taken into account two types of composite material, namely fiberglass and carbon fiber.

2 Wind Blades Overall Presentation

Wind blades are placed in sensitive aerodynamic devices category that need to subdue
some lift extreme forces. In the wind industry, wind turbine blades were manufactured
from steel, aluminum and wood and then development of technological processes
favored the emergence of composite materials, respectively fiberglass and carbon fiber.
Further in this study are presented, below, wind blades characteristics employed in a
RES project in isolated region respectively, fiberglass blades, carbon fiber blades,
blades with higher composition of carbon fiber and fiberglass. Fiberglass wind blades
have in composition up to 70%–75% glass by weight. These blades must fulfill
aerodynamic criteria and mechanical strictly requirements, such as high stiffness, tor-
sional rigidity and fatigue.
Blades lifecycle, statically and dynamically with temperatures oscillating, can pro-
vide functionality about 20 years. Wind blades of 35–40 m for standard 1.5-MW turbine
weighing 6–7 tons. Carbon fiber wind blades do not suffer from fatigue problem, if they
are properly designed and dimensioned. Because thermal stability of carbon fiber blades
can withstand very high temperatures. The use of carbon fiber for 60 m blades reduces
blade total mass by 38%, while the cost by 14% compared to 100% of fiberglass.
Wind blades with higher composition of carbon fiber and fiberglass are low in
weight but have an increased cost due to carbon fiber. The designing of carbon fiber
blades advantage is achieved in the direction of traction, therefore, it is done in parts
that are subject to high tensile stress, such as blade structure (spar cap). The docu-
mentation concerning the characteristics of wind blades materials has been made to
identify what combination of composite material is suitable for blade structure that can
be apply to the type of RES project in isolated region. In addition to this the research
team identifies suitable criteria for blades assembling on wind farm site in isolated
region. The criteria are presented in detail later in pages of the paper.

3 Wind Blades Assembling at the Wind Farm Site

Within the project was necessary an effective coordination of all parties involved for
blades assembly to the wind farm site. Due to difficult weather conditions the activities
undertaken have been hampered, making difficult the assembly process for the blades.
 Relieving Supply Chain Through Collaboration 203

Specifically, it has been analyzed important criteria of blades performance, suitable for
isolated region, the blades being identified to meet the unique technical parameters,
adaptable to wind turbine power. In this study it was found that the time used for the
stringent requirements of composite blade structure manufacturing correlated with time
to move orders in the logistics chain prolong the process of assembling blades at the
wind farm.

3.1 Restriction Degree of the Composite Material from Specialized

Teams
Currently a strong tendency of Romania’s green energy policy, adapting to all evo-
lutionary processes for implementing renewable energy projects has to face the com-
plexity of multidisciplinary design, namely a complex transfer of knowledge at the
administrative level, technical level and technological level. [6] While development
projects that support energy production from renewable sources are more complex, the
attention of specialized teams involved in the project adopt latest creative techniques
through Theory of Constraints-Thinking Processes (TOCTP) which provided adequate
solutions for research, development and implementation of this type of project. By
using TOCTP each team presents its design constraints from a technical standpoint and
with other teams involved have generated the most favorable alternative solutions.
Thus the project required complex knowledge and especially more experience in
several areas: specialized in mechanical structures and wind turbines (MSWT), spe-
cialized in electrical machines (EM), specialized in power electronics (PE), specialized
in automatic controlling strategies (ACS), and specialized in project management and
publicity (PM).

3.2 Generated Restrictions by Supply Chain Collaboration Links

In today technological context finding profitable alliances for project implementation
from renewable energy sources (RES) is a challenge both for compiling of competent
multidisciplinary teams in research, development and project implementation and also
choosing suppliers-partners that need to meet the stringent requirements. Seeking
profitable partnerships in supply chain to implement RES project required identification
of different suppliers. Activities such as transport, inventory, purchase order process-
ing, warehousing, material handling, packaging, and other complex activities involve a
collaborative logistics system that need to be monitored continuously to respond
challenges.

3.3 Wind Blades Classification

The type of wind blades depends on the turbine axis, namely horizontal or vertical.
Horizontal axis wind turbines have propeller blades inclined and can be sorted by the
number of blades, turbines, instead vertical axis can have different structures forms that
capture the wind. Up-wind blades require complex control systems to keep the blades
rotating in dealing with the wind, but it works more smoothly and deliver more power.
Most modern wind turbines are upwind type. Down-wind blades wind control
204 A. Badea et al.

deviation (left-right movement), it is oriented in the sense of wind direction. The effects
of wind make the blade to flex, resulting in fatigue, noise and low power.

3.4 Accessibility Location Degree

Wind farms locations and associated weather conditions have meant engineers to face
many challenges, like meet the design requirements and installed systems. Poor access
to location prevented delivery components, rocky terrain made it difficult to founding
and electrical grounding system and fog and rain led to water seeping into cable
connections.

3.5 Important Factors in Choosing Blades

Timely and frequent consultations with affected communities helped to identify
problems before they achieve the investment. Utilities, government agencies, envi-
ronmental organizations have worked with developers to ensure effective implemen-
tation of the project.

3.6 Exposure to Risk Factors in Choosing Blades

Wind blades can be overloaded due to excessive fatigue, wear, vibration and collapse
of external loads. The blades have aero loads imposed and they are made of composite
materials that must meet certain standardized properties. The wind blades should
ensure a good performance but to achieve this requirement the manufacturing stage
implies a complex process and involves finding superior resins quality fiber materials
and a higher level of automation for continuous improvement in reducing blade weight.

4 Using the AHP Methodology

Analytic Hierarchy Process (AHP) was developed in the 1980s by Thomas L. Saaty
and is a tool for prioritizing alternatives to a problem. Application of AHP method
divide a complex problem that requires solving the levels, more precisely criteria
(subcriteria) and alternatives, so as to be easily analyzed and compared independently.
AHP convert the comparisons, which are most of the time empirical, into numeric
values that are further processed and compared. The weight of each criteria allows the
assessment of each one of the elements inside the defined hierarchy. The capability of
transforming empirical data into mathematical models is the main distinctive contri-
bution of the AHP technique when contrasted to other comparing techniques.

4.1 Importance Scale Presentation

AHP uses a fundamental scale of absolute numbers that are compared criteria in
relation to criteria, sub-criteria and alternatives, Relative importance scale developed by
Saaty (1980) is a scale that captures individual preferences in quantitative and quali-
tative attributes (Table 1) [2].
 Relieving Supply Chain Through Collaboration 205

Table 1. Saaty relative importance scale

Saaty scale for comparing the Scalar Reciprocal Saaty scale for comparing the
relative importance of the numerical relative importance of the
criteria, sub criteria and weight criteria, sub criteria and
alternatives alternatives
Extremely high importance 9 1/9 (0,11) Extremely low importance
Importance between very 8 1/8 (0,125) Importance between very low
high and extremely high and extremely low
Very high importance 7 1/7 (0,143) Very low importance
Importance between 6 1/6 (0,166) Importance between
significant and very high insignificant and very low
Significant importance 5 1/5 (0,2) Insignificant importance
Importance between 4 1/4 (0,25) Importance between
moderately higher and much moderately lower and much
higher lower
Importance moderately 3 1/3 (0,33) Importance moderately
higher lower
Importance between equal 2 1/2 (0,5) Importance between equal
and moderately and moderately lower
Equal importance 1 1 Equal importance

The numbers 2, 4, 6 and 8 are intermediate values that can be used to represent law
shades in addition to those five basic assessments 1, 3, 5, 7 and 9.
Although the method has been criticized because of the link between verbal
descriptions and corresponding numerical scale, it certainly presents some advantages:
• pair comparisons users perceive as simple and convenient;
• it is very useful when the criteria are qualitative;
• can be managed discrepancies in relative judgments.
In this study are presented criteria for blades assembly at the wind farm site with
obvious sub-criteria which were been set relative weights of importance, which through
key comparisons matrix has lead the selection of the final decision.
The use of AHP methodology involves six steps.
Step 1
1. The scheme hierarchical composition of the problem analyzed;
AHP decision tree for blades assembly has 4 levels. The highest level describes the
general decision (goal), the middle level describes the criteria to be considered in
choosing the best alternative. The role of sub-criteria at the intermediate level involve
a more complex analysis in choosing the final decision, and the bottom level presents
the three alternatives (Fig. 1).
 206
A. Badea et al.

Fig. 1. AHP decision tree for blades assembly

 Relieving Supply Chain Through Collaboration 207

Step 2
2. Establish relative weights by Saaty scale and their argument values given for
comparison;
• 1 Restriction degree of the composite material from specialized teams
The comparison carried out between Restriction degree of the composite
material from specialized teams and Generated restrictions by supply
chain collaboration links can be argumented as:
– The criterion for Restriction degree of the composite material from spe-
cialized teams have importance between equal and moderately higher,
accounting for number 2 than criterion Generated restrictions by supply
chain collaboration links is 1/2 (0.5). This value is obtained as multidis-
ciplinary activities for choosing blades are the first important step in which
research is carried out on the properties of wind blades composite materials
instead criterion restrictions arising from supply chain collaboration links
represents only a portion whose value is between equal and moderately
lower as it is a link which must meet the requirements of technically teams
for choosing blades. Basically, a false start by multidisciplinary teams
involving many risks to the supply chain.
The comparison carried out between Restriction degree of the composite
material from specialized teams and Wind blades classification can be
argumented as:
– The criterion for Restriction degree of the composite material from spe-
cialized teams has an importance moderately higher, accounting for number
3 against Wind blades classification criterion with a weight of 1/3 (0.33),
namely importance moderately lower. The number 3 assigned to Restriction
degree of the composite material from specialized teams represent that
specialized tests and yielded data conducted by multidisciplinary teams for
choosing blades tests contribute to the choice of classifying blades.
The comparison carried out between Restriction degree of the composite
material from specialized teams and Accessibility location degree can be
argumented as:
– The criterion for Restriction degree of the composite material from spe-
cialized teams have an importance between moderately higher and much
higher with number 4 beside criterion Accessibility location degree with an
importance between moderately lower and much lower with a 1/4 (0,25).
The number 4 shows the importance given to specialized teams for
choosing blades, which by experts appointed to project management were
conducted feasibility studies for the choice of the wind blades location,
which ensure the effective realization of the project.
The comparison carried out between Restriction degree of the composite
material from specialized teams and Important factors in choosing blades
can be argumented as:
– Criterion Restriction degree of the composite material from specialized
teams have an importance between equal and moderately lower 1/2 (0,5)
beside criterion Important factors in choosing blades which have an
208 A. Badea et al.

importance between equal and moderately higher with number 2. This

number 2 given to criterion Important factors in choosing blades is due to
economic factors, namely project financing delay had an impact on the
project activities achievement.
The comparison carried out between Restriction degree of the composite
material from specialized teams and Exposure to risk factors in choosing
blades can be argumented as:
– Criterion Restriction degree of the composite material from specialized
teams have an importance between equal and moderately lower 1/2 (0,5)
beside criterion Exposure to risk factors in choosing blades which have an
importance between equal and moderately higher with number 2. The
argumentation given the weight 2 for criterion exposure to risk factors in
choosing blades is due to the risk factor blades fatigue strength which
present important indices that measure the performance of the blades. 1/2
weight is expressed through multidisciplinary teams that can identify risks
caused produce by weather conditions.
In addition numerical values were set by Saaty scale and also argu-
mentation data values for comparing next criteria in: Generated restrictions
by supply chain collaboration links, Wind blades classification, Accessi-
bility location degree, Important factors in choosing blades and Exposure
to risk factors in choosing blades. After has being justified the given
numerical values by comparing the decision criteria, the next step is to
calculate the criteria initial matrix. Criteria initial matrix indicates prefer-
ence criteria, namely presents scalar number given based on Saaty relative
importance scale, then comparisons are made to achieve the final decision
(Table 2).

Table 2. Scalar number

 Relieving Supply Chain Through Collaboration 209

After arguments were made and based on the scale values, the next step
involves calculating the matrix. Reciprocal numbers are thus calculated, 1/2
indicated 0,5 in gross matrix. (Table 3) On the diagonal matrix it is
assigned number 1 because comparing a criteria to himself returns an equal
comparison. After totals on columns were calculated it can be calculate the
normalized matrix. To interpret and give relative weights to each criterion,
it is necessary to normalize the previous comparison matrix. Normalization
is done by dividing each number from the table by the total amount on
column (Table 4).

Table 3. Gross matrix

The contribution of each criterion in the final decision, Blades

assembly at the wind farm site in isolated region is determined by cal-
culations using the priority vector (or eigenvector). Eigenvector indicates
the relative weights between each criterion and is obtained by calculating
the arithmetic average of all criteria, as illustrated in Table 5. It is noted that
the sum of all vectors is always equal to 1.
The results of comparisons in Criteria initial matrix shows the con-
tribution of each criterion to the purpose defined in the decision tree (Fig.2).
Data from Figure 2 and values vectors shows the contribution of the cri-
terion Generated restrictions by supply chain collaboration links with a
contribution of 28.1% to the decisional purpose, but also contributing close
enough the criterion Exposure to risk factors in choosing blades with a
value of 23 6%. With a minimal amount criterion Wind blades classifi-
cation contributes only 4.7% to the defined purpose.
210 A. Badea et al.

Table 4. Normalized matrix

Total sum on column 6,08 4,25 18 15,33 9,16 4,41
Normalizing criteria 1/6,08 2/4,25 3/18 4/15,33 0,5/9,16 0,5/4,41
0,5/6,08 1/4,25 4/18 4/15,33 4/9,16 2/4,41
0,33/6,08 0,25/4,25 1/18 0,33/15,33 0,33/9,16 0,25/4,41
0,25/6,08 0,25/4,25 3/18 1/15,33 0,33/9,16 0,33/4,41
2/6,08 0,25/4,25 3/18 3/15,33 1/9,16 0,33/4,41
2/6,08 0,5/4,25 4/18 3/15,33 3/9,16 1/4,41
Results
Restriction degree of the 0,1644 0,4705 0,1666 0,2609 0,0545 0,1133
composite material from
specialized teams
Generated restrictions by 0,0822 0,2352 0,2222 0,2609 0,4366 0,4535
supply chain collaboration
links
Wind blades classification 0,0542 0,0588 0,0555 0,0215 0,0360 0,0566
Accessibility location degree 0,0411 0,0588 0,1666 0,0652 0,0360 0,0748
Important factors in choosing 0,3289 0,0588 0,1666 0,1956 0,1091 0,0748
blades
Exposure to risk factors in 0,3289 0,1176 0,2222 0,1956 0,3275 0,2267
choosing blades

Step 3
3. The relative prioritization of the subcriteria is obtained by comparing
sub-criteria two by two depending on the decision criterion, in order to rank
them;
In this step is necessary to assess the relative weights of the sub-criteria from the third
level of the hierarchy, the process is similar to that used in the criteria initial matrix
(Figure 3). This process is executed as the previous step, but the difference does that
in this step is performed the first level of the hierarchy evaluation criteria (group) as
discussed above (Table 6).
Step 4
4. Pairwise comparison is made of decision criteria (criteria and sub-criteria)
gives them a global priority.
Global priority for each decision criterion is determined by the result of multiplying
each priority vector for each criterion on the second level of the decisional tree with
relative prioritization of sub-criteria from the third level of decisional tree (Table 7).
Step 5
5. Performance matrix is achieved by calculating the relative prioritization of the
alternative for each sub-criterion.
In this step is necessary to assess alternatives relative weights from fourth level of the
hierarchy, the process is similar to that has been used in criteria and sub-criteria group
for the purpose decision-making (Fig. 3). This process is executed as the previous
step, but the difference does that this step is performed assessment of each alternative
beside each sub-criterion.
 Relieving Supply Chain Through Collaboration 211

Table 5. Priority vector calculation criteria

Priority vector calculation criteria Results
Restriction degree (0,1644 + 0,4705 + 0,1666 + 0,2609 + 0,0545 + 0,1133)/6 0,205
of the composite
material from
specialized teams
Generated (0,0822 + 0,2352 + 0,2222 + 0,2609 + 0,4366 + 0,4535)/6 0,281
restrictions by
supply chain
collaboration
links
Wind blades (0,0542 + 0,0588 + 0,0555 + 0,0215 + 0,0360 + 0,0566)/6 0,047
classification
Accessibility (0,0411 + 0,0588 + 0,1666 + 0,0652 + 0,0360 + 0,0748)/6 0,073
location degree
Important factors (0,3289 + 0,0588 + 0,1666 + 0,1956 + 0,1091 + 0,0748)/6 0,155
in choosing
blades
Exposure to risk (0,3289 + 0,1176 + 0,2222 + 0,1956 + 0,3275 + 0,2267)/6 0,236
factors in
choosing blades

Fig. 2. Criteria initial matrix

212 A. Badea et al.

Fig. 3. Choosing high scor priority alternative

Table 6. The relative prioritization of subcriteria

Restriction degree of the composite material from Generated restrictions by supply chain collaboration
specialized teams links

Wind blades classification Accessibility location degree

Important factors in choosing blades Exposure to risk factors in choosing blades

 Relieving Supply Chain Through Collaboration 213

Table 7. Calculations for global prioritization of subcriteria

Decisional Priority Decisional The relative Global prioritization
criteria vector for subcriteria prioritization of
each of subcriteria subcriteria = Priority
decisional vector for each
criterion decisional
criterion  The
relative prioritization
of subcriteria
Restriction 0,205 Specialized in 0,177 0,0362
degree of the mechanical
composite structures
material from Specialized in 0,373 0,0764
specialized electrical
teams machines
Specialized in 0,258 0,0528
power
electronics
Specialized in 0,125 0,0256
automatic
controlling
strategies
Specialized in 0,060 0,0123
project
management
Generated 0,281 Quality raw 0,388 0,1090
restrictions by materials used
supply chain Component 0,230 0,0646
collaboration manufacturing
links Blades 0,163 0,0458
transportation
Blades on site 0,066 0,0185
integration &
installation
Operation & 0,103 0,0289
maintenance
Recycling 0,046 0,0129
Wind blades 0,047 Up-wind 0,833 0,0391
classification Down-wind 0,166 0,0078
Accessibility 0,073 Region 0,198 0,0144
location degree isolation
degree
Wind velocity 0,459 0,0335
The distance 0,113 0,0082
from electrical
grid
(continued)
214 A. Badea et al.

Table 7. (continued)
Decisional Priority Decisional The relative Global prioritization
criteria vector for subcriteria prioritization of
each of subcriteria subcriteria = Priority
decisional vector for each
criterion decisional
criterion  The
relative prioritization
of subcriteria
Location slope 0,164 0,0119
degree
Distance from 0,063 0,0045
manufacturing
to assembly
place
Important 0,155 Technical 0,238 0,0368
factors in factors
choosing Economic 0,559 0,0866
blades factors
Socio-political 0,064 0,0099
factors
Environmental 0,136 0,0210
factors
Exposure to 0,236 Energy 0,313 0,0738
risk factors in absorption to
choosing wind impact
blades Fatigue 0,459 0,1083
strength
Thermal 0,144 0,0339
conductivity
Resistance to 0,086 0,0202
humidity

Step 6
6. The final decision is obtained by choosing alternative priority with maximum
points and argumentation of the obtained solution (Fig. 3).

5 Buffer Dimensioning for Blades with Higher Composition

of Carbon Fiber and Fiberglass Through DBR Philosophy

DBR philosophy level orders delivery in supply chain through safety stocks that are
placed and dimensioned to protect the rhythm of each individual link in the chain. The
purpose of using DBR solution for dimensioning intermediary buffers has the role to
protect the links in supply chain against orders variation occurred within the chain, with
 Relieving Supply Chain Through Collaboration 215

the condition to keep the same rhythm at each link in the supply chain. To exemplify
this fact research team proposes a calculation model for dimensioning buffer, based
on the philosophy DBR [3] through target buffer (Tb) and status buffer (Sb), as follows
[4, 5]:
1. The maximum level of the buffer, called the Target buffer (Tb) is calculated
according to the formula (1)

MO
Tb ¼ DLT   SP ð1Þ
UDLT
Where:
DLT is delivery lead time
MO is medium order, on
UDLT measuring unit delivery lead time
SP is safety percentage
2. Status buffer (Sb) refers to the buffer stock level at a time.
Further buffer status is calculated for yellow in percentage rate (Sbyellow) (for-
mula 2), that will generate the order release, respectively for red buffer status (Sbred)
in percentage rate. Sbred has the significance of buffer level in which pulling is
generated „ROPE”, more specific, the entrance of the new stock in buffer (for-
mula 4). The status buffer for red (Sbred) is calculated based on the medium daily
order (MOdaily) according to the formula 3.

MO
Sbyellow ¼  100 ð2Þ
Tb
MO
MOdaily ¼ ð3Þ
7
MOdaily
Sbred ¼  100 ð4Þ
Tb

Where:
MOdaily is daily medium order
By monitoring target buffers and their colors addressed, is dimensioning interme-
diate inventory levels to maintain a normal constant of orders at each link in the chain.
In this case buffer dimensioning between F2–F3 with a target of 45 orders and a
minimum rate order of 30 considering that buffer replenishment takes one week
indicates that at the moment when the level of yellow buffer reach 66.66%, repre-
senting the net amount that is consumed during one week period equal to refill time
(DLT). Consequently, at this value the yellow buffer will generate the order release to
downstream link. The moment when the level of the red buffer reach 9.33% the
replenishment valve rope is pull (Rope). This is the time when the buffer must be
refilled, considering that DLT was respected from upstream link to the link that has
placed the order (Figure 4).
216 A. Badea et al.

Specifically, DBR method is designed to optimize a flow process, getting the

maximum capacity of the most constrained links (CCL) in the supply chain. [6] Then
the rhythm of CCL represents the drum for the rest of the system. The rope represents
the mechanism of releasing the raw material into the flow process, protecting the CCL
from being swamped with “work in progress”. The rope regulates the rate of inserting
the raw material into the flow process. The inserting rate is no faster than that impose
by the drum. The rope is connected with the drum with the help of the safety buffer that
protects the CCL from starving because of the work during the process. [7] DBR helps
to move important raw materials (carbon fiber and fiberglass) in supply chain in safety
stocks that are placed and dimensioned to protect the rhythm of each individual link in
the chain. The purpose of using DBR solution for dimensioning intermediary buffer has
the role to protect the links in supply chain against bottlenecks variation occurred
within the chain, with the condition to keep the movement rhythm of the project at each
link in the supply chain. DBR method helps at any time to supply raw materials for
creating security stocks, which are placed and sized to ensure fluency tests included in
the project development cycle phases.

Fig. 4. Buffer dimensioning

6 Conclusions

Using AHP algorithm identified the optimal combination of composite blade structure
so as to meet the unique technical parameters specific RES project. Using a hybrid
composite material for wind blades offers economic and environmental advantages.
Through the AHP method for obtaining the final decision were taken into account all
technical aspects, risk exposure, environment accessibility and not at least logistics
elements so selected raw materials for the manufacturing process to be available at the
appropriate time, even if they are not yet integrated into a logistic flow. Evolution of
intermediate and final results of the methodology AHP out clearly in evidence the
importance of logistical issues to select the final alternative.
 Relieving Supply Chain Through Collaboration 217

Philosophy Drum-Buffer-Rope (DBR), streamlined by sizing a buffer stock for

moving the fiberglass in the supply chain, which was reduced assembly time for the
blades at the wind farm site, but also the implementation time of the project.
The benefit of the two methods used, AHP and DBR, provides a collaborative plan for
renewable energy projects.

Acknowledgement. This paper was developed through the Partnerships in Priority Areas -
PN II, with the support of ANCS, CNDI - UEFISCDI, Project no. 36/2012.

References
1. Griffin, D.A.: Blade System Design Studies Volume I: Composite Technologies for Large
Wind Turbine Blades. Global Energy Concepts, LLC, Kirkland (2002)
2. Saaty, T.L.: Theory and Applications of the Analytic Network Process: Decision Making with
Benefits, Opportunities, Costs, and Risks. RWS Publications, Pittsburgh (2005)
3. Cox, J.F., Blackstone, J.H., Schleier, J.G.: Managing Operations: A Focus on Excellence, vol.
2. Éditeur North River Press, Great Barrington (2003). ISBN 0884271722, 9780884271727
4. Goldratt, E.M.: Production the TOC Way, Revised edn. North River Press, Great Barrington
(2003)
5. Goldratt, E.M., Goldratt, A.: TOC Insights into Operations. Goldratt’s Marketing Group
(2003)
6. Prostean, G., Badea, A.: Supply chain collaboration model based on drum-buffer-rope
philosophy. In: 6th International Workshop on Soft Computing Applications (2014)
7. Blackstone Jr., J.H.: A review of literature on drum-buffer-rope, buffer management and
distribution. In: Theory of Constraints Handbook. McGraw-Hill, New York (2010)
 New Learning and New Teaching in Higher Education.
Formal vs. Informal Learning

Camelia Ciuclea ✉ , Gabriela Prostean, and Diana Andone

( )

Politehnica University of Timisoara, Timisoara, Romania

{camelia.ciuclea,gabriela.prostean,diana.andone}@upt.ro

Abstract. Education engaged to the modern way of creating and using infor‐
mation, and focuses on becoming a solution, on answering to the needs of the
present. Quantity is no longer sufficient, as quality is demanded as well. Where
does this need of new learning and new teaching in higher education comes from?
What are the limitations? New learning and new teaching involves changing
mentality, it implies taking into consideration the practical experience, the align‐
ment with the reality we live in. It means learning about students’ understanding
and teaching accordingly. University teaching is unfortunately still based on the
theory transmitted in lectures or presented online. Teaching itself is seen as a
process of getting some skills about how to lecture, how to organize a small group,
how to manage the activity, how to set a homework, etc. rather than how to use
knowledge, how to approach a teaching problem. New learning and new teaching
is defined by certain characteristics: possessing teaching skills, focusing on
students’ goals, listening and learning from students, constantly evaluating their
own performance.
This paper provides related work in the field, study findings and conclusions.

Keywords: Learning · Teaching · Education · Formal learning · Informal

learning

1 Learning/Teaching

The heart of education consists in the exchange of information/a conversation between

learners. Although this is a constant that remains true, higher education became a place
that people of twenty or thirty years ago would not recognize. The explanation is simple:
education aligned to the modern way of creating and using information, and focuses on
becoming a solution, on answering to the needs of the present. Quantity is no longer
sufficient, as quality is demanded as well. Disciplinary boundaries are crossed and the
core of education is the need to instant access to information.
Where does this need of new learning and new teaching in higher education comes
from? What are the limitations? In order to answer the questions we must have a look
towards the way universities are seen, as engines for change and development.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_17
 New Learning and New Teaching in Higher Education 219

2 Delivering Education: What it Takes to be a New Educator?

Teachers must be competitive, they are required to be more businesslike. The adminis‐
trative effort should accompany the change and go beyond the bureaucracy that asso‐
ciates the dialogue between student and teacher. There are tensions to be solved because
of the requirement to do more with less resources, to work with a broad spectrum of
students (having various abilities and different background).
A university teacher is expected to: be able to redesign courses, know and use ICT
tools, find and apply suitable methods of teaching different groups of students, inspire
them, determine them to study for the pleasure of learning.
New learning and new teaching implies changing mentality, it involves taking into
consideration the practical experience, the alignment with the reality we live in. It means
learning about students’ understanding and teaching accordingly.
Learning and teaching are interchanging activities. ‘One learns by teaching; once
cannot teach except by constantly learning’ (Eble 1988, 9).
University teaching is unfortunately still preponderant based on the theory trans‐
mitted in lectures or presented online.
Teaching itself is seen as a process of getting some skills about how to lecture, how
to organize a small group, how to manage the activity, how to set a homework, etc. rather
than how to use knowledge, how to approach a teaching problem.
New learning and new teaching is defined by certain characteristics: possessing
teaching skills, focusing on students’ goals, listening and learning from students,
constantly evaluating their own performance.

3 Formal/Non-formal/Informal Learning

R. Säljö studied the phenomenon of learning and teaching. Regarding this subject it is
to be reminded that (according to him) there are many conceptions about education.
The three most important conceptions Säljö identified (that are related to the concepts
of new learning) are:
• Learning as acquiring facts, skills and methods
• Learning as making sense by relating information
• Learning as interpreting and understanding reality by comprehending and reinter‐
preting knowledge.
The original concept of non-formal education emerged in 1968 (Coombs 1968). The
context where it emerged was dominated by the idea that education was failing (e.g.
Illich 1973), in relation with developing countries but also in so-called Western (or
Northern) societies as well (e.g. Bowles and Gintis 1976).
There were, in the West, different forms of the reform. Between 1968 and 1986 non-
formal education was seen as the only solution for all the problems of education in the
society (Freire et al. 1972).
Many if not most aid agencies and governmental programs included non-formal
education in their portfolio of interventions, and there were substantial sums of money
220 C. Ciuclea et al.

spent on it (in fields such as: research centers, consultants, academics, publications and
reports, etc.).
Many considered non-formal education as the ‘ideal’ form of education, far better
than formal education. But there were others who have seen non-formal education as a
sub-system of education (characterized by being considerably inferior to formal
schooling). Some voices described it as a temporary ‘necessary evil’ in situations of
crisis (Pigozzi 1999).
The topic divided the world of education into two: on the one hand is formal educa‐
tion: ‘Formal education as used here is, of course, the highly institutionalized, chrono‐
logically graded and hierarchically structured ‘education system’, spanning lower
primary school and the upper reaches of the university’ (Coombs and Ahmed 1974: 8).
On the other hand is non-formal education. Non-formal education was defined as
every educational activity outside of formal: ‘Non-formal education … is any organized,
systematic, educational activity carried on outside the framework of the formal system
to provide selected types of learning to particular subgroups in the population, adults as
well as children’ (Coombs and Ahmed 1974: 8).
The meaning could be interpreted as every educational program provided by the
Ministry of Education or educational program like schooling provided by non-govern‐
mental agencies (NGOs). The definition could also comprise all the educational and
training activities of other Ministries (Health, Labour and Employment, Youth and
Sports and Culture, etc.).
The third element – informal education was defined as: ‘the lifelong process by which
every person acquires and accumulates knowledge, skills, attitudes and insights from
daily experiences and exposure to the environment – at home, at work, at play; from the
example and attitudes of family and friends; from travel, reading newspapers and books;
or by listening to the radio or viewing films or television. Generally, informal education
is unorganized and often unsystematic; yet it accounts for the great bulk of any person’s
total lifetime learning – including that of even a highly ‘schooled’ person’ (Coombs and
Ahmed 1974: 8). The concept is close to that of ‘experiential learning’.
The core of non-formal learning was the quality of being more ‘flexible’. According
to the public opinion, other ‘qualities’/setbacks include: less well qualified and trained
teachers, a simplified form of curriculum, different teaching-learning materials,
frequently part-time, more flexible dates of terms than the formal schools.

FORMAL FLEXIBLE PARTICIPATORY

EDUCATION SCHOOLING EDUCATION

The term ‘non-formal education’ is used to refer to both flexible schooling and
participatory education.
The three main types of learning/education are placed as it follows:
 New Learning and New Teaching in Higher Education 221

FORMAL NON-FORMAL INFORMAL

Related to the previous information, non-formal covers flexible schooling and

informal education covers participatory educational activities.
In order to better organize the research made by Coombs and Ahmed in their classic
studies, we could draw a distinction between education and learning (where the informal
learning is seen as incidental, unstructured, unintentionally, most extensive - Rogers
2003) this way:

NON-
FORMAL FORMAL PARTICIPATORY INFORMAL
EDUCATION EDUCATION EDUCATION LEARNING

The boundaries between each of these ‘sectors’ are very sensitive. The secret stays
in ‘learning’ as the keyword, as the original matter.
The purposeful and assisted learning, when we surrender our autonomy and accept
it is the first and most common type of education. When we control this and individu‐
alize/reshape it (in terms of choosing what to learn, when to learn/stop) we are engaging
in informal education. When engaging into a pre-existing learning programme but mould
it to our own features, we are engaging in non-formal education.

4 e-Learning – New Learning

The 1990s were dominated by e-learning (learning facilitated online through network
technologies). It emerged all over the world (Ravjee 2007) and appeared with new
vocabulary, new policies and different structures and budgets.
e-learning is an ICT practice present in universities through: e-mail, online journals,
networked libraries, creative software solutions for information management tasks in
teaching, research and administrative systems.
Moll et al. (2007) defined e-Learning as ‘flexible learning using ICT resources, tools
and applications, focusing on accessing information, interaction among teachers,
learners, and the online environment collaborative learning, and production of materials,
resources and learning experiences’.
eLearning is the use of technological interventions for teaching, learning and assess‐
ment (Mlitwa and Van Belle 2011), enabling learners to improve problem-solving skills
and helps educators to disseminate information more effectively. Millham et al. (2014)
stated eLearning is the key for the improvement of learners’ performance, engagement,
222 C. Ciuclea et al.

self- regulation, flexibility, interest and motivation because it includes active participa‐
tion and self-regulated learning, aspects leading to construction, learning-pace
adjustment and desired learning outcomes.
The Politehnica University of Timisoara (UPT) is well known for its opening towards
adopting and adapting to modern educational techniques (either formal, non-formal or
informal).
A large number of projects, collaboration, student contests, initiatives were and still
are taken by enthusiast teachers and employees willing to carry on the mission of deliv‐
ering new learning and new teaching (POSDRU Projects such as: DidaTech, eSTART,
Intreprindere Simulata, ViCaDis, Concurs IDMSC, Conferinte: OER, SMART,
SLERD, IEEE).
The “e-Learning” solution adopted and integrated by UPT since 1998 is a real
witness that higher education in Timisoara is connecting with the needs of its ‘customers’
(students, labour market) following the international trend of new learning, new
teaching.
Defining new learning and new teaching at the level of UPT is impossible without
mentioning the Distance Learning Center - a mean of catering to the needs of the emer‐
gent segment of (employed) adults who wanted access to higher learning, while still
keeping regular hours on their jobs (Ternauciuc 2014).
An attempt of establishing some defining points related to new learning/teaching
should include features like:
• Increased opportunities for assessment and self-assessment with a high degree of
objectivity;
• Adapting the learning style according to the needs of each individual student, as well
as greater freedom in the teacher’s teaching style;
• Temporal, geographical and organizational flexibility through creating asynchronous
online learning environments, where each participant in the educational process can
access the study materials and the evaluation tools at their own pace and from any
location;
• Improving the learning process by synthesizing larger amounts of information avail‐
able in a shorter time;
• Create new types of relationships among peers and between students and teachers
through the creation of virtual communities with a higher degree of interactivity and
informality;
• Decreasing personal inhibitions that commonly occur in a classic educational model
and stimulating creativity by removing formal barriers usually encountered in a
classroom.
An important evidence of the great impact of the implementation of the new learning
and new teaching can be translated after analyzing the data collected at the end of a case
study conducted at Politehnica University in 2014–2015. According to it, the use of new
methods and types of teaching resources lead to significant academic results: ‘The global
result is revealed in Fig. 1. We have the columns with the average notes from our eval‐
uation and the columns with multiannual grade (after the first and the second year of
 New Learning and New Teaching in Higher Education 223

study) at all disciplines. The video “generation”, even the multiannual is under the “print
generation”, obtain in our evaluation a +0.63’ (Onita et al. 2015).

Fig. 1. Evaluation average notes and multiannual grade

5 Discussions and Conclusion

‘As teachers we have the responsibility to assist students in understanding the basic
concepts and when it is necessary to deepen information upon a subject, the principles,
and particular facts and to give them the possibility to hear, try, see, make them think
and achieve the information’ (Onita 2015).
New learning and new teaching is a very popular and up to date concept. Combining
formal, non-formal and informal learning seems to be the solution for a successful
learning experience since learning is nowadays characterised by special needs that have
to be answered, so is the call of the labour market, for specialists trained in the use of
modern technologies that accompany the delivering of the knowledge.
New learning, new teaching are complementary to formal/traditional systems of
education, although not all academic environments are ready to embrace it. Yet, there
is plenty of room for positive examples and further considerations.
We can observe that the concept of new learning/teachings revolve around certain
topics such as education, change, future, collaboration, discussion or economy. Further
research in this direction could be made analysing the initiatives developed by important
academic centres in this direction with the purpose of extracting results, conclusion,
good practices that can be replicated in higher education.

References

Aspin, D., Chapman, J., Hatton, M., Sawano, Y. (eds.): International Handbook of Lifelong
Learning. Kluwer, London (2001)
Bowles, S., Gintis, H.: Schooling in Capitalist America. Basic Books, New York (1976)
Campbell, P., Burnaby, B. (eds.): Participatory Practices in Adult Education. Erlbaum, London
(2001)
224 C. Ciuclea et al.

Coombs, P.H.: World Educational Crisis: A Systems Approach. Oxford University Press, New
York (1968)
Coombs, P.H., Ahmed, M.: Attacking Rural Poverty: How Non-formal Education Can Help. John
Hopkins University Press, Baltimore (1974)
Coun Eur: Council of Europe Report Doc 8595 Non-formal Education (1999)
EU Memo: Memorandum on lifelong learning, Commission Staff Working Paper (2000)
Field, J., Leicester, M.: Lifelong Education. Routledge, London (2000)
Illich, I.: De-schooling Society. Penguin, Harmondsworth (1973)
Marton, F., Hounsell, D., Entwistle, N. (eds.): The Experience of Learning: Implications for
Teaching and Studying in Higher Education. Scottish Academic Press, Edinburgh (1997).
Laurillard, D.: Chap. 11
Marton, F., Säljö, R.: On qualitative differences in learning. I - outcome and process. Br. J. Educ.
Psychol. 46, 4–11 (1976)
Millham, R., Thakur, S., Malan, C.: Does self-regulating e-Learning assist in secondary school
preparation for engineering education? In: Proceedings of 2014 Zone 1 Conference of the
American Society for Engineering Education (2014)
Mlitwa, W., Van Belle, J.W.G.D.: Mediators for lecturer perspectives on learning management
systems at universities in the Western Cape, South Africa. In: 2011 Proceedings of the Pacific
Asia Conference on Information Systems, PACIS 2011, Brisbane. AIS Electronic Library,
Brisbane (2011)
Moll, I., Adam, F., Backhouse, J., Mhlang, E.: Status Report on ICTs and Higher Education in
South Africa (2007). http://www.judybackhouse.com/pdfs/saide_status_of_elearning_in_sa.
pdf
Mocker, D.W., Spear, G.E.: Lifelong Learning: Formal, Nonformal, Informal and Self-directed.
ERIC, Columbus, Ohio (1982)
Onita, M., Ciuclea, C., Vasiu, R.: The Value of Video Lectures in the Classroom-UPT Study Case.
Springer, Cham (2015)
Pigozzi, M.J.: Education in Emergencies and for Reconstruction: A Developmental Approach.
UNICEF, New York (1999)
Ramsden, P.: Learning to Teach in Higher Education. RoutledgeFalmer, Oxon (2003)
Ravjee, N.: The politics of e-Learning in South African higher education. Int. J. Educ. Dev. Inf.
Commun. Technol. (IJEDICT) 3(4), 27–41 (2007)
Rogers, A.: What Is the Difference? A New Critique of Adult Learning and Teaching. NIACE,
Leicester (2003)
Ternauciuc, A.: Analyzing the use of the CVUPT e-Learning platform. Scientific Bulletin of the
Politehnica University Timisoara (2014)
Integrating E-learning Platforms in Teaching/Learning. A
Critical Review Upon the Virtual Campus Usage in UPT

Camelia Ciuclea ✉ , Gabriela Prostean, Diana Andone, and Anca Draghici

( )

Politehnica University of Timisoara, Timisoara, Romania

{camelia.ciuclea,gabriela.prostean,diana.andone,
anca.draghici}@upt.ro

Abstract. E-learning is an emerging trend in education. As one of the most used

open source e-learning platform, Moodle became the solution that Politehnica
University Timisoara adopted and adapted in order to ensure the access to online
courses to the enrolled students. A generalized analysis on the platform, (compa‐
rative approach to the overall activity on the CVUPT learning management
system) based on two sources (the administrator’s web interface of the CVUPT,
and direct access to the database, where specially designed SQL queries were
employed to return results relevant to the use of the platform) led to some inter‐
esting results: the number of courses has more than doubled and the number of
active users has quadrupled (from September 2010 to September 2014). The
overall use of CVUPT has increased and each year, a large number of new
accounts are created. There is a dynamic growth in terms of integrating e-learning
platforms in teaching and Learning, especially in distance learning programmes,
as those run by Politehnica University Timisoara. Both students and tutors express
their constant interest in using CVUPT learning management system, either as a
main or auxiliary teaching/learning tool.

Keywords: Learning · Teaching · Education · Formal learning · Informal

learning

1 E-learning

The exact definition and explanation of the concept of eLearning has been an elusive
goal since the first mentioning of the term. Perhaps the clearest definition is also the
simplest: eLearning (or Electronic Learning) is “delivering learning by any means of
the electronic media” [5].
Using optical discs as a support for presentations or a projector connected to a laptop
during lectures can be viewed as e-learning, but the general meaning is that e-learning
states for online courses.
E-learning is an emerging trend in education [12] (Table 1).

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_18
226 C. Ciuclea et al.

Table 1. Types of learning/courses used in higher education.

Proportion of content Type of course Typical description
delivered online
0% Traditional Course with no online technology used –
content is delivered in writing or orally
1 to 29% Web Facilitated Course which uses web-based technology to
facilitate what is essentially a face-to-face
course. Uses a course management system
(CMS) or web pages to post the syllabus and
assignments, for example
30 to 79% Blended/hybrid Course that blends online and face-to-face
delivery. Substantial proportion of the content
is delivered online, typically uses online
discussions, and typically has some face-to-
face meetings
80+% Online A course where most or all of the content is
delivered online. Typically have no face-to-
face meetings

It is almost impossible to think about teaching and learning without associating it

with ICTs tools (Information and Communication Technologies). ICTs is empowering
educators and learners and transforms teaching and learning processes by shifting the
attention from a teacher-dominant aspect to a student-centered type of learning. The use
of ICTs in learning leads to increasing gains for students in terms of creativity, commu‐
nication, thinking, reasoning.
Whatever form it takes the e-learning has the following advantages [13]:
• Increased opportunities for assessment and self-assessment with a high degree of
objectivity;
• Adaptation of the learning style needed by each student, as well as greater freedom
in the teacher’s teaching style;
• Flexibility in terms of creating asynchronous online learning;
• Improving the learning process by synthesizing larger amounts of information avail‐
able in a shorter time;
• Creating new types of relationships among peers and between students and teachers
through the creation of virtual communities with a higher degree of interactivity and
informality;
• Decreasing of personal inhibitions that commonly occur in a classic educational
model and stimulating creativity by removing formal barriers usually encountered in
a classroom.
The most important role played by ICTs in education is connected with LMS
(Learning Management Systems). These systems allow students the access to course’s
content in text, image or sound format. Also they allow the interaction via messages,
forums, video-conference, chat, etc.
The configurable features that characterizes LMS allow creation add courses, regis‐
tering, monitoring, evaluation of students/teachers via Internet. From a technical point
 Integrating E-learning Platforms in Teaching/Learning 227

of view there are two major types of LMS: the one that provides commercial solutions
(Blackboard, WebCT) and the open source type (Moodle).
Moodle (Modular Object-Oriented Dynamic Learning Environment) has been iden‐
tified as the most used platform in higher education and also as the easiest to use [14].

2 E-learning Platforms. Moodle. Virtual Campus

As one of the most used open-source e-learning platform, Moodle became the solution
that Politehnica University Timisoara adopted and adapted in order to ensure the access
to online courses to the enrolled students (Fig. 1).

Fig. 1. Virtual Campus. Log-in page.

The Virtual Campus is a Moodle platform, whose features were reshaped and
enriched for the use in UPT. It allows the exchange of information between students
from different geographical points, via Internet, through synchronous and asynchronous
mechanisms of communication (chats, discussions forum, etc.).
It is easy to configure, it allows the users to create assessment processes and to
manage their tasks.
To better understand it, its main features are presented as it follows (Table 2):
228 C. Ciuclea et al.

Table 2. Activities and modules of the Moodle platform.

Activity Module Description
Creation Database Allows to build, display and search a bank of record
entries about any topic [11]; allows to share a collection
of data [2]
Organization Lessons Lessons represent a set of ordered topics summarizing
the instructional materials [2] and allow the access to
them through the respective link; [11]
Delivery Assignments Allow teachers to collect work from students [2]; allow
teachers to evaluate the student’s work and provide
feedback including grades, in a private mode [11]; allow
students to upload assignment files [11]
Workshops Represent a peer assessment activity with many options
[11]; allow students to submit their work via an online
text tool and attachments [11]
Communication Chats Allow synchronous conversation [11]
Forums Represent a communication tool where students and
teachers can exchange ideas by posting comments [11];
News Represent a special forum for general announcements
[11]; allow teachers to add posts and to send emails [11]
Collaboration Glossary Allows creating and maintaining a list of definitions [11];
represents a mechanism for collaborative activities that
can be restricted to entries made by the teacher [11]
Wikis Allow users to edit collaborative Web pages [2]; provide
space for collaborative work [11]
Assessment Choice Allows teachers to ask questions and specify multiple
choice answers [11]; represents a useful mechanism to
stimulate thinking about a topic [11]
Quiz Allows teachers to design and build quizzes with a
variety of questions, with different types of answers, such
as multiple choice, true/false, short answer [2]
Survey Feedback Allows teachers to create surveys to collect feedback [11]
a SCORM Represent specifications that enable interoperability,
Reusability
accessibility and reusability of the learning content [11];
represent tools that enable SCORM packages to be
included in the course [2]
External tools Enable interaction with compliant learning resources (eg.
Learning Tools Interoperability) and activities on other
Web sites [11]; provide access to new activities’ types or
materials [11]
a
This term is not included in the classification by Piotrowski [3].
 Integrating E-learning Platforms in Teaching/Learning 229

This classification is based on two classes of functionalities: resources and modules:

• The resources are instructional materials created in digital formats and then uploaded
to the Virtual Campus (VC) (Web pages, PowerPoint files, word documents, flash
animations, video and audio files).
• The modules are components created via Moodle in order to provide interaction
among students and teachers in terms of manipulation and content transformation.
Examples of modules: Database, Lessons, Assignments, Workshops, Chats, Forums,
News, Glossary, Wikis, Choice, Quiz, Survey, Feedback, SCORM (Sharable Content
Object Reference Model) and External tools [6].
Regarding the activities of the learning platforms Piotrowski, M. [10] presents a
classification based on six classes: Creation, Organization, Delivery, Communication,
Collaboration and Assessment.

3 Virtual Campus

The CVUPT platform became Plan B for Politehnica University Timisoara as it replaced
an in-house developed solution which met only partially the needs of a modern learning
management system (LMS). Although it provided basic access to course materials
(usually to resources in PDF format) and some communication tools (forums, blogs,
wikis) it almost failed to facilitate the learning process through grades management or
user activity tracking on the platform [18] (Fig. 2).
Based on an open-source LMS Moodle, CVUPT is the result of the work of a team
dedicated to e-Learning (the Multimedia Research Center of UPT).
September 2009 is the moment when CVUPT began functioning in as the learning
process management system for the Distance Learning Center that, by that time had two
specializations: Applied Informatics and Telecommunications Systems and Technolo‐
gies. Since the official policy of the program established that CVUPT was a closed
platform, without self-registration or guest access all the accounts and courses were
created by the E-Learning Centre (the technical team behind the platform: Andrei
Ternauciuc, Bogdan Dragulescu, Marian Bucos).
In 2010 there were introduced new sections dedicated to the Master level studies.
The UPT decided to provide blended-learning support to all the Master courses in the
university. This decision led to an exponential increase in user accounts (student and
teachers), courses, and activity on the platform, and therefore to a wider adoption of the
online medium as a valuable and necessary auxiliary tool to complement classical
teaching/learning scenarios.
Two different new specializations to Distance Learning were introduced in 2014:
Communication and Public Relations, and Manufacturing Engineering, and the use of
the Virtual Campus increased.
By the end of the 2013–2014 academic year, there were 1267 courses on CVUPT,
out of which 1022 dedicated to Distance Learning and Master Programs, the rest
comprising of additional courses requested from other specializations in UPT, a few
dedicated to various research projects, as well as a handful (10 to be more precise) of
230 C. Ciuclea et al.

Fig. 2. The virtual Campus of the Politehnica University Timisoara - CVUPT

social courses, aimed at encouraging communication between different parties, outside

of existingcourses (for instance, a forum for administrative issues regarding the Distance
Learning students and staff).
A generalized analysis on the platform, (comparative approach to the overall activity
on the CVUPT learning management system) based on two sources (the administrator’s
web interface of the CVUPT, and direct access to the database, where specially designed
SQL queries were employed to return results relevant to the use of the platform) led to
the following results (Fig. 3 and Table 3):
 Integrating E-learning Platforms in Teaching/Learning 231

Fig. 3. The evolution in number of users on CVUPT

Table 3. Number of users of CVUPT

CVUPT users 2010–2011 2011–2012 2012–2013 2013–2014
Active users 1522 2697 3951 6077
Total users 2906 4846 6332 10,206

From the perspective of the course resources the statistics show the following
(Table 4):

Table 4. Types of course resources

Resources 2010–2011 2011–2012 2012–2013 2013–2014
Resource 3203 - - -
Page - 1550 2828 4424
File - 2673 3814 5924
Folder - 165 411 751
URL - 192 339 620
Lesson 11 1 9 16
SCORM/AICC 4 2 4 10
232 C. Ciuclea et al.

The data that was analyzed proves that the activity on the CVUPT platform has
increased significantly in the four years period from the number of courses, to the number
of users and the expectations are that there will be a constant growth across all areas.
The number of courses has more than doubled (from 577 in 2011, to 1267 in 2014),
which can be explained, in part, by the growing number of non-mandatory courses
created at the request of willing teachers. Some of the courses also appeared as part of
the Masters programs in UPT.
The number of active users has quadrupled, from 1522 in 2011 to 6077 in 2014.
Also there are the now defunct courses, certain resources and activities that have not
been eliminated from the platform, courses that are contributing to the total number.
Whatever the case, the fact remains that overall use of CVUPT has increased and
each year, a large number of new accounts were/are created (Fig. 4).

Fig. 4. Number of courses created

In order to integrate e-learning platforms in teaching/learning there are some

improvements to be done [18]:
• The popularization of the less used tools in Moodle (either referring to course
resources and activities, or communication tools) by providing examples of concrete
use scenarios and comparisons with the more popular ones.
• An increased and constant support for the use of the platform by all types of users
(regular training seminars – face-to-face or at distance)
• The training of tutors in specific areas, such as the creation of calendar events, or the
configuration and use of the CVUPT gradebook.
• The identification of other factors and resources which might hinder the use of the
platform (such as usability, or mobile access), through further qualitative and quan‐
titative research.
 Integrating E-learning Platforms in Teaching/Learning 233

4 Discussions and Conclusion

There is a dynamic growth in terms of integrating e-learning platforms in teaching and

learning, especially in distance learning programmes, as those run by Politehnica
University Timisoara. Both students and tutors express their constant interest in using
CVUPT learning management system, either as a main or auxiliary teaching/learning
tool.
The progress of things will most probably lead to adopting the CVUPT at a larger
scale, to assist all the students enrolled by UPT, as Moodle proven to be accessible, user –
friendly, adaptive and widely known.
For its main six classes identified by Piotrowski, M. [15] (Creation, Organization,
Delivery, Communication, Collaboration and Assessment) the e-learning platform
Moodle represents the solution most preferred by the first 10 technical universities in
Romania (Fig. 5).

Fig. 5. The use of LMS in Romania

References

1. About CeL. http://elearning.upt.ro/despre-cel/c-18. Accessed June 2016

2. Alexander, B.: Web 2.0: a new wave of innovation for teaching and learning? Educ. Rev.
41, 32–44 (2006)
3. Andone, D., Vasiu, R., Robu, N.: Building a virtual campus for digital students. In: 2011 IEEE
Global Engineering Education Conference (EDUCON), pp. 1069–1073 (2011)
4. Aspin, D., Chapman, J., Hatton, M., Sawano, Y. (eds.): International Handbook of Lifelong
Learning. Kluwer, London (2001)
5. Berners-Lee, T., Hendler, J., Lassila, O.: The Semantic Web. Scientific American, p. 5 (2001)
234 C. Ciuclea et al.

6. Blin, F., Munro, M.: Why hasn’t technology disrupted academics’ teaching practices?
Understanding resistance to change through the lens of activity theory. Comput. Educ. 50(2),
475–490 (2008)
7. Campusul Virtual al Universității “Politehnica” din Timișoara - CVUPT (2016). http://
cv.upt.ro
8. Cole, J., Foster, H.: Using Moodle – Teaching with the Popular Open Source Course
Management System. O.R. Media, Editor, Sebastopol (2008)
9. Costa, C., Alvelos, H., Teixeira, L.: The use of Moodle e-learning platform. Elsevier Ltd.
(2012)
10. Ermalai, I., Dragulescu, B., Ternauciuc, A., Vasiu, R.: Building a module for inserting
microformats into moodle. Adv. Electr. Comput. Eng. 13(3), 23–26 (2013)
11. Moodle. Moodle (2012). http://moodle.org/. Accessed June 2016
12. Mihaescu, V.: The Development and Implementation of a MOOC Platform Model from a
Technical and Educational Perspective. Ed. Politehnica, Timisoara (2016)
13. Niculescu-Aron, G.I., Asandului, L., Marinescu Mazurencu, M., Mihaescu, C.: A cram of E-
learning advantages and disadvantages. Inf. Econ. 42, 6 (2007)
14. Paulsen, M.: Experiences with learning management systems in 113 European Institutions.
Educ. Technol. Soc. 6(4), 134–148 (2003)
15. Piotrowski, M.: What is an e-learning platform? In: Learning Management System
Technologies and Software Solutions for Online Teaching: Tools and Applications. I. Global,
Editor (2010)
16. Poulová, P., Šimonová, I., Janečka, P.: The use of a software application supporting individual
learning styles in eLearning. In: International Conference on Applied Computer Science,
Malta, p. 5 (2010)
17. Suvorov, R.: Using Moodle in ESOL writing classes. TESL-EJ 14(2), 1–11 (2010)
18. Ternauciuc, A.: E-learning 2.0 at the ‘Politehnica’ University of Timişoara. Sci. Bull. Politeh.
Univ. Timisoara 55(69), 23–26 (2010)
 Application of the Fuzzy-Pay-Off Method
in the Valuation of a Financial Instrument

Mihaela Crăciun1(&) and Luiela Magdalena Csorba2

1
Electrical and Power Engineering Faculty, Politehnica University Timisoara,
Timisoara, Romania
[email protected]
2
Department of Economics, Faculty of Economic Sciences, “Aurel Vlaicu”
University of Arad, Arad, Romania
[email protected]

Abstract. The current article aims to present the application of the

Fuzzy-Pay-Off Method within the valuation of a financial instrument (FI), more
specifically a patent. The Fuzzy-Pay-Off Method (FPOM) will be presented
alongside the mathematical model on which it is based on, also adding a short
description of what the concept of Fuzzy number represents. The advantages
and disadvantages of this valuation method will be outlined. The case study
accompanying this work will refer to the valuation of a patent using the FPOM
based on the analysis of a discounted cash flow (DCF). All work hypotheses are
based on three scenarios: an optimist one, a pessimist one and a realist one.
These scenarios are the basis upon which a research and development (R&D)
project shall be initiated. It is possible for the FPOM to be used as an extension
to the common DCF method.

Keywords: Fuzzy-Pay-Off Method (FPOM)
Financial Instruments (FI)

Discounted Cash Flow Method (DCFM)
Fuzzy control
Modeling and
optimization
Uncertainty

1 Introduction

This article was built upon the idea of valuating a specific financial instrument - the
patent. This type of valuation has proven itself to be a less tread upon niche amongst
most laborers in this domain. The Financial Instrument (FI) represents a contract that
generates certain rights and obligations between two sides, more precisely, collection of
capital or payments (in cash or by different means). An FI may also stand for being an
equity capital instrument [1].
The key to FI valuation is represented by the capacity to create a realistic projection
of recipient cash flows and/or any related payments in regard with owing the respective
FI. The most valuated FIs are: stocks, bonds, government securities and with a lower
rate of prevalence credit contracts.
The main reason behind the valuation of a FI would be a transaction intention,
either the purchase or the sale of the FI within an investment process. Transactions
can be public (ROMGAZ, TRANSELECTRICA, TRANSGAZ) or facilitated by

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_19
236 M. Crăciun and L.M. Csorba

auction conglomerates that deal with listings: investment banks, law firms or broker
agencies [2].
Another reason that calls for this type of valuation is financial reporting (the val-
uation is undertaken so as the FI would be registered in balance sheets) [3].
This aspect matters most for investment funds, and generally speaking for any legal
entity whose patrimony encompasses FIs.
Yet another reason for the valuation of a FI might be represented by case of dispute
(divorce, inheritance and so on).
There are rare cases in which the aim of the valuation of an FI is the granting of a
loan. This is only valid provided that the crediting institution accepts a FI as a security.
The result obtained from the valuation of a FI can be used independently or as a
component in the calculation of the final value of an enterprises’ valuation (alongside
real estate and mobile assets that contribute to the company’s patrimony). The valu-
ation of a company can be done both in a state of normal functioning of the firm or in
case of bankruptcy so as to restructure it [4].
It is easily understandable the reason why a FI valuation is necessary, the moment
any sort of transaction would be initiated both sides ought to know the market price of
the bartered object (the price being usually inferred from a recognized stock market
right at the time of the transaction or within a reasonable time-span before it) [5].
Most financial instruments, especially the ones exchanged at stock markets, are
valuated through automated models, using market analysis algorithms to determine the
price of assets. Be it one of the automated or classical valuating methods, each one of
them relies on the basic-most approaches, the market approach, the income approach
and the cost approach.
Within the income approach, the value of an FI can be determined through the use
of the Discounted Cash Flow Method (DCFM).

2 Related Works

2.1 The Discounted Cash Flow Method (DCFM)

The income approach may offer indications about the value through the conversion of
future cash flows into a single value [6].
The methods covered by the income approach are:
(a) The capitalization of income – to the income cumulated in a specific time-span the
rate including all risks is applied; that or the rate of total capitalization;
(b) DCF – to a series of cash flows from future time-spans a certain discount rate is
applied so as to convert their value to the present one [7];

X
n
CFi VT
ValueCF ¼ þ ; ð1Þ
i¼1 ð 1 þ aÞ i ð 1 þ aÞ n
 Application of the Fuzzy-Pay-Off Method 237

where:
• CFi = net cash flow at the shareholder’s disposal (NCFS) or net cash flow at the
enterprise’s disposal (NCFE);
• a = the discount rate;
• VT = Final/Residual Value;
(c) Different valuation models specific for option assessment.
DCFM is a method that allows the assessment of a certain type of value for an asset
capable to generate different forms of income throughout its expected life-span. The
method can be defined as the following [7].
This is a method included in the income approach, through which a discount rate is
applied to a number of expected future cash flows, so as to estimate their present value.
Cash Flow (CF) = income generated on a certain time-span by an asset, a group of
assets or by an enterprise.
Net cash flow at the firm’s disposal (NCFE) = a cash flow available for the reward
of shareholders (available as dividends) and creditors (credit and interests), after the
financing of enterprise exploitation activities and after all needed investments were
accomplished.
Net cash flow at the shareholder’s disposal (NCFS) = a cash flow available for the
reward of shareholders (dividends) after the financing of enterprise exploitation
activities were performed, after all needed investments and after the growth (or the
decrease) of financing through credits.
The discount rate represents how profitable would be to convert a sum of money or
a cash flow from the future into its current value.
Within this method each cash flow is updated through the use of a discount factor
(DF) which uses two variables: the size of the DF and the length of the future period
needed to generate the cash flow.

1
DF ¼ ; ð2Þ
ð1 þ k Þi

where:
• k = the discount rate;
• i = the future time-span in which the cash flow is generated.
As more distant in time from the present moment is the point at which the cash flow
appears, as low will the value of the DF be; furthermore, the present value will be as
small.
Thus, to conclude, the value of money decreases over time proportionately to the
amount of time that is to pass from the present moment.
238 M. Crăciun and L.M. Csorba

2.2 The Fuzzy Pay-Off Method (FPOM)

FPOM was first implemented by Mikael Collan, professor and researcher at the
Lappeenranta University of Technology Finland.
The method is based on Real Option Valuation (ROV), also known as Real Option
Analysis (ROA) using fuzzy logic.
Collan has worked within the IAMSR - Institute for Advanced Management
Systems Research, on the WAENO project, which focused on the valuation of a
gargantuan portfolio of industrial real estate investments. As a result of his research,
Collan publishes his dissertation thesis in 2004 [8].
Alongside Mikael Collan another team formed by Christer Carlsson - professor at
the Åbo Akademi University and Robert Fullér, PhD – researcher at the Institute for
Applied Mathematics - Óbuda University, Budapest, Hungary; has worked on a new
approach related to possible mean values and the variation of fuzzy numbers. The two
researcher’s results were used in the first attempt at a Black-Scholes model which used
fuzzy numbers. The said model was then later renamed as a hybrid fuzzy
Black-Scholes model, as it uses both fuzzy numbers and crisp numbers [9].
During the time of the research the team was joined by other researchers: Peter
Majlender, Francisco Alcaraz Garcia, and Markku Heikkilä – researchers at IAMSR
[10]. A derivate model that the team has later developed was hard to implement to
laborers in the field and was used only in research.
In 2008 Collan took part to the 12th Annual International Conference on Real Options
in Rio de Janeiro Brazil. He takes inspiration from the work of Scott Mathews from
Boeing Inc., after Mathews presents and explains the possible appliance of the
DATAR-Mathews method in ROV. The DATAR-Mathews method uses a simulation in
order to generate a distribution of the probabilities of certain cash flow scenarios that may
result, scenarios provided by project managers. After that, the probability of the weighted
mean value is calculated and then it is multiplied with the probability of a positive result
(% from a total possible of 100%), and thus obtaining the real option value.
Collan’s conclusion came as following: the Pay-Off distribution of a project’s
value, calculated through the use of fuzzy numbers, is no more than the fuzzy present
net value (fuzzy NPV) of the project. The mean value of all fuzzy NPV positive
numbers is equal to the mean possible value [11] of fuzzy NPV numbers. Thus the real
option value obtained from the fuzzy NPV is the possible mean value of the positive
fuzzy NPV multiplied with the ratio between the positive fuzzy NPV area and the total
fuzzy NPV area. This fact is described in Eq. 3 [12]:

R1
Að xÞdx
ROV ¼ 0R1  E ðA þ Þ; ð3Þ
Að xÞdx
1

where:
• A – represents the fuzzy NPV;
• E(A +) – possible mean value of positive fuzzy NPV;
 Application of the Fuzzy-Pay-Off Method 239

R1
• Að xÞdx – total fuzzy NPV area;
1
R1
• Að xÞdx – positive fuzzy NPV area.
0

FPOM is based on the theory of possibility (Lotfi Zadeh - 1978) while the
DATAR-Mathews method is based on the theory of probability. Operands of the theory
of possibility can be seen as a hyper-cautious version of the operators used by a
transferable belief model, the model being a modern development of the theory of
evidence [11].
Fuzzy numbers can be classified as triangular, trapezoidal or bell shaped.
Throughout this article only triangular fuzzy numbers (TFN) will be used (see Fig. 1).
TFN were chosen because they are easy to implement and lead to optimized results.
TFN are used in different domains and are implemented in various instruments, such as
Fuzzy Expert Systems [13].

Fig. 1. TFN (possible distribution), defined by three points {a, a, b} that describe the NPV of a
future project [14]

TFN is defined as a fuzzy number represented through three coordinates as follows:

A = (a1, a2, a3). This representation is interpreted as an appurtenance function this way
(see Fig. 2) [15]:
8
>
> 0 if x\a1
< xa1
if a1  x\a2
lð AÞ ð xÞ ¼ a a
a23 x1 : ð4Þ
>
> if a2  x\a3
: 3 2
a a
0 if x  a3

Fig. 2. The three point repre-

sentation of a fuzzy number and
its appurtenance function (4)
240 M. Crăciun and L.M. Csorba

If a a-cut type operation is applied, the Aa can be obtained by using Eq. (4) this
way:
ð aÞ ð aÞ
a1 a1 a3 a3
8a 2 ½0; 1 from a2 a1 ¼ a; a3 a2 ¼ a, we have/generates the following results:

ðaÞ ðaÞ
a1 ¼hða2  a1iÞa þ a1 and a3 ¼ ða3  a2 Þa þ a3 ; so
ðaÞ ðaÞ
Aa ¼ a1 ; a3 ¼ ½ða2  a1 Þa þ a1 ; ða3  a2 Þa þ a3 :

The appurtenance function of a TFN truncated at the a-a-z point, where 0  z

a, is given by relation (5):

0 if ta  a  z
ðAjzÞðtÞ ¼ : ð5Þ
AðtÞ else

Hence the TFN expected value is obtained in Eq. (6) [12]:

8
ba
>
> ða  aÞ [ 0 then EðA þ Þ ¼ a þ
>
> 6
< ðaaÞ3
a [ 0 [ ða  aÞ then EðA þ Þ ¼ a þ ba6 þ 6a2
E ðA þ Þ ¼ : ð6Þ
>
> 0 [ a then E ðA þ Þ ¼ ðabÞ3
>
> 6b2
:
a þ b\0 then EðA þ Þ ¼ 0

3 Case Study – Valuation of a Patent with the FPOM, Based

on DCF

This section will present a case study referring to the use of the FPOM based on the
DCF method, in the valuation process of an invention patent.
The patent as an element of intellectual property represents a protection title for an
invention; a title that confers its owner the exclusive right to exploit the invention
throughout the validity term of the document [16].
The validity term represents the time-span in which legal protection is offered to the
owner of the patent for his intellectual property, this period varies according to national
regulations implemented, from 3 years and a half up to 17 years.
In Romania, the normative act that regulates all legal aspects regarding invention
patents is law no. 64/1991 (republished on the 19 of August 2014) [17].
According to the standing orders for the application of the accountancy law no.
82/1991 (updated on the 20 of April 2016) [18], patents, licenses, factory marks and
trademarks, as well as other rights to similar industrial or intellectual property, are
registered in accountancy reports and balance sheets at their contribution values, their
cost of acquisition or their production costs.
In regard to invention patents, the valuation method can be selected according to
the status of the entity that uses the patent, the owner of the patent or the company that
uses it without owning legal rights to the invention registered.
 Application of the Fuzzy-Pay-Off Method 241

3.1 Improving Patent Valuation Using the FPOM

Assessing the value of a patent is usually a difficult task as there are great uncertainties
regarding the future of the invention and the precision at which mathematical estima-
tions can be performed. The FPOM represents an easily understandable and applicable
method to be used in the valuation process. This method is based on the use of value
scenarios (optimist and pessimist) as well as the prediction of a possible real value, the
realist option (RO). The method was conceived for the analysis of assets that suffer from
estimate difficulties and are more often than not presenting a high risk of uncertainty.
The valuation of patents is closely related to the valuation of R&D projects, often
using data collected from R&D valuators, data taken directly from the actual projects
where it was possible.
Patent valuation implies a strenuous work, as no specific method was ever
implemented for this type of valuation. Patents whose future cash flows can be easily
predicted can be assessed easier through common valuation methods. Because the CF
of a patent can be estimated rather hard the results of the valuation indicate much rather
a trend than a precise estimation.
The use of scenarios is the most spread out method for modeling any uncertain
future of projects. The idea that stands behind the usage of scenarios is the following:
using future scenarios thought after the possible states in which the market might
situate itself upon helps to approximate as accurate as possible the value of the future
cash flow. Creating scenarios for patents or patent portfolios can be accomplished on
the basis of known facts about the future. The information need not be necessarily
accurate as scenarios are intended to cover an area of possibility as wide as it can get.
All information used can be provided by: qualitative information gathered from an
existent patent, an intellectual property rights’ (IPR) analysis or form a different
management system [19]. Scenarios can be used to fill in any of the three valuation
approaches (market, income, cost). Usually three types of situations are defined for an
investment in a patent:
1. the optimist case;
2. the pessimist case;
3. the realist case.
Any result from the three situations can be achieved. The scenario approach is
compatible and can be used for the valuation of other types of assets (e.g. Corporate
assets) [20].
The FPOM is easily understandable, for it uses the scenarios to create the pay-off
distribution of the valuated asset. The method is especially designed to evaluate and
analyze assets that suffer from precise estimation difficulties of future cash flows,
offering a large array of possibilities for each and every uncertainty.

3.2 The Fuzzy Pay-Off Method (FPOM)

FPOM is an analysis method applied in cases where numerical information is found in
under the state of scenarios. The main idea that the FPOM is based upon is the creation
of a numerical distribution based on the three scenarios mentioned above [12].
242 M. Crăciun and L.M. Csorba

The work hypotheses are the following:

• Observing the fact that the realist scenario is the most probable one and granting it
the attention it deserves;
• Deciding the fact that the optimist and the pessimist scenario represent the upper
and the lower limits of the probability distribution, no values higher or lower than
the ones presented in these two scenarios will be taken into consideration during the
course of the study;
• Assuming the fact that the geometrical representation is a triangular fuzzy number.
After constructing the pay-off distribution, treating it as a fuzzy interval allows to
the performing of mathematical operations upon it. For example, by using the FPOM it
is possible to calculate the real option value of the assessed patent directly from the
pay-off distribution (POD). It would also be possible to calculate a unique value from
the POD, taking into consideration its shape and solving its mean possible value [21].
This mathematical calculation can be easily performed in the case of a triangular POD.
If the DCF method is used for the valuation of a patent, then, FPOM can be applied
in the assessment of a composed realist option (RO), or for the assessment of assets that
the scenarios base themselves on.
FPOM is efficient in the valuation of R&D projects, a valuation that is based on
Real Option Valuation.

3.3 Patent Valuation Using Three Scenarios Through the DCFM

The current section will refer to the creation of the three necessary scenarios needed for
the valuation of the patent using the DCFM. The values obtained in each scenario are
the basis for the appliance of the FPOM.
The most relevant cash flows for the patent are identified, afterwards to be iden-
tified in two categories: personal usage or licensing. The personal use of the patent
refers to its utility to the patent’s owner (be it a person or a company). Licensing the
patent represents all external cash flows (copyright income and sales) that the patent
generates. In reality cash flows can be generated by many other external sources. For
the simplification of the case study the only external sources of cash flows will be the
ones mentioned above, copyright income and sales.
In the approximation of the patent’s future cash flow, IPR managers will pick up all
relevant information that influence their own business and generate income from
licensing (licensing costs and copyrights). Besides production and existing licensing
contracts, managers have to estimate the effects of their own production, possible future
contracts, their own action on the IPR and the customers; for example replacing an
existing patent with a new one. The main reason is to observe the net effects that are
generated by a patent, either independently or as part of a larger entity (a patent
portfolio).
In case there is an great lack of data coming from the markets, cash flows can be
built upon an expert’s opinion and taking into consideration two variables: it’s size and
time that passes.
 Application of the Fuzzy-Pay-Off Method 243

Uncertainties related to future cash flows have a considerable influence and shall
not be omitted. The approach on the estimation’s uncertainties is based on the gathering
of information by the evaluator, being supported by the manager, while also building
three future cash flow scenarios: optimist, pessimist and realist. During the process of
information gathering, managers can transmit their data as averages, a fact that can lead
to false results in the valuation process.
After the gathering of expert’s opinions on cash flows, the three situations are
obtained. They are represented in Tables 1 and 2.
Scenarios can be estimated at their present value using adequate discount rates,
these being bigger for the company’s obtained profits than on their various costs. For
each of the three scenarios the PV and NPV formula is applied according to the
financial diagnosis analysis. Thus, in order to obtain the results from Tables 1 and 2,
Eqs. 7 and 8 are applied [22].

X
n
1
PV ¼ ; ð7Þ
i¼1 ð1 þ r Þi

Table 1. The prediction of the PV of cash flow and the PV of the income resulting from patent
in three scenarios (optimistic, pessimistic, realistic) in consecutive temporal milestones: year 0, 1,
2, 3, 4 and 5 of production

moment (t) 0 1 2 3 4 5
Cost of cash flow
Optimistic scenario 3,000 3,000 1,000 1,500 1,500 3,000
Realistic scenario 3,500 4,000 1,200 1,550 1,550 3,200
Pessimistic scenario 6,000 6,500 3,000 1,800 1,800 4,000

PV of cash flow, discount rate=5%

Optimistic scenario 3,000 2,857.143 907.029 1,295.756 1,234.054 2,350.578
Realistic scenario 3,500 3,809.524 1,088.435 1,338.948 1,275.189 2,507.284
Pessimistic scenario 6,000 6,190.476 2,721.088 1,554.908 1,480.864 3,134.105

License fee+ income from copyright

Optimistic scenario 0 0 20,000 30,000 50,000 50,000
Realistic scenario 0 0 10,000 18,000 27,000 35,000
Pessimistic scenario 0 0 0 3,000 5,000 5,000

Competitive advantage resulting from patent

Optimistic scenario 0 0 1,000 1,000 1,000 1,000
Realistic scenario 0 0 500 500 500 500
Pessimistic scenario 0 0 0 0 0 0

PV of the income resulting from the patent, discount rate=20%

Optimistic scenario 0 0 14,583 17,940 24,595 20,496
Realistic scenario 0 0 7,292 10,706 13,262 14,267
Pessimistic scenario 0 0 0 1,736 2,411 2,009
244 M. Crăciun and L.M. Csorba

Table 2. The prediction of the PV of cash flow and the PV of the income resulting from patent
in three scenarios (optimistic, pessimistic, realistic) in consecutive temporal milestones: year 6, 7,
8, 9, and 10 of production

moment (t) 6 7 8 9 10
Cost of cash flow
Optimistic scenario 1,500 1,500 1,500 2,000 2,000
Realistic scenario 1,600 1,600 1,700 2,200 2,200
Pessimistic scenario 2,000 2,000 2,000 2,500 2,500

PV of cash flow, discount rate=5%

Optimistic scenario 1,119.323 1,066.022 1,015.259 1,289.218 1,227.827
Realistic scenario 1,193.945 1,137.090 1,150.627 1,418.140 1,350.609
Pessimistic scenario 1,492.431 1,421.363 1,353.679 1,611.522 1,534.783

License fee+ income from copyright

Optimistic scenario 50,000 50,000 50,000 50,000 50,000
Realistic scenario 35,000 35,000 35,000 35,000 35,000
Pessimistic scenario 5,000 7,500 7,500 7,500 7,500

Competitive advantage resulting from patent

Optimistic scenario 1,000 1,000 1,000 1,000 1,000
Realistic scenario 500 500 500 500 500
Pessimistic scenario 0 0 0 0 0

PV of the income resulting from the patent, discount rate=20%

Optimistic scenario 17,080 14,233 11,861 9,884 8,237
Realistic scenario 11,889 9,907 8,256 6,880 5,733
Pessimistic scenario 1,674 2,093 1,744 1,454 1,211

NPV patent
Optimistic scenario 121,546 β
Realistic scenario 68,423 a
Pessimistic scenario -14,162 α

where:
• i – the period expressed in years;
• n – the final year;
• r – the discount rate.

NPV ¼ PV of income  PV of expenses: ð8Þ

The next step in the calculation of the real option value is represented by the pay-off
distribution based on the analysis of the DCF. The three hypotheses listed in this
 Application of the Fuzzy-Pay-Off Method 245

section will be taken into consideration. Applying the work hypothesises simplified
data is obtained, yet this data however simplified, does not reduce the credibility of the
result Thus, the pay-off distribution is obtained based on the triangular representation of
the three scenarios.
The pay-off distribution is constructed as following:
• NPVs are calculated on the ‘i’ time-spans;
• The line chart is drawn based on the three situations.
The result is presented in Fig. 3.

Fig. 3. The cumulative PV based on the three scenarios of the patent

4 Results and Their Interpretation

The study has started from the premise of a company that wishes to patent an invention
so as to diminish production costs. The company has a limited production capability
yet it wishes to use the advantage brought by the patent. The most important source of
income comes to be the licensing of their patents – income from copyrights, selling the
right to use their invention to other producers in the same industry.
In the DCF’s construction, both costs and possible income revenues are predicted for
a 10 year long time-span, modelled in the three scenarios. In the income present value
(PV) section, the discount rate is estimated to be around 20% (see Tables 1 and 2).
Choosing this rate influences the valuation process significantly. In the cost PV section,
the assessment of the patent is based on the company’s experience with similar patents.
In this case the discount rate for the maintenance cost is estimated to be around 5% (see
Tables 1 and 2). The choice for this rate also influences the valuation process majorly.
To create a maximum value for the NPV distribution, from the present value of all
income, taken from the optimist scenario, the cost value from the same scenario is
subtracted (see formula 8). The same way will the NPV distribution for the other two
scenarios be calculated, the resulting values being the minimal one (for the pessimist
scenario) and the one close to real life (for the realist scenario).
246 M. Crăciun and L.M. Csorba

In Table 2 the following values for the cumulative PV were obtained:

• maxim NPV = 121,546 in the optimistic scenario;
• minim NPV = −14,162 in the pessimistic scenario;
• realistic NPV = 68,423 in the realistic scenario.
As a follow up, a triangular NPV representation is created for the scenarios:
1. the realist value is given the maximum credibility degree;
2. the critical values (minimal and maximum values) are given a credibility factor that
tends to 0.
Thus, values smaller or larger than the limits of the considered numerical interval
are not considered to be part of the possible values of the NPV. Values in between the
realist and the maximum NPV, respectively, the ones between the realist and the
minimum NPV, are granted an intermediate credibility degree, in the [0, 1] interval.
The result of the NPV represents the PO distribution, presented in Fig. 4.
Valuating the patent using the FPOM is brought to the calculation of the real option
value; respectively to the calculation of the mean possible value (MEAN). This step
represents the key to the assessment of a financial instrument.
The real option value is calculated using formula (3). As part of the formula a
multiplication between a ratio of two integrals and formula (6) is to be made.
Because work is done according to three situations, within the FPOM, triangular
calculations were applied, calculations based on triangular fuzzy numbers. Firstly the
minimal, maximal and realist values that make up the NPV will be correlated and
defined as points {a, a, b}, those points will define the pay-off distribution (see Fig. 1).
After the correlation the values that exist are:
• a = realist NPV = 68,423
• a = minimal NPV = −14,162
• b = maximal NPV = 121,546

Fig. 4. PO distribution of the patent

 Application of the Fuzzy-Pay-Off Method 247

With the help of these values, the integrals from formula (3) will be calculated like
areas of two triangles; furthermore, expression six can also be determined with these
values and multiplied with the real option value. The total area is the (b-a) base
multiplied by the triangle’s height (value 1), divided by 2. The 67,854 will be obtained.
The positive area is calculated as a subtraction between the total area and the negative
one. The calculation of the negative area asks for the determination of the two catheti,
using the fundamental triangle similarity criterion. Hence, the 0.17 value will be
obtained, the negative area being 1,214. The positive area’s value being 66,640, its
ratio with the total area is 0.98.
The mean possible value is obtained from formula (6). The branch upon which the
function runs will be the a-a > 0 case, more specifically the first branch. The MEAN
value will be 91,041. Multiplying the area ratio of 0.98 with the mean possible value of
91,041 the final real option value will be achieved at 89,411. The patent’s value is
equal to the real option value that was obtained, 89,411 measuring units. This value is
close to the MEAN = 91,041 measuring units.
The main strongpoint of the FPOM is represented by its capacity to create graphical
comparisons between different temporal moments, a fact that gives the method its
dynamicity (see Figs. 5, 6 and 7).
Beside its dynamicity and advantages, the FPOM also presents some
disadvantages.
Advantages of the FPOM:
• Comparison between different temporal moments of the patent’s value gives FPOM
its high credibility.
• FPOM is an intuitive method to present a patent’s value.
• FPOM is a direct method based on fuzzy mathematics.
• FPOM gives continuous support to the one making decisions, allowing for
graphical comparisons of PO distributions while offering a single real option value.
• Any later modification leads to the updating of the given results.
• FPOM is a fast analysis.
• FPOM opens new horizons to decision makers who are interested in patent valu-
ation, as it does not conflict with the classical FI valuation methods, being much
rather a complementary approach, using existing DCF data.
Disadvantages of the FPOM:
• FPOM can generate errors during the creation of scenarios.
• FPOM outputs a narrow vision of reality, thus it is recommended for it to be use as a
control method for the classical approaches, than as an independent method of
valuation.
248 M. Crăciun and L.M. Csorba

Fig. 5. Patent’s PO distribution in year 8

Fig. 6. Patent’s PO distribution in year 9

Fig. 7. Patent’s PO distribution in year 10

 Application of the Fuzzy-Pay-Off Method 249

5 Conclusions

In this paper the valuation of a FI was chosen, more precisely the invention patent, as it
is considered this is a less tread upon niche available to evaluators.
The valuation model that was outlined in this article was the complementary Fuzzy
Pay-Off Method, based on the construction of a NPV distribution using three scenarios:
an optimist one, a pessimist one and a realist one. This approach can be used as an
extension to the DCF method.
It can be stated the field of statistics gain ground in the valuation community, thus
the regression method is more and more used in its various trends: linear regressions,
curved ones, logarithmic and even exponential.
This paper delved deep in statistics, as the FPOM method is based on NPV dis-
tributions in three possible situations.
The final conclusion would be that FPOM opens new horizons to valuators who are
interested in the assessment of invention patents or/and the assessment of enterprises. It
does not conflict with classical FI valuation approaches, using existing DCF data from
companies, complementing the works of the valuator.

Acknowledgement. This paper would not have been written without the help of several col-
leagues who offered their knowledge and unconditional support. Also special thanks to the family
and best friends who have always been there for us.

References
1. ANEVAR: Standardele de evaluare ANEVAR (2016). ISBN 978-973-0-21517-5
2. Intrumente Financiare – Pieţe. http://www.bvb.ro/FinancialInstruments/Markets/
AdvancedSearch. Accessed 02 July 2016
3. Huian, M.C.: Instrumente financiare: tratamente şi opţiuni contabile. Editura CECCAR,
Bucureşti (2008)
4. Crăciun, M., Raţiu, C., Bucerzan, D., Manolescu, A.: Actuality of bankruptcy prediction
models used in decision support system. Int. J. Comput. Commun. Control 8(3), 363–371
(2013). ISSN 1841-9836 (ISI)
5. Radu, S.C., Boajă, D.M.: Pieţe de capital. Instituţii şi instrumente financiare tranzacţionate.
Editura Universitară (2012)
6. Standardele de evaluare ANEVAR, SEV 100 - Cadrul general (2016)
7. Matiș, D., Manațe, D., Anghel, I.: Evaluarea instrumentelor financiare. Editura RISOPRINT,
Cluj-Napoca (2013)
8. Giga-Investments: Modelling the Valuation of Very Large Industrial Real Investments.
https://mpra.ub.uni-muenchen.de/4328/1/MPRA_paper_4328.pdf. Accessed 11 June 2016
9. Models for Inexact Reasoning. http://www.dia.fi.upm.es/*mgremesal/MIR/slides/Lesson%
201%20%28Fuzzy%20Sets%29.pdf. Accessed 15 June 2016
10. The home of the pay-off method. http://payoffmethod.com/the_method/. Accessed 21 June
2016
11. Possibility Theory and Statistical Reasoning. http://citeseerx.ist.psu.edu/viewdoc/download?
doi=10.1.1.449.5602&rep=rep1&type=pdf. Accessed 30 June 2016
250 M. Crăciun and L.M. Csorba

12. Collan, M., Fullér, R., Mezei, J.: A fuzzy pay-off method for real option valuation. J. Appl.
Math. Decis. Sci. 2009, 14 (2009). doi:10.1155/2009/238196. Article ID: 238196, Hindawi
Publishing Corporation
13. Geman, O.: A fuzzy expert systems design for diagnosis of Parkinson’s disease. In: E-Health
and Bioengineering Conference (EHB). IEEE (2011)
14. Manaţe, D., Crăciun, M., Horváth-Balta, E., Horváth-Balta, I.: A literature review fuzzy
pay-off-method – a modern approach in valuation. J. Econ. Bus. Res. XXI(1), 98–107
(2015). ISSN 2068-3537, E-ISSN (online) 2069-9476, ISSN-L = 2068-3537
15. Numere fuzzy triunghiulare. http://andrei.clubcisco.ro/cursuri/5master/ciblf/curs_fuzzy_2.
pdf. Accessed 01 July 2016
16. Deaconu, A.: Diagnosticul şi evaluarea intreprinderii. Editura Intelcredo Deva (2000)
17. Monitorul Oficial nr. 613/19 august 2014. http://www.cdep.ro/pls/legis/legis_pck.lista_mof?
idp=23978. Accessed 05 June 2016
18. Legea contabilitatii actualizata 2016 - legea 82/1991. http://contabilul.manager.ro/legea-
contabilitatii-actualizata.html. Accessed 07 June 2016
19. Litterman-Hillmer, G., Kuckartz, M.: SME tailor-designed patent portofolio analysis. Word
Pat. Inf. 31(4), 269–360 (2009)
20. Collan, M.: A DSS for supporting area development decision-making. In: IEEE International
Conference on Information Management and Engineering, Chengdu, PRC (2010)
21. Carlsson, C., Fuller, R.: On possibilistic mean value and variance of fuzzy numbers. Fuzzy
Sets Syst. 122, 315–326 (2001). doi:10.1016/S0165-0114(00)00043-9
22. Richard, A., Myers, S.C.: Principles of Corporate Finance, 10th edn. The McGraw-Hill
Companies. https://lednichenkoolgashulman.files.wordpress.com/2015/11/principles-of-
corporate-finance-10th-edition.pdf. Accessed 15 July 2016
Fuzzy Applications, Theory and Fuzzy
and Control
 Fuzzy Logic Control Strategy for a Three-Phase
Electric Arc Furnace

Loredana Ghiormez1 ✉ , Octavian Prostean1, Ioan Filip1, and Manuela Panoiu2

( )

1
Faculty of Automation and Computers, Politehnica University of Timisoara,
str. B-dul Vasile Parvan, No. 2, 300223 Timisoara, Romania
{loredana.ghiormez, octavian.prostean}@upt.ro,
[email protected]
2
Faculty of Engineering Hunedoara, Politehnica University of Timisoara,
str. Revolutiei, No. 5, 331128 Hunedoara, Romania
[email protected]

Abstract. The electric arc furnace is a nonlinear load, this nonlinearity being
given by the electric arc. In this paper is performed a study in order to control the
current of a three phase electric arc furnace. The control strategy is carried out in
Matlab/Simulink and is used a model of the electric arc that was developed by
the authors of this paper. The type of the controller used in this paper is a fuzzy
logic one. This fuzzy system has the error and the change of error as inputs and
the sending command to the execution element as output. The system is tested in
two ways. The first case is when is modified the value of the reference parameter.
In order to follow this reference parameter the controller should send command
to the execution element so the length of the arc will be modified and be obtained
the value of the set point parameter. The second case is when injecting distur‐
bances at the length of the arc and the controller should reject these disturbances
so the reference current is obtained again. The response of this system, tested in
a closed-loop, illustrates that this proposed system has good dynamics, small
overshot, good answering time, follow precisely the value of the reference param‐
eter and reject well injected disturbances.

Keywords: Electric arc furnace · Fuzzy logic controller · Current control · Arc
length

1 Introduction

Electric arc furnace is an equipment used to melt the scarp in order to obtain liquid steel.
This steel is used in many domains of our days like alimentary, chemically, construc‐
tions, marine and automobile industries. Besides these, the old steel can be melted many
times and used again. In this paper is studied a three-phase electric arc furnace this having
three graphite electrodes that are used to melt the metal loaded in the furnace tank. The
electric arc furnace is supplied by a special transformer. In this paper is studied an electric
arc furnace that has the capacity of 120 tones and it is supplied by a transformer that can
have voltage ranging within 550–960 V in the secondary windings of the transformer.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_20
254 L. Ghiormez et al.

Electric arc furnace creates power quality problems [2–4] such as generation of
harmonic voltage and currents, voltage flickers and consuming reactive power. These
kinds of problems need to be rejected in the electric arc furnace. So, it is needed to
maintain constant the current of the electric arc. In order to do this, the length of the arc
should be maintained constant. In the melting process of an electric arc furnace, the
scrap is randomly disposed in the furnace tank. Also, when the scrap starts to melt the
other solid metal loaded in the furnace can take different positions in the furnace tank,
so current fluctuations appear. On the other hand, it is possible to obtain the breakage
of the electrodes when the scrap is taking another position in the furnace tank. This is
the reason for using automated control of the electrode positioning system.
In the reference literature are found several control strategies for an electric arc
furnace [1]. Generally, these control strategies are used to modify the position of the
electrodes in order to obtain a new length of the electric arc. The movements of the
electrodes are made by separate hydraulic actuators for each of the three-phase of the
furnace transformer.

2 The AC Electric Arc Modeling

The electric arc supplied with alternative current (AC) is a nonlinear element [3]. In
Fig. 1 are presented the waveforms for the electrical voltage of the supply voltage, vs,
of an electric arc furnace and the voltage, varc, and current, iarc, of the electric arc. It can
be observed that current and voltage of the electric arc are in the same phase therefore
the electric arc has a resistive character. After the ignition of the electric arc voltage, vig,
the voltage of the electric arc is approximately constantly and the current of the electric
arc is varying. This is the reason why, the electric arc is considered a nonlinear element
of circuit [3].
In relation (1) is presented the mathematical model of the electric arc that was devel‐
oped by the author of this paper in [5]. This model was design by dividing the voltage–
current characteristic of the electric arc in three functioning zones.
The model presented in (1) is based on the relation between the arc length, the arc
voltage and the arc current. In the model presented in (1) the parameters are the
following:
– i1 and −i1 are the values of the electric arc current which correspond for the ignition
voltage of the electric arc in the two semi period of the alternative current;
– R1 is the line slope of the voltage-current characteristic of the electric arc and has
value 0.09 in this paper;
– v is the voltage of the electric arc;
– i is the current of the electric arc;
– vex is the extinction voltage of the electric arc;
– C and D are constants that determine the difference between the zones of the voltage–
current characteristic when the current is increasing or decreasing.
 Fuzzy Logic Control Strategy 255

Fig. 1. Variation of the curves for current and voltage of the electric arc and for the supply voltage

⎧ i ⋅ R , i < i and di ≥ 0 or i > − -i and di < 0

⎪ 1 1 1
⎪ −vex + C , i < −i1 and di < 0 or di ≥ 0
v=⎨ D + |i| (1)
⎪ v + C , i > i and di < 0 or di ≥ 0
⎪ ex D + |i| 1
⎩

The value of i1 is computed as the ignition voltage divided to R1, as illustrated in (2).
/
i1 = vig R1 (2)

The extinction voltage depends by arc length [6] as presented in relation (3).

vex = A + B ⋅ l (3)

The ignition voltage is computed as illustrated in relation (4).

vig = vex + C∕D (4)

In (3) variable A is a constant that has the value 40 V and B has the value 10 V on
centimeters. Constant A represents the sum of the anode and cathode of the drop voltage
and constant B stands for the drop voltage of the unit length [6].
In this paper the initial length of the arc is considered of 31 cm. In order to obtain
the rectifying character of the electric arc were chosen different values for constants C
and D for the two semi period of the alternative current. When the derivative of the
256 L. Ghiormez et al.

electric arc current is increasing, C = 190,000 W and D = 5000 A. When the derivative
of the current is decreasing, C = 39,000 W and D = 5000 A.
Electric arc has a nonlinear characteristic therefore the electric arc furnace is a
nonlinear element [7, 8].

2.1 Fuzzy Logic Control System for the Electric Arc Furnace
Fuzzy logic controller is used especially for the systems that are nonlinear [9]. In this
paper the fuzzy logic controller has two inputs and one output. The two inputs stands
for the error of the current, this being the difference between the set point and the obtained
value of the current, and the derivative of the error. The output of the fuzzy logic
controller is the command that is sent to the execution element. In this paper the execu‐
tion element is assume to be expressed as a proportional of the first order with a time
constant. In real situations, the execution element is compound by hydraulic actuators
that modify the electrodes in order to obtain different length of the arc. If the length of
the arc is kept constant this imply that the current of the electric arc will be constant
too [10].
The membership functions for the three variables are triangular and each of the vari‐
ables has seven membership functions. The inference engine is min-max and the defuz‐
zification method is the centroid one.
The closed-loop system will be tested in two cases. First case is when the value of
the reference arc current is modified in a specific range and the second case is when
disturbances are injected in the process. For both of the cases is presented the response
of the system without changing the value of the reference parameter or without injecting
disturbances in the process.

2.2 Tracking the Reference

In Fig. 2 is presented the block diagram of the closed-loop control system. The reference
arc current is set at 56.5 kA. This block diagram is tested first of all when the reference
arc current is set to a specific value and the fuzzy logic controller should send commands
to the execution element in order to be modified the length of the electric arc, so, reaching
the desired value and remain in the stable area. Then, this system is tested by modifying
the value of the reference arc current by a slider gain in order to see if the system can
follow this value by modifying the length of the arc each time that a new value is set.
Slider Gain block has initial value of 1 meaning that the value of the reference parameter
is multiplied by 1. This situation corresponds in the real situation when for the same
voltage from the secondary windings of the furnace transformer, are used different
lengths of the electric arc that are used to melt the scarp loaded in the furnace tank.
 Fuzzy Logic Control Strategy 257

Fig. 2. The block diagram of the closed-loop control system for the case when is modified the
value of the reference parameter

Fig. 3. The response of the closed-loop control system when is set a reference arc current
258 L. Ghiormez et al.

In Fig. 2 ARC block stand for the model of the electric arc that was presented in (1).
Rt is the source resistance and Rfc is the flexible cables resistance. Lt is the inductivity
of the source and Lfc is the inductivity of the flexible cables. These flexible cables are
used to connect the transformer to the electric arc furnace. The electric arc current has
a sinusoidal shape. In order to compare the reference arc current with the obtain arc
current RMS block is used that will return the effective value of electric arc current. EE
is the execution element that has a time constant of 0.75 s.
In Fig. 3 is presented the response of the system tested in the closed-loop control
system and is set a value of the reference parameter of 56.5 kA. One can notice that the
response of the system, which is the current of the electric arc, reaches the value of the
reference arc current and also reaches the stable state. In Fig. 2 was set the sample time
for the sinusoidal signal generated by the source block at 10,000 samples.
In Fig. 4 is illustrated the response of the system when is modified the reference arc
current. In real situation this corresponds to work with different lengths of the arc for
the same voltage from the secondary windings of the furnace transformer. One can notice
that the response of the system follows the value of the reference arc parameters which
is the current of the electric arc. This value is obtained in a small period of time and it
is very close to the reference value.

Fig. 4. The response of the closed-loop control system when is modifying value of the reference
parameter by a slider gain

In Fig. 5 is presented the variation for the lengths of the electric arc for the case
illustrated in Fig. 4. One can notice that when the current has to increase, the length of
 Fuzzy Logic Control Strategy 259

the arc decreases. Also, the length of the arc increases when the current of the electric
arc has to decrease.

Fig. 5. The variations of the length of the arc when the reference arc current is modified by a
slider gain

In Fig. 6 is presented the voltage-current characteristic obtained for the case when
modifying the reference arc current as presented in Fig. 4. One can notice that it is
obtained the dynamic characteristic of the electric arc because the electric arc has
different lengths.

2.3 Disturbances Rejection

The second case for testing of the closed-loop system is illustrated in Fig. 7 and corre‐
sponds for the case when are injected disturbances in the process, i.e., modifying the
length of the electric arc that will correspond in real situation with the moving of the
scrap in the furnace tank and obtaining different configuration of the scrap. In order to
do this, a Slider Gain is also used. This block has value zero that corresponds for the
initial arc length which is of 31 cm. Injecting disturbances in the process corresponds
in simulation for the length of the arc that can be modified within 10 cm up or down.
260 L. Ghiormez et al.

Fig. 6. The dynamical voltage-current characteristic of the electric arc for the case when the
reference arc current is modified by a slider gain

Fig. 7. The block diagram of the closed-loop control system for the case when are injected
disturbances in the process

In Fig. 8 it is presented the response of the system tested in the closed-loop system
and is set a value for the arc length of 31 cm and a value of electric arc current of 56.5 kA.
 Fuzzy Logic Control Strategy 261

One can notice that the response of the system, which is the current of the electric arc,
reaches the value of the reference arc current and also reaches the stable state. In Fig. 7
was set the sample time for the sinusoidal signal generated by the source block at 10,000
samples.

Fig. 8. The response of the closed-loop control system when is set a reference arc length

In Fig. 9 is presented the variation of the arc length for the case when are injected
disturbances in the process. One can notice that when the current has to increase, the
length of the arc decreases. Also, the length of the arc increases when the current has to
decrease.
In Fig. 10 is illustrated the response of the system for the case presented in Fig. 9.
In real situation this corresponds to the randomly configuration of the scarp in the furnace
tank, this scrap modifying its position until the solid metal is converted into liquid steel.
One can notice that the response of the system follow the value of the reference arc
parameters which is the current of the electric arc. This value is obtained in a small
period of time and it is very close to the reference value.
262 L. Ghiormez et al.

Fig. 9. The variations of the length of the arc when are injected disturbances by a slider gain

Fig. 10. The response of the closed-loop control system when are injected disturbances in the
process by a slider gain

In Fig. 11 is presented the voltage–current characteristic obtained for the case when
are injected disturbances in the process as presented in Fig. 9. One can notice that it is
obtained the dynamic voltage–current characteristic of the electric arc because the elec‐
tric arc has different lengths.
 Fuzzy Logic Control Strategy 263

Fig. 11. The dynamical voltage-current characteristic of the electric arc for the case when are
injected disturbances in the process by a slider gain

3 Conclusions

This research presented a fuzzy logic controller used to maintain constant the current of
the electric arc. In order to do this the fuzzy logic control system has as inputs the
deviation of electric arc current and the derivative of this error. The output stands for
the command sent to the execution element in order to modify the length of the electric
arc so obtaining the desired value of the reference parameter. To test the proposed closed-
loop control system, a model of the electric arc developed by the author of this paper
was used. This model is a hyperbolic model and it is based on the length of the arc, the
voltage of the electric arc and the current of the electric arc. The closed-loop control
system is tested in two ways. First case is when is modified the reference value of the
electric arc current and the second case corresponds for the case when are injected
disturbances in the process. Obtained results illustrates that the proposed control block
diagrams are good ones because the system has small overshoot, good answering time
and reach the stable state for both presented cases.

Acknowledgements. This work was developed through the Partnerships in Priority Areas - PN
II, with the support of ANCS, UEFISCDI, Project No. 36/2012, Code: PN-II-PT-
PCCA-2011-3.2-1519.
264 L. Ghiormez et al.

References

1. Hauksdottir, A.S., Gestsson, A., Vesteinsson, A.: Current control of a three-phase submerged
arc ferrosilicon furnace. Control Eng. Pract. 10(4), 457–463 (2002)
2. Hooshmand, R., Banejad, M., Esfahani, M.T.: A new time domain model for electric arc
furnace. J. Electr. Eng. 59(4), 195–202 (2008)
3. Panoiu, M., Panoiu, C., Sora, I., Osaci, M.: About the possibility of power controlling in the
three-phase electric arc furnaces using PSCAD EMTDC simulation program. Adv. Electr.
Comput. Eng. 7(1), 38–43 (2007)
4. Panoiu, M., Panoiu, C., Ghiormez, L.: Modeling of the electric arc behaviour of the electric
arc furnace. In: Proceedings of the 5th International Workshop on Soft Computing
Applications, Szeged, Ungaria, pp. 261–271 (2012)
5. Ghiormez, L., Panoiu, M., Panoiu, C.: Harmonics analysis of the 3-phase electric arc furnace
using models of the electric arc. In: Proceedings of the 3rd World Conference on Information
Technology - WCIT-2012, Barcelon, Spain, pp. 424–430 (2012)
6. Panoiu, M., Panoiu, C., Sora, I.: Modeling of three phase electric arc furnaces. Acta Electroteh.
48(2), 124–132 (2007)
7. Hui, Z., Fa-Zheng, C., Zhuo-qun, Z.: Study about the methods of electrodes motion control
in the EAF based on intelligent control. In: International Conference on Computer,
Mechatronics, Control and Electronic Engineering (CMCE), vol. 4, pp. 68–71. IEEE (2010)
8. Yan, L.I., et al.: Model predictive control synthesis approach of electrode regulator system
for electric arc furnace. J. Iron. Steel Res. Int. 18(11), 20–25 (2011)
9. Panoiu, M., Ghiormez, L., Panoiu, C.: Adaptive neuro-fuzzy system for current prediction in
electric arc furnaces. In: Soft Computing Applications, Advances in Intelligent Systems and
Computing, vol. 356, pp. 423–437. Springer, New York (2016)
10. Moghadasian, M., Alenasser, E.: Modelling and artificial intelligence-based control of
electrode system for an electric arc furnace. J. Electromagn. Anal. Appl. 3, 47–55 (2011)
 On Fuzzy Estimation of Channel Noise
Characteristics in MIMO Communication
Systems

V. Nicolau(&) and M. Andrei

Department of Electronics and Telecommunications,

“Dunarea de Jos” University of Galati,
47 Domneasca Street, 800008 Galati, Romania
{viorel.nicolau,mihaela.andrei}@ugal.ro

Abstract. In wireless communications, the signals are affected by channel

fading and additive noise generated by different Gaussian or non-Gaussian noise
sources. In practical situations, channel characteristics are only estimated with
more or less uncertainties, depending on the signal-to-noise ratio (SNR) of the
received signals. The noise characteristics, like power level, and distribution
parameters, are important for management system, regarding either transmitter
or receiver state. In this paper, fuzzy logic approach is used for power level
estimation of channel noise, based on a much smaller subset from the set of
estimated values of the received symbols determined by decoder. Some expert
rules are generated, and a fuzzy structure is proposed for SNR estimation.
Several fuzzy systems were tested, using different sets of membership functions
and rule bases. Simulation results and the output performances are presented.
Accurate estimation of SNR is obtained with lower computational complexity.

Keywords: Fuzzy logic
Estimation
Channel noise
Wireless

communication

1 Introduction

In general, communications through wireless channels are affected by time-varying

characteristics of propagation environment. A signal sent from a transmitter traverses
many different paths to its destination, so that multiple versions of the transmitted
signal reach the receiver [1]. On each path, a signal fading process is present, in which
the signal experiences different attenuations, time delays, and phase shifts on one or
more frequency components. The number of paths and their characteristics can change
in time [2]. In addition, the signals are affected by additive noise generated by different
Gaussian or non-Gaussian noise sources.
Orthogonal space-time block codes (OSTBC) are successfully used to exploit
spatial diversity in MIMO communication systems. These techniques provide full
diversity and reduce the complexity of the optimal maximum-likelihood (ML) decoder
[3]. However, for communications with high-data and low-error rates, at the receiver
the decoder must exactly know the full channel state information (CSI).

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_21
266 V. Nicolau and M. Andrei

In practical situations, the channel characteristics are changing frequently, and CSI
is only estimated with more or less uncertainties, depending on the signal-to-noise ratio
(SNR) of the received signals. There are two main groups of CSI estimation tech-
niques: pilot-aided estimation, and blind estimation.
In first case, CSI can be estimated using training based methods, but only with
penalty in performance. Supplemental training signals must be transmitted, as pilot
sequence or pilot symbols inserted into the data sequence [4].
The blind channel estimation techniques identify channel characteristics using only
received noisy signals, based on their second-order stationary statistics. Many blind or
semi blind channel estimation techniques were intensively studied in recent years
[3–7]. Noise level has direct influence on the channel estimation performance.
Low SNR generates big uncertainties in CSI, and decoder must know how to deal with
this. Several aspects of low SNR are treated in [8–10].
Low SNR can result from low-power regime of the transmitter, or high level of
noise in the communication channel. Hence, the noise characteristics, like power level,
and distribution parameters, are important for management system in data communi-
cations, regarding either transmitter or receiver state, especially for large data trans-
missions at high speed and low error probability.
Fuzzy logic has been widely used in MIMO communication channels. A survey of
fuzzy logic application in wireless communications is presented in [11]. Regarding
fuzzy estimation of channel characteristics, fading channel was studied in [12], and
channel estimation was investigated in [13]. Fuzzy estimation of SNR was studied for
both pilot-aided and blind estimation [14, 15].
In this paper, fuzzy logic approach is used for power level estimation of channel
noise, based on a much smaller subset from the set of estimated values of the received
symbols determined by decoder. Expert rules are generated starting from noise influ-
ence on the estimated symbols, and a fuzzy structure is proposed for SNR estimation.
Several fuzzy systems were generated, using different sets of membership functions
and rule bases. Simulation results and the output performances are presented. The
selection was made based on the mean square error and mean absolute error of the
fuzzy output. The output surface of the selected fuzzy estimator is also illustrated.
Accurate estimation of SNR is obtained with lower computational complexity.
The paper is organized as follows. Mathematical models are presented in Sect. 2. In
Sect. 3, some expert rules are generated based on the noise influence on the estimated
values of received symbols. The structure of the fuzzy system is detailed in Sect. 4.
Simulation results are presented in Sect. 5. Conclusions are pointed out in Sect. 6.

2 Mathematical Models

Consider a MIMO communication system with NT transmitting and NR receiving

antennas, as shown in Fig. 1. Data transmission is based on Alamouti space–time block
code (STBC) with binary phase-shift keying (BPSK) modulation.
The signals submitted by the NT antennas during a symbol period are denoted
xi, i = 1…NT, where i index represents the signal emitted by the i antenna. These
signals form the column vector x of size [NT, 1]:
 On Fuzzy Estimation of Channel Noise Characteristics 267

h11
h21 h12
x1 r1
h22
x2 hNr2 r2
Space- Space-
h2Nt
time h1Nt time
hNr1
encoder decoder
hNrNt
xNt rNr

Fig. 1. MIMO communication system with space-time block coding

x ¼ ½x1 ; x2 . . .; xNT T ð1Þ

The signals received by the NR antennas during a symbol period form the column
vector r of size [NR, 1]:

r ¼ ½r1 ; r2 . . .; rNR T ð2Þ

The MIMO channel, considered without memory and with flat fading, is modeled
by the channel matrix H, of size [NR, NT]. It contains the channel fading coefficients,
hjk, between the broadcasting antenna k and the receiving antenna j. For Rayleigh
slow-fading channel, which is considered in this paper, the fading coefficients are
constant during the transmission of an entire frame with L symbols, but they can
change from frame to frame. These coefficients are considered complex Gaussian
random variables, containing independent random variables for real and complex part,
with zero-mean identical distributions.
The linear input-output relationship of the MIMO channel is:

r ¼ H
x þ n; ð3Þ

where n is the column vector of additive noise produced by different Gaussian or

non-Gaussian noise sources, which affects the signals at the receiving antennas:

T
n ¼ g1 ; g2 . . . gN R ð4Þ

Additive White Gaussian Noise (AWGN) noise type is considered in the paper.
At every t moment of time, at the receiving antenna j, the received signal, denoted
rjt , is linear combination of all signals, with fading and added channel-noise:

X
NT
rjt ¼ htji
xti þ gtj ; ð5Þ
i¼1
268 V. Nicolau and M. Andrei

where gtj is circularly symmetric complex Gaussian noise, with zero-mean and r2
variance in the two time slots.
In space-time coding, for every symbol in the data sequence, each transmitting
antenna transmits a different version of the same input, generated by the space-time
encoder. During the transmission time of one symbol, T, the space-time encoder
generates NT modulated complex symbols, which form the column vector x, as input to
the MIMO wireless communication channel.
In this paper, Alamouti space–time block code with NT = NR = 2 is used. The
Alamouti encoder structure is shown in Fig. 2.

Encoder x1=[x1 -x2*]

[x1 x2]
Source Modulator x1 -x2*
[x1 x2] X=
x2 x1*
x2=[x2 x1*]

Fig. 2. Alamouti space-time block encoder

Encoder outputs are transmitted during two consecutive symbol periods through the
two transmitting antennas, denoted in this case Tx1 and Tx2. For each encoding
operation, the encoder takes the group of two modulated complex symbols [x1 x2], and
generates the codeword matrix X:

ð6Þ

A column of X matrix represents broadcasted symbols by all antennas during j time

slot, while i line contains the sequentially broadcasted symbols by i antenna.
For example, in the case of quadrature phase-shift keying (QPSK) modulation, the
modulator gets at the input two bits of information and produces two complex symbols
[x1 x2] into the constellation C = {1, j, −1, −j}, with xi 2 C. If BPSK modulation is
used, then constellation has real points, C = {1, −1}, and xi 2 C. Then, this symbols
are used to generate the codeword matrix X.
The signals received by j antenna at t and t + T time slots are rj,1 and rj,2:
(
rj;1 ¼ hj;1
x1 þ hj;2
x2 þ gj;1
ð7Þ
rj;2 ¼ hj;1
x2 þ hj;2
x1 þ gj;2

It is assumed that the receiver decoder estimates the maximum plausible sequence
of information transmitted. At the reception, squared Euclidean distance is used
between presumably received sequence and the received one:
 On Fuzzy Estimation of Channel Noise Characteristics 269

 2
XX nR  
 t X t i
nT
rj  hj;i xt  ð8Þ
t j¼1
 i¼1


The decoder uses a maximum plausibility algorithm, selecting the most likely look
symbols ^x1 and ^x2 . Considering a source of information without memory, x2 and x1
modulated symbols are independent to each other. Hence, it is possible separate
decoding of the two symbols:
2 !3
PNR     2 2

~rj;1  hj;1  þ hj;2 
^x1  7
2
  6 arg min
^x1 6 ^x1 2S 7
^x ¼ 6
¼6
j¼1
!7 ð9Þ
^x2 PR    2  2  7
4 N
2 5
arg min ~rj;2  hj;1  þ hj;2 
^x2 
^x2 2S j¼1

3 Fuzzy Rules of Channel Noise Influence on Received

Symbols

In this section, some expert rules are generated based on the noise influence on the
estimated values of received symbols. A MIMO communication system with two
transmitting and two receiving antennas is considered, over Rayleigh flat-fading
channel. Data transmission is based on Alamouti space–time block code with BPSK
modulation, having the constellation C with two real points, C = {1, −1}. The data
sequence transmitted by MIMO system contains N = 104 random bits of 0 and 1 values
with equal probability.
At the receiver, the received signals are affected by channel noise. Hence, the
decoder determines the estimated values of the received symbols, which are scattered in
the complex plane around the real values of the constellation points (centers). The noise
type and its power determine the spreading effect of the estimated symbols in the
complex plane. Inversely, by knowing the spreading effect of the noise and analyzing
the estimated symbol set, information about noise type and power can be obtained.
In decoding process, the channel coefficients (elements of H matrix) should be
known by the receiver, but in most cases they are only estimated. Knowing these
coefficients with uncertainties generates more spreading effects on the estimated
symbols. Therefore, the influence of the channel noise on the estimated received
symbols is studied separately for noise-free and noisy cases.
In the noise-free case, the transmission is considered to be affected only by the
fading process, in the absence of any Gaussian or non-Gaussian noise sources. If the
fading channel coefficients are perfectly known, then the symbols are estimated by the
decoder without any uncertainty, being assigned to the one of the constellation center.
If elements of H matrix are known with uncertainties, the estimated symbols are
scattered in the complex plane around the constellation points, forming point clouds
270 V. Nicolau and M. Andrei

around these centers, as illustrated in Fig. 3. In the absence of the noise, two situations
are considered at the receiver, knowing the H matrix with 25% uncertainty and 50%,
respectively.

Fig. 3. Noise-free estimated symbols and H matrix known with 25% and 50% uncertainties

The estimated symbols are represented as points with grey color in the complex
plane. The ordinate axe represents the decision boundary in estimating process at the
receiver. It can be observed that the more precisely is known the H matrix, the more
compact are the estimated points around real values of constellation centers, which are
denoted c0 and c1 in the negative and positive half-plane. Even for high grade of
uncertainties, the decoder works well with no receiving errors.
In the noisy case, the data transmission is affected by the fading process and
additive noise produced by different Gaussian or non-Gaussian noise sources. Different
situations are considered with transmission affected by Additive White Gaussian Noise
(AWGN) and impulsive noise. To illustrate the spreading effect of estimated symbols
due to channel noise, the channel coefficients are considered to be known without
uncertainty.
If data transmission is affected by AWGN, the noise power has direct influence on
estimated values of received symbols and on their spreading in the complex plane. Two
situations are illustrated in Fig. 4, for two SNR values: SNR1 = 10 dB and SNR2 =
7 dB, respectively. The clouds of points which are placed around real values of
constellation centers are well separated only for high SNR values.
In the presence of AWGN, estimation errors appear at the decoder stage. The wrong
classified points are placed around decision boundary, being represented with black
dots in the figure. The error numbers, Nerr1 = 14 and Nerr2 = 62, correspond to the bit
error rate (BER) of BER1 = 0.0014 and BER2 = 0.0062, respectively.
It can be observed that as noise power increases and SNR decreases, the estimated
points are more dispersed in the complex plane and the estimation process is more
difficult. Even for high SNR values the spreading effect is bigger and the noise
influence is more evident than fading influence.
Rule No. 1: The noise influence on the spreading effect of estimated symbols is
more evident than fading influence, even for high SNR values.
 On Fuzzy Estimation of Channel Noise Characteristics 271

Fig. 4. Estimated symbols for AWGN noise and for two SNR values: 10 and 7 dB

In general, the distance between two points, x and y, into a d-dimensional space can
be defined as general Minkovski distance:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u d
u X
dðx; yÞ ¼ t
p
jxi  yi jp ; x; y 2 <d ; ð10Þ
i¼1

where for p = 2 Euclidean distance is obtained, which is considered in this paper.

The distance from every estimated symbol ^xi in the complex plane to the nearest
center ck is denoted di:

di ¼ minðd ð^xi ; c0 Þ; d ð^xi ; c1 ÞÞ ð11Þ

These distances generate a new random series, S, with N positive elements, which
characterizes the spreading effect of the points in the complex plane:

S ¼ fdi jdi  0; i ¼ 1. . .Ng ð12Þ

All the estimated points placed closed to one center have small distance value. These
points put no problem to the decoder and they give no additional information about the
spreading effect of the points in the complex plane. Therefore, all the points with
distance to the nearest center smaller or equal than a threshold, denoted p, can be
removed from the new series. In addition, by removing these points, the new series is
much smaller, and computational complexity of the proposed algorithm for noise
estimation heavily decreases.
The new subset of S is:

SS ¼ fdi jdi [ p; i ¼ 1. . .NS ; NS \ Ng; SS  S ð13Þ

Rule No. 2: The spreading effect in the complex plane can be analyzed based on a
smaller subset of the estimated points, those with the distance to the nearest center
bigger than the chosen p threshold.
272 V. Nicolau and M. Andrei

Fig. 5. Estimated symbols with distance bigger than the threshold

The threshold can be a fixed chosen value or can be an adaptive parameter based on
previous noise estimations. The fixed threshold can be chosen as a fraction of the
distance between centers and decision boundary:

p ¼ w
d ðc0 ; 0Þ ¼ w
d ðc0 ; c1 Þ=2; w 2 ð0; 1Þ; ð14Þ

where w is the weighting factor. In this paper, w = 0.5 was chosen.

For example, for the two previous situations affected by AWGN (with SNR 10 and
7 dB, respectively), the estimated symbols with the distance to the nearest center bigger
than the threshold are illustrated in Fig. 5. Their distance subsets, denoted SS1 and SS2,
have NS1 = 1409 and NS2 = 3426 elements.
The rest of the elements, which are the most of the estimated points in the complex
plane, are placed inside the threshold distance, closer to their corresponding centers.
As noise power increases, there are points in the complex plane which are much
closer to the wrong center, being wrong classified by the decoder. For some, the
distance to the wrong center is even smaller than the threshold. Two such points appear
for SNR = 7 dB (figure right) in the negative half plane. They were drawn in the figure
for supporting the idea, but they are not included in the subset. The random distance
points of the two subsets, SS1 and SS2, are illustrated in Fig. 6, and their statistical
characteristics are represented in Table 1. For bigger noise level, there are more esti-
mated points which are spread beyond the p threshold (NS2 > NS1), and in addition, the
spreading effect is larger.

Fig. 6. Estimated symbols with distance bigger than the threshold

 On Fuzzy Estimation of Channel Noise Characteristics 273

Rule No. 3: The power level of the channel noise can be characterized by the
number of elements and statistical characteristics of the subset, which is generated by
the estimated symbols with distance to the nearest center bigger than threshold.
As a result, two common-sense expert rules can be derived.
Rule No. 4: If noise power increases then the number of subset elements, NS, also
increases.
Rule No. 5: If noise power increases then the mean (l), and standard deviation (r)
of distance elements in the subset increase.
For the fuzzy estimator of channel noise, two input fuzzy variable are generated
taking into account the number of subset elements and their standard deviation.
The influence of the value ranges is reduced if relative input variables are con-
sidered. Therefore, the number of elements is weighted with the number of data
sequence transmitted by MIMO system, representing the rate of points, RP, with dis-
tance to the nearest center greater than threshold:

NS
Rp ¼ ¼ ; Rp 2 ½0; 1 ð15Þ
N
For the second fuzzy variable, as the mean of SS is not near zero value, coefficient of
variation, CV, also known as relative standard deviation, can be computed as the ratio
of the standard deviation (r) to the mean (l):
r
CV ¼ ¼ ; CV 2 ½0; 1 ð16Þ
l

4 Fuzzy System for Channel Noise Estimation

In this section, a fuzzy estimator of channel noise (FECN) is proposed to characterize

the noise of the communication channel. It is a Mamdani-type fuzzy system, with two
inputs and one output. The structure of FECN is shown in Fig. 7. The inputs give
information about the random distance points of the subset obtained from estimated
symbols: rate of points (RP) and coefficient of variation (CV).

Fig. 7. Structure of fuzzy system for channel noise estimation

274 V. Nicolau and M. Andrei

The membership functions of FECN inputs are illustrated in Fig. 8. Their limits are
chosen based on ranges of relative input variables.

Fig. 8. Membership functions of FECN inputs

The fuzzy output estimates the SNR into a range of practical values. It has five
membership functions, represented in Fig. 9. The membership functions of inputs and
output are denoted: Very-Small (VS), Small (S), Medium-Small (MS), Medium (M),
Medium-Big (MB), Big (B), and Very-Big (VB).

Fig. 9. Membership functions of FECN output

The knowledge base characterizes the channel noise based on fuzzy inputs. It was
generated using expert rules deduced in previous section. The complete rule base has
35 fuzzy rules, an example being represented in Table 2.

Table 1. Statistical characteristics of the random distance subsets SS1 and SS2
SNR (dB) NS N = 104 estimated symbols, p = 0.35355
l = mean of SS r = standard deviation of SS
10 1409 0.45995 0.11031
7 3426 0.51728 0.15706
 On Fuzzy Estimation of Channel Noise Characteristics 275

Table 2. Rule base example of FECN

Rate of points Coefficient of variation
VS S M B VB
VS VB B B MB MB
S B MB MB MB MB
MS MB MB MB MS MS
M MS MS MS S S
MB S S S S VS
B S S S S VS
VB S S S VS VS

5 Simulation Results

For simulations, a MIMO communication system with 2 transmitting and 2 receiving

antennas is considered over Rayleigh flat-fading channel. Data transmission is based on
Alamouti space–time block code with BPSK modulation. The transmitted data
sequence contains N = 104 random bits of 0 and 1 values with equal probability. In this
paper, the transmission is considered to be affected by AWGN with different levels of
noise power.
Simulations are run for six different values of SNR: 3, 5, 7, 10, 12, and 15 dB. For
every SNR value, the new subset SS used for fuzzy input generation has different
number of elements NS, which is represented in Table 3. It can be observed that as the
noise power increases (and SNR decreases), then the number of subset elements also
increases.

Table 3. Number of elements NS of the subset SS

SNR (dB) 3 5 7 10 12 15
NS 6249 4873 3426 1409 630 119

For every noise power level, the fuzzy system estimates the SNR, and the result is
compared with real noise level. Different FECN structures have been tested, with
various numbers and shapes of membership functions. The selection was made based
on the mean square error (mse) and mean absolute error (mae) of the fuzzy output,
compared with the real SNR. The performance indices of the best four fuzzy systems
are represented in Table 4.

Table 4. Performance indices of the best 4 fuzzy systems

FECN1 FECN2 FECN3 FECN4
mse 0.6736 0.8890 1.4511 5.8438
mae 0.6983 0.7817 0.7967 1.6401
276 V. Nicolau and M. Andrei

Output surfaces of the first two selected fuzzy estimators are illustrated in Fig. 10.

Fig. 10. Output surfaces of the first two selected fuzzy estimators

6 Conclusions

In this paper, fuzzy logic approach is used for power level estimation of channel noise,
based on a smaller subset from the set of estimated values of the received symbols.
Expert rules are generated, and a fuzzy structure is proposed for SNR estimation.
Several fuzzy systems were tested, using different sets of membership functions and
rule bases. Accurate estimation of SNR is obtained with lower computational
complexity.

References
1. Pottie, G.J.: System design issues in personal communications. IEEE Pers. Commun. Mag. 2
(5), 50–67 (1995)
2. Foschini Jr., G.J., Gans, M.J.: On limits of wireless comm. in a fading environment when
using multiple antennas. Wirel. Pers. Commun. 6, 311–335 (1998)
3. Via, J., Santamaria, I.: Correlation matching approaches for blind OSTBC channel
estimation. IEEE Trans. Signal Process. 56(12), 5950–5961 (2008)
4. Mahdi, H.F., Ismail, M., Nordin, R.: Literature survey and comprehensive study on channel
estimation of Wimax system. Int. J. Adv. Comput. Technol. (IJACT) 5(15), 31–47 (2013)
5. Zhou, L., Zhang, J.-K., Wong, K.M.: A novel signaling scheme for blind unique
identification of Alamouti space-time block-coded channel. IEEE Trans. Signal Process. 55
(6), 2570–2582 (2007)
6. Zaier, A., Bouallegue, R.: Blind channel estimation enhancement for MIMO-OFDM systems
under high mobility conditions. Int. J. Wirel. Mob. Netw. (IJWMN) 4(1), 207–214 (2012)
7. Kumar, R., Saxena, R.: MIMO capacity analysis using adaptive semi blind channel
estimation with modified precoder and decoder for time varying spatial channel. Int. J. Inf.
Technol. Comput. Sci. 10, 1–19 (2012)
 On Fuzzy Estimation of Channel Noise Characteristics 277

8. Beko, M., Xavier, J., Barroso, V.A.N.: Further results on the capacity and error probability
analysis of noncoherent MIMO systems in the low SNR regime. IEEE Trans. Signal Process.
56(7), 2915–2930 (2008)
9. Lozano, A., Tulino, A., Verdu, S.: Multiple-antenna capacity in the low-power regime. IEEE
Trans. Inf. Theory 49(10), 2527–2544 (2003)
10. Rao, C., Hassibi, B.: Analysis of multiple-antenna wireless links at low SNR. IEEE Trans.
Inf. Theory 50(9), 2123–2130 (2004)
11. Erman M., Mohammed A., Rakus-Andersson E.: Fuzzy logic applications in wireless
communications. In: Proceedings of European Society for Fuzzy Logic and Technology
(IFSA/EUSFLAT), pp. 763–767 (2009)
12. Niranjan, M., Sinhal, A.: Fuzzy modeling for quality estimation of fading channel. Int.
J. Comput. Appl. 51(18), 11–16 (2012)
13. Byun, Y.B., Takama, Y., Hirota, K.: Combined channel estimation and data decoding based
on fuzzy logic. IEEE Trans. Instrum. Meas. 51(2), 342–346 (2002)
14. Sastry, K.S.: Fuzzy logic based adaptive modulation using non data aided SNR estimation
for OFDM system. Int. J. Eng. Sci. Technol. 2(6), 2384–2392 (2010)
 Fuzzy-Expert System for Investments in Winemaking

Marius M. Balas ✉ , Emanuela V. Muller, and Ramona Lile

( )

“Aurel Vlaicu” University of Arad, Arad, Romania

[email protected], [email protected], [email protected]

Abstract. The paper presents an expert system dedicated to investment decisions

in winemaking. The system was implemented in two versions: a Matlab fuzzy
inference system, which is illustrated for the calculus of the feasibility and a
standalone application illustrated for the calculus of the costs, which was devel‐
oped with the help of a new visual fuzzy-expert system development shell,
conceived in Aurel Vlaicu University.

Keywords: Fuzzy-expert system · Decision making · Investment feasibility ·

Winemaking

1 Introduction

Decision making is a cognitive process that selects a particular course of an action among
different alternatives. Decision making is usually a complex reasoning process which
may be based on explicit or tacit assumptions and may have rational and irrational sides,
with a heavy psychological load [1]. This means that decision making is very often
affected by uncertainty. Wrong investment decisions may be very costly and human
solutions are often producing controversies, so in spite of difficulties, rigorous computer
support is very necessary.
The statistic calculus that is largely used in economic decisions does not offer precise
representations for the imprecise knowledge involved in investment decision making.
Statistics rely on large amount of data, which are usual missing in the majority of
investment decisions from many causes: such decisions are rare, non-repetitive, context
related, secret, etc. The typical mathematic tool addressing uncertainty in such cases is
the fuzzy logic, by means of the fuzzy-expert system.
Expert systems are intended to operate the same way as human experts and the if-
then rules of their rule base may be formulated by specialists without IT skills.
Expert system is a mature concept dating from ’70s [2–4] and the fuzzy version
issued in ’80s [5–7]. New significant fundamental ideas are difficult to imagine in this
matter, so our interest is focused on two tenure research items:
(a) New applicative domains: we approached winemaking, on behalf of the eight
centuries of experience issued from the vineyards of our region, Arad, Romania;
(b) The refinement of the software tools: besides the traditional Fuzzy Inference System
Tool-kit of Matlab we successfully experienced a new visual fuzzy-expert system
development shell, programmed in our university.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_22
 Fuzzy-Expert System for Investments in Winemaking 279

2 The Role of Expert Systems in Economics

Economics is a social science studying how choices are made on allocating scarce
sources to satisfy unlimited wants. More often, when a choice has to be made, it is more
likely not to consider all aspects regarding a capital allocation decision or to foresee all
possibilities of certain risks, but instead to use a holistic view and qualitative judgements.
The best tool to solve these problems is the fuzzy logic [8–14]. It has been widely
recognized by economists that economic behavior generally involves both elements of
stochastics and/or fuzziness [15]. The objects which economic theory deals with are
replete with all sorts of perceptions or emotions. While a rich literature on economic
behavior under a stochastic environment has developed during the past decade, it is
notable that less systematic attempts have been made to-date, to investigate economic
behavior under fuzziness (Chen et al. 1983).
Fuzzy set theory introduced by L.A. Zadeh (1965) can be used in uncertain economic
decision environments to deal with vagueness. Fuzzy numbers can be used, such as cash
amounts and interest rates in the future, for conventional decision-making models.
In the past decade there were some noticeable works concerning the implication of
fuzzy sets in the world of economics, more specific in finance. There are several works
to be mentioned that are based study: Ward [8] developed fuzzy present worth analysis
by introducing trapezoidal cash flow amounts; Buckley [9] proposed a fuzzy capital
budgeting theory in the mathematics of finance; Kaufmann and Gupta [10] used the
fuzzy numbers to the discount rate by which they derived a fuzzy present worth method
for investment alternative selection; Kuchta (2001) considers net present value as a
quadratic 0–1 programming problem in a fuzzy form; Kahraman et al. [15] derived the
formulas for the analyses of fuzzy present value, fuzzy equivalent uniform annual value,
fuzzy final value, fuzzy benefit-cost ratio and fuzzy payback period in capital budgeting;
Dimitrovski and Matos [13] gave the techniques for comparing and ordering fuzzy
numbers of independent numbers and dependent numbers and they examined fuzzy case
with partial correlation.
There are three areas where fuzzy set theory is applied in economics: economic
modeling, which gives expression to the fact that parameters are not estimated accurately
even though structural equations are obtained from theoretical considerations; economic
forecasting where a solution is given to multiple model problem; and economic policy
where the final policy is chosen by the decision maker from a range of optimum policies
depending on a risk parameter (Chang 1977). This paper aims to combine all three areas
to give a reliable solution for this type of investment.
There are several fuzzy investment analysis techniques that can be employed to
measure the rentability of an investment. Relying on the particular problem and asso‐
ciated data, there can be certain advantages to apply one technique or another [14].
280 M.M. Balas et al.

3 Estimating the Investment Feasibility (Matlab)

Our objective is to estimate the feasibility an investment in vineyards and vinery. The
feasibility of the investment is estimated (strong points, weak points, opportunity, risks,
perspectives). The feasibility is expressed based on cost and profit [16]. Specific expert
knowledge must be considered to determine the feasibility of a vinery, such as: planting
costs, winemaking costs, ecological factors, agro technical factors, technological
factors, production and sales [17, 18] (Figs. 1 and 2).

Planting
Cost
Winemaking

Ecology Fezability

Agro Technical
Profit
Technology

Production+Sales
a) Structure

b) Matlab implementation: fis. files (Fuzzy Inference System) embedded in Simulink

Fig. 1. The feasibility expert system.

 Fuzzy-Expert System for Investments in Winemaking 281

Business Maturity

a) The structure of the Planting controller

b) The linguistic variable Business_Maturity (zero, small medium, great)

Linguistic Linguis-
Signification
variables tic labels
Zero Starting from scratch
1.Business Maturity Small First years
(input) Medium 10-20 years
Great Traditional
Zero Outside vineyards, no expertize
2. Terrain Small Outside vineyards with expertize
(input) Medium Inside vineyards, no expertize
Great Inside vineyards with expertize
Zero No laborers
3. Human Re- Small Unskilled laborers
sources (input) Medium Skilled laborers in restricted number
Great Skilled laborers in great number
Zero Avoid this zone
Small Not very recommendable to plant in this zone
Planting (output)
Medium Recommended to plant
Great Highly recommended to plant

c) The input and output variables

Fig. 2. The Planting fuzzy controller.

282 M.M. Balas et al.

The inference’s result is a Feasibility index, bounded in the [0…1] interval.

The classic Fuzzy Logic Tool-box of Matlab [19] is widely used when testing
algorithms [20, 21]. We will continue to use it, in order to extend and to improve the
Wine-making application. However, when realizing portable standalone applica‐
tions, dedicated software is obligatory [22]. The following version of the Wine-
making expert system was realized in this respect.

4 Estimating the Cost (BES)

At the Aurel Vlaicu University of Arad, R. Boran conceived and programmed a visual
fuzzy-expert system development shell named BES (Basic Expert System). The appli‐
cations are built with a small number of basic blocks [23]:
• start block (green);
• single choice block (yellow);
• decision block (white);
• integer block (turquoise);
• display block (orange);
• buffer block (brown), able to memorize data during the execution;
• computing bloc (magenta);
• stop block (red).
We illustrate the BES implementation of the Winemaking expert system by the costs
estimation section, which is an important issue in financial software [24].
The algorithm starts with a greed start block and ends with a red stop block. Further,
A single choice block is used for selecting the region and variety of grapes, a decision
block selects between the types of grapes and what costs does the user want to know,
an integer block allow the user to introduce the planting area. Message blocks, a buffer
blocks and computing blocks are also used (Figs. 3, 4 and 5).

Fig. 3. The Winemaking BES expert system

 Fuzzy-Expert System for Investments in Winemaking 283

Fig. 4. The choice of the region and of the grapes

Fig. 5. The costs estimation

284 M.M. Balas et al.

5 Discussion

The paper is presenting the implementations of a winemaking investment decision

expert system in Matlab (FIS) and as a standalone application.
The final goal of this research is to transfer the ability of the fuzzy expert systems to
solve economic issues into the winemaking domain. Since the feasibility of an invest‐
ment has many ambiguous sides, fuzzy expert system are most suitable to solve the
problem, if appropriate expert knowledge is available.

Acknowledgements. This work was co-funded by European Union through European Regional
Development Funds Structural Operational Program “Increasing of Economic Competitiveness”
Priority axis 2, operation 2.1.2. Contract Number 621/2014.

References

1. Ellsberg, D.: Risk ambiguity and the Savage axioms. Quart. J. Econ. 75, 643–669 (1961)
2. Malita, M., Malita, M.: Bazele inteligentei artificiale. Logici propozitionale. Ed. Tehnica,
Bucuresti (1987)
3. Poole, D., Mackworth, A., Mackworth, R.: Artificial Intelligence. Goebel Oxford University
Press, New York (1998)
4. Negnevitsky, M.: Artificial Intelligence: A Guide to Intelligent Systems. Addison Wesley
Longman, Boston (2005)
5. Zadeh, L.A.: The role of fuzzy logic in the management of uncertainty in expert systems.
Fuzzy Sets Syst. 11(1–3), 197–198 (1983)
6. Buckley, J.J., Siler, W.: Fuzzy Expert Systems and Fuzzy Reasoning. Wiley, Hoboken (2005)
7. Torra, V.: Hesitant fuzzy sets and decision. Int. J. Intell. Syst. 25(6), 395–407 (2010)
8. Ward, T.L.: Discounted fuzzy cash flow analysis. In: Proceedings of 1985 Fall Industrial
Engineering Conference, pp. 476–481. Institute of Industrial Engineers (1985)
9. Buckley, J.J.: The fuzzy mathematics of finance. Fuzzy Sets Syst. 21, 257–273 (1987)
10. Kaufmann, A., Gupta, M.M.: Fuzzy mathematical models in engineering and management
science. Elsevier, Amsterdam (1988)
11. Chiu, C.-Y.: Fuzzy cash flow analysis using present work criterion. Eng. Econ. 39(2), 113–
138 (1994)
12. Kuchta, D.: Fuzzy capital budgeting. Fuzzy Sets Syst. 111, 367–385 (2000)
13. Dimitrovski, A., Matos, M.: Fuzzy present worth analysis with correlated and uncorrelated
cash flows. In: Kahraman, C. (ed.) Fuzzy Engineering Economics with Applications. Springer,
Heidelberg (2008)
14. Kahraman, C., Çağrı Tolga, A.C.: Fuzzy Investment Planning and Analyses in Production
Systems. Springer, Heidelberg (2010)
15. Kahraman, C., Ruan, D., Tolga, E.: Capital budgeting techniques using discounted fuzzy
versus probabilistic cash flows. Inf. Sci. 142(1–4), 57–76 (2002)
16. Emhjellen, K., Emhjellen, M., Osmundsen, P.: Cost Estimates and Investment Decisions.
Foundation for Research in Economics and Business Administration, Bergen (2001)
17. Botezatu, A. (2012). http://www.startups.ro/tutoriale/cum-sa-iti-faci-o-podgorie-si-o-crama.
Accessed 24 Apr 2014
18. Chetreanu, D.: Plantarea viței de vie. http://www.cartiagricole.ro/infiintare-plantatie-
viticola/. Accessed 24 Apr 2016
 Fuzzy-Expert System for Investments in Winemaking 285

19. Sivanandam, S., Sumathi, S., Deepa, S.: Introduction to Fuzzy Logic using MATLAB.
Springer, Berlin (2007)
20. Fasanghari, M., Montazer, G.A.: Design and implementation of fuzzy expert system for
Tehran stock exchange portfolio recommendation. Expert Syst. Appl. 37(9), 6138–6147
(2010)
21. Balas, M.M., Csorba, L.M.: An application of the fuzzy logic in selecting a production line,
as an investment decision. J. Econ. Bus. Res., 1–12 (2010). Aurel Vlaicu University of Arad,
Arad, Romania
22. Ruiz-Mezcua, B., Garcia-Crespo, A., Lopez-Cuadrado, J.L., Gonzalez-Carrasco, I.: An expert
system development tool for non AI experts. Expert Syst. Appl. 38(1), 597–609 (2011)
23. Boran, R.: Basic Expert System. Diploma work, Aurel Vlaicu University of Arad (2016).
Coordinated by M.M. Balas
24. Idri, A., Abran, A., Khoshgoftaar, K.: Fuzzy case-based reasoning models for software cost
estimation. In: Proceedings of the 11th International Workshop on Software Measurements,
Montreal, Canada, pp. 93–101 (2001)
 Closed Loop Blood Glucose Regulation of Type
1 Diabetic Patient Using Takagi-Sugeno Fuzzy
Logic Control

Anirudh Nath1(&), Rajeeb Dey1, and Valentina E. Balas2

1
Department of Electrical Engineering, National Institute of Technology,
Cachar District, Silchar 788010, Assam, India
[email protected],[email protected]
2
Department of Automatics and Applied Software, Faculty of Engineering,
“Aurel Vlaicu” University of Arad, Arad, Romania
[email protected]

Abstract. This paper presents a Takagi-Sugeno (TS) Fuzzy Logic based

intelligent control algorithm for maintaining plasma glucose regulation for Type
1 Diabetic Mellitus (T1DM) patient in presence of known meal disturbance and
parametric variations. The mathematical model considered here to describe the
interaction between glucose and insulin, is modified Bergman minimal model.
Simulation results of the plasma glucose concentration and insulin infusion rate
validate the effectiveness of the designed control law.

Keywords: Takagi-Sugeno (TS)
Fuzzy logic control (FLC)
Bergman

minimal model
Type 1 diabetes
Blood glucose

1 Introduction

Diabetes is one of the most challenging and wide-spread diseases around the world.
Huge amount of money is invested every year in order to find effective treatment, alone
the United States spent about $132 billion USD in 2002 [1]. There are mainly two
forms of the disease, namely, Type 1 Diabetes Mellitus and Type 2 Diabetes Mellitus.
Type 1 Diabetes Mellitus is basically a class of chronic and autoimmune disease which
is characterized by b-cell destruction resulting in severe hyperglycemia. Insufficient or
negligibly small amount of insulin production results in failure in maintaining the blood
glucose level in the safe range (70–180 mg/dL) [2]. The long-term consequences may
be renal failure, neural damage, cardiac arrest, lower-limb amputations, retinopathy and
diabetic related blindness [3].
Enormous research have been carried out for control and management of this
disease. Globally, one of the aspects of research is towards the development of efficient
Artificial Pancreas system (APS), where the development of efficient control algorithm
is a key research issue. The key features of the mathematical model considered in this
work which is a modification of the Bergman minimal model (based on Intravenous
Glucose Tolerance Test (IVGTT)) [4], is that the glucose is measured subcutaneously
and insulin is injected intravenously.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_23
 Closed Loop Blood Glucose Regulation 287

The control problem of regulating the blood glucose level for T1DM patient has
been solved using both classical and various variants of PID control schemes [9, 10]. In
[11] H-infinity control has been attempted to make the controller robust in presence of
parametric uncertainty and disturbance attenuation. In [12, 13] different sliding mode
control (SMC) techniques have been applied for regulating blood glucose in the
presence of parametric uncertainty and unmodelled dynamics. But the control signal in
SMC is affected by the chattering phenomenon and further research is going on to
eliminate chattering in the control signal. Model predictive control algorithms are
proposed in [14, 15] which are based on both simple and complex dynamical models of
glucose-insulin regulatory system. MPC solves the problem of blood glucose regula-
tion but require identification of linear and nonlinear time series models which needs to
be validated against clinical data, and also the calculation of the control law at an given
instant requires prediction of future values of blood glucose level which may result in
inappropriate control signals if the predictions are not good enough. Neural Network
based control is designed in [16] which requires large amount of input-output data for
training of the network and may face the problem of computational burden. Fuzzy logic
control is essentially an expert knowledge based system [5] and finds a wide range of
applications in the field of Biomedical engineering [6–8]. Different fuzzy logic con-
troller based on Mamdani inference system have been designed for the blood glucose
regulation in T1DM patients in [17–20]. In order to reduce the computational burden
which arise from the need of defuzzification of the fuzzy output in Mamdani based
FLC, a TS FLC scheme is proposed here to keep the blood glucose level in the normal
range (70 mg/dL–120 mg/dL) and avoid any instance of post prandial hyperglycemia
(>180 mg/dL) or hypoglycaemia (<70 mg/dL). The proposed controller is robust with
respect to parametric variations of the model and have a simple design structure and is
computationally more efficient than its fuzzy counterparts which are based on Mamdani
inference system.
The paper is organized as follows. In Sect. 2, the Bergman minimal model and
modified minimal model representing the T1DM patient are briefly presented. The
adopted model is nonlinear dynamical in nature. In Sect. 3, a TS FLC is proposed for
studying the closed loop performance of glucose control in T1DM patients. The details
about the controller synthesis are also discussed. Then, in Sect. 4, the simulation results
for modified minimal model with both nominal parameters and with parameter varia-
tions are presented. Finally, several conclusions are drawn in Sect. 5.

2 Glucose-Insulin Dynamical Model

Many mathematical models have evolved over the years to describe the dynamics of
glucose-insulin regulatory system. Bergman minimal model is probably the first model
to depict the dynamics of glucose-insulin regulatory system intravenously [21]. In the
last couple of decades, an attempt has been made to develop more realistic situations of
T1DM patients by considering the glucose-insulin dynamics subcutaneously [24–27].
However, the evolution of the models have not taken into account the control oriented
modelling approach. In this paper, we have considered simple mathematical model for
288 A. Nath et al.

T1DM patients yet realistic than that of Bergman minimal model. We describe the
details of modelling in the succeeding subsections.

2.1 Bergman Minimal Model IVGTT

Bergman’s minimal model is basically a single compartmental model which assumes
the body as a compartment with basal concentrations of glucose and insulin. There are
two minimal models, one of which describes the dynamics of glucose clearance and
uptake in the body, and the other describes the delay in the active insulin (in the remote
compartment) which accounts for the glucose utilization by tissues and glucose
dynamics of the liver (uptake and production). The dynamics of the above regulatory
system is described by the following nonlinear ordinary differential equations [21]:

dGðtÞ
¼ ðp1 þ X ðtÞÞGðtÞ þ p1 Gb ð1Þ
dt
dXðtÞ
¼ p2 ðtÞXðtÞ þ p3 ðI ðtÞ  Ib Þ ð2Þ
dt

The following figure describes the dynamics graphically (Fig. 1):

Liver

Production Uptake

effect

Blood Glucose
Remote Compartment of Active Insulin Glucose Compartment
Insulin Dose

Flow
effect
Clearance Uptake

Fig. 1. Graphical representation of the minimal model [21]

The insulin minimal model which represents the insulin kinetics is given by the
following differential equations:

dIðtÞ
¼ p6 ½GðtÞ  p5  þ t  p4 ½I ðtÞ  Ib  ð3Þ
dt
 Closed Loop Blood Glucose Regulation 289

For T1DM patient, the pancreatic response is replaced by external insulin infusion
term given below where term related to the pancreatic insulin release is omitted and
external insulin infusion is introduced.

dIðtÞ U ðt Þ
¼ p4 I ðtÞ þ ð4Þ
dt VI

2.2 Modified Bergman Minimal Model

The fisher meal disturbance model [22] is introduced in the modelling of T1DM patient
in order to address the rapid changes in the postprandial plasma glucose concentration
profile just after the meal, which describes the appearance of glucose in the blood
plasma. The mathematical model describing the meal disturbance is given as follows:

DðtÞ ¼ B:expðdrate :tÞ ð5Þ

D(t) represents the rate of glucose in mg/dL entering into the blood plasma. As
suggested in the literature, derivative of D(t) is considered in the modelling in place of t.

dDðtÞ
¼ drate :DðtÞ ð6Þ
dt

Since the modern day continuous glucose monitor (CGM) devices are based on
subcutaneous glucose measurements, the controller input is chosen as subcutaneous in
place of intravenous. This choice makes the modified minimal model more realistic for
the design of closed loop control algorithms for the development of Artificial Pancreas
(AP). The model introduced here is adopted from GIM (a T1DM simulator), which
takes a delay of about 10 min for the glucose measurement to appear in the subcuta-
neous depot from the blood plasma [23].

dGsc ðtÞ 1
¼ ðGðtÞ  Gsc ðtÞÞ ð7Þ
dt T
where G(t) denotes the subcutaneous glucose concentration (mg/dL) and T denotes the
time constant. The open loop response of the modified Bergman minimal model rep-
resenting a T1DM patient with very high basal glucose concentration (Gb = 200
mg/dL) and zero Ib and assuming that no external insulin is infused and a meal
disturbance of a rate of 10 mg/dL/min at time, t = 100 min with the following
parameters are given below [21] (Figs. 2 and 3):

p1 ¼ 0:028735; p2 ¼ 0:028344; p3 ¼ 5:035  105 ; Gb ¼ 200; Ib ¼ 0

290 A. Nath et al.

Fig. 2. Fisher disturbance model [12]

Fig. 3. Open loop response of plasma glucose concentration of modified minimal model

3 Closed Loop Control: Design Strategy

The modified minimal model represents the virtual patient which has exogenous
intravenous insulin infusion as the input and plasma glucose concentration as the
output. The meal disturbances act as additional disturbance input to the system. The
main objective of the controller is to generate appropriate control signal based on the
subcutaneous glucose measurements so that the blood glucose level is maintained in
the safe range [70 mg/dL–180 mg/dL] in the presence of known meal disturbances. It
is assumed that controller is acting in the closed loop system for some time and a basal
concentration of plasma glucose, Gb and basal insulin concentration, Ib is established.

3.1 Controller Synthesis

In this section, fuzzy logic control for blood glucose regulation of T1DM patient is
described briefly. The structure of the controller is based on zero order Takagi-Sugeno
 Closed Loop Blood Glucose Regulation 291

(TS) fuzzy logic architecture with two inputs, error in subcutaneous glucose concen-
tration, e(t) which is the difference between the target glucose level and the current
subcutaneous glucose concentration at that instant, and the second input is the
derivative of Gsc(t). The output of the controller is the exogenous insulin infusion rate.
Both the inputs have the same normalized universe of discourse [−1, 1]. Three
membership functions are used to partition the input universe of discourse of both the
inputs and the type of membership functions are chosen as Gaussian functions:
Negative, Zero and Positive for both the inputs. The universe of discourse for the
output fuzzy variable is [−2, 2] and the range of insulin infusion rate is 0–10 U/min and
the output membership functions are chosen as constants namely, Negative Big (NB),
Negative Small (NS), Zero (Z), Positive Big (PS), Positive Small (PB). The input
membership functions are presented in Fig. 4. A total of nine fuzzy IF–THEN rules are
defined to calculate the output and are presented in Table 1. MIN–MAX laws are used
to derive the appropriate fuzzy output for a given set of inputs and Weighted Average
method of defuzzification is used to calculate the final output of the TS fuzzy logic
controller. Three scaling factors namely, GE, GCE and GU are used to scale the input
and output variables and these are nothing but nonlinear gains which are tuned
manually.

Fig. 4. Membership functions for error and derivative of Gsc(t)

4 Results

In this section, the results from the simulation studies are presented to validate the
proposed non-linear intelligent closed loop control algorithm based on TS fuzzy logic
control paradigm. The simulation studies are classified into two broad cases, (i) Sim-
ulation studies with nominal model parameters and (ii) Simulation studies with realistic
model with known parametric variations.
The following assumptions are made for the simulation studies:
• The nominal non-linear modified minimal model has fixed parameters for controller
design [14].
292 A. Nath et al.

Table 1. Fuzzy rule base

e cy
N Z P
N NB NS Z
Z NS Z PS
P Z PS PB

Table 2. Parameters/variables for modified Bergman minimal model [21]

Parameter/Variable Unit Description
T Min Time
G(t) mg/dL Plasma glucose concentration
Gb mg/dL Basal glucose concentration
X(t) 1/min Effect of active insulin
I(t) mU/L Plasma insulin concentration
Ib mU/L Basal insulin concentration
D(t) mg/min Rate of glucose entering blood
Gsc ðtÞ mg/dL Subcutaneous glucose concentration
p1 1/min Glucose clearance rate independent of insulin
p2 1/min Rate of clearance of active insulin
p3 L/min2.mU Glucose clearance rate dependent on insulin
p4 1/min Decay rate of blood insulin
p5 mg/dL Threshold for blood glucose concentration
p6 mU.dL/L.mg.min Rate of pancreatic release after glucose bolus
T 1/min Time lag between subcutaneous depot and plasma

• Parametric variations in the considered model for known ranges have been intro-
duced. The ranges of parameter variation are adopted from [16] which effectively
represent the worst case of parametric variations which may arise due to
inter-patient and intra-patient variability.
• The amount and timing Fisher meal disturbances is assumed to be exactly known.
• Initial value of plasma glucose concentration is very high 200 mg/dL.

4.1 Case I: Simulation Studies with Nominal Model Parameters

The nonlinear modified minimal model is used to represent the glucose- insulin
interactions of the T1DM patient with nominal parameters provided in Table 3. The
trajectory of plasma glucose concentration in T1DM patient with nominal parameters
after applying the control action is shown in Fig. 5a. A set point of 81 mg/dL has been
used. The simulation result depicts that the designed control logic could track the
set-point, thus avoiding the events of both postprandial hyperglycemia and hypogly-
caemia. The control signal is shown in Fig. 5b. depicts the nature of external insulin
infusion intravenously. The control signal remains steady throughout the simulation
 Closed Loop Blood Glucose Regulation 293

Table 3. Different parameters for realistic modified Bergman minimal model

S. no. p1 p2 p3 p4
1 0.0284735 (nominal) 0.028344 (nominal) 5.035  10−5 (nominal) 0.0925 (nominal)
2 0.0143675 (−50%) 0.018991 (−33%) 2.517  10−5 (−50%) 0.0537 (−42%)
3 0.0431025 (+50%) 0.037698 (+33%) 7.552  10−5 (+50%) 0.1315 (+42%)
4 0.0215513 (−25%) 0.024092 (−15%) 3.776  10−5 (−25%) 0.0740 (−20%)
5 0.0355188 (+25%) 0.032596 (+15%) 6.293  10−5 (+25%) 0.1111 (+20%)

time except overshoots are observed. The reason for the overshoot in control signal is
for compensating the meal disturbance effects.

4.2 Case II: Simulation Studies with Known Parameter Variations

In this study four sets of parameters are presented in Table 3. The main objective of
this study is to analyze the robustness of the proposed control algorithm with respect to
parametric variations which may arise due to inter-patient and intra-patient variability.

Fig. 5. a. Closed loop response of plasma glucose concentration for nominal parameters.
b. External insulin infusion rate
294 A. Nath et al.

(a)

(b)

(c)

(d)

Fig. 6. Closed loop response of plasma glucose concentration in presence of meal disturbance
and parametric variations
 Closed Loop Blood Glucose Regulation 295

The structure of the controller is kept constant for parametric variations without any
changes in the structure of the controller. After introducing meal disturbances at three
different times of a day, the plasma glucose concentration is maintained not only in the
safe range [70–180 mg/dL] but also the events of postprandial hyperglycemia and
hypoglycemia are avoided for all the four different cases considered. The results for the
above referred cases are placed in Fig. 6a, b, c and d respectively.

5 Conclusion

In this work TS fuzzy logic based closed loop control algorithm for maintaining plasma
glucose concentration within postprandial safe range of (70–180 mg/dL) is proposed.
Simulation results reveal the effectiveness of the closed loop control algorithm in terms
of prevention of the occurrences of severe postprandial hyperglycemia and hypo-
glycemia. The proposed control algorithm is robust with respect to a wide range of
parametric variations of the modified Begman minimal model.

References
1. Takahashi, D., Xiao, Y., Hu, F., Lewis, M.: A survey of insulin-dependent diabetes-part I:
therapies and devices. Int. J. Telemed. Appl. 2008, 2 (2008)
2. González, A.A., Voos, H., Darouach, M.: Glucose-insulin system based on minimal model:
a realistic approach. In: Proceedings of the 2015 17th UKSIM-AMSS International
Conference on Modelling and Simulation, pp. 55–60. IEEE Computer Society (2015)
3. Al-Fandi, M., Jaradat, M.A., Sardahi, Y., Al-Ebbini, L., Khaleel, M.: Intelligent control of
glucose concentration based on an implantable insulin delivery system for type I diabetes. In:
Applied Electrical Engineering and Computing Technologies (AEECT), IEEE Jordan
Conference, pp. 1–6 (2011)
4. Bergman, R.N., Phillips, L.S., Cobelli, C.: Physiologic evaluation of factors controlling
glucose tolerance in man: measurement of insulin sensitivity and beta-cell glucose sensitivity
from the response to intravenous glucose. J. Clin. Invest. 68(6), 1456–1467 (1981)
5. Wang, L.X.: A Course in Fuzzy Systems and Control. Prentice-Hall, Upper Saddle River
(1997)
6. Meier, R., Nieuwland, J., Zbinden, J.M., Hacisalihzade, S.S.: Fuzzy logic control of blood
pressure during anesthesia. IEEE Control Syst. Mag. 12(6), 12–16 (1992)
7. Duckstein, L., Bilinowska, A., Verroust, J.: Fuzzy classification of patient state with
application to electro-diagnosis of peripheral polyneuropathy. IEEE Trans. Biomed. Eng. 42
(8), 786–792 (1995)
8. Buisson, J.C., Garel, A.: Balancing meals using fuzzy arithmetic and heuristic search
algorithm. IEEE Trans. Fuzzy Syst. 11(1), 68–78 (2003)
9. Marchetti, M.B.G., Jovanovic, L., Zisser, H., Seborg, D.E.: An improved PID switching
control strategy for type 1 diabetes. IEEE Trans. Biomed. Eng. 55(3), 857–865 (2008)
10. Li, C., Hu, R.: PID control based on BP neural network for the regulation of blood glucose
level in diabetes. In: Proceedings of the 7th IEEE International Conference on Bioinfor-
matics and Bioengineering, Boston, pp. 1168–1172 (2007)
296 A. Nath et al.

11. Chee, F., Savkin, A.V., Fernando, T.L., Nahavandi, S.: Optimal H-infinity insulin injection
control for blood glucose regulation in diabetic patients. IEEE Trans. Biomed. Eng. 52(10),
1625–1631 (2005)
12. Abu-Rmileh, A., Garcia-Gabin, W., Zambrano, D.: Internal model sliding mode control
approach for glucose regulation in type1 diabetes. Biomed. Signal Process. Control 5, 94–
102 (2010)
13. Shtessel, Y., Kaveh, P.: Blood glucose regulation using higher-order sliding mode control.
Int. J. Robust Nonlinear Control 18, 557–569 (2008)
14. Lynch, S.M., Bequette, B.W.: Model predictive control of blood glucose in type 1 diabetics
using subcutaneous glucose measurements. In: Proceeding of the American Control
Conference, Anchorage, May 2002, pp. 4039–4043 (2002)
15. Hovorka, R., Canonico, V., Chassin, L., Haueter, U., Massi-Benedetti, M., Federici, M.,
et al.: Nonlinear model predictive control of glucose concentration in subjects with type1
diabetes. Int. J. Control 25, 905–920 (2004)
16. Ali, S.F., Padhi, R.: Optimal blood glucose regulation of diabetic patients using single
network adaptive critics. Optim. Control Appl. Methods 32(2), 196–214 (2009)
17. Ibbini, M.S., Masadeh, M.A.: A fuzzy logic based closed-loop control system for blood
glucose level regulation in diabetics. J. Med. Eng. Technol. 29(2), 64–69 (2005)
18. Maleki, A., Geramipour, A.: Continuous control of blood glucose in TIDM using fuzzy logic
controller in insulin pump: a simulation study. In: 2nd International Conference on Control,
Instrumentation and Automation, pp. 122–127 (2011)
19. Yasini, S., Naghibi-Sistani, M.B., Karimpour, A.: Active insulin infusion using fuzzy-based
closed-loop control. In: Proceedings of 3rd International Conference on Intelligent System
and Knowledge Engineering, Xiamen, November 2008, pp. 429–434 (2008)
20. Campos-Delgado, D.U., Hernandez-Ordonez, M., Femat, R., Gordillo-Moscoso, A.: Fuzzy-
based controller for glucose regulation in type-I diabetic patients by subcutaneous route.
IEEE Trans. Biomed. Eng. 53(11), 2201–2210 (2006)
21. Friis-Jensen, E.: Modeling and Simulation of Glucose-insulin Metabolism, Dissertation,
Technical University of Denmark (2007)
22. Fisher, M.E.: A semi closed-loop algorithm for the control of blood glucose levels in
diabetics. IEEE Trans. Biomed. Eng. 38(1), 57–59 (1991)
23. Dalla Man, C., Raimondo, D.M., Rizza, R.A., Cobelli, C.: GIM, simulation software of meal
glucose-insulin model. J. Diabetes Sci. Technol. 1(3), 323–330 (2007)
24. Makroglou, A., Li, J., Kuang, Y.: Mathematical models and software tools for glucose-
insulin regulatory system and diabetes: an overview. Appl. Numer. Math. 56, 559–573
(2006)
25. Dalla Man, C., Rizza, R.A., Cobelli, C.: Meal simulation model of the glucose-insulin
system. IEEE Trans. Biomed. Eng. 54(10), 1740–1749 (2007)
26. Sorensen, J.T.: A physiologic model of glucose metabolism in man and its use to design and
assess improved insulin therapies for diabetes (1985)
27. Dalla Man, C., Toffolo, G., Basu, R., Rizza, R.A., Cobelli, C.: A model of glucose
production during a meal. In: Proceedings of the IEEE EMBS Conference, pp. 5647–5650
(2006)
 Experimental Implementation of Fuzzy Vision-Based
Tracking Control of Quad-Rotor

Hemjyoti Das1 ✉ , Aniket Samrat Mazumdar1, Rajeeb Dey1, and Lintu Roy2
( )

1
Electrical Engineering, National Institute of Technology, Silchar, India
[email protected], [email protected]
2
Mechanical Engineering, National Institute of Technology, Silchar, India
[email protected]

Abstract. This paper presents a novel approach to detect and track an object
using a quadrotor-UAV. The proposed system mainly consists of two parts- (i)
Object detection and tracking using histogram backprojection and CAMSHIFT
tracker, (ii) Fuzzy Proportional and Fuzzy Proportional-Derivative controller for
controlling the drone. We implemented our algorithm using ROS (Robot Oper‐
ating System), OPENCV library and MATLAB programming environment.

Keywords: Camshift tracker · Quadrotor · Fuzzy logic · Fuzzy-PD controller

1 Introduction

In the recent years, there has been a lot of ongoing research in the field of Aerial Robotics.
The quad-rotor which is a four rotor micro-aerial vehicle has got immense popularity
among various research groups due to its ease of handling compared to the fixed wing
vehicles. Various algorithms related to path planning, machine learning, computer
vision, control systems, etc. are now being implemented on drones to increase the wide
range of applications of drones.
A wide variety of research is being done by various universities and research groups
on the vision based control of drones like, in [1] a camera based navigation system for
the Parrot AR Drone have been designed using visual SLAM, Extended Kalman Filer
and PID Controller. In [2], an image based visual servoing (IBVS) controller using
OpenTLD tracker and PD controller have been developed.
Quadrotors are also used nowadays for tracking of Marine Animals as, in [4] Hue
and Saturation values for calculating the Histograms for object identification and
tracking have been used. In [5], Hough transform was used for detecting players in an
indoor soccer game using quadrotor. Similarly in [3], Hough transform is used for
autonomous object tracking. In [6], a new control approach is designed using switching
control which utilizes Edge Detection and Optical Flow algorithm.
In [7], a quad-rotor UAV is simulated for tracking a moving ground object in the
presence of noise. In [8], a new algorithm called dagger algorithm is proposed for high
speed autonomous flight through dense forest environment. Nowadays landmarks have
become very effective tool for localization as discussed in [9].

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_24
298 H. Das et al.

Fuzzy controllers are also being used in different systems like in [10], they have
implemented Fuzzy-PD+I controller for controlling a Magnetic Ball Levitation System.
In [14], they successfully implemented a fuzzy controller for target tracking on a Pelican
quad-rotor running on an Intel Atom Board PC.
The main aim of this experiment is to autonomously detect an object and track it
using Fuzzy based Controller. The paper is organized as follows: Sect. 2 describes the
hardware platform and the software architecture used in the experiment, Sect. 3 discusses
the object detection and tracking algorithm. In Sect. 4, controller design for the Parrot
AR Drone is discussed. Finally the experimental results are shown in Sect. 5.

2 Experimental Platform

For the realization of this project, the Parrot AR drone quad-rotor is used. The AR-Drone
2.0 [13] is an electrically powered quad-rotor intended for augmented reality games. It
is preferred to other drones due to its cheap price, robustness of hardware, ability to
execute basic control commands through its on-board softwares and compatibility with
Robot Operating System (ROS). The AR Drone’s SDK functions can be translated easily
into ROS nodes, which allows us to integrate and compare our project with existing
research.

2.1 Hardware Description [13]

1. Support structure and body: The body and support structure of the Parrot AR Drone
2.0 quad-rotor are made out of plastic and carbon fiber respectively because of their
lightweight and resistance to mechanical wear out. This keeps the quad-rotor’s
weight to approximately 380 g or 420 g, depending on the hull of the quad-rotor.
2. Four high-efficiency brushless motors: The Parrot AR Drone 2.0 quad-rotor uses
four electric rotors, each with power of 14.5 W and a dedicated 8 MIPS AVR micro‐
controller to ensure better flight. Speeds of up to 10 m/s can be reached.
3. Sensors: The Parrot AR Drone 2.0 consists of the following sensors: 3 axis gyro‐
scope, 3 axis accelerometer, 3 axis magnetometer, an ultrasonic sensor and a barom‐
eter. It has two camera modules- the frontal camera has a 720p sensor with maximum
30 fps recording and the bottom camera has a QVGA sensor with upto 60 fps
recording.
4. Control board: The Parrot AR Drone 2.0 control board consists of the following:
1 GHz and 32 bit ARM Cortex A8 processor, a Wi-Fi b/g/n device which creates an
Ad-hoc network, a dedicated digital signal processing (DSP unit) TMS320DMC64X
running at 800 MHz.
5. Battery: The standard provided along with the Parrot AR Drone 2.0 is 1000 mAh,
11.1 V Lithium Polymer battery. It provides power to the quadrotor for about 10 min.
6. Coordinate System: The Euler angle representation of the drone is shown in Fig. 1
to represent the angular orientation. The origin of the body frame is located in the
quadrotor center of gravity. Its X- axis is pointing along the front camera, the Z- axis
is pointing vertically upward from the base and the Y- axis is directed sideways along
 Experimental Implementation of Fuzzy Vision-Based Tracking 299

the body frame as shown in Fig. 1. The pitch, roll and the yaw directions are also
shown in Fig. 1 where θ, Φ and ψ denotes pitch, roll and yaw respectively.

Fig. 1. Parrot AR Drone 2.0 and its coordinate system

2.2 Software Platform

The control board of the Parrot AR Drone 2.0 runs a GNU/Linux with kernel version
2.6.32 with versatile command line program BusyBox. The quadrotor is controlled via
WLAN through set of commands sent from the ground to the quadrotor. After
connecting the battery to the quadrotor, an ad-hoc network appears through which
communication is done by connecting to its unencrypted access point on address
192.168.1.1. Once this communication is established, quadrotor is led over several
channels with a predefined set of operations. There are three main channels for inter‐
facing the drone. We can publish various commands like takeoff, land, velocity, etc.
over the command channel. The sensory data and the drone status are received using the
navdata channel at a frequency of 30 Hz. The frontal and the bottom camera publishes
the image using stream channel. In our project, we have used a few third party softwares
as described below:-
1. Robot Operating System: Robot Operating System (ROS) [15] is used for accessing
the different functionalities of the quadrotor like position, orientation, velocity,
acceleration etc. It is a collection of software frameworks for robot software devel‐
opment. ROS contains a distributed framework of processes (called nodes) that
communicate via message passing, grouped into easily shared and distributed pack‐
ages. Information flows into and out of an ROS node via messages (called topics)
and services (higher-level abstractions operated through messages to implement
return values).
2. OpenCV: OpenCV (Open Computer Vision) library [17] is used for basic image
handling and processing.
3. cv_bridge: The ROS library cv_bridge is used to provide an interface between ROS
and OpenCV.
300 H. Das et al.

4. ardrone_autonomy [11]: This is the driver used for interfacing the quadrotor to ROS.
It is an autonomous flight library for the AR Drone 2.0 based on its official AR Drone
SDK version 2.0.1 built on top of the node-ar-drone library. This package is used to
plan and execute missions by describing the path, attitude and orientation the drone
must follow, as compared to direct control of speed.
5. MATLAB Robotics Systems Toolbox [18]: It is used to connect MATLAB with
ROS. It helps to work with ROS messages and also exchange data with publishers
and subscribers in the ROS network.
6. MATLAB Fuzzy Logic Toolbox: It is used for designing fuzzy inference system in
MATLAB.

3 Object Detection and Tracking

3.1 Histogram-Backprojection [19]

This method calculates the histogram of an image to denote the probabilities of colour
which appear in the image. To calculate the histogram we first transform the image
obtained from RGB image formant to HSV as it is more robust to environmental changes.
Then we separated the Hue channel which has got a range of 0 to 180. It is followed by
conversion to a greyscale image from where the histogram is calculated and then
normalized. Finally the backprojection is calculated using the inbuilt function “calc‐
BackProject” in OpenCV.
Backprojection calculates the percentage of a particular colour in the image using
histogram and modifies the intensities of each pixels in proportion to its weightage in
the image. For example, if a particular pixel, say red has got 30% weightage, then we
modify all the intensities of the red pixels to (0.3 * 255). Here 255 is maximum value
of the intensity as it is an 8 bit image (Fig. 2).
 Experimental Implementation of Fuzzy Vision-Based Tracking 301

Fig. 2. (a) Example of an RGB image, (b) Its separated Hue component, (c) The histogram of
the Hue component, (d) Backprojection of the image

3.2 Meanshift Algorithm [16]

Its uses the backprojection technique to obtain the data points of the detected object.
And then the meanshift tracker tries to cluster these points by shifting each data point
to the average of its neighboring data-point. It is summarized as below (Figs. 3 and 4):-
302 H. Das et al.

Fig. 3. Flowchart for the meanshift algorithm

Fig. 4. Tracking window sliding to get maximum mass density

Let the initial estimate be x. We then define a Gaussian Kernel Function K(xi − x)
where it is defined as

K(xi − x) = e−c(‖xi −x‖)

2
(1)

The weighted mean of the tracked object is given by

 Experimental Implementation of Fuzzy Vision-Based Tracking 303

∑
x ∈N(x) K(xi − x)xi
m(x) = ∑i (2)
xi ∈N(x) K(xi − x)

where xi are the pixel points of the image inside the tracking window. N(x) denotes the
neighboring points of x and K determines the weighted mean of the density in the
window. In [20], they defined the difference [x − m(x)] as meanshift and so the meanshift
algorithm repeats its estimation till m(x) converges to x.

3.3 Camshift Tracker

The Camshift Tracker [12, 21] stands for Continuously Adaptive Meanshift Tracker. It
is an advanced version of Meanshift Tracker with an adaptive region sizing step. Unlike
the meanshift tracker, it varies the size of the window until it converges. It uses the
colour information of the object in the form of histogram to track the object using adap‐
tive clustering method.
The camshift tracker uses the hue information to track the location and also find the
size of the object. After acquiring the image in the very frame from drone camera, the
current position and the size of the object is initialized. And then in subsequent frames,
the search window scans the area near the tracked object (from the previous frame) till
it converges to a maximum density of the desired pixel points. It is robust in nature as
it just uses the colour information which does not change with the motion of the object
unlike the Feature descriptors which changes with orientation. It is summarized as
below:-
1. We first calculate the moments
∑ ∑
M00 = I(x, y) M10 = xI(x, y)
x,y x,y

∑
M01 = yI(x, y) (3)
x,y

Here I = f(x, y) gives the probability density of the particular colour obtained after
backprojection at the point (x, y). Here M00 is zero order moment, M01 is the first order
moment about y axis and M10 is the first order moment about x axis
2. Then we obtain the centroid (xc, yc)

M10 M01
xc = , yc = (4)
M00 M00

The tracking window W is now centered about (xc, yc)

3. Change the width of the tracking window. As the maximum discrete probability
density of the greyscale is 256, so the relation between the window size w and zero
order moment is set as
304 H. Das et al.

√
M00
w=2 (5)
256
This algorithm is more robust as it can change the tracking size of the window with
the distance of the object from the camera.

4 Controller Design

In our experiment, we have used the Parrot AR Drone quadrotor. The complexity of this
drone model is simplified by the internal controller running in the drone which tries to
maintain the desired set-points of roll, pitch, yaw rate and vertical speed. The various
states of the drone such as translational speeds, angular velocity, acceleration, actual
angles can be subscribed through the ‘navdata’ topic of ardrone_autonomy package [11].
So we have designed a compatible outer loop controller such that it gives the desired
translational speed which is published to the A.R. Drone.
In our experiment, we have designed two different outer-loop controllers:- Fuzzy
Proportional Controller and Fuzzy Proportional-Derivative controller. Once the target
gets detected as discussed in the previous Sect. 3, the controller tries to reduce the
difference between the centroid of the tracked object and the centroid of the image plane
which is located at [320, 240] in pixels dimensions. The drone camera we are using has
a maximum pixel dimensions of 640 × 480. So the X-coordinate of the image plane
centroid is 640/2 = 320. Similarly the Y-coordinate is 240. For tracking purpose we
have used the linear velocity along the Y axis and Z axis in the body frame of reference
Vy and Vz.
The linear velocity in X axis of the drone (Fig. 1) and the angular velocity along the
Z axis of the body frame are kept zero which acts as its set-point for the internal
controller. They are denoted by Vx, Φz as shown in Fig. 5.

Vx = 0
Φz = 0

Drone
Output
Set-Point
(320, 240) +

Fig. 5. Controller diagram

This simple Image Based Visual Servoing technique to track the object is based upon
the local coordinate system of the camera plane. After getting the x and y coordinates
of the tracked object (XC, YC) we calculate its distance from the centroid of the camera
plane which acts as a set-point for the controller.
 Experimental Implementation of Fuzzy Vision-Based Tracking 305

Fig. 6. Input membership function for error in X axis

The fuzzy proportional controller takes the error as input and gives the required
velocity commands. Similarly the fuzzy proportional-derivative controller takes error
and the change in error as the inputs. We have designed 2 controllers for the errors in X
axis and Y axis of the camera plane. In our experiment, we implemented our algorithm
using the fuzzy logic toolbox in MATLAB.

4.1 Fuzzy Proportional Controller

As the scaling factor in a conventional controller is fixed, so the system is not very robust.
So, we applied a Fuzzy Proportional Controller which transforms the input error non-
linearly to the output command velocity. In our experiment, we have used a Mandani-
Type Fuzzy Inference System. We have used Gaussian Membership function for the
input variable as it gives a much smoother response. We have defined the Fuzzy Input
Sets as Negative Large (NL), Negative Medium (NM), Negative Small (NS), Zero (Z),
Positive Small (PS), Positive Medium (PM), Positive Large (PL). The range of the input
variable is [−240, 240] for the error in Y axis and [−320, 320] for the error in X axis as
the maximum number of pixels in Y axis and X axis of the camera plane is 480 and 640
respectively and we are calculating the error in pixel dimension. The output fuzzy set is
similar to the input fuzzy set but the range is [−0.25, 0.25]. The maximum range of the
output velocity command which can be published to the ardrone_autonomy node is [−1,
1]. As we were performing the experiments in an indoor environment so we chose a
much lower range for the fuzzy output which is [−0.25, 0.25]. The output obtained is
published to the ardrone_autonomy package to get the required movements. The rule
base is defined as:-
(i) if (the error is Negative Big), then (the controller output is Negative Big)
(ii) if (the error is Negative Medium), then (the controller output is Negative Medium)
(iii) if (the error is Negative Small), then (the controller output is Negative Small)
(iv) if (the error is Zero), then (the controller output is Zero)
(v) if (the error is Positive Small), then (the controller output is Positive Small)
(vi) if (the error is Positive Medium), then (the controller output is Positive Medium)
306 H. Das et al.

(vii) if (the error is Positive Big), then (the controller output is Positive Big)
The fuzzy inference systems calculates the activation of each membership function
and then aggregates the output using the max operation. It is followed by the deffuzifi‐
cation process using the centroid method.
The membership functions are shown below. As seen from Fig. 7, the membership
function is sparsely distributed in the range of −0.15 till 0.15 as the drone’s movement
is very sluggish in this range (Figs. 6 and 8).

Fig. 7. Output membership function for Vx

Fig. 8. Generated output surface

4.2 Fuzzy Proportional-Derivative Controller

The Fuzzy-PD Controller uses the error e (t) and the change in error {e (t) − e (t − 1)}
to calculate an output variable. The Fuzzy Input Set are defined as Negative (N), Zero
(Z) and Positive (P). The Fuzzy Output Set are defined as Negative Big (NB), Negative
Medium (NM), Negative Small (NS), Zero (Z), Positive Small (PS), Positive Medium
(PM) and Positive Big (PB). The range of the input error is same as the previous case,
whereas the range for the change in Error is [−480, 480] for the Y Axis and [−640, 640]
for the X Axis. The given dimensions are in pixels. We have used AND connector for
 Experimental Implementation of Fuzzy Vision-Based Tracking 307

combining the inputs and the centroid of area method to defuzzify the outputs. The
centroid method is given below.

∫X μA (x)xdx
OutputX =
∫X μA (x)dx

Here μA is aggregated Output Membership Function corresponding to the x- axis

coordinate of the membership graph (Figs. 9, 10, 11 and Table 1).

Fig. 9. Input membership function for error in X axis

Fig. 10. Input membership function for change in error in X axis

308 H. Das et al.

Fig. 11. Generated fuzzy surface for the output

Table 1. Rule base for Fuzzy PD system

Error Change in error
N Z P
N NB NM NS
Z PS Z NS
P PS PM PB

5 Experimental Results

We have conducted our experiments on the Parrot AR Drone. Here we compared the
outputs of the Fuzzy-P and the Fuzzy-PD controller. The vision system and the control
algorithm were running in a 64 bit Linux Processor which was acting as the ground
station. We first executed the C++ node for the camshift tracker using OpenCV library
and linked it to ROS using cv_bridge package [22]. It generates the coordinates of cent‐
roid of the tracked object. We have used the forward facing camera of the drone for
tracking the ball. It is then used to calculate the errors in the individual axis which in
turn is fed to controller. The output command for the velocity given to the Drone lies
between −0.25 to 0.25 which is published to the ardrone_autonomy [11] node. We used
MATLAB Robotics System Toolbox [18] to create a virtual master for ROS system. It
is also used for publishing and subscribing messages across various topics from the
drone. In our experiment we generated a similar random motion of the ball for both the
Fuzzy-P and Fuzzy-PD controllers to track it (Figs. 12, 13, 14, 15 and 16).
Experimental Implementation of Fuzzy Vision-Based Tracking 309

Fig. 12. Real time experimental setup

Fig. 13. Extracted histogram of the coloured ball

Fig. 14. Ball detection using camshift tracker

310 H. Das et al.

Fig. 15. Error in X axis of image plane vs. no of samples

Fig. 16. Error in Y axis of image plane vs. no of samples

As seen from the above diagram, the errors generated by the Fuzzy-P controller is
more unstable in nature compared to the Fuzzy-PD controller. This is because the Fuzzy-
PD controller takes into account the changes in error and tries to create a damping effect
for the quadrotor.
As already discussed, we have used the forward looking camera for tracking the ball.
So the velocity command along the Y axis of the body frame (quad-rotor frame of refer‐
ence) is used to minimize the error in X axis of the image plane. Similarly the velocity
command along the Z axis of the body frame of reference is used for the error in Y axis
of image plane. The coordinate system of the quadrotor is shown in Fig. 1.
The velocity commands generated by the controller are sent to the Parrot AR Drone
using a publisher node being run in MATLAB. As seen from Figs. 17 and 18, the sudden
spike in the velocity command are present in the Fuzzy-P controller, but the Fuzzy-PD
controller tries to reduce the effect of sudden vibrations in the motion.
 Experimental Implementation of Fuzzy Vision-Based Tracking 311

Fig. 17. Velocity command along Y axis of body frame

Fig. 18. Velocity command along Z axis of body frame

The actual velocity provided by the drone is highly prone to noises and moreover
the velocity state in body frame are obtained by integrating their accelerations. So to
compare the output of these two controllers, we plotted their accelerations which is
acquired from the navdata topic. The acceleration provided by the Fuzzy-PD controller
is more stable and linear in nature compared to that of a fuzzy-P controller as seen from
Figs. 19 and 20.
312 H. Das et al.

Fig. 19. Acceleration along Y axis of body-frame

Fig. 20. Acceleration along Z axis of body-frame

6 Conclusion

In this experiment, a visual servoing mechanism has been implemented on a quadrotor

to track an object on the basis of its hue histogram. We have implemented the camshift
tracker for tracking a coloured ball. We have also designed and compared the outputs
of Fuzzy-P and Fuzzy-PD controller. Both the methods were successful in tracking the
ball but the fuzzy-pd controller was more efficient than the fuzzy-p controller as besides
minimizing the error, it also tries to minimize the changes in error which creates a
damping effect. In the future, we are expanding it for 3D tracking of objects using image
depth and also implementing a neural network based object tracking system.

References

1. Engel, J., Sturm, J., Daniel, C.: Camera-based navigation of a low-cost quadrocopter. In: 2012
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 2815–
2821. IEEE (2012)
 Experimental Implementation of Fuzzy Vision-Based Tracking 313

2. Pestana, J., Sanchez-Lopez, J.L., Campoy, P., Saripalli, S.: Vision based gps-denied object
tracking and following for unmanned aerial vehicles. In: 2013 IEEE International Symposium
on Safety, Security, and Rescue Robotics (SSRR), pp. 1–6. IEEE (2013)
3. Arifin, F., Daniel, R.A., Widiyanto, D.: Autonomous detection and tracking of an object
autonomously using AR.Drone quadcopter. Jurnal Ilmu Komputer dan Informasi 7(1), 11–17
(2014)
4. Selby, W., Corke, P., Rus, D.: Autonomous aerial navigation and tracking of marine animals.
In: Proceedings of the Australian Conference on Robotics and Automation (ACRA) (2011)
5. Ferreira, F.T., Cardoso, J.S., Oliveira, H.P.: Video analysis in indoor soccer using a quadcopter
(2015)
6. Gomez-Balderas, J.-E., Flores, G., Carrillo, L.R.G., Lozano, R.: Tracking a ground moving
target with a quadrotor using switching control. J. Intell. Rob. Syst. 70(1–4), 65–78 (2013)
7. Razinkova, A., Cho, H.-C.: Tracking a moving ground object using quadcopter UAV in a
presence of noise. In: 2015 IEEE International Conference on Advanced Intelligent
Mechatronics (AIM), pp. 1546–1551. IEEE (2015)
8. Ross, S., Melik-Barkhudarov, N., Shankar, K.S., Wendel, A., Dey, D., Bagnell, J.A., Hebert,
M.: Learning monocular reactive UAV control in cluttered natural environments. In: 2013
IEEE International Conference on Robotics and Automation (ICRA), pp. 1765–1772. IEEE
(2013)
9. Jayatilleke, L., Zhang, N.: Landmark-based localization for unmanned aerial vehicles. In:
2013 IEEE International Systems Conference (SysCon), pp. 448–451. IEEE (2013)
10. Nath, A., Samantaray, J., Chaudhury, S., Roy, B.K.: Magnetic ball levitation system control
using sliding mode control and fuzzy PD+I control: a comparative study. In: International
Conference on Energy, Power and Environment, Shillong, India, 12–13 June 2015
11. Monajjemi, M.: ardrone autonomy: a ROS driver for ardrone 1.0 & 2.0 (2012). Autonomy
Lab, Simon Fraser University. https://github.com/AutonomyLab/ardrone_autonomy
12. Yuan, X.: Continuously adaptive mean SHIFT. http://eric-yuan.me/continuously-adaptive-
shift
13. Parrot A.R. Drone 2.0. https://en.wikipedia.org/wiki/Parrot_AR.Drone
14. Olivares-Mendez, M.A., Mondragon, I., Campoy Cervera, P., Mejias, L., Martinez, C.: Aerial
object following using visual fuzzy servoing. In: Proceedings of the 1st Workshop on
Research, Development and Education on Unmanned Aerial Systems (RED-UAS 2011), pp.
61–70. Centro Avanzado de Tecnologías Aeroespaciales (CATEC) (2011)
15. Robot Operating System Documentation. http://wiki.ros.org/
16. Mean Shift Algorithm. https://en.wikipedia.org/wiki/Mean_shift
17. Opencv Documentation. http://docs.opencv.org/2.4.12/
18. Matlab Robotics System Toolbox. http://in.mathworks.com/products/robotics/
19. Histogram Backprojection documentation. http://docs.opencv.org/2.4/doc/tutorials/imgproc/
histograms/back_projection/back_projection.html
20. Fukunaga, K., Hostetler, L.D.: The estimation of the gradient of a density function, with
applications in pattern recognition. IEEE Trans. Inf. Theory 21(1), 32–40 (1975)
21. Allen, J.G., Xu, R.Y.D., Jin, J.S.: Object tracking using camshift algorithm and multiple
quantized feature spaces. In: Proceedings of the Pan-Sydney Area Workshop on Visual
Information Processing, pp. 3–7. Australian Computer Society, Inc. (2004)
22. cv_bridge Documentation. http://wiki.ros.org/cv_bridge
 Fuzzy Fractional Derivative: A New Definition

Marzieh Najariyan1, Mehran Mazandarani2(&),

and Valentina Emilia Balas3
1
Department of Applied Mathematics,
Ferdowsi University of Mashhad, Mashhad, Iran
[email protected]
2
Department of Electrical Engineering,
Ferdowsi University of Mashhad, Mashhad, Iran
[email protected], [email protected]
3
Department of Automatics and Applied Software,
Faculty of Engineering, University Aurel Vlaicu, Arad, Romania
[email protected]

Abstract. In this paper, a new definition of fuzzy fractional derivative is pre-

sented. The definition does not have the drawbacks of the previous definitions of
fuzzy fractional derivatives. This definition does not necessitate that the diam-
eter of the fuzzy function be monotonic, and it does not refer to derivative of
order greater than the one that is considered. Moreover, the fractional derivative
of a fuzzy constant is zero based on the definition. Restrictions associated to
Caputo-type fuzzy fractional derivatives are expressed. Additionally, general-
ized Hukuhara difference and generalized difference of perfect type-2 fuzzy
numbers are defined. Furthermore, using two examples the advantages of the
new definition compared with the others are borne out.

Keywords: Fuzzy fractional derivative
Modified Riemann-Liouville

derivative
Caputo-type derivative
Type-2 fuzzy numbers

1 Introduction

“Can the meaning of derivatives with integer orders be generalized to derivatives with
non-integer orders?” (Leibniz 1695). Leibniz might not even have imagined that a
question similar to his would be posed 300 years later: “Can the meaning of fuzzy
derivatives with integer orders be generalized to fuzzy derivatives with non-integer
orders?”.
Fuzzy fractional calculus has increasingly attracted researchers’ attention in the past
years including:
(1) Definition of fuzzy fractional derivatives and integrals and examining their fea-
tures. See [1–4].
(2) Solving fuzzy fractional differential equations. See [5–8].
(3) Application of fuzzy fractional calculus in real world. See [9, 10].

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_25
 Fuzzy Fractional Derivative: A New Definition 315

Since the last two cases have to do with the type of derivative and integral; and how
they are defined, defining fuzzy fractional integral and derivative are of great impor-
tance. In this paper, we restrict our focus on fuzzy fractional derivatives. Before, dealing
with fuzzy fractional derivatives, it is to pinpoint that the most widely known and
applied fuzzy derivatives include Hukuhara derivative (H  derivative) [11], strongly
generalized Hukuhara derivative (SGH  derivative) [12], generalized Hukuhara
derivative (gH  derivative) [13], and generalized derivative (g  derivative) [13]
among which g  derivative is the most general, i.e. other derivatives are special forms
of g  derivative. Fuzzy fractional derivatives, presented to date, are restricted to fuzzy
Riemann-Liouville fractional derivative [14], Caputo-type fuzzy fractional derivatives
based on Type-1 Fuzzy Sets (T1FSs) [15–17], and Type-2 Fuzzy Sets (T2FSs) [9], see
Fig. 1.

Fig. 1. The brief background of the development of fuzzy derivatives.

316 M. Najariyan et al.

In crisp context Riemann-Liouville fractional derivative of a constant is not zero;

neither is fuzzy Riemann-Liouville fractional derivative. Likewise, in fuzzy fractional
differential equations in which the derivative is regarded as fuzzy Riemann-Liouville
fractional derivative, the initial conditions are considered as fractional derivatives
unlike the ordinary differential equations. Such a derivative comes with another
downside: if fuzzy function is differentiable in the sense of fuzzy Riemann-Liouville
fractional derivative, then the length of the closure of the support of the fuzzy function
is compulsory to be monotonic. That is, the diameter of the fuzzy function,
Diamðf ðxÞÞ ¼ f a ðxÞ  f a ðxÞ, must be monotonic.
First presented by Mazandarani and Kamyad [15], Caputo-type fuzzy fractional
derivative that is based on Hukuhara difference necessitates that the length of the
closure of the support of the fuzzy function be non-decreasing. Additionally, the dif-
ferentiability of the order b, 0\b\1, of the fuzzy function depends on the first order
differentiability of the function. The aforementioned drawbacks associated with
Caputo-type fuzzy fractional derivative are also true of Caputo-type fuzzy fractional
derivative defined based on type-2 fuzzy sets [9]. This paper implies that some results
obtained based on Caputo-type fuzzy fractional derivative that are presented in [9, 15]
are incorrect. This paper then is about to modify these definitions so that it would
comply with the obtained results.
In [15, 16], Caputo-type fuzzy fractional derivative of order b, 0\b\1, was
defined based on SGH  derivative. This fuzzy fractional derivative has two restric-
tions: 1. It necessitates that the length of the closure of the support of the fuzzy function
be monotonic; 2. If the fuzzy function is not SGH  differentiable, then it is not
Caputo-type fuzzy fractional differentiable either.
Based on gH  derivative, Caputo-type fuzzy fractional derivative was presented in
[17]. In spite of the fact that this derivative, unlike the previous definitions, does not
necessitate the monotonicity of the diameter of the fuzzy function, the differentiability
of the order b, 0\b\1, of the fuzzy function depends on the first order gH 
differentiability of the function.
Because the generalizing a concept or definition implies the maintenance of its
characteristics in the generalized version, none of the fuzzy fractional derivatives
presented so far could be an appropriate answer to the question mentioned in the
beginning of this paper. Some of the features of fuzzy derivative with integer order are:
(1) Derivative of a fuzzy constant is zero.
(2) Existence of a derivative of a fuzzy function is not dependent the existence of a
derivative of the function in higher order.
(3) In general, it does not necessitate that the diameter of the fuzzy function be
monotonic.
This paper aims at presenting a definition of fuzzy fractional derivative named
modified Riemann-Liouville fuzzy fractional derivative that does not have the draw-
backs of the previous definitions. In other words, it inherits the features of fuzzy
derivative mentioned above. In 2006, Jumarie presented a definition of fractional
 Fuzzy Fractional Derivative: A New Definition 317

derivative called modified Riemann-Liouville derivative [18] that lacked the restric-
tions of Riemann-Liouville and Caputo-type derivative. In other words, in this defi-
nition the derivative of order b, 0\b\1, of a constant is zero; moreover, the definition
does not refer to derivative of order greater than the one that is considered. Further-
more, like initial conditions considered in ordinary differential equations, the initial
conditions are considered in fractional differential equations, in which the derivative is
regarded as Jumarie’s definition. The definition of modified Riemann-Liouville fuzzy
fractional derivative is based on an integration of Jumarie’s definition of fractional
derivative, g  difference [13], and g  derivative.

2 Basic Concepts

This section presents a theorem and some definitions which will be used in this paper.
Throughout this paper, the set of all real numbers is denoted by R, the set of all type-1
fuzzy numbers (T1FNs) on R by E1 ; and the set of all perfect type-2 fuzzy numbers
(T2FNs) on R by E2 . The well-known a  level set of the fuzzy set A is denoted by Aa .
The right and left end-points of Aa are denoted by Aa and Aa , respectively. The
triangular T1FN u 2 E1 is characterized using triple ða; b; cÞ, a  b  c, as u ¼ ða; b; cÞ.
Definition 2.1. Let u; v 2 E1 . If there exists a w 2 E1 such that u ¼ v þ w, then w is
called Hukuhara difference (H  difference) of u and v; and is denoted by u H v.
Definition 2.2 [13]. Let u; v; w 2 E1 . The generalized Hukuhara difference (gH
difference) of u and v is defined as:
(
gH u ¼ vþw
u v¼w, or ð2:1Þ
u þ ðwÞ ¼ v

Definition 2.3 [13]. The generalized difference (g  difference) of two type-1 fuzzy
numbers u and v is given by its a  level sets as
!
h g ia [  c gH c 
u v ¼ cl ½u ½ v ; a 2 ½0; 1 ð2:2Þ
ca

where clð
Þ denotes the closure of the set.

Note 2.1. In [19], it has been shown that the g  difference between two type-1 fuzzy
numbers is not always a type-1 fuzzy number. Then, a modified version of g 
difference was proposed in [19]. However, in this paper we consider g  difference
based on Definition 2.3.
Definition 2.4 [20]. Let ~u; ~v 2 E2 . If there exists a w
~ 2 E2 such that ~
u ¼ ~v þ w
~ , then w
~
is called type-2 Hukuhara difference (H2  difference) of ~ u and ~v; and is denoted by
~u H2 ~v.
318 M. Najariyan et al.

Definition 2.5. Let ~u; ~v; w~ 2 E2 . The type-2 generalized Hukuhara difference (gH2 
difference) of ~u and ~v is defined as:
(
gH2 ~u ¼ ~v þ w
~
~u ~v ¼ w
~, or ð2:3Þ
~u þ ð~ wÞ ¼ ~v

Definition 2.6 [12]. Let f : ða; bÞR ! E1 be a fuzzy function and x0 2 ða; bÞ,
then:f ð xÞ is strongly generalized Hukuhara differentiable (SGH  differentiable) at the
0
point x0 , in the first form, if there exists an element fSGH ðx0 Þ 2 E1 such that for h [ 0
sufficiently near zero, there exist f ðx0 þ hÞ f ðx0 Þ; f ðx0 Þ H f ðx0  hÞ and the limits
H

f ðx0 þ hÞ H f ðx0 Þ f ð x 0 Þ H f ð x 0  hÞ 0
lim ¼ lim ¼ fSGH ð x0 Þ ð2:4Þ
h!0 h h!0 h

or
f ð xÞ is SGH  differentiable at the point x0 , in the second form, if there exists an
0
element fSGH ðx0 Þ 2 E1 such that for h [ 0 sufficiently near zero, there exist
f ðx0 Þ f ðx0 þ hÞ; f ðx0  hÞ H f ðx0 Þ and the limits
H

f ðx0 Þ H f ðx0 þ hÞ f ð x 0  hÞ H f ð x 0 Þ 0
lim ¼ lim ¼ fSGH ð x0 Þ ð2:5Þ
h!0 h h!0 h

Definition 2.7 [13]. Let f : ða; bÞR ! E1 be a fuzzy function and x0 2 ða; bÞ, then
f ð xÞ is generalized Hukuhara differentiable (gH  differentiable) at the point x0 if there
0
exists an element fgH ðx0 Þ 2 E1 such that for h sufficiently near zero the following limit
exists

f ðx0 Þ gH f ðx0 þ hÞ 0
lim ¼ fgH ð x0 Þ ð2:6Þ
h!0 h
h i
Theorem 2.1 [13]. Let f : ða; bÞR ! E1 be such that ½f ð xÞa ¼ f a ð xÞ; f a ð xÞ .
Suppose that the real-valued functions f a ð xÞ; f a ð xÞ are differentiable with respect to x,
uniformly in a 2 ½0; 1. Then, the function f ð xÞ is gH  differentiable at a fixed x 2
ða; bÞ if and only if one of the following two cases holds:
(1) f 0a ð xÞ; f 0a ð xÞ are increasing and decreasing as functions of a, and f 01 ð xÞ; f 01 ð xÞ,
(2) f 0a ð xÞ; f 0a ð xÞ are decreasing and increasing as functions of a, and f 01 ð xÞ; f 01 ð xÞ.
Furthermore,
h ia h n o n oi
0
fgH ð xÞ ¼ min f 0a ð xÞ; f 0a ð xÞ ; max f 0a ð xÞ; f 0a ð xÞ ð2:7Þ

df a ð xÞ df a ð xÞ
where f 0a ð xÞ and f 0a ð xÞ mean dx and dx , respectively.
 Fuzzy Fractional Derivative: A New Definition 319

3 Restrictions Associated to Caputo-Type Fuzzy Fractional

Derivatives

In this section, three definitions of Caputo-type fuzzy fractional derivatives presented in

[15–17] are recalled. Then, the restrictions of each definition are expressed.
Definition 3.1 [15]. Let f : ½a; bR ! E1 be continuous and Lebesque integrable
Rx f ðtÞHf ðaÞ
fuzzy-valued function, GðxÞ ¼ Cð1bÞ
1
ðxtÞb
dt. Function f ð xÞ is the Caputo-type
a
fuzzy fractional differentiable of order 0\b\1 at x0 2 ða; bÞ if there exists an element
c b
D f ðx0 Þ 2 E1 such that for h [ 0 sufficiently near zero the following limits exist

Gðx0 þ hÞ H Gðx0 Þ Gðx0 Þ H Gðx0  hÞ c b

lim ¼ lim ¼ D f ð x0 Þ ð3:1Þ
h!0 h h!0 h

or

Gðx0 Þ H Gðx0 þ hÞ Gðx0  hÞ H Gðx0 Þ c b

lim ¼ lim ¼ D f ð x0 Þ ð3:2Þ
h!0 h h!0 h

Note 3.1. The Caputo-type fuzzy fractional derivative based on type-2 fuzzy sets
which has been presented in [9] is similar to Definition 3.1 except that the difference
used therein is H2  difference.
Remark 3.1. If fuzzy function f ð xÞ is Caputo-type fuzzy fractional differentiable based
on Definition 3.1, then there exists f ðtÞ H f ðaÞ, t 2 ½a; xÞ,which implies that f ð xÞ has
non-decreasing length of the closure of the support. Hence, the results obtained in [15]
which correspond to function with non-increasing length of the closure of the support
are not true. There is a similar reason which can be stated for the Caputo-type fuzzy
fractional derivative based on type-2 fuzzy sets, Definition 3.6 in [9]. Moreover, these
definitions are based on the assumption of differentiability of fuzzy function.
In the following, the corrected version of Definition 3.1 is presented.
Definition 3.2. Let f : ½a; bR ! E1 and its derivative be continuous and Lebesque
Rx f ðtÞgHf ðaÞ
integrable fuzzy-valued function, GðxÞ ¼ Cð1bÞ
1
ðxtÞb
dt. The function f ð xÞ is the
a
Caputo-type fuzzy fractional differentiable of order 0\b\1 at x0 2 ða; b if there
exists an element c Db f ðx0 Þ 2 E1 such that for h sufficiently near zero the following
limit exists

Gðx0 þ hÞ gH Gðx0 Þ c b
lim ¼ D f ð x0 Þ ð3:3Þ
h!0 h

Theorem 3.1. Let f : ½a; bR ! E1 be hCaputo-type ifuzzy fractional differentiable in

the sense of Definition 3.2 and ½f ð xÞa ¼ f a ð xÞ; f a ð xÞ . Then,
320 M. Najariyan et al.

c a h i
Db f ð xÞ ¼ c Db f a ð xÞ; c Db f a ð xÞ ð3:4Þ

 a h i
provided that f ðtÞ gH f ðaÞ ¼ f a ðtÞ  f a ðaÞ; f a ðtÞ  f a ðaÞ ; 8t 2 ða; b
or
c a h i
Db f ð xÞ ¼ c Db f a ð xÞ; c Db f a ð xÞ ð3:5Þ

 a h i
provided that f ðtÞ gH f ðaÞ ¼ f a ðtÞ  f a ðaÞ; f a ðtÞ  f a ðaÞ ; 8t 2 ða; b, where
D f ð xÞ and c Dbf a ð xÞ mean the well-known Caputo-type derivative of real-valued
c b a

functions f a ð xÞ and f a ð xÞ, respectively.

Proof. The proof is straightforward and hence omitted.
Theorem 3.2. Let f : ½a; bR ! E1 be hCaputo-type ifuzzy fractional differentiable in
the sense of Definition 3.2 and ½f ð xÞa ¼ f a ð xÞ; f a ð xÞ . Then, it is called Caputo-type
h ia h i
0
fuzzy fractional derivative in the first form if fgH ð xÞ ¼ f 0a ð xÞ; f 0a ð xÞ , 8x 2 ½a; b.
Moreover,
c a h i
Db f ð xÞ ¼ c Db f a ð xÞ; c Db f a ð xÞ ð3:6Þ

and it is called Caputo-type fuzzy fractional derivative in the second form if

h ia h i
0
fgH ð xÞ ¼ f 0a ð xÞ; f 0a ð xÞ ; 8x 2 ½a; b. Moreover,

c a h i
Db f ð xÞ ¼ c Db f a ð xÞ; c Db f a ð xÞ ð3:7Þ

Note 3.2. The solution of Example 1 in [15] based on Definition 3.1 is incorrect,
whereas it is correct based on Definition 3.2 and Theorem 3.2.
Note 3.3. The Caputo-type fuzzy fractional derivative based on type-2 fuzzy sets
presented in [9] can be corrected in the same fashion as in Definition 3.1. The corrected
definition and theorem are left to the appendix.
Definition 3.3 [16]. Let f : ½a; bR ! E1 be continuous and Lebesque integrable
fuzzy-valued function. Then f ð xÞ is said to be Caputo’s H-differentiable when

Zx 0
b 1 fSGH ðtÞ
c
D f ð xÞ ¼ dt 0\b\1 ; x[a ð3:8Þ
Cð1  bÞ ðx  tÞb
a
 Fuzzy Fractional Derivative: A New Definition 321

0
Due to the fact that SGH  differentiability, fSGH ðxÞ, necessitates that the fuzzy
function f ð xÞ must have monotonic length of the closure of the support, Caputo’s
H-differentiability introduced in Definition 3.3 also does the same. Furthermore, it is
easy to see that if the fuzzy function f ð xÞ is not SGH  differentiable, then it is not
Caputo’s H  differentiable either.
0
Definition 3.4 [17]. Let f : ½a; bR ! E1 and fgh ð xÞ be continuous and Lebesque
integrable fuzzy-valued function. The fuzzy gH  fractional Caputo differentiability of
fuzzy-valued function f ð xÞ of order 0\b\1 is defined as:

Zx 0
b 1 fgH ðtÞ
c
D f ð xÞ ¼ dt 0\b\1 ; x[a ð3:9Þ
Cð1  bÞ ðx  tÞb
a

Although the fuzzy gH  fractional Caputo differentiability does not have some
restrictions of Definitions 3.1 and 3.3, it may exist provided that the fuzzy function f ð xÞ
is gH  differentiable on the interval ½a; x, x 2 ða; b.
Readers should bear in mind that presenting a new definition of derivative, as the
generalization of another one, means that the new definition inherits the properties of
previous definition of derivative. Accordingly, inheriting properties of fuzzy derivative
with integer order by fuzzy fractional derivative means that fuzzy fractional derivative
necessitates the following requirements: 1. The fuzzy function must not have neces-
sarily monotonic length of the closure of the support; 2. The fuzzy fractional derivative
must not rely on differentiability of function in higher orders.

4 Modified Riemann-Liouville Fuzzy Fractional Derivative

In this section, it is attempted to present a new definition of fuzzy derivative with

non-integer order as the generalized definition of fuzzy derivative with integer order.
In 2006, Jumarie introduced a new definition of derivative with non-integer order
which is called modified Riemann-Liouville derivative [18]. This definition possesses
some advantages; it does not refer to derivative of the order greater than the one that is
considered; additionally, it removes the effect of initial value of the function. The
Jumarie’s definition is presented below.
Definition 4.1 [18]. Let f : ½0; bR ! R be continuous and Lebesque integrable
real-valued function. The modified Riemann-Liouville fractional derivative of function
f ð xÞ of order b; 0\b\1, is denoted by mRL Db f ð xÞ and defined as:

Zx
b 1 d f ðtÞ  f ð0Þ
mRL
D f ð xÞ ¼ dt 0\b\1 ; x[0 ð4:1Þ
Cð1  bÞ dx ðx  tÞb
0
322 M. Najariyan et al.

The definition of modified Riemann-Liouville fuzzy fractional derivative based on

an integration of Definition 4.1, g  difference, and g  derivative is introduced below.
Definition 4.2. Let f : ½0; bR ! E1 be continuous and Lebesque integrable
Rx f ðtÞgf ð0Þ
fuzzy-valued function, GðxÞ ¼ Cð1bÞ
1
ðxtÞb
dt. The function f ð xÞ is modified
0
Riemann-Liouville fuzzy fractional differentiable of order 0\b\1 at x0 2 ð0; b if
there exists an element mRL Db f ðx0 Þ 2 E1 such that for h sufficiently near zero the
following limit exists

Gðx0 Þ g Gðx0 þ hÞ mRL b

lim ¼ D f ð x0 Þ ð4:2Þ
h!0 h

Whereas Definitions 3.1 to 3.4, modified Riemann-Liouville fuzzy fractional dif-

ferentiability of a fuzzy function does not necessitate that the fuzzy function be dif-
ferentiable of the order greater than the one that is considered. Moreover, modified
Riemann-Liouville fuzzy fractional derivative of order b; 0\b\1; of a constant is
zero.
Theorem 4.1. Let f : ½0; bR ! E1 be differentiable in thehsense of fuzzy i modified
a a a
Riemann-Liouville fractional differentiability, and ½f ð xÞ ¼ f ð xÞ; f ð xÞ . Then, for
x 2 ð0; b
mRL a h i
Db f ð xÞ ¼ mRL Db f a ð xÞ; mRL Db f a ð xÞ ð4:3Þ

 a h a
i
provided that f ðtÞ g f ðaÞ ¼ f a ðtÞ  f a ð0Þ; f a ðtÞ  f ð0Þ ; 8t 2 ð0; b,
or
mRL a h i
Db f ð xÞ ¼ mRL Db f a ð xÞ; mRL Db f a ð xÞ ð4:4Þ

 a h i
provided that f ðtÞ g f ð0Þ ¼ f a ðtÞ  f a ð0Þ; f a ðtÞ  f a ð0Þ ; 8t 2 ð0; b, where
D f ð xÞ and mRL Dbf a ð xÞ mean the well-known modified Riemann-Liouville frac-
mRL b a

tional derivative of real-valued functions f a ð xÞ and f a ð xÞ, respectively.

Proof. The proof is straightforward and hence omitted.
Definition 4.3. Let ~f : ½0; bR ! E2 be continuous and Lebesque integrable type-2
Rx ~f ðtÞg2~f ð0Þ
~
fuzzy number-valued function, GðxÞ ¼ Cð1bÞ
1
ðxtÞb
dt. The function ~f ð xÞ is modi-
0
fied Riemann-Liouville type-2 fuzzy fractional differentiable of order 0\b\1 at x0 2
ð0; b if there exists an element mRL Db~f ðx0 Þ 2 E2 such that for h sufficiently near zero
the following limit exists
 Fuzzy Fractional Derivative: A New Definition 323

~ ð x0 Þ g 2 G
G ~ ð x 0 þ hÞ
lim ¼ mRL Db~f ðx0 Þ ð4:5Þ
h!0 h

where g2  difference of G ~ ðx0 Þ and G~ ðx0 þ hÞ, G ~ ðx0 þ hÞ, is similar to Defi-
~ ðx0 Þ g2 G
nition 2.3 and hence it is left to the appendix.
Table 1 embodies some fuzzy derivatives with their characteristics which have
been presented till now. The table tries to shed light on the answer that we are seeking
for the main question.

5 Example

In this section, using two purposeful examples we bear out the advantages of the new
definition of fuzzy fractional derivative compared with the others.

Table 1. Comparison of some features of fuzzy derivatives with integer and non-integer order
Fuzzy derivatives The difference The status of Dependency on the Derivative
used in diameter of fuzzy differentiability of of a
derivative function higher order constant
Integer H  derivative H  difference Non-decreasing No Zero
order gH  derivative gH  difference Monotonicity is No Zero
not necessary
g  derivative g  difference Monotonicity is No Zero
not necessary
Fractional Riemann-Liouville H  difference Monotonic No Not Zero
order Definition 3.1 H  difference Non-decreasing Yes Zero
Definition 3.2 gH  difference Monotonicity is Yes Zero
not necessary
Definition 3.3 H  difference Monotonic Yes Zero
Definition 3.4 gH  difference Monotonicity is Yes Zero
not necessary
Modified g  difference Monotonicity is No Zero
Riemann-Liouville not necessary

Example 1. Consider the following triangular fuzzy function which has been adopted
from [21]:
 3 3 
x x 2x3
f ð xÞ ¼ ; þ x þ 3; þ4 x 2 ½0; 2 ð5:1Þ
3 3 3

Figure 2 represents the function f ð xÞ. The derivatives of the right and left
end-points of the a  level sets of f ð xÞ are as
324 M. Najariyan et al.

(
df a ðxÞ
dx ¼ ð2  aÞx þa
2
df a ðxÞ x 2 ½0; 2 ; 0  a  1 ð5:2Þ
dx ¼ x þ a
2

a df a ðxÞ
The functions d fdxðxÞ and dx have been shown in Fig. 3. As is seen, the function f ð xÞ
is neither gH  differentiable nor SGH  differentiable in the interval ½0; 1. Then fuzzy
fractional derivative does not exist in the sense of any of Definitions 3.1 to 3.4 exists.
However, the modified Riemann-Liouville fuzzy fractional derivative of the
function exists, and it can be obtained as
8
>
>
> D f ð xÞ ¼ 3CCð4b
mRL b a ð 4Þ
Þx
3b a
þ Cð2b Þx
1b
0x2
< qffiffiffiffiffiffiffiffiffiffiffiffiffi
D f ð xÞ ¼ 3CCð4b
mRL b a ð4Þ
Þx
3b
þ Cð2b
1
Þx
1b
0\x  2C Cð4bÞ
ð2bÞ ð5:3Þ
>
> qffiffiffiffiffiffiffiffiffiffiffiffiffi
>
: mRL Db f a ð xÞ ¼ ð2aÞCð4Þ x3b þ a x1b Cð4bÞ
3Cð4bÞ Cð2bÞ 2  x [ 2Cð2bÞ

which has been illustrated in Fig. 4 for b ¼ 34.

2
Example 2. Consider the function f ð xÞ ¼ kðx  1Þ3 ; x 2 ½0; 2; in which k ¼ ð1; 2; 3Þ
is the triangular fuzzy number. Figure 5 shows the function with its a  level sets. The
modified Riemann-Liouville fuzzy fractional derivative of order b ¼ 13 of the function
is as follows:

 
x3 x3 3
Fig. 2. Representation of the a–level sets of fuzzy function f ð xÞ ¼ 3 ; 3 þ 3; 2x3 þ 4
considered in example 1. The blue and red lines show left and right end-points of the a–level
sets of the function, respectively, and the black line corresponds to the level a = 1.
 Fuzzy Fractional Derivative: A New Definition 325

Fig. 3. The derivatives of the left and right end-point of the a– level sets of fuzzy function f ðxÞ
in Example 1. The blue and red lines showleft and right end-point of the a–level sets,
respectively, and the black line corresponds to the level a = 1.

Fig. 4. The a– level sets of modified Riemann-Liouville fuzzy fraction derivative of order b ¼ 34
of fuzzy f ðxÞ function in example1. The blue and red lines show left and right end-point of the a–
level sets, respectively, and the black lines corresponds to the level a = 1
326 M. Najariyan et al.

Fig. 5. Representation of the a–level sets of fuzzy function f ð xÞ ¼

 2 2 2

ðx  1Þ3 ; 2ðx  1Þ3 ; 3ðx  1Þ3 considered in Example 2. The blue and red lines show left
and right end-point of the a– level sets of the function, respectively, and the black line
corresponds to the level a = 1.

8
> ð1 þ aÞCð53Þ 1
< D3 f a ð xÞ ¼
1
mRL
Cð43Þ
ð x  1Þ 3 0x2
ð5:4Þ
>
: mRL D13 f a ð xÞð xÞ ¼ ð3aÞC ðÞ 5 1

Cð43Þ
ð x  1Þ
3 3 0x2

Figure 6 shows functions mRL D3 f a ð xÞ and mRL D3 f a ð xÞ. It should be noted that none
1 1

of Definitions 3.1 to 3.4 is capable of obtaining the fractional derivative of the function.

Fig. 6. The a– level sets of modified Riemann-Liouville fuzzy fraction derivative of order b ¼ 13
of fuzzy function f ðxÞ in Example 2. The blue and red lines show left and right end-point of the
a–level sets, respectively, and the black lines corresponds to the level a = 1
 Fuzzy Fractional Derivative: A New Definition 327

6 Conclusions

In this paper, the proposed definitions in [9, 15] have been modified such that so as to
conform with the results obtained in these papers. Generalized type-2 Hukuhara dif-
ference (gH2  difference), type-2 generalized difference (g2  difference), and mod-
ified Riemann-Liouville fuzzy fractional derivative based on type-1 and type-2 fuzzy
sets were also defined. Moreover, restrictions associated with fuzzy fractional deriva-
tives presented so far were mentioned. To handle such restrictions, new definitions of
fuzzy derivative with non-integer order for type-1 and type-2 fuzzy number-valued
functions were set forth.
To obtain fuzzy fractional derivative more easily, a useful theory concerning the
Definition 4.2 was offered as well. Drawing on two examples, it was demonstrated how
effective and capable modified Riemann-Liouville fuzzy fractional derivative can be,
compared to other derivatives.
This question may now arise as to whether the meaning of fuzzy derivative with
integer orders could successfully be generalized to fuzzy derivative with non-integer
orders or not. On the one hand, unlike the previous definitions, modified Riemann-
Liouville fuzzy fractional derivative could maintain some of the features of fuzzy
derivative with integer order. On the other hand, if the function is not continuous at the
origin, e.g. Lnð xÞ, obtaining its fractional derivative using modified Riemann-
Liouville derivative may not be possible [22]. This is directly inherited by modified
Riemann-Liouville fuzzy fractional derivative too. Accordingly, the research could
claim to have reached this goal, to a great extent. Nonetheless, more extensive research is
to be conducted. The future works which can be done in this area are 1. Defining the
fractional fuzzy derivative based on granular derivative [23], 2. Investigating type-2
fuzzy fractional differential equations, on the space of triangular perfect quasi type-2
fuzzy numbers [24], based on the new definition presented in this paper, 3. Examining
fuzzy fractional dynamical systems optimal control problem similar to what has been
done in [25, 26].

Acknowledgments. The authors really appreciate the helpful comment and enthusiastic support
of Professor Guy Jumarie, from Department of Mathematics, University of Québec at Montréal,
Montréal, Canada.

Appendix

Definition 1. The generalized difference of two perfect type-2 fuzzy numbers ~u and ~v
~ 2 E2 which is given by its a  level and ~
(g2  difference) is w a  plane [9] sets as
follows

h g ia [  !
gH
~ a~a ¼ ~u ~v ¼ cl ½~uc~a ½~vc~a
2 2
½w ; a; ~
a 2 ½0; 1 ð1Þ
~a
ca
328 M. Najariyan et al.

Definition 2. Let ~f : ½a; bR ! E2 and its derivative be continuous and Lebesque
Rx ~f ðtÞgH2~f ðaÞ
~
integrable fuzzy-valued function, GðxÞ ¼ Cð1bÞ
1
ðxtÞb
dt. The function ~f ð xÞ is the
a
Caputo-type fuzzy fractional differentiable of order 0\b\1 at x0 2 ða; b if there
exists an element c Db~f ðx0 Þ 2 E2 such that for h sufficiently near zero the following
limit exists

~ ðx0 þ hÞ gH2 G
G ~ ðx0 Þ
lim ¼ c Db~f ðx0 Þ ð2Þ
h!0 h

References
1. Lupulescu, V.: Fractional calculus for interval-valued functions. Fuzzy Sets Syst. 265, 63–85
(2015)
2. Malinowski, M.T.: Random fuzzy fractional integral equations – theoretical foundations.
Fuzzy Sets Syst. 265, 39–62 (2015)
3. Van Hao, N.: Fuzzy fractional functional integral and differential equations. Fuzzy Sets Syst.
280, 58–90 (2015)
4. Agarwal, R.P., Lakshmikantham, V., Nieto, J.J.: On the concept of solution for fractional
differential equations with uncertainty. Nonlinear Anal. Theory Methods Appl. 72, 2859–
2862 (2009)
5. Salahshour, S., Ahmadian, A., Senu, N., Baleanu, D., Agarwal, P.: On analytical solutions of
the fractional differential equation with uncertainty: application to the basset problem.
Entropy 17(2), 885–902 (2015)
6. Takači, D., Takači, A., Takači, A.: On the operational solutions of fuzzy fractional
differential equations. Fract. Calc. Appl. Anal. 17(4), 1100–1113 (2014)
7. Jafarian, A., Golmankhaneh, A.K., Baleanu, D.: On fuzzy fractional laplace, transformation.
Adv. Math. Phys. (2014). http://dx.doi.org/10.1155/2014/295432
8. Van Hoa, N.: Fuzzy fractional functional differential equations under caputo
gH-differentiability. Commun. Nonlinear Sci. Numer. Simul. 22(1–3), 1134–1157 (2015)
9. Mazandarani, M., Najariyan, M.: Type-2 fuzzy fractional derivatives. Commun. Nonlinear
Sci. Numer. Simul. 19(7), 2354–2372 (2014)
10. Ahmadian, A., Senu, N., Larki, F., Salahshour, S., Suleiman, M., Islam, S.: A legendre
approximation for solving a fuzzy fractional drug transduction model into the bloodstream.
Advances in Intelligent Systems and Computing, vol. 287, pp. 25–34 (2014)
11. Puri, L.M., Ralescu, D.: Differentials of fuzzy functions. J. Math. Anal. Appl. 91, 552–558
(1983)
12. Bede, B., Gal, S.G.: Almost periodic fuzzy-number-valued functions. Fuzzy Sets Syst. 147,
385–403 (2004)
13. Bede, B., Stefanini, L.: Generalized differentiability of fuzzy-valued functions. Fuzzy Sets
Syst. 230, 119–141 (2013)
14. Salahshour, S., Allahviranloo, T., Abbasbandy, S.: Solving fuzzy fractional equations by
fuzzy Laplace transforms. Commun. Nonlinear Sci. Numer. Simul. 17, 1372–1381 (2012)
15. Mazandarani, M., Kamyad, A.V.: Modified fractional euler method for solving fuzzy
fractional initial value problem. Commun. Nonlinear Sci. Numer. Simul. 18(1), 12–21
(2013)
 Fuzzy Fractional Derivative: A New Definition 329

16. Salahshou, S., Allahviranloo, T., Abbasbandy, S., Baleanu, D.: Existence and uniqueness
results for fractional differential equations with uncertainty. Adv. Differ. Equ. (2012). doi:10.
1186/1687-1847-2012-112
17. Allahviranloo, T., Armand, A., Gouyandeh, Z.: Fuzzy fractional differential equations under
generalized fuzzy caputo derivative. J. Intell. Fuzzy Syst. 26(3), 1481–1490 (2014)
18. Jumarie, G.: Modified riemann-liouville derivative and fractional taylor series of
non-differentiable functions further results. Comput. Math Appl. 51(9–10), 1367–1376
(2006)
19. Gomes, L.T., Barros, L.C.: A note on the generalized difference and the generalized
differentiability. Fuzzy Sets Syst. 280, 142–145 (2015)
20. Mazandarani, M., Najariyan, M.: Differentiability of type-2 fuzzy number-valued functions.
Commun. Nonlinear Sci. Numer. Simul. 19(3), 710–725 (2014)
21. Bede, B.: Mathematics of Fuzzy Sets and Fuzzy Logic. Studies in Fuzziness and Soft
Computing, vol. 295. Springer, Heidelberg (2013)
22. Atangana, A., Secer, A.: A note on fractional order derivatives and table of fractional
derivatives of some special functions. Abstr. Appl. Anal. (2013). http://dx.doi.org/10.1155/
2013/279681
23. Mazandarani, M., Pariz, N., Kamyad, A.V.: Granular differentiability of
fuzzy-number-valued functions. IEEE Trans. Fuzzy Syst. (2017, in press). doi:10.1109/
TFUZZ.2017.2659731
24. Najariyan, M., Mazandarani, M., John, R.: Type-2 fuzzy linear systems. Granul. Comput. 2
(3), 175–186 (2017). doi:10.1007/s41066-016-0037-y
25. Najariyan, M., Farahi, M.H.: A new approach for the optimal fuzzy linear time invariant
controlled system with fuzzy coefficients. J. Comput. Appl. Math. 259, 682–694 (2014)
26. Najariyan, M., Farahi, M.H.: A new approach for solving a class of fuzzy optimal control
systems under generalized Hukuhara differentiability. J. Franklin Inst. 352, 1836–1849
(2015)
 Solving First Order Linear Fuzzy Differential
Equations System

Marzieh Najariyan1, Mehran Mazandarani2(&),

and Valentina Emilia Balas3
1
Department of Applied Mathematics, Ferdowsi University of Mashhad,
Mashhad, Iran
[email protected]
2
Department of Electrical Engineering, Ferdowsi University of Mashhad,
Mashhad, Iran
[email protected], [email protected]
3
Department of Automatics and Applied Software, Faculty of Engineering,
University Aurel Vlaicu, Arad, Romania
[email protected]

Abstract. This paper aims at solving first order linear fuzzy differential
equations system by an approach called quasi-level-wise system. Some com-
parative examples show while some other approaches fail to obtain possible
system solutions, the proposed approach is able and effective. Moreover, how
the linear fuzzy differential equations system may arise in applications is
explained and inverted pendulum system is given as an example. Through the
example, it is also demonstrated how helpful this fuzzy linear model can be,
compared to the crisp linear model.

Keywords: Fuzzy differential equations
Ordinary differential equations

system
Quasi-level-wise system approach
Inverted pendulum system

Fuzzy Linear Time-Invariant System
Fuzzy linear model

1 Introduction

When modeling a dynamic system using differential equations, the main goals of
obtaining solutions of the differential equations are predicting and analyzing the
behaviors associated to that dynamic system. Considering the system model as Fuzzy
Differential Equations [FDEs] means the initial conditions and/or parameters of the
system are uncertain. Such an uncertainty stems from the unsharpness of boundaries. In
simple terms, uncertainty is of fuzzy nature, and obtaining solutions of FDEs
demonstrate the possible behaviors of the system in the presence of the uncertainties.
Since vanishing uncertainty reduces FDEs to crisp differential equations, and due to the
fact that in a large number of dynamic systems, the initial conditions and/or parameters
are fuzzy, not certain, such FDEs have attracted growing attention in recent years.
Some approaches including approach based on extension principle [11], differential
inclusion [4], and fuzzy derivative definition [3, 8, 13] have already been proposed for
solving FDEs. FDEs differ in some ways from crisp differential equations. For instance,
an FDE could have more than one solution [3]. In addition, each of the approaches
© Springer International Publishing AG 2018
V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_26
 Solving First Order Linear Fuzzy Differential Equations System 331

mentioned may obtain solutions that are different from those of the other two, this is
while one same FDE has been applied to all the approaches.
For FDEs to prove practically useful, two key points are to be taken into account:
1. Uncertainty is to be appropriately considered in differential equations corresponding
to a dynamic system. In other words, considering uncertainty must not lead to the
representation of a mathematical model whose general structure and inherent
properties are different from those of the real world system. As a way of illustration,
consider the model of a dynamic system with decreasing exponential behavior as
x_ ðtÞ ¼ k2 xðtÞ. Assume “k 2 ” is “about −2”, i.e. it is known but not certain. If
uncertainty is considered as the triangular type-1 fuzzy number ð3; 2; 1Þ, unlike
the dynamic system, the output of FDE ~x_ ðtÞ ¼ ð3; 2; 1Þ~xðtÞ will not be invari-
ably stable and will definitely end up in solutions that are not consistent with the
natural behavior of system.
In this paper, it is always assumed that uncertainty has properly been taken into
account.
2. Solutions of FDEs are to be obtained using an appropriate approach. As a matter of
fact, the approach applied, is to result in the solutions of FDEs such that:
(a) With vanishing uncertainty, the obtained solutions corresponded to those obtained
by the certain differential equations of the dynamic system.
(b) The solutions correspond to the natural behavior of the system.
Research aims at analyzing and obtaining solutions of FDEs, and accordingly
applying the results in practice. In the majority of real applications, researchers deal
with dynamic systems that are modeled by a nonlinear differential equations system. In
many cases, for some reasons such as the simplicity of the system behavior analysis,
controller design, stability analysis and so on, linear approximation of the nonlinear
model is obtained around the operating point. Such an approximation results in a linear
differential equations system including more than one state variable and is referred to as
linear state space equations. Therefore, from a practical point of view, we deal with
dynamic systems whose mathematical model is presented as the linear differential
equations system around the operating point.
The point is compared to the enormous volume of the literature carried out con-
cerning linear one-dimensional FDEs, little research has been performed in the field of
first order linear FDEs system, see [6, 15, 16]. The present paper demonstrates that the
proposed approaches in [6, 15, 16] fail to obtain correct solutions of first order linear
FDEs system; i.e. possible solutions of the dynamic system [9]. Such approaches result
in unstable solutions for a stable first order linear FDEs system. In other words, they
cannot preserve the inherent properties of the system like stability, controllability, and
the like. Moreover, the presented approaches in [6, 15, 16] are only applicable in
differential equations system in which either just the initial conditions or only the
coefficients matrix are considered as fuzzy numbers.
As a result, the present paper aims at proposing an approach named
quasi-level-wise system which, contrary to the previous approaches, preserves the
inherent system properties and achieves the possible solutions of the system. Fur-
thermore, this approach enables us to obtain solutions of first order linear FDEs system
332 M. Najariyan et al.

in which not only initial conditions, but also coefficients matrix are fuzzy numbers, and
force functions are fuzzy functions. It is noteworthy that almost all studies concerning
linear FDEs have ignored two questions: 1. how linear FDEs arise in practical appli-
cations; 2. whether the linear model of a dynamic system could be regarded as the
fuzzy linear model of the system through direct fuzzifying of its parameters. This paper
briefly addresses these issues, demonstrating how first order linear FDEs emerge while
linear model of inverted pendulum system is being obtained, and how effective they
can be.

2 Preliminaries

This section presents some definitions and theorems which will be used in the rest of
the paper. Throughout this paper, the set of all real numbers is denoted by R, and the
set of all Type-1 Fuzzy Numbers [T1FNs] on R by E1 . The transpose of matrix A is
denoted by AT .
Definition 2.1 [17]. A T1FS u 2 E1 ; u : R ! ½0; 1 is called a T1FN if it is normal,
fuzzy convex, upper semi-continuous and compactly supported fuzzy subsets of the
real numbers.
The a-cut of u 2 E1 ; ua ; is a closed and bounded interval, ½ua ; 
ua , where ua and 
ua
a
are called the left and right end points of u . The T1FN u 2 E1 can be characterized in a
parametric form by the pair ðua ; ua Þ, 0  a  1. The functions ua and  ua have the
following properties:
(a) ua is a monotonically increasing left continuous function,
(b) ua is a monotonically decreasing left continuous function,
(c) ua  ua ; a 2 ½0; 1.
The sum of two T1FNs u and v, u þ v, and the multiplication between a real number
k and a T1FN can be defined respectively as [3]:

½ u þ v  a ¼ fx þ y j x 2 ua ; y 2 v a g ¼ ua þ v a ;

½kua ¼ fkxjx 2 ua g ¼ kua

where ua þ va is the sum of two intervals, and kua is the usual product of a number and
a subset of R.
_
Consider the Linear Time-Invariant [LTI] system XðtÞ _
¼ AXðtÞ, where XðtÞ ¼ dXðtÞ

 dt ,
X : ½0; tf  ! Rn , and A ¼ aij nn ; aij 2 R. The eigenvalues of A, ki, i = 1, …, n, are
the roots of the characteristic polynomial detðkI  AÞ, where detð
Þ means determinant,
and I is the n  n identity matrix [9].
In the following, we recall similarity transformations and the well-known Jordan
canonical form of a matrix. For more details see [1, 7, 14].


Theorem 2.1. Let A ¼ aij nn be a real matrix whose real eigenvalues, ki, i = 1,…, n,
are distinct, and vi is the eigenvector of A corresponding to the eigenvalue ki. Then,
 Solving First Order Linear Fuzzy Differential Equations System 333

there exists a basis fv1 ; v2 ; . . .; vn g in Rn such that the modal matrix P ¼ ½v1 ; v2 ; . . .; vn 
is invertible, and J = P−1AP = diag(k1, k2,…,kn), where diag(k1, k2,…,kn) denotes the
diagonal matrix whose diagonal entries that start in the upper left corner are
k1, k2,…,kn.


Theorem 2.2. Let A ¼ aij nn be a real matrix whose real eigenvalues are ki ; i ¼
1; . . .; n where k1 is of multiplicity k. Then, there exists a basis fv1 ; v2 ; . . .; vn g in Rn
such that the modal matrix P ¼ ½v1 ; v2 ; . . .; vn  is invertible, and J ¼ P1 AP ¼
diag Jp1 ðk1 Þ; . . .; Jpa ðk1 Þ; kk þ 1 ; . . .; kn , where


½v1 ; . . .; vk  ¼ v1 ; n11 ; n12 ; . . .; n1p1 1 ; v2 ; . . .; va ; na1 ; na2 ; . . .; napa 1 in which
vi ; i ¼ 1; . . .; a and nipi 1 are the independent and generalized eigenvectors corre-
sponding to eigenvalue k1 . Eigenvectors vk þ 1 ; . . .; vn are obtained from kk þ 1 ; . . .; kn ,
respectively. Additionally, Jordan blocks Jpi ðk1 Þ are in the following from:
2 3
k1 1 0


0
6 .. .. 7
6 0 k1 . . 7
6 .. 7
Jpi ðk1 Þ ¼ 6
6 0 . 1 07
7
6 .. .. 7
4 . . k1 15
0


0 k1


Theorem 2.3. Let A ¼ aij nn be a real matrix whose k þ 1 distinct eigenvalues
k1;2 ¼ r1  jx1 ; k3;4 ¼ r3  jx3 ; . . .; kk;k þ 1 ¼ rk  jxk are complex, and the others
are real and distinct. Then, there exists a basis fv1 ; v2 ; . . .; vn g in Rn such that the modal
matrix P ¼ ½v1 ; v2 ; . . .; vn  is invertible, and J ¼ P1 AP ¼ diagðJ1 ; J3 ; . . .;
Jk þ 1 ; kk þ 2 ; . . .; kn Þ, where ½v1 ; . . .; vk þ 1  ¼ ½Reðv1 Þ; Imðv1 Þ; Reðv3 Þ; Imðv3 Þ; . . .;
Reðvk þ 1 Þ; Imðvk þ 1 Þ in which Reðvi Þ and Imðvi Þ denote the real and imaginary part of
the eigenvector vi , respectively. Eigenvectors vk þ 2 ; . . .; vn are obtained from
kk þ 2 ; . . .; kn , respectively. Moreover, Jordan blocks Ji are in the following from:
 
ri xi
Ji ¼
xi ri


Definition 2.2 [5, 18]. Let A ¼ aij nn be a real matrix. The matrix measure corre-
sponding to A, is denoted by lð AÞ, and defined as:

kI þ hAki 1
lð AÞ ¼ lim ð2:1Þ
h!0 h

where k
ki is a matrix induced norm, and I is the n  n identity matrix.

lð AÞ, is well defined
Remark 2.1 [5, 18]. The matrix measure, 8 9for any matrix induced
<   P n  =
norm. It is characterized by lð AÞ ¼ max Re ajj þ aij  ; j ¼ 1; . . .; n; for the
j : i¼1 ;
i6¼j
334 M. Najariyan et al.

8 9
n  =
>
< >
P  
matrix induced 1-norm and by lð AÞ ¼ max Reðaii Þ þ aij ; i ¼ 1; . . .; n; for the
i >
: j¼1 >
;
i6¼j

matrix induced /-norm.

_
Remark 2.2 [5, 18]. Consider the LTI system XðtÞ ¼ AXðtÞ with the initial condition
X ð0Þ. The bounds of the solution of LTI system are given as:

kX ð0ÞkelðAÞðt0 tÞ  kXðtÞk  kX ð0Þkelð AÞðtt0 Þ ð2:2Þ

where k
k is a norm.

3 The Solutions of First Order Linear FDEs System

Solving a class of first order linear FDEs system based on the quasi-level-wise system
approach is introduced in this section. A quasi-level-wise system is a complex number
representation of the FDEs where in addition to the formal structure of the FDEs, the
inherent properties of the FDEs are preserved. In constructing a quasi-level-wise
system each fuzzy function, ~xðtÞ, is replaced by the complex representation as
ya ðtÞ þ i ya ðtÞ. Hereafter, for the sake of simplicity the triangular type-1 fuzzy
number-valued functions, ~xðtÞ ¼ ðx0 ðtÞ; x1 ðtÞ; x0 ðtÞÞ, are considered, where the left and
right end points of the support of ~xðtÞ are denoted by x0 ðtÞ and x0 ðtÞ, respectively.
Additionally, x1 ðtÞ denotes the fuzzy function ~xðtÞ in the level a ¼ 1.

3.1 One-Dimensional First Order Linear FDEs

Consider the following one-dimensional first order linear FDE:


~x_ ðtÞ ¼ a~xðtÞ þ ~f ðtÞ ~xð0Þ ¼ ~x0 t 2 0; tf ð3:1Þ



where ~x_ ðtÞ ¼ d~xdtðtÞ, ~x; ~f : 0; tf ! E1 are the fuzzy functions on the real interval 0; tf ,
~x0 2 E1 is the initial condition, and a 2 R. According to obtained results in [4], it can
be concluded that the solution to FDE (3.1) is equivalent to solving the following
quasi-level-wise system:

y_ a ðtÞ þ i y_ a ðtÞ ¼ a ya ðtÞ þ ia ya ðtÞ þ f a ðtÞ þ i f a ðtÞ ya ð0Þ ¼ xa0 ; ya ð0Þ ¼ xa ð0Þ ð3:2Þ

Or equivalently,
    a   a 
y_ a ðtÞ a 0 y ðt Þ f ðt Þ
y_ a ðtÞ ¼ a þ a ya ð0Þ ¼ xa0 ; ya ð0Þ ¼ xa ð0Þ ð3:3Þ
0 a y ðtÞ f ðt Þ
|fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl} |fflfflfflffl{zfflfflfflffl}
Y_ a ðtÞ A Y a ðtÞ f a ðtÞ
 Solving First Order Linear Fuzzy Differential Equations System 335

Hereafter, by an asterisk, we denote the equivalent matrix or functions in moving

from the quasi-level-wise system to its equivalent ordinary differential equations sys-
tem. The solution
n of FDE (3.1),oi.e. ~xðtÞ ¼ ðx0 ðtÞ; x1 ðtÞ; x0 ðtÞÞ, can be determined
n as
x0 ðtÞ ¼ min y0 ðtÞ; y1 ðtÞ; y0 ðtÞ , x1 ðtÞ ¼ y1 ðtÞ, and x0 ðtÞ ¼ max y0 ðtÞ; y1 ðtÞ;
t2½0;tf  t2½0;tf 

y ðtÞg, where y ðtÞ and y ðtÞ are the solutions of system (3.3) for a ¼ 0, and y1 ðtÞ for
0 0 0

a ¼ 1.
Now, consider FDE (3.1) in which the coefficient of ~xðtÞ, a, is a type-1 fuzzy
number as follows:


~x_ ðtÞ ¼ ~a~xðtÞ þ ~f ðtÞ ~xð0Þ ¼ ~x0 t 2 0; tf ð3:4Þ

For obtaining the solution of (3.4), ~xðtÞ ¼ ðx0 ðtÞ; x1 ðtÞ; x0 ðtÞÞ, what is needed is to
consider the fuzzy coefficient, ~a, and other functions involved in (3.4) only in levels
a ¼ 0; 1. As a result, the quasi-level-wise systems can be constructed as:
8 a
< y_ 1 ðtÞ þ i y_ 1 ðtÞ ¼ a y1 ðtÞ þ ia y1 ðtÞ þ f ðtÞ þ i f ðtÞ
a a a a a a a

y_ ðtÞ þ i y_ 2 ðtÞ ¼ aa ya2 ðtÞ þ iaaya2 ðtÞ þ f ðtÞ þ i f a ðtÞ
a a a
ð3:5Þ
: 2a
y2 ð0Þ ¼ ya1 ð0Þ ¼ xa0 ; ya2 ð0Þ ¼ ya1 ð0Þ ¼ xa ð0Þ

Or equivalently,

Y_ ia ðtÞ ¼ Ai Yia ðtÞ þ f a ðtÞ yai ð0Þ ¼ xa0 ; yai ð0Þ ¼ xa ð0Þ; i ¼ 1; 2 ð3:6Þ
h iT
where Yia ðtÞ ¼ yai ðtÞ yai ðtÞ , A1 ¼ diagðaa ; aa Þ and A2 ¼ diagð
aa ; 
aa Þ. Accordingly,
the solution
n of FDE (3.4) ocan be characterized as x0 ð t Þ ¼
n
mint2½0;tf  y01 ðtÞ; y01 ðtÞ; y11 ðtÞ; y02 ðtÞ; y02 ðtÞ , x1 ðtÞ ¼ y11 ðtÞ, and x0 ðtÞ ¼ max y01 ðtÞ;
t2½0;tf 

y1 ðtÞ; y1 ðtÞ; y2 ðtÞ; y2 ðtÞg:

0 1 0 0

As is seen, in addition to the formal structure of FDEs (3.1) and (3.4), the
quasi-level-wise system has preserved their inherent properties

3.2 First Order Linear FDEs System with Uncertain Initial Conditions
Consider the FDEs system as follows:
(

X~_ ðtÞ ¼ AX
~ ðt Þ þ F
~ ðt Þ ~ ð 0Þ ¼ X
X ~0 t 2 0; tf ð3:7Þ
Y~out ðtÞ ¼ CX~ ðtÞ


where X ~ F
~ ðtÞ ¼ ½~x1 ðtÞ; . . .; ~xn ðtÞT , X; ~ : 0; tf ! E1n are the fuzzy functions on the real



interval 0; tf , X ~0 2 E1n is the initial condition, A ¼ aij ; aij 2 R, and

 nn
C ¼ cij mn , cij 2 R. Due to the fact that there may exist interactions between func-
tions ~xi ðtÞ and ~xj ðtÞ, i 6¼ j, constructing the quasi-level-wise system corresponding to
336 M. Najariyan et al.

FDEs system (3.7) may not be possible directly. Therefore, we try to construct a
diagonal quasi-level-wise system corresponding to the system (3.7) based on the
similarity transformation. Then, the solution to quasi-level-wise system can be obtained
using the diagonal system. Consequently, the solution to the FDEs system (3.7) can be
determined from that of quasi-level-wise system. It is noteworthy that using the
diagonal quasi-level-wise system has some advantages such as use of the obtained
results corresponding one-dimensional first order linear FDEs and preserving the
inherent properties of the system observable in the diagonal representation.
Assume the quasi-level-wise system corresponding to the FDEs (3.7) is as:


Y_ a ðtÞ ¼ BY a ðtÞ þ F a ðtÞ þ iF
 a ðt Þ 0a
Y a ð0Þ ¼ X a0 þ iX t 2 0; tf ð3:8Þ
h iT
where Y a ðtÞ ¼ ya1 ðtÞ þ i ya1 ðtÞ ya2 ðtÞ þ i ya2 ðtÞ. . .yan ðtÞ þ i yan ðtÞ . The solution to system
(3.8) can be obtained in two ways called direct method and indirect method.
Direct Method. Consider system (3.8). Let Y a ðtÞ ¼ PZ a ðtÞ where Z a ðtÞ ¼

a T

z1 ðtÞ þ i za1 ðtÞ za2 ðtÞ þ i za2 ðtÞ. . .zan ðtÞ þ i zan ðtÞ , and P ¼ pij nn is the modal matrix
obtained from the matrix A of the system (3.7). Using Theorems 2.1, 2.2, and 2.3 leads
to the construction of the diagonal representation corresponding to quasi-level-wise
system (3.8) as follows:

Z_ a ðtÞ ¼ J  Z a ðtÞ þ ðP Þ1 F a ðtÞ Z a ð0Þ ¼ ðP Þ1 Y a ð0Þ ð3:9Þ

T
where Z a ðtÞ ¼ za1 ðtÞ za1 ðtÞ za2 ðtÞ za2 ðtÞ. . .zan ðtÞ zan ðtÞ , J  ¼ ðP Þ1 Bd P , P ¼

2 3
p11 gðp11 Þ p11 gðp11 Þ


p1n gðp1n Þ p1n gðp1n Þ
6 .. .. 7
4 . . 5 ; and gð xÞ ¼
pn1 gðpn1 Þ pn1 gðpn1 Þ


pnn gðpnn Þ pnn gðpnn Þ 2n2n

x x 0
0 x\0
Owing to the fact that matrix B is unknown in general, Bd is unknown - d as the
index of B stands for direct. Hence, J  ¼ ðP Þ1 Bd P symbolically shows the rela-
tionship between J  ; Bd and P . Based on the definition of quasi-level-wise system and

Theorems 2.1, 2.2, and 2.3, matrix J  is obtained from J ¼ P1 AP, i.e. J  ¼ ðP1 APÞ .
For instance, if all eigenvalues are distinct, then J ¼ P1 AP ¼ diagðk1 ; k2 ; . . .; kn Þ and
J  ¼ diagðk1 ; k1 ; k2 ; k2 ; . . .; kn ; kn Þ.
By solving differential equations systemn(3.9), the solutiono to FDEs system (3.7)
can be characterized as x0i ðtÞ ¼ min y0i ðtÞ; y1i ðtÞ; y0i ðtÞ , x1i ðtÞ ¼ y1i ðtÞ, and
t2½0;tf 
n o
x0i ðtÞ ¼ max y0i ðtÞ; y1i ðtÞ; y0i ðtÞ , i ¼ 1; . . .; n where Y a ðtÞ ¼ P Z a ðtÞ.
t2½0;tf 
In the following, an example is presented to illustrate the proposed approach.
 Solving First Order Linear Fuzzy Differential Equations System 337

Example 3.1. Consider the following FDEs system

        
~x_ 1 ðtÞ 2 5 ~x1 ðtÞ ~x1 ð0Þ ð7; 8; 9Þ
¼ ¼ t 2 ½0; 1 ð3:10Þ
~x_ 2 ðtÞ 0 2 ~x2 ðtÞ ~x2 ð0Þ ð7; 8; 9Þ

The FDEs system (3.10) is similar to Example 4.6 in [15]. Using the approach
proposed in [15], for solving FDEs system (3.10) results in obtaining the solutions of
following differential equations system:
2 3 2 32 a 3 2 3 2 3
x_ a1 ðtÞ 0 2 5 0 x1 ðtÞ xa1 ð0Þ 8  ð 1  aÞ
6 x_ a1 ðtÞ 7 6 2 0 0 5 76 xa1 ðtÞ 7 6 xa1 ð0Þ 7 6 8 þ ð1  aÞ 7
6 a 7¼6 76 7 6 a 7¼6 7
4 x_ ðtÞ 5 4 0 0 0 2 54 xa2 ðtÞ 5 4 x ð0Þ 5 4 8  ð1  aÞ 5 ð3:11Þ
2 2
x_ a ðtÞ 0 0 2 0 xa2 ðtÞ xa2 ð0Þ 8 þ ð 1  aÞ
2
t 2 ½0; 1

Differential equations system (3.11) presents the system with four eigenvalues −2,
−2, 2 and 2. Although the FDEs system (3.10) is invariably stable, the solutions
represent an unstable system. See Figs. 1 and 2.
On the basis of the proposedn approach in thiso paper, the solutions to (3.10) are
obtained as x0i ðtÞ ¼ max y0i ðtÞ; y1i ðtÞ; y0i ðtÞ , x1i ðtÞ ¼ y1i ðtÞ, and x0i ðtÞ ¼
t2½0;2
n o h iT

min y0i ðtÞ; y1i ðtÞ; y0i ðtÞ , i ¼ 1; 2, where ya1 ðtÞ ya1 ðtÞ ya2 ðtÞ ya2 ðtÞ ¼ P za1 ðtÞ za1 ðtÞ;
t2½0;2

za2 ðtÞ za2 ðtÞT , and P is extracted from modal matrix P as follows:
2 3
  1 0 0 0
1 0 60 1 0 0 7
P¼ ; P ¼6

40
7;
0 0:2 0 0:2 0 5
0 0 0 0:2

Fig. 1. The solution of Example 3.1, ~x1 ðtÞ based on th approach proposed in [15].
338 M. Najariyan et al.

Fig. 2. The solution of Example 3.1, ~x2 ðtÞ based on th approach proposed in [15].

T
and Z a ðtÞ ¼ za1 ðtÞ za1 ðtÞ za2 ðtÞ za2 ðtÞ is the solution to the following differential
equations system:
2 3 2 32 a 3 2 3 2 3
z_ a1 ðtÞ 2 0 1 0 z1 ðt Þ za1 ð0Þ 8  ð1  aÞ
6 z_ a1 ðtÞ 7 6 0 2 0 76 za1 ðtÞ 7 6 za1 ð0Þ 7 6 7
6 a 7¼6 1 76 7 6 a 7 ¼ ðP Þ1 6 8 þ ð1  aÞ 7
4 z_ ðtÞ 5 4 0 0 2 0 54 za2 ðtÞ 5 4 z ð 0Þ 5 4 8  ð1  aÞ 5
2 2
z_ a ðtÞ 0 0 0 2 za ðtÞ za2 ð0Þ 8 þ ð 1  aÞ
2
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} 2
J
t 2 ½0; 2
ð3:12Þ

As expected, the stability property of FDEs system (3.10) has been preserved.
Figures 3 and 4 show ~x1 ðtÞ and ~x2 ðtÞ, respectively.
Indirect Method. Consider system (3.8). Let P1
Y a ðtÞ ¼ Z a ðtÞ where the matrix
Z ðtÞ is similar to that in direct method, and P1 ¼ qij nn is the inverse of modal
a

matrix P. Using Theorems 2.1, 2.2 and 2.3 leads to the construction of the diagonal
representation corresponding to quasi-level-wise system (3.8) as follows:

Fig. 3. The solution of Example 3.1, ~x1 ðtÞ based on the quasi-level-wise system approach.
 Solving First Order Linear Fuzzy Differential Equations System 339

   
Z_ a ðtÞ ¼ J  Z a ðtÞ þ P1 F a ðtÞ Z a ð0Þ ¼ P1 Y a ð0Þ ð3:13Þ

T    1
where Z a ðtÞ ¼ za1 ðtÞza1 ðtÞza2 ðtÞza2 ðtÞ. . .zan ðtÞzan ðtÞ , J  ¼ ðP1 Þ Bid ðP1 Þ ,
2 3
q11 gðq11 Þ q11 gðq11 Þ


q1n gðq1n Þ q1n gðq1n Þ
 6 .. .. 7
ðP1 Þ ¼ 4 . . 5 , and
qn1 gðqn1 Þ qn1 gðqn1 Þ


qnn gðqnn Þ qnn gðqnn Þ 2n2n

x x 0
gð x Þ ¼ .
0 x\0
Similar to the previous method, Bid is unknown - id as the index of B stands for
indirect - that is, J  can be determined from J ¼ P1 AP.
Even though matrix J  is the same in the two methods, matrices Bid and Bd are not
the same in general. Consequently, by solving differential equations system (3.13),
the solution to FDEs system (3.7) can be characterized as x0i ðtÞ ¼
n o n
min y0i ðtÞ; y1i ðtÞ; y0i ðtÞ , x1i ðtÞ ¼ y1i ðtÞ, and x0i ðtÞ ¼ max y0i ðtÞ; y1i ðtÞ; y0i ðtÞg, i ¼
t2½0;tf  t2½0;tf 
a
 1  1 a
1; . . .; n where Y ðtÞ ¼ ðP Þ Z ðtÞ.
Example 3.2. Consider the FDEs system presented in Example 3.1. Using the indirect
method corresponding to the quasi-level-wise system of FDEs (3.10) leads to:

2 3 2 3 2 a 32 a 3 2 3
z_ a1 ðtÞ 2 0 1 0 z1 ðt Þ z1 ð0Þ 8  ð 1  aÞ
6 z_ a1 ðtÞ 7 6 0 2 0 1 7 6 a 76 a 7   6 7
6 a 7¼6 7 6 z1a ðtÞ 76 z1a ð0Þ 7 ¼ P1  6 8 þ ð1  aÞ 7t
4 z_ ðtÞ 5 4 0 0 2 0 5 4 z2 ðt Þ 5 4 z2 ð0Þ 5 4 8  ð 1  aÞ 5
2
z_ a ðtÞ 0 0 0 2 a
z ðtÞ a
z2 ð0Þ 8 þ ð 1  aÞ
2
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} 2
J
2 ½0; 2
ð3:14Þ

where
2 3
    1 0 0 0
1 0 1 0  1  6 0 1 0 07
P¼ ; P1 ¼ ; P ¼6
40
7
0 0:2 0 5 0 5 05
0 0 0 5

The solutions to the FDEs system (3.10) are obtained as x0i ðtÞ ¼
n o n o
max y0i ðtÞ; y1i ðtÞ; y0i ðtÞ , x1i ðtÞ ¼ y1i ðtÞ, and x0i ðtÞ ¼ min y0i ðtÞ; y1i ðtÞ; y0i ðtÞ ,
t2½0;2 t2½0;2
h iT  
T
a a a a 1  1 a
i ¼ 1; 2, where y1 ðtÞ y1 ðtÞ y2 ðtÞ y2 ðtÞ ¼ ðP Þ z1 ðtÞ za1 ðtÞ za2 ðtÞ za2 ðtÞ . The
solutions obtained using the indirect method coincide those obtained by the direct
method. It should be noted that the two solutions obtained using the direct and indirect
methods may not be generally the same, as each other.
340 M. Najariyan et al.

Fig. 4. The solution of Example 3.1, ~x2 ðtÞ based on the quasi-level-wise system approach.

Remark 3.1. It is expected that using the quasi-level-wise system approach may
present more than one solution for FDEs (3.7). This is not very uncommon because,
according to [3], even one-dimensional FDEs are likely to have two solutions.
Nonetheless, this is regarded as an advantage for the approach. What follows will
demonstrate that using this approach can give us the chance of selecting solutions. Such
a selection could be made wisely and based on the intended application.
Example 3.3. Consider the FDEs system

        
~x_ 1 ðtÞ 1 3 ~x1 ðtÞ ~x1 ð0Þ ð7; 8; 9Þ
¼ ¼ t 2 ½0; 5 ð3:15Þ
~x_ 2 ðtÞ 2 4 ~x2 ðtÞ ~x2 ð0Þ ð7; 8; 9Þ

Figures 5 and 6 show the solutions to FDEs system (3.15) obtained using the direct
and indirect methods. As is seen, the two solutions obtained by the two methods have
not coincided. In this example, as the figures depict, if one decides to analyze the
system with precaution, they will use the solutions obtained using the indirect method.
As a way of illustration, assume someone wants to design a controller for the system,
with precaution and the possible dynamics of the system with the presence of uncer-
tainty in the initial conditions are of importance to him. Then the solutions obtained
using the indirect method can prove more helpful.

3.3 First Order Linear FDEs System with Uncertain Coefficients

and Initial Conditions
Consider FDEs system (3.7) in which the coefficients matrix, A, is a fuzzy matrix, i.e.


~ ¼ ~aij
A , ~aij 2 E1 , as follows:
nn
(

~_ ðtÞ ¼ A
X ~X~ ðt Þ þ F
~ ðt Þ ~ ð 0Þ ¼ X
X ~0 t 2 0; tf ð3:16Þ
~ ~
Yout ðtÞ ¼ CX ðtÞ
 Solving First Order Linear Fuzzy Differential Equations System 341

Fig. 5. The left and right end points of the solution of Example 3.3, ~x1 ðtÞ, at the level a = 0,
using direct and indirect methods. The red and green lines correspond to x01 ðtÞ, and the blue and
yellow lines correspond to x01 ðtÞ obtained by indirect and direct method, respectively.

The approach for constructing the quasi-level-wise system in order to solve FDEs
system (3.16) is a combination of what were done for FDE (3.4) and FDEs system
(3.7). In other words, it is necessary to consider functions involved in (3.16) only in
levels a ¼ 0; 1, and then, to obtain the solutions of (3.16) using the direct or indirect
method. Owing to the fact that there are so many situations, i.e. less than or equal to

Fig. 6. The left and right end points of the solution of Example 3.3, ~x2 ðtÞ, at the level a = 0,
using direct and indirect methods. The red and green lines correspond to x02 ðtÞ, and the blue and
yellow lines correspond to x02 ðtÞ, obtained by indirect and direct method, respectively.
342 M. Najariyan et al.

2nn , by which the coefficients matrices can be constructed in level a ¼ 0, the com-
plexity and computational costs of obtaining all solutions corresponding to the situa-
tions rise. Therefore, a criteria should be looked for in order to choose some of the
matrices which affect the left and right end points of the solutions in level a ¼ 0.
Assume n the
set of coefficients
n matrices hthati can be constructed
o h i in level ao¼ 0 is
^
^ ¼ bij  ^ a a a a  a a
S1 ¼ B b 2 aij ; aij ; A ¼ aij
nn  ij
; A ¼  aij i; j ¼ 1; . . .; n .

  nn    nn 
Moreover, let S2 ¼ D ¼ dij nn lðDÞ ¼ max l B ^ ; D; B^ 2 S1 , S3 ¼ fE ¼

    
eij nn lðE Þ ¼ max l B ^ ; E; B ^ 2 S1 g, and S4 ¼ S2 [ S3 . According to
Remark 2.2, the matrices belong to set S4 , as the effective matrices, can be candidates
for obtaining the solutions. Assume the number of matrices included in set S4 is m. As a
result, the quasi-level-wise systems corresponding to FDEs system (3.16) are consid-
ered as:


Y_ ja ðtÞ ¼ BYja ðtÞ þ F a ðtÞ þ iF
 a ðt Þ 0a
Yja ð0Þ ¼ X a0 þ iX t 2 0; tf ð3:17Þ
h iT
where Yja ðtÞ ¼ ya1;j ðtÞ þ iya1;j ðtÞ ya2;j ðtÞ þ iya2;j ðtÞ. . .yan;j ðtÞ þ i yan;j ðtÞ , j ¼ 1; . . .; m, and
B 2 S4 . Consequently, n the solution o to FDEs system (3.16) can be determined n as:
xi ðtÞ ¼ min yi;j ðtÞ; yi;1 ðtÞ; yi;j ðtÞ , xi ðtÞ ¼ yi;1 ðtÞ, and xi ðtÞ ¼ max y0i;j ðtÞ;
0 0 1 0 1 1 0
t2½0;tf ;j t2½0;tf ;j

y1i;1 ðtÞ; y0i;j ðtÞg, i ¼ 1; . . .; n; j ¼ 1; . . .; m, where Yja ðtÞ ¼ Pj Zja ðtÞ for direct method or

 1 a
Yja ðtÞ ¼ ðP1 Þj Zj ðtÞ for indirect method.

Example 3.4. Consider the radioactive decay model presented in [6] as

          
~x_ 1 ðtÞ ~k1 ~0 ~x1 ðtÞ ~r ~x1 ð0Þ ð995; 1000; 1005Þ
_~x2 ðtÞ ¼ ~k1 ~k2 ~x2 ðtÞ
þ ~
0 ~x2 ð0Þ
¼
ð0; 0; 0Þ ð3:18Þ
t 2 ½0; 20

where ~k1 ¼ ð0:2; 0:3; 0:4Þ, ~k2 ¼ ð0:02; 0:03; 0:04Þ, ~0 ¼ ð0; 0; 0Þ, and ~r ¼ ð4:9; 5; 5:1Þ.
Figures 7 and 8 show the solutions of FDEs system (3.18) which have been obtained
based on the quasi-level-wise system approach under direct method.
Note 3.1. Using the indirect method leads to solutions which coincide the solutions
obtained by the direct method.
Remark 3.2. From a practical point of view, no radioactive substance could be
divergently. Put another way, it must be finite and convergent. Therefore, it is expected
that the solutions of FDEs system (3.18) that are the possible dynamics of the system
are also stable. As quasi-level-wise system approach preserves the inherent properties
of the system, what was expected as the possible solutions has come true in Figs. 7 and
8. It is noteworthy that, based on the approach proposed in [6], it is concluded that the
dynamic system with the considered uncertainty does not decay for a 6¼ 1. Simply put,
a minor uncertainty will lead to instability, i.e. as time passes, the absolute value of
 Solving First Order Linear Fuzzy Differential Equations System 343

Fig. 7. The solution of Example 3.4, ~x1 ðtÞ, based on the quasi-level-wise system approach,
direct method.

Fig. 8. The solution of Example 3.4, ~x2 ðtÞ, based on the quasi-level-wise system approach,
direct method.

radioactivity for a 6¼ 1 will rise in an exponential and unbounded way. But this is
definitely not the case with the real dynamic system. Accordingly, the solutions cannot
represent the possible dynamics of the system, which is the drawback of the approach.
Example 3.5. Consider the following FDEs system:

        
~x_ 1 ðtÞ ~0 ~b ~x1 ðtÞ ~x1 ð0Þ ð1:5; 2; 2:5Þ
¼ ~ ¼ t 2 ½0; 10 ð3:19Þ
~x_ 2 ðtÞ b ~0 ~x2 ðtÞ ~x2 ð0Þ ð0:5; 0; 0:5Þ

where ~b ¼ ð1:8; 2; 2:2Þ and ~0 ¼ ð0; 0; 0Þ. FDEs system (3.19) is similar to Example 5.4 in
[16] except that the initial conditions in the latter had been considered as crisp numbers,
344 M. Najariyan et al.

i.e. singleton fuzzy numbers. Using the approach proposed in [16] for solving FDEs
system (3.19) leads to obtaining the solution of following differential equations system:
2 3 2 32 a 3
x_ a1 ðt; lÞ 0 0 0 bðl; aÞ x1 ðt; lÞ
6 x_ a1 ðt; lÞ 7 6 0 0 bðl; aÞ 0 76 xa1 ðt; lÞ 7
6 a 7 6 76 a 7
4 x_ ðt; lÞ 5 ¼ 4 bðl; aÞ 0 0 0 54 x ðt; lÞ 5 ð3:20Þ
2 2
x_ a ðt; lÞ 0 bðl; aÞ 0 0 xa2 ðt; lÞ
2

where bðl; aÞ ¼ ð1  lÞð2  0:2ð1  aÞÞ þ lð2 þ 0:2ð1  aÞÞ, l 2 ½0; 1. The solution
of differential equations system (3.20) has been presented in [16] as:
2 3 2 M þ cos bðl;aÞ N þ cos b2ðl;aÞ N  sin b2ðl;aÞ
3
N  sin b2ðl;aÞ 2 xa ð0; lÞ 3
xa1 ðt; lÞ 2
6 xa1 ðt; lÞ 7 6 N  sin b2ðl;aÞ 7
1
cos bðl;aÞ
6 a 7 6 M þ N þ cos b2ðl;aÞ N  sin b2ðl;aÞ 76 a 7
6 x1a ð0; lÞ 7
4 x ðt; lÞ 5 ¼ 6 7
2
ðl;a Þ cos bðl;aÞ 54 x ð0; lÞ 5
2 4 Nþ sin b
2 M þ sin b2ðl;aÞ M þ cos b2ðl;aÞ M þ 2 2
xa2 ðt; lÞ N þ sin bðl;aÞ
M þ sin bðl;aÞ
M þ cos b2ðl;aÞ M þ cos b2ðl;aÞ xa2 ð0; lÞ
2 2
ð3:21Þ
bðl;aÞt bðl;aÞt bðl;aÞt bðl;aÞt
þe
where M; N denote e 4 ; e e4 , respectively. The solution shows an
a a
unstable system, lim xi ðt; lÞ ! 1 or lim xi ðt; lÞ ! 1. Nevertheless, FDEs system
t!1 t!1
(3.20) is marginally stable. In other words, the obtained results based on the proposed
approach express that the dynamic system (3.20), even by the smallest uncertainty,
behaves in the same way as an unstable system, and the inherent properties of the
dynamic system are lost. However, this is not the case. Moreover, the proposed approach
in [16] works well only with situations in which the initial conditions and force functions
are certain that is highly restrictive. The solutions obtained by the quasi-level-wise
system approach under direct method have been shown in Figs. 9 and 10. As is expected,
the trajectories in the figures show FDEs system (3.19) is not unstable.

Fig. 9. The solution of Example 3.5, ~x1 ðtÞ, based on the quasi-level-wise system approach. The
red, green and blue lines show the trajectories x01 ðtÞ, x11 ðtÞ, and x01 ðtÞ, respectively.
 Solving First Order Linear Fuzzy Differential Equations System 345

Fig. 10. The solution of Example 3.5, ~x2 ðtÞ, based on the quasi-level-wise system approach. The
red, green and blue lines show the trajectories x02 ðtÞ, x12 ðtÞ, and x02 ðtÞ, respectively.

4 Applications

Even though numerous works in the literature have been published in the field of fuzzy
differential equations, few have dealt with FDEs in applications; see for example [2, 10,
12]. Moreover, a comprehensive and useful explanation might not be found about how
the linear FDEs may arise in the applications. As a matter of fact, it is still unknown
how fuzzy linear differential equations arise in practice. Almost all over the literature,
linear model parameters are directly fuzzified and; thus, the model is regarded as a
fuzzy linear differential equations one. But this does not invariably hold true. To
clarify, note that almost all dynamic systems are modeled using nonlinear differential
equations. Such models are linearized around the operating point for different purposes.
As a result, linear differential equations system is considered as the system model.
Now, if a parameter of the linear differential equations system is fuzzified, then it
could be concluded that such an uncertainty has originated from the nonlinear differ-
ential equations and/or through the linearization process. In other words, linear model
uncertainty stems from either the uncertainty existing in the nonlinear system and/or
the process of linearization. In principle, a nonlinear model is not expected to lead to a
fuzzy linear differential equations system - fuzzy linear model - i.e. through the direct
fuzzification of the linear differential equations system - linear model - the result cannot
be claimed to be a linear approximation of the nonlinear differential equations related to
the dynamic system, see Figs. 11 and 12.
In the following, two situations are mentioned in which the fuzzy linear differential
equations may arise:
(1) Linearization of the model of nonlinear dynamic systems with uncertain operating
point and/or uncertain parameters.
346 M. Najariyan et al.

Fig. 11. The relationship between fuzzy nonlinear and linear model of a nonlinear system with
considering uncertainty in the nonlinear model.

Fig. 12. The relationship between nonlinear and fuzzy linear model of a nonlinear system with
considering fuzzification in the linearization process.

In linearization of a nonlinear model of dynamic system it is almost always covertly

assumed that the operating point and/or system parameters are certain. If this is not the
case, then the linearization “may” result in the emergence of the fuzzy linear differ-
ential equations system. To illustrate, assume it is desirable to linearize the nonlinear
differential equations of an aircraft around the operating point. The aircraft Mach
number at an altitude of “around 11000 meters” is also known to be “approximately
1.2”. Assuming that all other parameters are certain, linearization “may” lead to a fuzzy
linear model.
 Solving First Order Linear Fuzzy Differential Equations System 347

(2) Linearization process

In this case, ignoring or fixing the value of some parameters, functions and the like,
because of their small size or being around a certain value, “may” lead to a fuzzy linear
model.
What follows is an example for explaining how the fuzzy linear model may arise in
the applications, and how helpful they can be in the analysis of systems.
Figure 13 shows the inverted pendulum system with its features represented in
Table 1. In Fig. 13, hðtÞ is the pendulum angle from the vertical axis - in radian -, xðtÞ
is the cart position coordinate - in meter -, and F is the force applied to the cart.

Fig. 13. The inverted pendulum on a cart

Table 1. The inverted pendulum system parameters.

Mass of the cart (M) 1 kg
Mass of the pendulum (m) 0.3 kg
Friction of the cart (b) 0.1 N/m s
Distance from the center of gravity to the pivot (l) 0.5 m
Mass moment of inertia of the pendulum (I) 0.008 kg m2
the acceleration of gravity (g) 9.8 N/kg

The nonlinear differential equations of the inverted pendulum system are written as:
2 3 2 32 3 2 3
x_ 1 ðtÞ 0 1 0 0 x 1 ðt Þ 0
6 x_ 2 ðtÞ 7 1 60 bðI þ ml2 Þ 2
ðmlÞ g cos x3 ðtÞ þ ðI þ ml2 Þmlx24 ðtÞ 0 76 x2 ðtÞ 7 6 Dðx3 ðtÞÞ 7
7 6 7 6 I þ ml2
6 7¼ 6 þ 7F
4 x_ 3 ðtÞ 5 Dðx3 ðtÞÞ 4 0 0 0 1 54 sin x3 ðtÞ 5 4 0 5
ml cos x3 ðtÞ
x_ 4 ðtÞ 0 mlb cos x3 ðtÞ ðM þ mÞmgl  ðmlx4 ðtÞÞ2 cos x3 ðtÞ 0 x 4 ðt Þ Dðx3 ðtÞÞ

ð4:1Þ

where x1 ðtÞ ¼ xðtÞ, x3 ðtÞ ¼ hðtÞ, and Dðx3 ðtÞÞ ¼ ðM þ mÞI þ Mml2 þ
ð1  cos x3 ðtÞÞm l . The velocity of the cart, x_ ðtÞ, and the angular velocity of the
2 2 2

pendulum, h_ ðtÞ, are denoted by x2 ðtÞ and x4 ðtÞ, respectively. Assuming hðtÞ
0,
348 M. Najariyan et al.

h_ 2 ðtÞ
0, cos hðtÞ
1, and sin hðtÞ
hðtÞ, in most of the literature, the linear model
of system (4.1) is presented in the following:
2 3 2 32 3 2 3
x_ 1 ðtÞ 0 1 0 0 x1 ðtÞ 0
6 x_ 2 ðtÞ 7 1 60 bðI þ ml2 Þ ðmlÞ2 g 7
0 76 x2 ðtÞ 7 6 ðM þ mÞI þ Mml2 7
76 6 I þ ml2
6 7 6 7F
4 x_ 3 ðtÞ 5 ¼ ðM þ mÞI þ Mml2 4 0 0 0
þ
1 54 x3 ðtÞ 5 4 0 5
ml
x_ 4 ðtÞ 0 mlb ðM þ mÞmgl 0 x4 ðtÞ ðM þ mÞI þ Mml2

ð4:2Þ

In the literature in this case, for obtaining linear model (4.2) the exact values, i.e.
“1, hðtÞ and 0”, replace “cos hðtÞ, sin hðtÞ, and h_ 2 ðtÞ”, respectively.
But the key point is that, cos hðtÞ
1 means cos hðtÞ is “approximately equal to 1”
and not always exactly equal to “1”. Likewise sin hðtÞ
hðtÞ and h_ 2 ðtÞ
0 mean
“approximately equal to hðtÞ”, and “approximately equal to 0”. As a matter of fact,
linear system (4.2) has been obtained by translating “cos hðtÞ is approximately equal to
1” as “cos hðtÞ is equal to 1” and the same has been done for the other two.
Now, we are going to consider the “approximately equal to 1”, “approximately
equal to hðtÞ”, and “approximately equal to 0” as the triangular fuzzy numbers
 
~1 ¼ ða; 1; 1Þ, ~hðtÞ ¼ h0 ðtÞ; h1 ðtÞ; h0 ðtÞ , and ~0 ¼ ð0; 0; cÞ for obtaining the fuzzy
linear system as follows:
2 3 2 32 3 2 3
~x_ 1 ðtÞ 0 1 0 0 ~x1 ðtÞ 0
6 ~x_ 2 ðtÞ 7 6 0 ~a22 ~a23 07 6 ~x2 ðtÞ 7 6 ~ b 7
6 7 6 76 7 þ 6 2 7F
4 ~x_ 3 ðtÞ 5 ¼ 4 0 0 0 1 54 ~x3 ðtÞ 5 4 05
ð4:3Þ
~x_ 4 ðtÞ 0 ~a42 ~a43 0 ~x4 ðtÞ ~
b4
 
bðI þ ml2 Þ bðI þ ml2 Þ bðI þ ml2 Þ ðmlÞ2 g ðmlÞ2 g
where ~a22 ¼ d1 ; d1 ; d2 , ~
a23 ¼ d1 ; d1 ;
ðmlÞ2 ga þ ðI þ ml2 Þmlc

ðmlÞ2 c
d2 Þ, ~a42 ¼ d2 ; d1 ; d1
mlbc mlb mlb
, ~a43 ¼ ðM þ mÞmgl
d2 ; ðM þ mÞmgl
d1 ;
 
ðM þ mÞmgl
Þ, ~b2 ¼ I þ ml2 ; I þ ml2 ; I þ ml2 , ~b4 ¼ ml ; ml ; mla , d1 ¼
d1 d2 d1 d1 d1 d1 d2
ðM þ mÞI þ Mml , and d2 ¼ ðM þ mÞI þ Mml þ ð1  a Þm l .
2 2 2 2 2

In this example, we consider “approximately equal to 1” as ~ 1 ¼ ð0:99; 1; 1Þ, which

means hðtÞ
0 is considered a value between −8.109 to 8.109 °, i.e. 0  8:109, and ~ 0
as ~0 ¼ ð0; 0; 102 Þ.
Figures 14, 15, 16, 17, 18 and 19 show the solutions to fuzzy linear system (4.3),
which have been obtained using the quasi-level-wise system approach. The initial
conditions, ½~x1 ð0Þ; ~x2 ð0Þ; ~x3 ð0Þ; ~x4 ð0ÞT , in Figs. 14, 15, 16 and 17 have been consid-
ered as ½ð0; 0; 0Þ; ð0; 0; 0Þ; ð0:1; 0:1; 0:1Þ; ð0; 0; 0ÞT , which mean with the cart
still in its original position, the pendulum is released when deviated −0.1 rad or
−5.73 ° from its vertical position. Furthermore, the initial conditions in Figs. 18 and 19
 Solving First Order Linear Fuzzy Differential Equations System 349

Fig. 14. The trajectory of the cart piostion coordinate, ~xðtÞ, based on the fuzzy linear model. The
red, green and blue lines show the trajectoris x0 ðtÞ, x1 ðtÞ, and x0 ðtÞ, respectively. The black line is
the trajectory of the cart position coordinate obtained based on the nonlinear model.

Fig. 15. The trajectory of the velocity of cart, ~x_ ðtÞ, based on the fuzzy linear model. The red,
green and blue lines show the trajectoris x_ 0 ðtÞ, x_ 1 ðtÞ, and x_ 0 ðtÞ, respectively. The black line is the
trajectory of the cart velocity obtained based on the nonlinear model.

have been considered as ½ð0; 0; 0Þ; ð0; 0; 0Þ; ð0:1; 0:1; 0:1Þ; ð0; 0; 0ÞT . In all the given
figures, the force is 0.2, and the solutions of nonlinear system (4.1) have been shown by
the black line. Additionally, the output of linear model (4.2), which is the same as the
output of fuzzy linear model (4.3) in level a ¼ 1, has been shown by the green line.
350 M. Najariyan et al.

Fig. 16. The trajectory of the inverted pendulum angle, ~hðtÞ, based on the fuzzy linear model.
The red, green and blue line show the trajectoris h0 ðtÞ, h1 ðtÞ, and h0 ðtÞ, respectively. The black
line the trajectory of the inverted pendulum angle based on the nonlinear model.

_
Fig. 17. The trajectory of the angular velocity of inverted pendulum, ~ hðtÞ, based on the fuzzy
_0 1
linear model. The red, green and blue line show the trajectoris h ðtÞ, h_ ðtÞ, and h_ ðtÞ,
0

respectively. The black line the trajectory of the angular velocity of inverted pendulum based on
the nonlinear model.

What figures suggest is that in some cases, as time goes by, significant differences
are observed between the dynamic behavior of nonlinear model, black line, and linear
model, green line. As a result, if the system is to be analyzed using the outputs of linear
model, significant error might arise that can even prove misleading. This is, however,
not the case with the outputs of the fuzzy linear model.
 Solving First Order Linear Fuzzy Differential Equations System 351

Fig. 18. The trajections of the cart position coordinate, ~xðtÞ, based on the fuzzy linear model.
The red, green and blue lines show the trajections x0 ðtÞ, x1 ðtÞ, and x0 ðtÞ, respectively. The black
line is the trajectory of the cart position coordinate based on the nonlinear model.

Fig. 19. The trajectory of the inverted pendulum angle, ~ hðtÞ, based on the fuzzy linear model.
The red, green and blue lines show the trajections h0 ðtÞ, h1 ðtÞ, and h0 ðtÞ, respectively. The black
line is the trajectory of the inverted pendulum angle based on the nonlinear model.

As is seen, based on the outputs of fuzzy linear model, one could make a more
appropriate and precise prediction of the dynamic behavior of inverted pendulum
system, compared to those of linear model, and this precise analysis and prediction
capability is the very main objective of obtaining the mathematical model of dynamic
systems. Thus, fuzzy linear model presents more information related to the system
outputs and is then richer and more resourceful.
352 M. Najariyan et al.

5 Conclusion

Having presented quasi-level-wise system approach for solving fuzzy first order linear
differential equations system, this paper showed that, preserving the inherent system
properties and formal structure, such an approach, unlike the approaches presented in
[6, 15, 16], is capable of obtaining possible system solutions. The reasons for pre-
senting such an approach are summarized as:
1. For the direct analysis of nonlinear model of dynamic systems in the presence of
uncertainty, namely fuzzy nonlinear differential equations system, the first step is to
make developments in fuzzy linear differential equations systems.
2. For making use of fuzzy linear model in simulation, stability analysis, controller
design, and so forth, such an approach seems necessary.
The example of inverted pendulum system illustrated how the linearization process
may result in fuzzy linear differential equations, and how practical and effective this
fuzzy linear model can be in the analysis of the dynamic system. The implications of
this work can shed some light on studies concerning:
1. Controller design for dynamic system whose linear model is considered as Fuzzy
Linear Time-Invariant system [20].
2. Optimization while dealing with FLTI system.
3. Solving fuzzy linear time-varying system.
4. Development of FLTI system using concepts related to type-2 fuzzy differential
equations [8] or type-2 fuzzy fractional differential equations [10].
5. Investigating Type-2 Fuzzy Differential Equations (T2FDEs) and fuzzy fractional
differential equations [8, 10, 19, 21].
6. Examining Type-2 fuzzy linear systems based on the proposed approach [22].

References
1. Antia, H.M.: Numerical Methods for Scientists and Engineers. Springer, Berlin (2002)
2. Ahmadian, A., Senu, N., Larki, F., Salahshour, S., Suleiman, M., Islam, M.S.: A legendre
approximation for solving a fuzzy fractional drug transduction model into the bloodstream.
Adv. Intell. Syst. Comput. 287, 25–34 (2014)
3. Bede, B.: Mathematics of Fuzzy Sets and Fuzzy Logic. Studies in Fuzziness and Soft
Computing, vol. 295. Springer, Heidelberg (2013)
4. Chalco-Cano, Y., Román-Flores, H.: Some remarks on fuzzy differential equations via
differential inclusions. Fuzzy Sets Syst. 230, 3–20 (2013)
5. Desoer, C., Haneda, H.: The measure of a matrix as a tool to analyze computer algorithms
for circuit analysis. IEEE Trans. Circuit Theory 19, 5 (1972)
6. Fard, O.S., Ghal-Eh, N.: Numerical solutions for linear system of first-order fuzzy
differential equations with fuzzy constant coefficients. Inf. Sci. 181, 4765–4779 (2011)
7. Gopal, M.: Modern Control System Theory. New Age International, New Delhi (1993).
Automatic control
8. Mazandarani, M., Najariyan, M.: Differentiability of type-2 fuzzy number-valued functions.
Commun. Nonlinear Sci. Numer. Simul. 19, 710–725 (2014)
 Solving First Order Linear Fuzzy Differential Equations System 353

9. Mazandarani, M., Najariyan, M.: A note on “A class of linear differential dynamical systems
with fuzzy initial condition”. Fuzzy Sets Syst. 265, 121–126 (2014). doi:10.1016/j.fss.2014.
05.018
10. Mazandarani, M., Najariyan, M.: Type-2 fuzzy fractional derivatives. Commun. Nonlinear
Sci. Numer. Simul. 19, 2354–2372 (2014)
11. Mizukoshi, M.T., Barros, L.C., Chalco-Cano, Y., Román-Flores, H., Bassanezi, R.C.: Fuzzy
differential equations and the extension principle. Inf. Sci. 177(17), 3627–3635 (2007)
12. Najariyan, M., Farahi, M.H.: Optimal control of fuzzy controlled system with fuzzy initial
conditions. Iran. J. Fuzzy Syst. 10, 21–35 (2013)
13. Najariyana, M., Farahi, M.H.: A new approach for the optimal fuzzy linear time invariant
controlled system with fuzzy coefficients. J. Comput. Appl. Math. 259, 682–694 (2014)
14. Williams, R.L., Lawrence, D.A.: Linear State-Space Control Systems. Wiley, Hoboken
(2007)
15. Xu, J., Liao, Z., Hu, Z.: A class of linear differential dynamical systems with fuzzy initial
condition. Fuzzy Sets Syst. 158, 2339–2358 (2007)
16. Xu, J., Liao, Z., Nieto, J.J.: A class of linear differential dynamical systems with fuzzy
matrices. J. Math. Anal. Appl. 368, 54–68 (2010)
17. Zimmermann, H.J.: Fuzzy Set Theory and its Applications. Springer, Boston (2001)
18. Zahreddine, Z.: Matrix measure and application to stability of matrices and interval
dynamical systems. Int. J. Math. Math. Sci. 2003, 75–85 (2003)
19. Mazandarani, M., Kamyad, A.V.: Modified fractional euler method for solving fuzzy
fractional initial value problem. Commun. Nonlinear Sci. Numer. Simul. 18(1), 12–21
(2013)
20. Najariyan, M., Farahi, M.H.: A new approach for solving a class of fuzzy optimal control
systems under generalized Hukuhara differentiability. J. Franklin Inst. 352, 1836–1849
(2015)
21. Van Hoa, N.: Fuzzy fractional functional differential equations under Caputo
gH-differentiability. Commun. Nonlinear Sci. Numer. Simul. 22(1–3), 1134–1157 (2015)
22. Najariyan, M., Mazandarani, M., John, R.: Type-2 fuzzy linear systems. Granul. Comput. 2
(3), 175–186 (2017). doi:10.1007/s41066-016-0037-y
 An LMI Based Integral SMC for Tracking Control
of a New 4-D Conservative Chaotic System

Jay Prakash Singh ✉ , Rajeeb Dey, and Binoy Krishna Roy

( )

National Institute of Technology Silchar, Silchar 788010, Assam, India

[email protected], [email protected],
[email protected]

Abstract. The objectives of the paper are (i) to develop a new 4-D conservative
chaotic system and (ii) to design an LMI based integral SMC for tracking control
of the proposed system. The proposed system has hidden chaotic orbits with no
equilibrium point and transient chaotic behaviour. The proposed controller is
designed for controlling the transient chaotic behaviour of the system. The gain
of the sliding surface is determined using LMI toolbox in MATLAB. The
performance of the proposed controller is evaluated in terms of low chattering
and small tracking error.

Keywords: New chaotic system · Conservative chaotic system · Hidden chaotic

orbits · No equilibria · LMI based SMC

1 Introduction

The dynamical systems are more interesting because of their complexity and nonlinearity
[1]. The chaotic dynamical systems have aperiodically complex behaviour and hence,
observed in many fields like communication, signal processing, economics, robotic, etc.
[2, 3]. Motivated with the various applications, many new chaotic systems are reported
with various characteristics [4]. The higher dimensional chaotic systems are more inter‐
esting and effective compared with the low dimensional systems [5, 41, 42].
In the present era, chaotic systems are divided into two groups, namely (i) self-
excited attractors and (ii) hidden attractors [6–10]. The available chaotic systems like
Lorenz system [11], Rossler [12], Lu system [13, 14] have self-excited attractors. The
chaotic systems with (i) no equilibria [15], (ii) only stable equilibria [16, 17] and (iii)
plane or line of equilibria are called the hidden attractors [18–21]. Usually chaotic
systems are of dissipative nature [22]. Very few chaotic systems are reported with
conservative nature [22–25]. Very recently two conservative 3-D chaotic systems with
no equilibria are reported [24, 25]. But, developing a conservative chaotic system with
no equilibrium point and transient chaotic behaviour is a new problem.
The transient chaotic behaviour is an interesting property of a chaotic system [26].
In this, the chaotic system exhibits chaotic behaviour for a definite time interval and
then the response goes to either a fixed point or becomes periodic [26].
A part of research in the field of chaos theory is also devoted to the control and
synchronisation of chaotic systems. In the past decade, various control schemes have

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_27
 An LMI Based Integral SMC for Tracking Control 355

been developed for controlling a chaotic or a hyperchaotic system, such as active control
[14], observer based control [27], adaptive control [28], backstepping control [29], time
delay feedback control [30], etc. Many SMC techniques are developed for the control
of a chaotic system. These are like fuzzy SMC [31], neural network SMC [32], adaptive
SMC [33], terminal SMC [34], nonsingular terminal SMC [35], integral SMC [36], PI
SMC [37, 38], LMI based SMC [39], second order SMC [40], etc. Most of the SMC
techniques use user dependent sliding surface gains [31–34]. The above mentioned SMC
schemes are designed for a chaotic or a hyperchaotic system with countable number of
equilibrium points. But, designing of a robust chattering free first order SMC for a no
equilibrium point conservative chaotic system is still an open problem.
This paper presents a new 4-D conservative chaotic system. The new system has no
equilibrium point and transient chaotic behaviour for the chosen value of the parameter.
An LMI based integral sliding mode control is proposed for tracking control problem
of the proposed system. The controller is designed for controlling the transient chaotic
behaviour of the system. It is shown that the desired trajectory tracking is achieved with
less steady state error and a very small chattering.
Organisation of this chapter is as follows: in Sect. 2, dynamics of the new 4-D
conservative chaotic system is presented. The dynamical behaviours of the new system
using numerical techniques are presented in Sect. 3. Section 4 describes the proposed
control strategy. The results and discussion of the transient chaos control problem is
given in Sect. 5. Finally, in Sect. 6, summary of the paper is derived as conclusions.

2 Dynamics of the New 4-D Conservative Chaotic System

The dynamics of the new 4-D conservative chaotic system proposed here is described
as below:

ẋ 1 = x2 + x1 x3
ẋ 2 = −x1 + x2 x3 + x4
ẋ 3 = 1 − x12 + x22 (1)
ẋ 4 = −ax1

where a is the positive constant parameter and x1 , x2 , x3 , x4 are the state variables of
system (1). Here all the simulations of system (1) are carried out in MATLAB-14a
simulation environment using ode − 45 with initial conditions x(0) = (0.2, 0.2, 0.2, 0.2)T
and step size Δt = 0.01.
System (1) is invariant under the coordinate transformation
(x1 , x2 , x3 , x4 ) → (−x1 , −x2 , x3 , −x4 ). Thus, the system has symmetry about x3 axis.
It is not hard to verify that the system has no equilibrium point. This can be verified
by equating derivative of each state variables of (1) to zero and observe that there is no
particular solution.
356 J.P. Singh et al.

3 Numerical Findings

In this section, some numerical simulations are shown to validate the transient chaotic
behaviour of system (1).
The Lyapunov exponents of the system are shown in Fig. 1. It is seen from Fig. 1
that all the Lyapunov exponents are conversing to zero with some initial positive and
negative values. Thus, system (1) has transient chaotic behaviour and is a conservative
chaotic system. The divergence of system (1) is ∇v = 0 + 2x3. Thus, if the average of
the x3 is zero, then the system is conservative which is verified from Fig. 2. The time
response of x3(t) state of system (1) for transient chaotic behaviour and periodic behav‐
iour are shown in Fig. 3. It is seen from Fig. 3 that the system has transient chaotic
behaviour from t = 0 to 1000 time unit.

0.6 L1
L2
0.4 L
3
L
4
LEs

0.2

-0.2
0 200 400 600 800 1000 1200 1400 1600 1800 2000
t

Fig. 1. Lyapynov exponents of the system with a = 0.003 and x(0) = (0.2, 0.2, 0.2, 0.2)T .

1.2

1
0.8
average of x3

0.6

0.4
0.2

-0.2
0 100 200 300 400 500 600 700 800 900 1000
t

Fig. 2. Average of the x3 state of system (1).

The phase portraits of the system with a = 0.003 for transient chaotic behaviour and
periodic behaviour are shown in Fig. 4. It is seen from Fig. 4 that the system has chaotic
behaviour for the initial transients and periodic behaviour after 1000 time unit.
 An LMI Based Integral SMC for Tracking Control 357

3 0.2

2 0

1 -0.2
1900 1920 1940 1960 1980 2000
0
x3

-1 0.5

0
-2
-0.5
800 850 900 950 1000
-3
0 200 400 600 800 1000 1200 1400 1600 1800 2000
t

Fig. 3. Transient chaotic (red colour) and periodic (blue colour) behaviours of x3 state of system
(1) with a = 0.003.

2
2
1.5

1 1

0.5

0 0
x2

3
x

-0.5
-1
-1

-1.5
-2
-2
-2 -1 0 1 -2 -1 0 1 2
x (a) x (a)
1 2

Fig. 4. Phase portraits of the system during transient chaotic behaviour (red) and periodic
behaviour (blue) with a = 0.003.

The normalised autocorrelation plot of x3(t) state of the system during transient
chaotic behaviour and periodic behaviour are shown in Fig. 5. The autocorrelation of
any signal is maximum at origin and for a chaotic signal it decreases aperiodically with
time whereas for periodic signal it decreases periodically. It is seen from Fig. 5 that the
system has chaotic behaviour during transient chaotic region and periodic behaviour
during steady state response.
358 J.P. Singh et al.

1 1

0.8 0.8

0.6 0.6

Auto (x3)
Auto (x3)

0.4 0.4

0.2 0.2

0 0
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4 0.5
t (a) t (b)

Fig. 5. Normalised autocorrelation plot of the x3 signal of system (1) for: (a) transient chaotic
and (b) periodic behaviours.

4 Control of Transient Chaos

In this section an LMI based integral sliding mode control strategy is presented for
controlling the transient chaotic behaviour of the new conservative 4-D system.

4.1 LMI Based Integral SMC

The controlled system (1) can be rewritten as:

ẋ 1 = x2 + x1 x3 + u1
ẋ 2 = −x1 + x2 x3 + x4 + u2
ẋ 3 = 1 − x12 − x22 + u3 (2)
ẋ 4 = −ax1 + u4

where u1, u2, u3 and u4 are the control inputs. The above controlled system (2) can be
written in matrix form as:

ẋ = Ax + Bu + Bg (3)

with,

⎡ 0 1 0 0⎤ ⎡1 0 0 0⎤ ⎡ x1 x3 ⎤
⎢ −1 0 0 1⎥ ⎢0 1 0 0⎥ ⎢ x2 x3 ⎥
A=⎢ B=⎢ g=⎢
0⎥ 0⎥ 2⎥
, ,
0 0 0 0 0 1 1 − x1 − x2
2
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ −a 0 0 0⎦ ⎣0 0 0 1⎦ ⎣ 0 ⎦

where A is the system matrix, B is the identity matrix and g consists of system nonli‐
nearities and a constant term.
 An LMI Based Integral SMC for Tracking Control 359

Now, control objective is to design a suitable sliding mode control so that the state
[ ]T [ ]T
x = x1 , x2 , x3 , x4 tracks the desired trajectories xd = xd1 , xd2 , xd3 , xd4 . The tracking
error vector is defined as
e = x − xd (4)
Using (3), error dynamics (4) can be rewritten as

ė = Ax + Bg + Bu − ẋ d (5)

Now, integral sliding mode surface s(e, t) ∊ R4 is designed to achieve the desired
objective. The surface s(e, t) is defined as
t
s(e, t) = e(t) + (A − BL)e(𝜏)d𝜏 (6)
∫0

where L is the gain matrix and satisfies the following condition:

(i) (A − BL) must be stable, i.e. eigenvalues of (A − BL) must be negative.
Here, the value of the gain matrix L is determined using the feasible solution of the
LMI defined in Theorem 1.

Theorem 1. Suppose there exist matrices P, Q > 0 with appropriate dimensions and
following LMI is satisfied:
{
Q>0
AQ − BP + QAT − PT BT < 0 (7)

Then the sliding mode dynamics (6) is asymptotically stable to the equilibrium point
at origin. The value of L can be obtained as L = PQ−1.
Proof: Suppose there exists a positive definite matrix Q which satisfies the following
inequality

Ak Q + QATk < 0 (8)

where Ak = A − BL. Inequality (8) can be rewritten as

(A − BL)Q + Q(A − BL)T < 0 (9)

Now consider a positive definite matrix P as P = LQ, then (9) can be modify as

AQ − BP + QAT − PT BT < 0 (10)

The value of matrices Q and P can be obtained from the feasible solutions of the
LMI defined in (7). The value of the gain matrix L can be obtained as L = PQ−1.
The existence of the sliding mode must satisfy the conditions, ṡ = 0. We have
360 J.P. Singh et al.

ṡ = B(g + Le) + Bu + A(x + e) − ẋ d = 0 (11)

The tracking of the desired trajectory for system (1) is determined using the following
Theorem 2.

Theorem 2. The states of chaotic dynamics (2) follow the desired trajectories if the
sliding surface is defined as in (6), the value of the gain matrix L is determined using
LMI defined in Theorem 1 and the control input is chosen as given in (12).
[ ] [ ]
u = − g + Le − [B]−1 A(x + e) − ẋ d − [B]−1 𝜌 tanh(s) (12)

Proof: Consider a Lyapunov function candidate as v(s) = 1∕2sT s. Its time derivative can
be written as

̇
v(s) = sT [B(g + Le) + Bu + A(x + e) − ẋ d ] (13)

Now using the control input in (12), (13) can be written as

̇
v(s) = sT ṡ < −𝜌‖s‖ (14)

Since ρ is a positive constant, v(s)

̇ in (14) is a negative definite function. Hence, the
states of the chaotic system (2) converge on the sliding surface (6) and remain on it.
Therefore, x = xd and e → 0.

5 Results and Discussion

In this, section numerical simulation results for tracking control of chaotic system (2)
are presented. The numerical simulations are carried out using ode − 45 in MATLAB
simulation environment. The initial conditions used for simulating system dynamics (2)
are considered as x(0) = (0.5, 0.5, 0.5, 0.5)T . The control law gain ρ is selected as ρ = 5.
The feasible value of the gain matrix L obtained using the LMI Toolbox in
MATLAB-14a is given in (15).

⎡ 0.5 0 0 −0.0015 ⎤
⎢ 0 0.50 0 0.50 ⎥
L=⎢ ⎥ (15)
0 0 0.50 0
⎢ ⎥
⎣ −0.0015 0.50 0 0.50 ⎦

The desired signals are chosen as xd1 = xd2 = xd3 = xd4 = sin (2.41t).
The tracking responses are shown in Fig. 6. Tracking errors of the states are shown
in Fig. 7. The integral sliding surfaces designed for tracking are shown in Fig. 8. The
control inputs are shown in Fig. 9. It is seen from Fig. 7 that tracking of the desired
signals is achieved in less than 1 s. It is noted from Fig. 7 that the steady state tracking
errors are in magnitude of order 10−4. It is observed from Fig. 8 that the chattering is
negligible.
 An LMI Based Integral SMC for Tracking Control 361

2 2

1 1

0 0
1

2
x

x
-1 -1
-2 -2
0 1 2 3 4 5 0 1 2 3 4 5
t (a) t (b)
2 2
1

0 0

x4
3
x

-1

-2 -2
0 1 2 3 4 5 0 1 2 3 4 5
t (c) t (d)

Fig. 6. Tracking responses of the states of system (2) when control input is activated at t = 0.

0.6 -4
0.6 -4
x 10 x 10
2 5
0.4 0.4
0 0
0.2
e2

0.2 -5
1

-2
e

4.5 4.6 4.7 4.8 4.9 4.5 4.6 4.7 4.8 4.9
0 0

-0.2 -0.2
0 1 2 3 4 5 0 1 2 3 4 5
t (a) t (b)
0.6 -4
0.6 -4
x 10 x 10
5 5
0.4 0.4
0 0
0.2 0.2
3

e4

-5
e

-5
4.5 4.6 4.7 4.8 4.9 4.5 4.6 4.7 4.8 4.9 5
0 0

-0.2 -0.2
0 1 2 3 4 5 0 1 2 3 4 5
t (c) t (d)

Fig. 7. Responses of the tracking error of states of system (2).

362 J.P. Singh et al.

0.6 -4 0.6 -4
x 10 x 10
2 5
0.4 0.4
0 0
0.2 0.2

2
-5
s1

s
-2 4.5 4.6 4.7 4.8 4.9
4.5 4.6 4.7 4.8 4.9 5
0 0

-0.2 -0.2
0 1 2 3 4 5 0 1 2 3 4 5
t (a) t (b)
0.6 -4 0.6 -4
x 10 x 10
5 5
0.4 0.4
0 0
0.2 0.2

4
-5
3

-5

s
s

4.5 4.6 4.7 4.8 4.9 5 4.5 4.6 4.7 4.8 4.9
0 0

-0.2 -0.2
0 1 2 3 4 5 0 1 2 3 4 5
t (c) t (d)

Fig. 8. Integral sliding surfaces.

2 2

0 0
u1

2
u

-2
-2
0 1 2 3 4 5 0 1 2 3 4 5
t (a) t (b)
1 2
0
0
u3

-1
u

-2
-3 -2
0 1 2 3 4 5 0 1 2 3 4 5
t (c) t (d)

Fig. 9. Required control input used for sinusoidal tracking of states of system (2).

6 Conclusions

In this chapter a new 4-D conservative chaotic system is proposed. The proposed system
has transient chaotic behaviour for the chosen parameter. The system has no equilibrium
point. The transient chaotic behaviour of the system is controlled to track the desired
trajectories. An LMI based SMC is proposed for tracking the desired sinusoidal signals.
The proposed control technique has resulted in small steady state tracking error and
negligible chattering. More dynamical analyses of the proposed system are the future
scope of this paper.
 An LMI Based Integral SMC for Tracking Control 363

References

1. Lakshmanan, M., Rajasekar, S.: Nonlinear Dynamics, Integrability, Chaos and Patterns.
Springer, Heidelberg (2003)
2. Andrievskii, B.R., Fradkov, A.L.: Control of chaos: methods and applications. II.
Applications. Autom. Remote Control 65(4), 505–533 (2004)
3. Zhou, Y., Hua, Z., Pun, C.M., Chen, C.L.P.: Cascade chaotic system with applications. IEEE
Trans. Cybern. 45(9), 2001–2012 (2014)
4. Pham, V., Volos, C., Wang, X., Vaidyanathan, S., Kapitaniak, T., Wang, X., Kapitaniak, T.:
A chaotic system with different families of hidden attractors. Int. J. Bifurc. Chaos 26(8), 1–9
(2016)
5. Shen, C., Yu, S., Lu, J., Chen, G.: Designing hyperchaotic systems with any desired number
of positive Lyapunov exponents via a simple model. IEEE Trans. Circuits Syst. I Regul. Pap.
61(8), 2380–2389 (2014)
6. Dudkowski, D., Jafari, S., Kapitaniak, T., Kuznetsov, N.V., Leonov, G.A., Prasad, A.: Hidden
Attractors in Dynamical Systems. North-Holland, Elsevier B.V. (2016)
7. Leonov, G.A., Kuznetsov, N.V., Vagaitsev, V.I.: Hidden attractor in smooth Chua systems.
Phys. D 241(18), 1482–1486 (2012)
8. Leonov, G.A., Kuznetsov, N.V., Vagaitsev, V.I.: Localization of hidden Chuas attractors.
Phys. Lett. A 375(23), 2230–2233 (2011)
9. Leonov, G.A., Kuznetsov, N.V.: Hidden attractors in dynamical systems: from hidden
oscillations in Hilbert-Kolmogorov, Aizerman, and Kalman problems to hidden chaotic
attractor in chua circuits. Int. J. Bifurc. Chaos 23(01), 1330002–1330071 (2013)
10. Jafari, S., Sprott, J.C., Nazarimehr, F.: Recent new examples of hidden attractors. Eur. Phys.
J. Spec. Top. 224(8), 1469–1476 (2015)
11. Lorenz, E.N.: Deterministic nonperiodic flow. J. Atmos. Sci. 20(2), 130–141 (1963)
12. Rössler, O.E.: An equation for continuous chaos. Phys. Lett. A 57(5), 397–398 (1976)
13. Lü, J., Chen, G., Cheng, D., Celikovsky, S.: Bridge the gap between the Lorenz system and
the Chen system. Int. J. Bifurc. Chaos 12(12), 2917–2926 (2002)
14. Singh, P.P., Singh, J.P., Roy, B.K.: Synchronization and anti-synchronization of Lu and
Bhalekar-Gejji chaotic systems using nonlinear active control. Chaos, Solitons Fractals 69,
31–39 (2014)
15. Jafari, S., Sprott, J.C., Reza, S.M., Golpayegani, H.: Elementary quadratic chaotic flows with
no equilibria. Phys. Lett. Sect. A Gen. At. Solid State Phys. 377(9), 699–702 (2013)
16. Kingni, S.T., Jafari, S., Simo, H., Woafo, P.: Three-dimensional chaotic autonomous system
with only one stable equilibrium: analysis, circuit design, parameter estimation, control,
synchronization and its fractional-order form. Eur. Phys. J. Plus 129(76), 1–16 (2014)
17. Molaie, M., Jafari, S., Sprott, J.C., Golpayegani, S.M.R.H.: Simple chaotic flows with one
stable equilibrium. Int. J. Bifurc. Chaos 23(11), 1350188–1350195 (2013)
18. Jafari, S., Sprott, J.C.: Simple chaotic flows with a line equilibrium. Chaos, Solitons Fractals
57, 79–84 (2013)
19. Wang, X.: A chaotic system with different shapes of equilibria. Int. J. Bifurc. Chaos 26(04),
1650069–1650075 (2016)
20. Jafari, S., Sprott, J.C., Molaie, M.: A simple chaotic flow with a plane of equilibria. Int. J.
Bifurc. Chaos 26(06), 1650098–1650104 (2016)
21. Singh, J.P., Roy, B.K.: A novel hyperchaotic system with stable and unstable line of equilibria
and sigma shaped Poincare map. IFAC-Papers On-Line 49(1), 526–531 (2016)
22. Sprott, J.C.: Elegant Chaos, Algebraically Simple Chaotic Flows. World Scientific Publishing
Co. Pvt. Ltd., Singapore (2010)
364 J.P. Singh et al.

23. Vaidyanathan, S., Volos, C.K.: A novel conservative jerk chaotic system with two cubic
nonlinearities and its adaptive backstepping control. In: Vaidyanathan, S., Volos, C. (eds.)
Advances and Applications in Chaotic System. Studies in computational Intelligence, pp. 85–
108. Springer, Cham (2016)
24. Vaidyanathan, S., Volos, C.: Analysis and adaptive control of a novel 3-D conservative no-
equilibrium chaotic system. Arch. Control Sci. 25(3), 333–353 (2015)
25. Vaidyanathan, S., Pakiriswamy, S.: A 3-D novel conservative chaotic system and its
generalized projective synchronization via adaptive control. J. Eng. Sci. Technol. Rev. 8(2),
52–60 (2015)
26. Sabarathinam, S., Thamilmaran, K.: Transient chaos in a globally coupled system of nearly
conservative Hamiltonian Duffing oscillators. Chaos, Solitons Fractals 73, 129–140 (2015)
27. Solak, E., Morgul, O., Ersoy, U.: Observer-based control of a class of chaotic systems. Phys.
Lett. A 279(1–2), 47–55 (2001)
28. Handa, H., Sharma, B.B.: Novel adaptive feedback synchronization scheme for a class of
chaotic systems with and without parametric uncertainty. Chaos, Solitons Fractals 86, 50–63
(2016)
29. Peng, C.C., Chen, C.L.: Robust chaotic control of Lorenz system by backstepping design.
Chaos, Solitons Fractals 37(2), 598–608 (2008)
30. Yamashita, Y.: Continuous-time delayed feedback control for chaotic systems. IFAC-Papers
OnLine 48(2002), 199–204 (2015)
31. Lin, T.-C., Chen, M.-C., Roopaei, M.: Synchronization of uncertain chaotic systems based on
adaptive type-2 fuzzy sliding mode control. Eng. Appl. Artif. Intell. 24(1), 39–49 (2011)
32. Chen, Y., Hung, L., Chao, S.: RBF network with sliding-mode control for chaos. In: Tan, H.,
Zhou, M. (eds.) Advances in Information Technology and Education. CCIS 2011, pp. 266–
273. Springer, Heidelberg (2011)
33. Liu, L., Pu, J., Song, X., Fu, Z.: Adaptive sliding mode control of uncertain chaotic systems
with input nonlinearity. Nonlinear Dyn. 76, 1857–1865 (2014)
34. Xiang, W., Huangpu, Y.: Second-order terminal sliding mode controller for a class of chaotic
systems with unmatched uncertainties. Commun. Nonlinear Sci. Numer. Simul. 15(11), 3241–
3247 (2010)
35. Aghababa, M.P., Feizi, H.: Nonsingular terminal sliding mode approach applied to
synchronize chaotic systems with unknown parameters and nonlinear inputs. Chin. Phys. B
21(6), 060506 (2012)
36. Toopchi, Y., Wang, J.: Chaos control and synchronization of a hyperchaotic Zhou system by
integral sliding mode control. Entropy 16(12), 6539–6552 (2014)
37. Singh, J.P., Singh, P.P., Roy, B.K.: Synchronization between hyperchaotic systems using PI
based sliding mode control, pp. 424–428. IEEE ICCSP, Chennai (2014)
38. Singh, J.P., Singh, P.P.: Synchronisation of Lu and Bhalekar-Gejji chaotic systems using
sliding mode control, pp. 20–24. ICICES, Chennai (2014)
39. Mobayen, S.: An LMI-based robust controller design using global nonlinear sliding surfaces
and application to chaotic systems. Nonlinear Dyn. 79, 1075–1084 (2015)
40. Mobayen, S., Baleanu, D., Tchier, F.: Second-order fast terminal sliding mode control design
based on LMI for a class of non-linear uncertain systems and its application to chaotic systems.
J. Vib. Control 79, 1–14 (2016)
41. Singh, J.P., Roy, B.K.: Analysis of an one equilibrium novel hyperchaotic system and its
circuit validation. Int. J. Control Theory Appl. 8(3), 1015–1023 (2015)
42. Singh, J.P., Roy, B.K.: A novel asymmetric hyperchaotic system and its circuit validation.
Int. J. Control Theory Appl. 8(3), 1005–1013 (2015)
 Tracking Control with Vibration Suppression
of a Two-link Flexible Manipulator
using Singular Perturbation with Composite
Control Design

Kshetrimayum Lochan1(B) , Rajeeb Dey1 , Binoy Krishna Roy1 ,

and Bidyadhar Subudhi2
1
Department of Electrical Engineering, National Institute of Technology,
Silchar 788010, Assam, India
[email protected], [email protected], [email protected]
2
Department of Electrical Engineering, National Institute of Technology,
Rourkela 769008, Orrisa, India
[email protected]

Abstract. In this chapter, the mathematical modelling of an assumed

modes method modelled two-link flexible manipulator moving in a hori-
zontal plane is derived using singular perturbation. The system dynam-
ics is divided into slow subsystem and fast subsystem. Considering
the separated dynamics, a composite controller is designed for tracking
the desired joint angle and tip deflection suppression. A linear matrix
inequality based sliding mode control (LMI-SMC) is proposed for the
slow subsystem and linear matrix inequality based state feedback con-
troller (LMI-SFC) is proposed for the fast subsystem. Simulation results
validate the effectiveness of the proposed control technique.

Keywords: LMI based SMC · LMI based SFC · Tracking control of a

two-link flexible manipulator · Vibration suppression

1 Introduction

With the advancement of technology, many manual works are replaced by robots
in the field of medicine, industries, military, aerospace and so on [1,2]. In the field
of robotics, many anthropomorphic manipulators are assigned to carry and move
objects in specialised sectors [3]. For faster operational speed and lesser energy
consumption, many rigid and bulky manipulators are replaced by flexible manip-
ulators [4]. In recent decades, flexible manipulators have acquired and received
much attention in modelling the dynamics and its control for completing the task
faster [5,6]. Many control methods are exploited to deal with payload variation
and modelling uncertainties [7]. Some problems are solved for controlling pay-
load variations at the free end [8]. Moreover, control of tip position of a flexible
manipulator using an adaptive model predictive approach is done in [9]. Some

c Springer International Publishing AG 2018
V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9 28
366 K. Lochan et al.

control approaches are reported for controlling a two-link flexible manipulator

(TLFM) like backstepping [10], SMC techniques [11,12], fuzzy logic control [13],
etc. However, many challenges like high non-linearity, uncertainties and elastic
deformations are hardly overcome [14].
Reference [15] discussed the control problems by applying singular perturba-
tion and applied to flexible manipulators. In [16], it is shown that the frequen-
cies of the elastic deflections are much more higher than the rigid counterparts
and hence a flexible manipulator system can be divided into two parts. A slow
subsystem and a fast subsystem. References [17,18] adopted computed torque
method for controlling the slow subsystem and pole placement for controlling
the fast subsystem. Variable structure control is used for the slow subsystem and
Lyapunov based control, virtual force control are designed for the fast subsys-
tem in [19,20], respectively. However, the choice of the controller parameters in
the above reported papers are user dependent. Hence, there is a wider scope to
improve this issue where the control parameters are computed through a process
which is feasible and optimal.
In this chapter, we propose a robust control technique which includes Linear
Matrix Inequality (LMI) with Sliding Mode Control (SMC) for the slow subsys-
tem and LMI with state feedback strategy for the fast subsystem. The simulation
results are shown where the desired trajectories are tracked and deflections are
suppressed properly.
The rest of the paper is organised as follows: Sect. 2 describes the mathe-
matical modelling of a two-link flexible manipulator. Model decomposition by
singular perturbation is given in Sect. 3. Designing of a controller, and results
and discussion are given in Sects. 4 and 5, respectively. Finally, Sect. 6 describes
the conclusions of this paper.

Fig. 1. Schematic diagram of a planar two-link flexible manipulator.

 Tracking Control with Vibration Suppression of a Two-Link 367

2 Mathematical Modelling of a Two-Link Flexible

Manipulator
Considering a TLFM is rotated in a horizontal plane by two motors, let (X0 , Y0 )
represents the generalised coordinate frame in Fig. 1. (Xi , Yi ) be the inertial
frame and (X̂i , Ŷi ) be the moving frame associated with the ith link of the rigid
body. A payload of mass Mp is attached at the tip of the manipulator. θi is the
joint angle made by each ith link. The total displacement made by the flexible
manipulator is given as
y i = θ i li + u i (1)
where ui is the elastic deflection. li is the length of ith link. The elastic deflection
is given as
n
ui (x, t) = ηij (t)ξij (x) (2)
j=1

where ηij (t) is the j modes and ξij (x) is the j th mode shape of the ith link.
th

Thus, the generalised Lagrangian co-ordinate can be given as q = [θiT , ηij

T
(t)].
The dynamic equation of a TLFM based on Lagrange’s assumed modes method
[21] can be written as
       
θ̈i θi θ̇i I
N (θij , ηij ) + h(θ, θ̇, ηij , η̇ij ) + K +D = τ (3)
η̈ij ηij η̇ij 0

which can be further simplified as

 
I
N (q)q̈ + h(q, q̇) + Kq + Dq̇ = τ (4)
0

where q = [θ1 , θ2 , η11 , η12 , η21 , η22 ]T = [θT η T ]T , N (q) is the inertia matrix,
h(q, q̇) is the centrifugal and coriolis force, K is the positive definite stiffness
matrix, D is the positive definite damping matrix and τ is the joint torque. All
the matrices are of approximate dimensions and two modes are considered here.

3 Model Decomposition by Singular Perturbation

Considering only the first two modes of a link of a TLFM, the dynamic model
can be written as
         
θ̈ hs (θ, θ̇, η, η̇) D1 θ̇ K1 θ I
N + + + = τ (5)
η̈ hf (θ, θ̇, η, η̇) D2 η̇ K2 η 0

where θ = [θ1 θ2 ]T is the joint angle, η = [η1T η2T ]T , ηi = [ηi1 ηi2 ]T is the modes
of the ith link and i = 1, 2. The mass matrix is a positive definite and its inverse
is written as    −1
T
B11 B12 Ms Msf
N −1 = B = = (6)
B21 B22 Msf Mf
368 K. Lochan et al.

−1
where Ms = (B11 − B12 B22 B21 )−1 , B11 ∈ R2X2 , B12 ∈ R2X4 , B21 ∈ R4X2 ,
B22 ∈ R 4X4
and K1 = 0. We can rewrite (5) as

θ̈ = −B11 (D1 θ̇ + hs + K1 θ) − B12 (D2 η̇ + hf + K2 η) + B11 τ (7)

η̈ij = −B21 (D1 θ̇ + hs + K1 θ) − B22 (D2 η̇ + hf + K2 η) + B21 τ (8)

For a singular perturbation, we define a new state variables η = δ and Ks = K2
and  is the perturbation parameter. 1 is the smallest value of the stiffness
constant matrix K. The singularly perturbed model of a TLFM, given in (7)
and (8), can be described as

θ̈ = −B11 (D1 θ̇ + hs ) − B12 (D2 δ̇ + hf + Ks δ) + B11 τ (9)

η̈ = −B21 (D1 θ̇ + hs ) − B22 (D2 δ̇ + hf + Ks δ) + B21 τ (10)

With the composite control strategy, the total torque can be divided as

τ = τs + τf (11)

where τs and τf are the torques for the slow and fast subsystems, respectively.
For identifying the slow subsystem, we set  = 0 in (10) and is obtained as
−1
δ̄ = −Ks−1 B̄22 (B̄21 D̄1 θ̄˙ + B̄21 h̄s + B̄22 h̄f − B̄21 τs ) (12)

where overbar indicates the value of the variable when  = 0.

Applying two-time scale perturbation technique, the slow and the fast subsys-
tems can be obtained. The slow subsystem is described as

θ̄¨ = (B̄11 − B̄12 B̄22

−1
B̄21 )(−D̄1 θ̄˙ − h̄s + τs ) (13)

Using (6), the slow subsystem can be written as

θ̄¨ = M̄s−1 (−D̄1 θ̄˙ − h̄s + τs ) (14)

In order to derive a boundary layer correction, a fast time scale, γ = √t and

√ 
boundary layer correction terms x1 = δ − δ̄ and x2 = δ is defined.
The fast subsystem is described as

ẋf = Af xf + Bf τf (15)
     
0 I 0 x1
where Af = ; Bf = ; xf =
−B̄22 Ks 0 B̄21 x2
which corresponds to the linear system parameters of the slow subsystem.
 Tracking Control with Vibration Suppression of a Two-Link 369

4 Designing of Controller
The dynamics of the slow subsystem of a TLFM, given in (14) can be written
in state space form as

ẇ = A(w)w + B(w)u + d(t) (16)

˙ u = τ , d = [0T dT ]T ∈ R4X1 .
where w = [θ̄T θ̄˙T ] = [w1T w2T ]T , w1 = θ̄, w2 = θ̄, s l
   
0 I 0
A(w) = , B(w) = (17)
0 −Ms−1 D1 −Ms−1

Here, the system matrix A(w) and input matrix B(w) are written as
   
A11 A12 0
A(w) = , B(w) = (18)
A21 A22 B2

where A11 = 02X2 , A12 = I2X2 , A21 = 02X2 , A22 = −Ms−1 D1 ∈ R2X2 , B2 =
−Ms−1 ∈ R2X2 .
Now, the state space model of (16) can be written as

ẇ1 = A11 w1 + A12 w2 (19)

ẇ2 = A21 w1 + A22 w2 + B2 u + dl (20)

where dl is the matched uncertainties and is bounded as |dl (t)| = |d˙l (t)| ≤ ll

4.1 LMI Based SMC for Slow Subsystem

Here, a normal SMC using LMI is designed for the trajectory tracking of the
slow subsystem of a TLFM. Consider the wdi ∈ R2X1 be the desired trajectories,
then trajectories tracking error dynamics is defined as

e1 = w1 − wd1
(21)
e2 = w2 − wd2

where wd1 = [θ1d θ2d ]T , wd2 = [θ̇1d θ̇2d ]T .

Assumption 1. Consider the wdi is consistent with the plant system matrix.
Let the dynamics of the desired trajectories wdi is continuous and is excited by
uw . It is described as:

ẇd1 = A11 wd1 + A12 wd2 (22)

ẇd2 = A21 wd1 + A22 wd2 + uw (23)
Now, the normal sliding surface is defined as

s(e) = e2 + λe1 (24)

370 K. Lochan et al.

where λ > 0 is the constant matrix which is determined here using LMI toolbox.
The necessary condition for the existence of sliding surface is s(e) = 0, the
equivalent sliding mode dynamics is described as

e2 = −λe1 (25)

Using (18), (19), (22) and (23), the equivalent sliding mode dynamics for e1 is
described as
ė1 = (A11 − A12 λ)e1 (26)

Theorem 1. Suppose, there exist matrices β, X, Y > 0 with appropriate dimen-

sions and satisfy the following LMI:
 
A11 X − A12 Y + XAT11 − Y T AT12 X
Q= <0 (27)
X −α

Then, the error dynamics (26) is asymptotically stable. The value of λ can be
obtained as λ = Y X −1 .

Proof. Considering a Lyapunov function candidate as


V1 (e1 ) = eT1 Re1 V̇1 (e1 ) = eT1 Rė1 + ėT1 Re1 (28)

where R is positive definite matrix. Using (26), (28) can be written as

V̇1 (e1 ) = eT1 [R(A11 − A12 λ) + (A11 − A12 λ)T R]e1 (29)

Suppose the following inequality holds

R(A11 − A12 λ) + (A11 − A12 λ)T R ≤ −α−1 (30)

then
V̇1 ≤ eT1 (−α−1 )e1
(31)
≤ −(μmin (α−1 ))e1 2
where μmin is the minimum eigenvalue of α−1 . The Lyapunov function (31) can
be simplified as
V̇1 ≤ −γV1 (e1 ) (32)
where γ = (μmin (α−1 )/μmax (R)). Since α is the positive definite matrix, the
quantity γ is the positive scalar. Now, consider X = R−1 and pre multiplying
and post multiplying X in (30) we get

A11 X − A12 λX + XAT11 − AT12 (λX)T ≤ −XαX −1 (33)

Considering Y = λX in (33) and using Shur complement [22], the LMI in (28) is
satisfied. Thus, the equivalent sliding mode error dynamics (26) is asymptotically
stable if the LMI (27) is feasible.
 Tracking Control with Vibration Suppression of a Two-Link 371

The existence of the manipulator dynamics (16) on the sliding surface (34) is
achieved using the following Theorem 2.
Theorem 2. Consider, the sliding surface defined in (24) and suppose the value
of λ is determined using LMI in (27), and if the control torque input for the slow
subsystem of TLFM dynamics (16) is defined as

τs = Ms Ps + hs + D1 θ̄˙ − ρtanh(s) (34)

where Ps = −cs e2 + z̈d , tanh is the tan hyperbolic function and ρ is the positive
gain. Then, the trajectories of the flexible manipulator dynamics converge to the
sliding surface (24) in finite time and remain on it.

Proof. Consider another Lyapunov function candidate as

1 T
V2 (s) = ss (35)
2
Taking the derivative (35) and using (24), it can be written as
η 1
V̇2 (s) = − √ ≤ −ηV32 (36)
2
where η > 0 and ρ = λll + √η2 . Hence, the joint trajectories of the flexible
manipulator (3) follow the desired trajectories using the control input torque
proposed in (34).

4.2 LMI Based State Feedback Controller (SFC) for Fast Subsystem

Here, in this section an LMI based state feedback controller (SFC) is designed
for regulating the fast subsystem dynamics of the TLFM to zero. Suppose the
matrices Af and Bf in (15) are controllable then a suitable state feedback based
controller can be designed as below.

τf = −kxf (37)

where k ∈ R2X8 is the positive definite matrix and its value is determined using
the feasible solution of a LMI obtained from Theorem 3.
Theorem 3. Suppose there exist matrices P, S > 0 with appropriate dimensions
and following LMI is satisfied:

P >0
(38)
Af P − Bf S + P ATf − S T BfT < 0

Then, the fast subsystem is asymptotically stable along the equilibrium trajectory
x̄ defined in (12). The value of k in (37) can be obtained as k = SP −1 .
372 K. Lochan et al.

Proof. Suppose there exist a positive definite matrix P which satisfies the fol-
lowing inequality
Ac P + P ATc < 0 (39)
where Ac = (Af − Bf k) is the closed loop state matrix of (15). The inequality
(38) can be written as

(Af − Bf k)P + P (Af − Bf k)T < 0 (40)

Now, consider a positive definite matrix S which is described as S = kP ,
then (40) can be modified as

Af P − Bf S + P ATf − S T BfT (41)

Thus, using (41), the LMI in (38) can be obtained. In (41) two matrices P and
S are unknown whose value can be obtained from the feasible solution of LMI
defined in (38). The gain of the SFC input k can be obtained as k = SP −1 .

5 Results and Discussion

The proposed control technique with singular perturbation modelling method is
validated on a TLFM. The parameters of the manipulator are given in Table 1.
All the simulations are solved using ode-45 solver in MATLAB-14a simulation
environment with fixed step size of 0.001. The desired trajectories for both the
joints are considered as:
π 7 −3t 7
θ1d = − e 2 + e−2t (42)
4 6 19

0.6
0.7
0.5
0.6
0.4
0.5
(rad)

(rad)

0.4 0.3
θ2, θ2d
1d

0.3 0.2
θ1, θ

0.2
0.1
0.1 θ1 θ
2
θ 0 θ2d
0 1d

−0.1 −0.1
0 2 4 6 0 2 4 6
t (s) (a) t (s) (b)

Fig. 2. Joint trajectory tracking: (a) Link-1 and (b) Link-2.

 Tracking Control with Vibration Suppression of a Two-Link 373

−4
x 10
0.05 2
(rad)

δ12 (rad)
0 0
11
δ

−0.05 −2
0 2 4 6 8 10 0 2 4 6 8 10
x 10
−3 t (s) (a) −4 t (s) (b)
2 x 10
2
(rad)

δ22 (rad)
0
0
−2
21
δ

−4 −2
0 2 4 6 8 10 0 2 4 6 8 10
t (s) (c) t (s) (d)

Fig. 3. Modes of the respective link: (a) Mode-1 of the link-1 (b) Mode-2 of the link-1
(c) Mode-1 of the link-2 (d) Mode-2 of the link-2.

0.1

0.05
u1 (mm)

−0.05

−0.1
0 1 2 3 4 5 6 7 8 9 10
−3
t (s) (a)
x 10
5
(mm)

−5
2
u

−10
0 1 2 3 4 5 6 7 8 9 10
t (s) (b)

Fig. 4. Tip deflection of the links: (a) Link-1 (b) Link-2.

Table 1. The parameters of the physical system of TLFM.

Parameter Link-1 Link-2

Length 0.201(m) 0.2(m)
Density 0.9059(kg/m) 0.3072(kg/m)
Link Moment of Inertia 0.17043(kg/m2 ) 0.0064387(kg/m2 )
Inertia of Link w.r.t joints 0.00249(kg/m2 ) 0.0008416(kg/m2 )
Gear ratio 100 50
Maximum Rotation (±90, ±90)deg (±90, ±90)deg
Drive Torque Constant 0.119(N m/A) 0.0234(N m/A)
374 K. Lochan et al.

1.5 −3
x 10
2
1
0
0.5 −2
1
s

9.6 9.8 10
0

−0.5
0 2 4 6 8 10
t (s) (a)
1.5
−4
x 10
5
1
0
0.5
s2

−5
9.6 9.8 10
0

−0.5
0 2 4 6 8 10
t (s) (b)

Fig. 5. Sliding surfaces of the slow subsystem: (a) Link-1 (b) Link-2.

5
0
(Nm)

−5
s1

−10
τ

−15
0 1 2 3 4 5 6 7 8 9 10
t (s) (a)
5
0
τs2 (Nm)

−5
−10
−15
0 1 2 3 4 5 6 7 8 9 10
t (s) (b)

Fig. 6. Required torque of the slow subsystem: (a) Link-1 (b) Link-2.

π 7 −3t 7
− e 2 + e−2t
θ2d = (43)
6 5 11
The initial conditions for simulating the manipulator dynamics are given as
q = (0.05, 0, 0, 0, 0, 0, 0, 0)T . The external disturbance d(t) added to the manip-
ulator is d(t) = [0.1sin(πq) sin(π q̇)]T . The value of different constants obtained
from the LMI is given below:
 
11.8095 0
λ=
0 11.8095

 
−0.164 14.081 30.513 49.255 3.045 −0.405 0.222 −4.960
k=
0.166 −2.133 −4.816 −10.295 −0.505 −0.565 −0.277 0.713
 Tracking Control with Vibration Suppression of a Two-Link 375

10

0
(Nm)

−10
τf1

−20

−30
0 1 2 3 4 5 6 7 8 9 10
t (s) (a)
2
(Nm)

−2
τf2

−4
0 1 2 3 4 5 6 7 8 9 10
t (s) (b)

Fig. 7. Required torque of the fast subsystem: (a) Link-1 (b) Link-2.

20

0
τ1

−20

−40
0 1 2 3 4 5 6 7 8 9 10
t (s) (a)
10

0
2
τ

−10

−20
0 1 2 3 4 5 6 7 8 9 10
t (s) (b)

Fig. 8. Composite control torque inputs of the flexible manipulator: (a) Link-1 (b)
Link-2.

The other constants used for simulating SMC of the slow subsystem are
ρ1 = 10, ρ2 = 10. These parameter values are chosen by trial and error to
maintain a good tracking performance and low control input. Figure 2 shows the
joint trajectory tracking of links. It is seen that manipulator track the desired
trajectories properly within a small time of 2 s for both the links. The obtained
two modes of each link are shown in Fig. 3. The tip deflection in Fig. 4 is the
combination of the modes with its mode shapes. It is seen in Fig. 4 that the tip
deflection of the first link and the second link is suppressed within 0.05mm and
0.005mm, respectively. The proposed sliding surfaces for the slow subsystem are
shown in Fig. 5. Figures 6 and 7 show the proposed control inputs for the slow
subsystem dynamics and fast subsystem dynamics, respectively. The composite
376 K. Lochan et al.

control inputs which is the combination of slow and fast subsystems control
inputs are shown in Fig. 8. It is apparent from the Fig. 8 that the maximum
torque requirement for the first link and second link are in bound of [−35, 10]Nm
and [−15, 1]Nm, respectively.

6 Conclusions

The dynamics of a TLFM is obtained using assumed modes method and sepa-
rated into slow and fast subsystems describing the rigid and flexible dynamics
using singular perturbation technique in this chapter. A composite control for
tracking the desired trajectories and tip deflection suppression is proposed for
a two-link flexible manipulator. A linear matrix inequality based sliding mode
control is designed for tracking the control problem of slow subsystem in order
to remove the uncertainties due to nonlinear structure. In addition, a linear
matrix inequality based state feedback control is designed for regulating the
control problem of fast subsystem. The effectiveness of the proposed control
technique is validated in simulation environment in the presence of an exter-
nal disturbance. Simulation results reflect that the objectives of the chapter are
successfully achieved.

References
1. Hua, C.: Teleoperation over the internet with/without velocity signal. IEEE Trans.
Instrum. Meas. 60, 4–13 (2011)
2. Hua, C., Yang, Y.: Bilateral teleoperation design with/without gravity measure-
ment. IEEE Trans. Instrum. Meas. 61, 3136–3146 (2012)
3. Hua, C.: A new coordinated slave torque feedback control algorithm for network-
based teleoperation systems. IEEE/ASME Trans. Mechatron. 18, 764–774 (2013)
4. Zhang, X., Xu, W., Nair, S.S., Chellaboina, V.: PDE modeling and control of a
flexible two-link manipulator. IEEE Trans. Control Syst. Technol. 13, 301–312
(2005)
5. Wai, R., Lee, M.: Intelligent optimal control of single-link flexible robot arm. IEEE
Trans. Ind. Electron. 51, 201–220 (2004)
6. Zhang, L., Liu, J.: Observe-based partial differential equation boundary control
for a flexible two-link manipulator in task space. IET Control Theory Appl. 6,
2120–2133 (2012)
7. Pradhan, S.K., Subudhi, B.: Real-time adaptive control of a flexible manipulator
using reinforcement learning. IEEE Trans. Autom. Sci. Eng. 9, 237–249 (2012)
8. Zhang, L., Liu, J.: Adaptive boundary control for flexible two-link manipulator
based on partial differential equation dynamic model. IET Control Theory Appl.
7, 43–51 (2013)
9. Pradhan, S.K., Subudhi, B.: Nonlinear adaptive model predictive controller for a
flexible manipulator. IEEE Trans. Control Syst. Technol. 22, 1754–1768 (2014)
10. Yongjiu, Y., Yang, Y., Yuan, Y., Xiaozhao, F.: Back-stepping control of two-link
flexible manipulator based on an extended state observer. Adv. Space Res. 56,
2312–2322 (2015)
 Tracking Control with Vibration Suppression of a Two-Link 377

11. Lochan, K., Roy, B.K., Subudhi, B.: SMC controlled chaotic trajectory trajec-
tory tracking of two-link flexible manipulator with PID sliding surface. IFAC-
PapersOnLine 49, 219–224 (2016)
12. Lochan, K., Suklabaidya, S., Roy, B.K.: Sliding mode and adaptive sliding mode
control approaches of two link flexible manipulator. In: Advances in Robotics, Goa,
India, vol. 1, pp. 15–20 (2015)
13. Subudhi, B., Morris, A.S.: Soft computing methods applied to the control of a
flexible robot manipulator. Appl. Soft Comput. 9, 149–158 (2009)
14. Pereira, E., Aphale, S.S., Feliu, V., Moheimani, S.O.R.: Integral resonant control
for vibration damping and precise tip-positioning of a single-link flexible manipu-
lator. IEEE/ASNE Trans. Mechatron. 16, 232–240 (2010)
15. Kokotovic, P.V.: Application of singular perturbation techniques to control prob-
lems. SIAM Rev. 26, 501–550 (1984)
16. Fraser, A.S., Daniel, R.W.: Perturbation Techniques for Flexible Manipulators.
Kluwer, Norwell (1991)
17. Khorrami, F., Ozguner, U.: Perturbation methods in control of flexible link manip-
ulators. In: Proceedings of IEEE on Robotics & Automation, vol. 1, pp. 310–315.
IEEE Press, New York (1988)
18. Siciliano, B., Book, W.J.: A singular perturbation approach to control of light-
weight flexible manipulator. Int. J. Rob. Res. 7, 79–90 (1988)
19. Mirzaee, E., Eghtesad, M., Fazelzadeh, S.A.: Maneuver control and action vibration
suppression of a two-link flexible arm using a hybrid variable structure/lyapunov
control design. Acta Astronaut. 67, 1218–1232 (2010)
20. Lee, S.H., Lee, C.W.: Hybrid control scheme for robust tracking of two-link flexible
manipulator. J. Intell. Robot. Syst. 34, 431–452 (2002)
21. Subudhi, B., Morris, A.S.: Dynamic modelling, simulation and control of a manip-
ulator with flexible links and joints. Robot. Auton. Syst. 41, 257–270 (2002)
22. Boyd, S.P., Ghaoui, L.E., Feron, E., Balakrishnan, V.: Linear Matrix Inequalities
in System and Control Theory. Studies in Applied and Numerical Mathematics,
vol. 15. SIAM, Philadelphia (1994)
Computational Intelligence in Education
 Distributive Fairness in Educational
Assessment
Psychometric Theory Meets Fuzzy Logic

Paul Hubert Vossen(&)

SQUIRE Research Institute, Kiefernweg 1a, 97996 Niederstetten, Germany

[email protected]

Abstract. If we accept that there might be a non-trivial relationship between

team performance and performance of team members in non-trivial learning
settings, then we need to consider seriously the question, what kind of rela-
tionship that may be. This paper will explore various possibilities by proposing
and analysing candidate theories and models. Moreover, we will consider dif-
ferent ways to test empirically, which theory does a better job of describing and
explaining teachers’ judgements of team and student performance. To this end,
we shall also discuss the important distinction between analytic and holistic
assessment. This research will combine Poppers’/Miller’s research paradigm of
Falsifiability and Suppes’/Narens’ research paradigm of Meaningfulness.

Keywords: Project-based learning
Team-Mate-Dilemma
Teacher

judgement
Student assessment
Team assessment
Classroom-Based
Assessment
Analytic assessment
Holistic assessment
Scoring
Rating

Distribution
Aggregation
Split-join-invariance
Distributive fairness
Additive
model
Multiplicative model
Exponential model
Ecological validity

1 Introduction

If we assume that there is a lawlike relationship between team performance and per-
formance of team’s members in an educational setting, i.e. a relationship based on
rational, logical and empirical principles and constraints, then we need to consider the
question, what kind of relationship this may be. This paper will explore various pos-
sibilities by proposing and analysing candidate theories and models1. Our research
proposal will be synthetic in that it combines the virtues of Poppers’ and Miller’s

1
Following the customs of applied research, we won’t distinguish sharply between theories and
models, and use both terms as quasi-synonyms. Note, though, that theories intend to make claims
about reality in the form of laws (confirmed relationships) or hypotheses (tentative relationships),
whereas models as such are more or less detached from reality. In some fields, however, theories and
models have come up almost simultaneously. Furthermore, models tend to be quite specific, whereas
theories may be more like broad, sweeping generalizations. Finally, some models are of a more
pragmatic nature and used successfully to tell us what to do in certain circumstances, thereby
becoming part of our world: such models may be called self-fulfilling theories. In short: theories are
about truth and knowing, models are about language and doing.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_29
382 P.H. Vossen

research paradigm with those of Suppes and Narens. The first one is a syntactic one
mainly aimed at the falsification of theories [1, 2], the latter is a semantic one mainly
focussed on the meaningfulness of theories [3, 4]. We think that such a blended
epistemological approach is more fruitful than either one on its own, as is unfortunately
mostly the case in social or behavioural research, leading to poor research designs and
outcomes.
First, however, we should clarify some methodological and epistemological points.
The field of educational assessment, like the entire field of educational practice and
research, is strongly influenced, sometimes even determined by historical, cultural, and
social forces [5, 6]. Therefore, in contrast to natural sciences, we cannot assume, or
expect, to find law like empirical regularities, independent of when, where and what we
observe to test our theories and models.
In particular, the observed kinds of relationship between team performance and
member performance will certainly not be the result of a rational process of deter-
mining or choosing the most adequate and powerful alternatives for educational
assessment. Rather it will be the unpredictable, mostly unsteady outcome of historical,
cultural and social developments with a narrow bandwidth of applicability, i.e. eco-
logical validity.
Therefore, it is of foremost importance to put together a portfolio of plausible
relationships, based on rational, logical and empirical considerations. These relation-
ships should differ in crucial respects so that we can easily recognise them when they
occur and inform practitioners as well as researchers about the implicit features and
limits of the adopted theories and used models.
Furthermore, it will be worthwhile to look from a dual perspective to our research
problem. The first, primary perspective is the one stated above: given the overall
performance of a team, how may we define and calculate in a meaningful way the
component performances of its members? This quest we have come to call the
Team-Mate-Dilemma [7], because a non-negligible cost regarding resources and
planning will be attached to any solution which differs from the equilibrium solution,
i.e. an equal distribution.
The second, dual perspective is the following: given the individual performances of
each and every team member, how may we define and calculate the combined per-
formance of the team? Given this dual perspective, it is also quite natural to ask a third
question, namely: how will the initially given team performance be related to the finally
calculated team performance?
Let us take up each of these questions one by one, starting with the aggregation and
the distribution problems, moving on to the all-important split-join-invariance property
or principle, and finally delving into the details of three distributive fairness models: the
additive, the multiplicative and the exponential model.

2 Aggregation

The aggregation problem has a much longer tradition than the distribution problem if
we just think of it in a more abstract way: how can we aggregate some parts into a
representative whole? Several branches of mathematics, both pure and applied, offer a
 Distributive Fairness in Educational Assessment 383

plethora of answers. Quite recently (2009) a comprehensive handbook, as well as a

habilitation treatise, appeared, in which the authors describe in detail all kinds of
aggregation principles, rules, and methods [8, 9]. Empirical researchers will mostly be
acquainted with the simple aggregation measures of statistics: modus, median and
mean. Some will also know that there are distinct sorts of means, like the arithmetic
mean (AM), the geometric mean (GM) and the harmonic mean (HM). Here we will
only need the weighted arithmetic and geometric mean.
In the context of educational assessment, aggregation methods may be used to
represent a set of core quality measurements, e.g. collected using n-point scales, by a
single score or rate. Such measurements may refer to individual students, teams of
students, classes, teachers, faculties or even larger units of analysis [6]. Here, we are
specifically interested in the aggregation of measurements on student level into an
aggregated measurement on the team level: the team score. Sometimes, we may have
one or more quality measurements at the team level only, i.e. not attributable to any
single student. We may combine such measurements with the measurements per stu-
dent. One way or another we will end up with a single overall team score.
This overall team score is usually the baseline and starting point for subsequent
procedures of teachers or assessors in the context of formative or summative assess-
ment. In the case of formative assessment, team scores together with all collected
evidence – qualitative, as well as quantitative –, will be used during feedback and
debriefing sessions with the teams to enhance reflective learning in their projects. In the
case of summative assessment, team scores together with student-specific data will be
used to derive individual team member scores which in turn form the basis for formal
grading, which may greatly differ from country to country.
Although issues of formative and summative assessment are important and chal-
lenging on their own, we will not deal with them in this paper. Instead, our goal will be
to disentangle problems of distribution and aggregation.

3 Distribution

The problem of distribution appears much later and less formally in the research
literature. The reason is probably that its relevance and application is usually con-
strained to such rather ethical or political issues as the fair distribution of power and
rights, or goods and commodities. The theoretical formulation goes like this: how can
we distribute a given total over a given number of constitutive parts according to some
rational or consensual principles?
Although the mathematical theory of partitions of sets or numbers comes to mind,
this theory obviously doesn’t apply to our problem, i.e. the Team-Mate-Dilemma. That
is because partitions lack some of the defining characteristics of teams like (a) the
existence of interactions between team members and (b) the achievement of a whole
team is not usually the simple sum of the individual achievements.
In the context of educational assessment, in particular, the Team-Mate-Dilemma,
publications specifically dealing with the distribution problem as such are unfortunately
quite rare, e.g. [10, 11]. Several authors circumvent a more thorough discussion and
analysis by focussing immediately and exclusively on one simple assessment technique
384 P.H. Vossen

called peer reviews. The technique of peer review puts most of the burden and
responsibility in the hands of the students and raises more problems of its own than that
it solves the underlying problem (see e.g. [12] as well as Appendix B in [13]).
In this paper we try to find a principled solution to the problem by linking it to the
general theory of distributive justice applied to fairness in assessment (judgement),
formalising it precisely in the context of the Team-Mate-Dilemma and combining it
with the notion of aggregation, which is nothing but the dual of distribution. We turn
now to the core of the matter: The Split-Join-Invariance criterion.

4 Split-Join-Invariance

We want to work out an approach to solving our core problem, i.e. the
Team-Mate-Dilemma, by combining both aspects of the problem as outlined before,
i.e. distribution (or splitting) and aggregation (or joining) of assessment scores. We will
refer to this combination as the Split-Join-Invariance principle of team assessment.
More precisely, we will rephrase our central problem in the following way: How shall
we distribute the assessed quality of team performance over team members in a rea-
sonable way such that when aggregating team members’ performance assessments
correspondingly, we will get back the original team performance judgement? (Fig. 1).

Fig. 1. The Split-Join-Invariance criterion

Thus we want to force the score distribution – the splitting of team score y in the
student scores si(y) – and subsequent score aggregation – the joining of the student
scores si(y) into the aggregated team score A(si(y)) – to be closely related via distri-
bution parameters di and aggregation parameters ai. If so, we will achieve coherence
and consistency of team and student assessment before and after these operations:
y = A(si(y)). Such a distribution which has aggregation as a natural i.e. dual coun-
terpart we like to call a fair distribution. Of course, whether (and how) we can do this
will depend critically on the splitting functions si(y) = s(y, di) with weights di and the
aggregation function A() with weights ai for student i. Furthermore, we want the
 Distributive Fairness in Educational Assessment 385

distribution and aggregation weights to be logically related, e.g. at least

order-preserving or order-inversing, depending upon how we process the distribution
weights. Also, we will generally assume that the following unanimity criterion will
hold: if di is a constant d, then si(y) = s(y, d) will be equal to the team score y; ai will be
a constant a; and A(si(y)) = A(y,…y) = y. Thus this criterion says, that if we have no
reason to assume distinct distribution weights, then all member scores will be equal to
the team score, all aggregation weights will be equal to a constant a and the aggregated
score will be equal to the original team score y.
The emerging theory of distributive fairness or equity in educational assessment
[14–18] is of course conceptually related to theories of similar names in ethics, law,
sociology, politics and other social disciplines, e.g. [19, 20]. They give us a rough
answer to the concept of fairness in testing, evaluation and assessment in education.

5 Theories of Distributive Fairness of Team Assessment

We have been able to reconstruct three formally distinct theories of distributive fairness
for our purposes. Each theory offers a coherent and consistent answer to our research
problem and allows the derivation of empirical hypotheses (simple predictions) which
we may test (i.e. falsify) in the Popperian sense [1, 2]. Of course, the derived
hypotheses must differ on some crucial points so that they can be used to decide
between the rival theories, i.e. which one applies (better) in a given case.
All three theories start from the assumption that we may measure competence or
performance on a continuous two-sided bounded scale, e.g. the standardised scale from
0 (representing a total lack of competence or worst performance) to 1 (representing full
competence or best performance on the agreed upon goals and requirements of
assessment). The scale is thought of as an ordinal scale, although we will see that we
may view the scale used in the third theory as an interval scale.

5.1 Additive Theory

The first theory conforms to standard arithmetic with its four defining operations and
some simple constructions to define distribution and aggregation. We will call this the
additive theory of distributive fairness. In fact, distribution is defined by only allowing
arithmetical addition (i.e. reward) and subtraction (i.e. penalty) on a team performance
measure. Moreover, the distribution may be done in many ways by stretching or
shrinking a given distribution within certain bounds.
We will show that we may solve our core problem by a modest technical
assumption (i.e., the Cauchy equation, see [21:8]2). Aggregation is then defined by a
weighted arithmetic mean where the weights should add to one. However, those
aggregation weights may be chosen independently from the rewards and penalties for
the distribution. This feature of our first theory strikes us as counterintuitive, but may
nevertheless appear acceptable in some applications.

2
This notation has to be read: [reference:page].
386 P.H. Vossen

5.1.1 The Basics

Given student scores xk = sk(y) = s(y, dk) as a result of the distribution of a team score
y we want to find a function A such that y = A(x1,…,xn), where n is the team size.
A shall be additive, or even quasi-additive, i.e. of the form:
Xn  Xn
y ¼ u1 a uðxk Þ where
k¼1 k
a ¼1
k¼1 k
ð1Þ

So team score y will be the weighted sum of xk, where the weights ak should add up
to 1, however, before adding we allow a transformation by a suitable function u and
after adding we “undo” this transformation again by applying the inverse of u. The
result is called a quasi-arithmetic mean, or more precisely, a u-arithmetic mean.
It may look as if we have made things now even more complicated, but with a
rather innocuous assumption about the function u and just a little bit of high school
algebra we will show that A happens to be the usual arithmetic mean, i.e. our simple
assumption will, in fact, eliminate u. Moreover, the distribution parameters dk shall be
such that there weighted sum equals 0 (some will be positive, some will be negative).
The student scores xk will be the sum y + dk of the team score y and the distribution
weights dk (so some student scores will be lower than the team score, others will be
higher). It explains fully why we call it the additive model.
Here is the simple proof. Our assumption is the so-called Cauchy equation for the
function u:

uðx þ yÞ ¼ uðxÞ þ uðyÞ ð2Þ

This functional equation restricts u to such functions which are additive in their
single argument, i.e. the value of u for the sum of scores x and y equals the sum of the
values of u for scores x and y. Under rather weak regularity conditions it turns out that
only functions of the form bx can do this [21]. With this knowledge we can go ahead now:
Xn 
y ¼ u1 a uðxk Þ
k¼1 k
Xn
¼ b1 ak bxk ð3Þ
Xn k¼1
¼ a x
k¼1 k k
Xn
¼ a k ð y þ ð xk  yÞ Þ
Xk¼1
n
¼ ð a k y þ a k dk Þ ð4Þ
Xk¼1
n
 Xn 
¼ k¼1
a k y þ k¼1
a k d k

P
Thus we see that nk¼1 ak = 1, i.e. the aggregation weights should add up to 1, and
Pn
k¼1 ak dk = 0, i.e. the weighted average of the distribution weights should equal 0. It
follows that once we have chosen arbitrarily the first n-1 distribution and aggregation
parameters, an and dn are also fixed:
 Distributive Fairness in Educational Assessment 387

Xn1 Xn1
an ¼ 1  a
k¼1 k
and dn ¼ a1
n ad
k¼1 k k
ð5Þ

One caveat: so that all student scores will always lie within the range of valid
scores, i.e. [0, 1], the distribution weights dk ought to be restricted to the interval [−y,
1 − y]. Shouldn’t this be the case we ought to multiply all distribution weights with a
constant factor so that the smallest and largest dk will fall in this range. This rescaling
will leave the weighted sum of all distribution weights 0 and will respect the rank order
and even the relative distances between the individual rewards and penalties.

5.2 Multiplicative Theory

The second theory conforms to standard mathematics as well, here analysis. However,
it is based on the concept of weighted geometric mean, forced upon us by a remarkable
result of János Aczél on the synthesis of measure judgments [21:106ff]. Let’s call this
the multiplicative theory of distributive fairness, for reasons which will become clear
below. The weighted geometric mean appears to be the only solution to a particular
functional equation involving several variables and some natural regularity conditions
like the unanimity condition, which indeed apply by assumption. Although it solves
only the aggregation part of our problem, it can readily be extended to solve the
distribution part, too.

5.2.1 The Basics

Again we start from the general split-join-invariance principle: given student scores
xk = sk(y) as a result of the distribution of a given team score y we want to find a
function A such that y = A(x1,…,xn), where n is the team size. This time however we
are looking for a solution of the following form:
Yn
y¼a k¼1
ðrk xk Þak with a; ak ; rk [ 0 ð6Þ

Again, this formulation follows from results from Aczél in [21]. We will start with
the concept of a multiplicative function A which satisfies the following functional
equation [21:27]:

AðrxÞ ¼ Aðr Þ  Að xÞ ð7Þ

This functional equation can be generalised to a multidimensional version of the

following form [21:151]:

Aðr1 x1 ; . . .; rn xn Þ ¼ Rðr1 ; . . .; rn Þ  Aðx1 ; . . .; xn Þ ð8Þ

This equation states that if we multiply the scores x1, … xn with coefficients r1, …,
rn, then the aggregated score Aðr1 x1 ; . . .; rn xn Þ of the new rescaled scores will be a
rescaling of the old aggregated score Aðx1 ; . . .; xn Þ by a factor Rðr1 ; . . .; rn Þ which is
itself a function of the rescaling factors r1, …, rn. In fact, if we assume that A also obeys
388 P.H. Vossen

the unanimity condition Aðx; . . .; xÞ ¼ x (aggregating n equal scores yields that score
again) and a = 1, then we see for xi = x = 1:

Aðr1 ; . . .; rn Þ ¼ Rðr1 ; . . .; rn Þ  Að1; . . .; 1Þ ¼ Rðr1 ; . . .; rn Þ ð9Þ

Thus, in the case of unanimity, R is, in fact, identical to A, so (8) is indeed a

generalisation of (7). Furthermore, because we are dealing here with positive numbers
on a bounded interval, i.e. [0, 1], it can be proved that only the weighted geometric
mean solves this functional equation, i.e.:
Yn Xn
Aðx1 ; . . .; xn Þ ¼ ðr x Þa k
k¼1 k k
with ak ; rk [ 0 and k¼1
ak ¼ 1 ð10Þ

For simplicity, we will set all rk to 1 in the following, so that the question becomes
now how we can express the xk in the team score y and distribution weights dk, so that
the above product will always equal y. The simplest solution we can think of is the
following:

 ak P n
ak=
Yn 1= Yn ak=
Aðx1 ; . . .; xn Þ ¼ y dk ¼ y dk ¼ y k¼1 dk ¼ y ð11Þ
k¼1 k¼1

1= P
n
ak= ¼ 1. Again, as with the additive
with xk ¼ sðy; dk Þ ¼ y dk and such that dk
k¼1
model, we are free to choose the first n-1 distribution and aggregation parameters, but
then an and dn are fixed as follows:
Xn1 an
an ¼ 1  a
k¼1 k
and dn ¼ Pn1 ak ð12Þ
1 k¼1 dk

As the distribution weights can be smaller or larger than 1, the xk can be smaller or
bigger than y, but contrary to the additive theory we do not have to worry about the xk
becoming smaller than 0 or greater than 1. We might be tempted to rescale the dis-
tribution and aggregation by the same factor because that would leave the sum of the
quotients ak =dk at 1, however, the sum of the ak wouldn’t be 1 anymore, so that we
would lose the unanimity condition, which wouldn’t be a good idea.

5.3 Exponential Theory

The third theory uses a simple version of fuzzy logic and fuzzy arithmetic. It requires a
re-definition of the addition of competence or performance scores, which takes into
account that the scale is bounded below and above, mimicking the well-known “addi-
tion” of probabilities. We will call this the exponential theory of distributive fairness for
reasons which will become clear upon looking at the resulting formulas. The remarkable
thing is that all other required operations and constructs follow by straightforward
definitions solely based on this new addition operation. Distribution and aggregation can
 Distributive Fairness in Educational Assessment 389

then be introduced such that there is again a well-defined correspondence between

distribution weights and aggregation weights without producing counterintuitive results.
Furthermore, for this third theory, it is relatively easy to introduce variations in the
aggregation weights with corresponding effects on the distribution weights such that we
can account for various distribution schemes, still obeying the split-join-invariance
criterion. Other modifications or extensions of this basic theory can be introduced, too.
For instance, we may add a team achievement factor different from the size of a team (so
that over- or underachievement of the team given the size of a team can be modelled).
Alternatively, we may add the degree of impact of distribution weights (so that the
extent of deviation from an indifferent distribution can be controlled). In fact, this third
theory is better thought of as a parametrized family of theories than as a single theory.

5.3.1 The Basics

This third theory has emerged from our research on fair distribution in educational
assessment with the goal of solving the Team-Mate-Dilemma. It is well-grounded in
fuzzy logic in that it treats the underlying scoring scale [0, 1] as a lattice of truth values
in a multivalued logic instead of the unit interval on the real line. Therefore we will
begin with a list of definitions of new operations, the most important one being a
re-definition of the addition of scores, which will have far-reaching theoretical and
formal consequences, as we will see. Here is that basic definition:

x  y ¼ x þ y  x
y ¼ 1  ð 1  xÞ
ð 1  yÞ ð13Þ

The underlying notion is that competence or performance scores may partially

overlap given the set of competence or performance factors or criteria which they
represent. Therefore, just taking the sum of x and y would over represent those
overlapping factors. They would actually be “counted” twice. Thus we should subtract
a certain amount of both x and y to compensate for this double-counting. As in
probability calculus, we have adopted a relatively natural-looking way of dealing with
this overshooting: subtract the product of scores as an approximation (proxy) of what x
and y may have in common.
Simple as it is, this fuzzy addition operation has interesting and surprising conse-
quences. All other operations and constructs we will now propose, follow more or less
directly from this single new operation. These auxiliary operations are: the difference
between two scores (14), the sum of any number of scores (15), scalar operation of a
rate – any natural number (16) or nonnegative real number (17) – on a score, the
“inverse” of this last operation (18) giving rise to a fuzzy version of the modus ponens
(19), and a way of specifying an interval scale of scores (20).

x  y ¼ 1  ð1  xÞ
ð1  yÞ1 ð14Þ
Yn
nk¼1 xk ¼ 1  k¼1 ð1  xk Þ ð15Þ

n  x ¼ 1  ð 1  xÞ n for any integer n0 ð16Þ

390 P.H. Vossen

r  x ¼ 1  ð 1  xÞ r for any real r0 ð17Þ

1 1
r!x¼  x ¼ 1  ð 1  xÞ r for any r[0 ð18Þ
r
r  ð r ! xÞ ¼ x for any r[0 ð19Þ

gð x Þ  gð x 0 Þ ¼ f ð x Þ  f ð x 0 Þ iff gð x Þ ¼ a  f ð x Þ  b ð20Þ

With these operations, we can now easily define the exponential family of dis-
tributive fairness models, as the procedure is quite similar as in the preceding case. We
define xk ¼ sk ð yÞ ¼ dk ! y, i.e. the score that is required to yield y if multiplied with
the scalar dk. Then the following holds:

Aðx1 ; . . .; xn Þ ¼ nk¼1 ak  xk
¼ nk¼1 ak  sk ð yÞ
Yn
¼1  k¼1 ð1  sk ð yÞÞak
Yn  1=
ak
¼1  k¼1 ð1  yÞ k d
Yn ak= ð21Þ
¼1  k¼1
ð 1  y Þ dk
Pan
k=
dk
¼1  ð1  yÞ k¼1
¼1  ð1  yÞ
¼y

P
n
a
where dk ¼ 1, as in the previous, multiplicative, model.
k=
k¼1
As in the multiplicative model, we need a restriction on the aggregation weights,
namely, that they should add up to 1. Otherwise, it could happen that the unanimity
criterion would not be satisfied anymore. Moreover, we may impose a restriction on the
quotients ak =dk ¼ 1=n, where n is again the team size, so that the distribution weights
can be expressed as fractions or multiples of the team size: dk = n∙ak. Thus distribution
and aggregation weights may be coupled in a very reasonable way! Finally, because of
these restrictions, once n − 1 aggregation weights ak are fixed, all other parameters will
be fixed, too.

5.4 Validation of the Three Theories

Although the first two theories of distributive fairness appear to be very straightforward
and simple, they are in fact not only too permissive (first theory) or too restrictive
(second theory), they also don’t allow any interesting modifications or extensions and
 Distributive Fairness in Educational Assessment 391

suffer from counterintuitive results (at least the first one). In contrast, the third theory of
distributive fairness appears to fully work as required and allows a lot of useful,
meaningful modifications and extensions (without sacrificing the split-join-invariance
criterion). Thus, from a rational point of view, one would expect that institutions and
teachers have adopted the third theory far more often than the first two.
However, as explained in the introduction, decisions about the adoption of edu-
cational methods including models of fair assessment will not be taken in a cleanroom,
but rather in a context of historical, cultural and social practices which are handed over
from generation to generation of teachers. Also, despite the moderate complexity and
certain elegance of the third theoretical model as shown here, it does not belong to the
standard package of assessment tools taught to lecturers today. Thus, from a realistic
point of view, we may rather expect to find many more cases in which variants of the
first or second theory have been adopted as cases in which variants of the third theory
of distributive fairness have been applied.
Thus depending upon the approach of the assessors and assessing institutions, i.e. a
more formalistic approach versus a more humanistic approach, we may predict the
adoption of entirely different models of distributive fairness in the practice of educa-
tional assessment. However, this kind of prediction is not very interesting; it is not even
a scientific prediction in the Popperian sense because it can only corroborate what we
already know. The humanistic point of view, which emphasises the historical, cultural
and social forces at work in all human affairs, will clearly dominate the formalistic
approach, insofar as they are distinct and incommensurable. Therefore, if we want to
discriminate the proposed theories of distributive fairness, in particular, the hypothe-
sised duality between distribution and aggregation as a guiding principle for dis-
tributive fairness, then we have to look for a prediction which is falsifiable in the right
context.

5.4.1 First Approach Focussing on Holistic Assessment

The context should be such, that the evaluation regime, in particular, the distribution
and aggregation of assessment scores in case of team-based assessment, is not a priori
(i.e. de jure or de facto) fixed. That is, it should not be dictated by either the empirical
approach (by conjecture associated with the first theory) or the rational approach (by
conjecture associated with the latter two theories). We can only think of such a context
in the case of a truly holistic approach to assessment as described in detail in [22]. In
such a case team performance scoring and scoring of team member performance will
not be made by prescribed assessment principles, rules, and criteria, but results from an
intuitive, unobservable judgment processes of teachers who may or may not be able to
articulate and justify their judgment in hindsight.
Unfortunately, nowadays it might be very hard to find teachers who not only in
principle favour a holistic approach to assessment, but also adopt and use it in their
classrooms3. The reason is the extensive regulation and application of analytical

3
Note, that in this research project we are exclusively interested in Classroom-Based Assessment of
Team Assignments (CBA-TA), not in e.g. large-scale testing for research purposes.
392 P.H. Vossen

approaches to educational assessment, due to the heavy involvement of governmental

and para-governmental educational bodies, pushing standardisation and accountability
of educational testing, evaluation, and assessment to ever newer heights [23]. In such a
context, a genuinely falsifiable test of the rival theories will be hard to conceive of:
there will be nothing to falsify in a Popperian sense.

5.4.2 Second Approach Comparing Analytic with Holistic Assessment

Therefore, we are thinking now of an alternative research design to demonstrate the
empirical viability, utility, usability and superiority of the second or third theory over
the first theory of distributive fairness of assessment. The basic idea is quite simple. Get
pairs of teachers involved in a field experiment possibly consisting of several in-depth
case studies, in which they follow both assessment strategies, i.e. both an analytic and a
holistic strategy. One teacher follows the standard analytic approach (according to
regulations and conventions of their institution with only marginal adjustments by the
teacher himself), and the other one follows an alternative holistic approach (completely
in the hands of the teacher). The alternative approach shall obey a minimal set of rules
which the teachers are free to define as they like as long as they refrain from specifying
assessment criteria and distribution or aggregation rules.
Thus, we will have a single assessment situation (same teachers, same students,
same lessons, same assignments), but two different assessment regimes (analytic versus
holistic). The holistic assessment regime will give us just the assessment scores for
teams and team members, and a rank order of the students according to their contri-
bution to and involvement in the team’s project, but no explicitly fixed rules how all
that is related.
Our task will be to find out, which of our theories is clearly better able to explain
how the scores and ranks fit together: the additive, the multiplicative or the exponential
theory. The rank order will be a means of checking predictions or fitting of the the-
oretical models. It may be specified by the teacher based on his or her observations, by
members of a team (e.g. through peer reviews), or both.

5.5 By Way of Conclusion

Here is a list as well as a table of basic similarities and differences between the three
theories as we have been able to work out for the moment:
• the additive has a counter-intuitive independence of distribution and aggregation
weights, the other theories don’t;
• the additive theory is typically related to the empirical-humanistic approach;
• both multiplicative and exponential theories are generally linked to the
rational-formalistic approach;
• the exponential theory offers opportunities to incorporate additional features (at the
cost of a moderate increase in complexity), the others do not (Table 1).
 Distributive Fairness in Educational Assessment 393

Table 1. Theoretical predictions regarding three distributive fairness models of assessment

Theoretical predictions Distributive fairness models
Additive Multiplicative Exponential
(arithmetic) (geometric) (logic)
pffiffiffi
xk = y + dk xk ¼ dk y xk ¼ dk ! y
Assessment Empirical Analytic OK Less probable Improbable
approach (humanistic) Holistic OK Less probable Improbable
Rational Analytic Counterintuitive OK OK
(formalistic) Holistic Counterintuitive OK OK

References
1. Popper, K.: Objective Knowledge. Clarendon Press, Oxford (1972)
2. Popper, K., Miller, D.: Popper Selections. Princeton University Press, Princeton (1984)
3. Suppes, P.: Set Theoretical Structures in Science: Problems of Representation and
Invariance. CSLI, Stanford (2001)
4. Narens, L.: Theories of Meaningfulness. Lawrence Erlbaum Associates, Mahwah (2002)
5. Madaus, G.F., O’Dwyr, L.M.: A short history of performance assessment: lessons learned.
Phi Delta Kappan 80(9), 688–695 (1999)
6. Scheerens, J., Glas, C., Thomas, S.: Educational evaluation, assessment, and monitoring.
Swets & Zeitlinger, Lisse, Netherlands (2003)
7. Vossen, P.H.: Educational assessment engineering - a pattern approach. In: Balas, V.E., et al.
(eds.) Soft Computing Applications, Proceedings of the 6th International Workshop Soft
Computing Applications (SOFA 2014). Advances in Intelligent and Soft Computing, vol.
356. Springer, Berlin (2015)
8. Grabisch, M.: Aggregation Functions. Cambridge University Press, Cambridge (2009)
9. Saminger-Platz, S.: Aggregation functions, constructions, characterizations, and functional
equations. Habilitation Treatise, Linz University (2009)
10. Hayes, J.H., Lethbridge, T.C., Port, D.: Evaluating individual contribution toward group
software engineering projects. In: International Conference on Software Engineering,
Portland (2003)
11. Lambert, S.C., Carter, A.J., Lightbody, M.: Taking the guesswork out of assessing individual
contributions to group work assignments. Issues Account. Educ. 29(1), 169–180 (2014)
12. van den Berg, I., Admiraal, W., Pilot, A.: Peer assessment in university teaching: evaluating
seven course designs. Assess. Eval. High. Educ. 31(1), 19–36 (2006)
13. Michaelsen, L., Knight, A., Fink, L.: Team-based learning: a transformative use of small
groups: appendix B. Westport, Conn., Praeger (2002)
14. Camilli, G.: Ongoing issues in test fairness. Educ. Res. Eval. 19(2–3), 104–120 (2013)
15. Dorans, N., Cook, L.: Fairness in Educational Assessment and Measurement. Routledge,
London (2016)
16. Gipps, C., Stobart, G.: Fairness in assessment. In: Wyatt-Smith, C., Cumming, J. (eds.)
Educational Assessment in the 21st Century, Chap. 6, 1st edn. Springer, Berlin (2009)
17. Lam, T.: Fairness in performance assessment. ERIC Digest. Ericdigests.org (1995). http://
www.ericdigests.org/1996-4/fairness.htm. Accessed 30 Apr 2016
18. Mislevy, R., Haertel, G., Cheng, B., Ructtinger, L., DeBarger, A., Murray, E., et al.: A
“conditional” sense of fairness in assessment. Educ. Res. Eval. 19(2–3), 121–140 (2013)
394 P.H. Vossen

19. Maiese, M.: Distributive justice. In: Burgess, G., Burgess, H. (eds.) Beyond Intractability.
Conflict Information Consortium, University of Colorado, Boulder, June 2003. http://www.
beyondintractability.org/essay/distributive-justice. Accessed 30 Apr 2016
20. Sahbaz, F.: Study of the principles of distributive justice in entrepreneurial teams. Master
Thesis, University of Oslo (2013). http://urn.nb.no/URN:NBN:no-43044. Accessed 30
Apr 2016
21. Aczél, J.: A Short Course on Functional Equations. D. Reidel, Dordrecht (1987)
22. Sadler, D.R.: Indeterminacy in the use of preset criteria for assessment and grading. Assess.
Eval. Higher Educ. 34(2), 159–179 (2009)
23. Sadler, D.R.: Evaluating complex outcomes: the shortcomings of criteria-based schemes and
what to do instead. Keynote address, University of Alberta, 6 p. (2009)
 E-Learning Web Based Grid by Utilising P2P
Technology in Presentation Broadcasting

Rafid Al-Khannak(&)

Department of Computing, Buckinghamshire New University,

High Wycombe, UK
[email protected]

Abstract. This paper expands the knowledge and implementation for elec-
tronic learning “e-learning”, by proposing the use of grid computing base on
peer-to-peer (P2P) technology in the context of web based learning. In partic-
ular, the advantages of the utilisation of grid computing which can be offered for
presentation broadcasting application is shown and approved. Moreover,
architecture for a web based learning grid based is outlined, and the notion of a
grid learning object is introduced as a first step towards the realisation of this
novel concept.

1 Introduction

E-learning has been a topic of interest in the recent years [1], mainly due to the fact that
increased scheduling flexibility as well as tool support can be offered now at a wide
affordable level, both with respect to technical prerequisites and pricing. At the same
time, the learning content can be made available even in remote places and without the
need to travel to the site where content is delivered. As a result, many e-learning
platforms and systems have been developed and commercialised; these are based on
user-server, on peer-to-peer, or, more recently, on Web service architectures [2], with a
major drawback having their limitations in scalability, availability and distribution of
computing power as well as storage capabilities. Thus, e-learning is currently deployed
mostly in areas where high requirements in any of these aspects are not mission critical,
which excludes its exploitation, for example, in most natural sciences or medical areas.
Web based learning that was developed by profit and non-profit institutions offer
various asynchronous (activities use technologies such as blogs, wikis, and discussion
boards, archived and allow participants to contribute when time allows) and syn-
chronous features (activities occur with all participants joining in at once, as with a chat
session or a virtual classroom or meeting) [1]. Presentation Broadcasting based on
webcasting technology gives solutions to the problems concerning the far distance
between presenters and audiences and the high cost to bring them together [4]. The
problems and goals of the work are the same, which are to deliver lectures to students
anywhere, through the Internet as a live event. The solutions are not only expected to
be able to deliver slide and audio-video but also to give a low cost solution.
As a typical example, where currently e-learning systems reach their limits, con-
sider a Presentation Broadcasting where Encoder system build the contents (audio,

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_30
396 R. Al-Khannak

video and slide) and streaming server send these contents. Up to now, it is vastly
impossible to have large number of audience without a high budget for providing a
large Internet bandwidth (considering a lot of Internet provider still blocking multicast
protocol) [3]. With the advanced functionality of a web based learning grid, unlimited
number of students could be provided with lecture through the Internet by using web
browser. Basically as before, the web based learning system could support the learner
with delivering lecture to students anywhere through the Internet as a live event and to
automatically record the lecture so that students can view it again afterwards.

2 Fundamentals

There are many possibilities that make up web based learning. Some are more critical
than others. In this section we introduce the fundamentals that are relevant to our work,
which fall into two categories: presentation broadcasting and P2P grid. First of all, let’s
go over what the presentation broadcast would include, and then consider how critical
and difficult each element is to achieve.
A. Presentation Broadcasting
Ideal Presentation broadcasting includes the following communication channels, in
order of content importance:
(1) Slides: Displays with text, graphics, images, including animation and special
effects. (“Slides” are not just used to mean overheads, but also to represent the
richer presentation graphics.)
(2) Audio: The speaker’s voice (“the talk”) and audio such as the sound tracks and
sound effects.
(3) Video: The “talking head” of the speakers and video clips (snapshots).
(4) Back-channels: These provide the speaker with feedback about the audience’s
status with the chat, questions, comments, and a way of recording action items,
among the other things.
(5) Slide comments and “the talk” script: When a presentation is viewed
“on-demand” from a server, having the talk in the script form allows the viewer to
read versus and hear the talk [4].
The slides and videos are two complementary visual portions of the presentation.
Several ways of dealing with slides and videos should be possible. They could be
displayed in individual windows, which the user can size and place according to their
own desire. Ideally, this would include the ability to use two monitors. In the case of a
single monitor, the presenter should be able to size and place the video and the slides
relatively to each other (see Fig. 1). For example, at some point, the video window may
be entirely hidden to give the slide the full display. At another time, the video window
may be full screened, with the slides window occupying the top-right corner, like these
are commonly seen in a news broadcast [5].
It is important to assert that the only critical media in a presentation are the pre-
sentation graphics and audio. In order to add life to the presentation, and a feeling of
presence with the presenter, it is highly desirable to display some kind of talking head
 E-Learning Web Based Grid by Utilising P2P Technology 397

Fig. 1. E-learning system components (1) View, microphone (2), webcam (3), audiences (4) and
presenter

no matter, how low the resolution or frame rates are. Occasional still shots improve the
feeling of a presentation by an order of magnitude over one without any headshots at
all. (Table 1 summarizes the important elements of presentation broadcasting).

Table 1. Important elements of presentation broadcasting

Critical media:
Audio (the talk)
Slides (presentation graphics)
Important for adding feeling of presence:
talking head (quality and frame rate not critical)

Live presentation broadcasting is delivered by using unicast and multicast methods

but for on-demand presentation broadcasting only unicast is usually used.
Unicast
Unicast is a classic approach to live streaming. In general, the setup is to use one or
more servers to manage the request from users to receive a stream (see Fig. 2). The
size/capacity of an Internet line is denoted by the bandwidth.
For example, the shared Internet line from the media server is 100 Mbps and then
an audio stream at 100 Kbps may have 1000 simultaneous listeners as a maximum. The
bandwidth needed is proportional to the size of audience [6].
This can be represented as follows:

Bandwidth Needed ¼ Size of Audience X Streaming Quality: ð1Þ

398 R. Al-Khannak

User 1

User 2

User 3

Streaming Media
Server User 4

User ...

Fig. 2. Unicast network

Doubling the quality requires a doubling or tripling in the bandwidth. The audience
mostly requires tripling the bandwidth. This causes many problems, such as the costs,
which are seen in this occasion. In other words, as the quality of the broadcast and the
size of the audience increases, the costs increases. Secondly, the scaling problem is that
since all streams are sent to the Internet from one source or perhaps two, this
approaches quick reaches to its upper limit. Bottleneck problems and “Server busy”
messages occur. So, it is difficult and it explains why doubling the bandwidth capacity
can double the price, making it a “mission impossible.”
Multicast
The multicast’s main idea, is that one may address a stream and not to be sent to a
specific machine, instead to a multicast Internet Protocol (IP) number. Any information
that has been sent to the IP number will be forwarded to all the signed up as listeners.
The stream is not duplicated by the transmitter but instead it is replicated inside the
network [7].
With multicast broadcast, the server does not create a connection to any users. The
server delivers the content to a Class D IP address1 on the network, and any user on the
network can receive it. This conserves network bandwidth. For example, a multicast
delivery to 100 users consumes only as much bandwidth as delivery to one user.
However, many networks do not support multicast delivery by default. To support the
multicast delivery, the network routers switches between the server and the users,
which must be configured to transmit Class D IP addresses and interpret multicast
information packets [6].
Although a multicast IP number is not a physical entity, it has to be supported by
hardware and software. The major questions are: Who enables it? How is it enabled?
What would happen if virtually anyone could use it, and who pays for it? It can be

1
A class D IP address is called a multicast group address. It is converted to an Ethernet address by
placing its lower 23 bits into a fixed Ethernet address. The mapping is not unique, requiring
additional filtering by one of the protocol modules.
 E-Learning Web Based Grid by Utilising P2P Technology 399

User 1

User 2

User 3
Switch

Streaming Media
User 4
Server

User ...

Fig. 3. Multicast network

costly for an internet service provider (ISP) to multicast-enable their network but when
done, the ISP can then send multiple streams to all the customers on this “local
network” offering services similar to a cable network. The following figure shows an
example of a content being distributed as a multicast stream by using a switch (Fig. 3).
Even when multicast works in both of the ways described above, there are still
issues to deal with, such as congestion problems. Furthermore, multicast introduces a
strict one-way type of communication, which involves other challenges such as how to
gather statistics [3].
B. Web Services (WS)
In essence, WS are independent software components that use the Internet as a com-
munication and composition infrastructure. They abstract from the view of specific
computers and provide a service-oriented view by using a standardized stack of pro-
tocols. To specify the operations supported by a WS, the Web Services Description
Language (WSDL) is used [8]. The Simple Object Access Protocol (SOAP) is used to
exchange structured data over the Web by building upon the HTTP protocol [9].
To discover new services on the Web or to publish existing ones, the Universal
Description Discovery and Integration Protocol (UDDI) is employed [10].
More complex Web services can be composed out of existing services using
BPEL4WS2 [11].
In Fig. 4, the typical steps of an invocation of a Web service are shown. In a first
step, suppose that a user needs to find a Web service which provides a specific
functionality. This is done by contacting a UDDI registry (step 1), which returns the
name of a server (service provider) where an appropriate Web service is hosted (step 2).
Since the user still does not know how to invoke the desired service, a WSDL

2
The Business Process Execution Language for Web Services (abbreviated to BPEL4WS) is a
notation for specifying business process behavior based on Web Services. Processes in BPEL4WS
export and import functionality by using Web Service interfaces exclusively.
400 R. Al-Khannak

3. Which operations? How to invoke?

4. Description (WSDL)
1. Where is a Web service
with operation X? (UDDI) 5. Call operation X (SOAP over HTTP)
2. Server A (UDDI) 6. Result (SOAP over HTTP)

UDDI Registry Client Web Server A

Fig. 4. Invocation steps of a web service

description will be requested which contains the name and the parameters of the
operations of the service (steps 3 and 4). The user is now able to invoke the service
using the SOAP protocol, which essentially puts the data in an envelope and sends it
over the Web by using HTTP. The service provider receives the request and executes
the desired operations on behalf of that user. The results are finally sent back to the user
by using SOAP over HTTP again (step 6) [12].
The aforementioned view of making e-learning offers available as Web services
have recently been developed [2]. Various functionalities of an e-learning system can
be decomposed and be made available as Web services individually. The decompo-
sition is based on an appropriate modelling of the processes, objects, and their inter-
actions as they occur in such a platform; finally, the results need to be mapped to
distinct services (and corresponding providers) as desired. E-learning functions result
from adequate service compositions.
C. P2P Grid
P2P grid is a distribution network which consists of individual users but appears as a
single coherent system. P2P creates a “virtual software overlay” on the existing internet
to enable collaboration and sharing of users’ resources. It capitalizes on the existing
user-machine infrastructure but nevertheless requires a degree of willingness from the
users to share their processing power and hard-disk capacity [13].
Without P2P, the users receive a stream but do not send anything. However, a user
is both capable of sending and receiving. Thus, the above solutions place the full
burden of transmitting to the end users on the broadcaster. At the same time, the end
users’ sending capacity is unused, standing in an idle mode [14].
With P2P live streaming, the idea is to use some of the idle sending capacity at the
end users. This reduces the burden on the broadcaster and may be represented as
follows:

Bandwidth needed for the broadcaster

ð2Þ
¼ Streaming Quality X Audience SizeUsed Idle End user Bandwidth:

A broadcaster could send the stream once to one end user, who would then forward
the stream to another user, who again would forward the stream, and so on. Theo-
retically, the broadcaster could have an audience of millions of people while only
 E-Learning Web Based Grid by Utilising P2P Technology 401

User 3

User 1

User 4

User 5

P2P Server
User 2

User ...

Fig. 5. P2P network in theory

sending the stream once. For a number of technical reasons, however, such an approach
with a long chain of users forwarding the stream to each other does not work [15].
Using a grid for downloading, is much easier than using it for live streaming. If a
sender turns off the PC, receivers will need time to find another sender, so it will only
extend the time you need to download. However, it could involve that you lose the
signal completely until another available user is found in the live streaming case.
Furthermore, it is not obvious how to apply the grid for live streaming: once a receivers
cannot receive the first segment then the middle and last 5 min drop as well simulta-
neously from several senders, since a first, middle and last part do not exist. The only
thing that will exist is the signal that is transmitted live. This explains why previous
technologies that tried to exploit the users’ idle bandwidth used the unreliable tree
structure (see Fig. 5).
The key of using the grid technology for broadcasting is Octoshape’s3 Live
Streaming technology, which consists of several components to obtain the full benefit
from the grid approach. In the subsequent, some of the techniques used will be
explained. The video signal is transmitted as a live stream. From the live stream,
several data streams are constructed. However, no data streams are identical. For the
purpose of this example, assume the live stream is a 400 Kbps signal and each of the
data streams has a size of 100 Kbps. In the Octoshape solution, a large number of
unique data streams are constructed. In the example above, 12 data streams are pic-
tured. Now, if an end user receives any four of the 12 different data streams of 100
Kbps, hemay use these four data streams to construct the original live stream, and thus
the video can play in real time at the end user (see Fig. 6) [16].

3
Octoshape is a peer to peer streaming media server and User which uses peer to peer grid technology
to minimize the bandwidth for any broadcaster to stream any material.
402 R. Al-Khannak

Movie Live Stream Data Streams Data Streams Movie

Fig. 6. Octoshape’s live streaming technology

3 Web Based Learning Grid

During the last decade the Internet has become a very useful broadcasting media,
particularly for on-demand services such as archive play-out. For live video and audio
streaming, the Internet is costly today. It is simply too expensive to provide streams to
large audiences simultaneously. To this end, only a few broadcasters can afford it, to
invest in large and expensive servers and networks. It is well known, that the total
bandwidth required, and thus the cost, increases proportionally to the size of the
audience. As the number of streams increases, the broadcaster has to pay more, and is
therefore penalized for his own success. In addition, investments in the streaming
server infrastructure grow rapidly with the audience size. Large streaming server plants,
capable of providing streaming services to several tens of simultaneous users, may cost
several hundred thousand of euros.
Complex applications, which are using a lot of bandwidth and handling large data
sets, have been systematically ignored in the context of e-learning until now, mainly
due to technical feasibility problems and prohibitively cost intensive. This will be
demonstrated in the remainder of this paper, P2P grid which, can close this gap and
enable new types of e-learning applications, such as presentation gridcasting. Data
could be distributed on a grid as soon as desktop computers of learners join the P2P
grid network. This particularly pertains to P2P environments where, the idle resource is
the user which in the same time uses the gridcast services, since there are computers
which, are not working to a maximum capacity of all the time. Therefore, the creation
of web based learning grids is proposed in which P2P grid functionality is integrated
into e-learning systems.
Next, is to outline the architecture for web based learning grid. To demonstrate the
technical feasibility, the architecture will be kept as simple as possible. It contains a
presentation broadcasting as well as P2P technology, which are both based on WS and
grid services (GS). In Fig. 7, presenter distributes audio, video from media encoder and
synchronizing slide script to streaming media server. It does not deliver the slides to
streaming media server, but to a web server as a HTML format. To display the pre-
sentation, internet browser with a media player is needed to view the audio, video
stream and synchronise the slides. The presentation broadcasting interacts transparently
 E-Learning Web Based Grid by Utilising P2P Technology 403

Audio & Video

capture hardware

Gridcast server User 1

User 2

Encoder Streaming media server

User 3

User 4
File server

User 5
Presenter

Web server User 6

User 7

Chat server
User 8

User 9
DB server

User ...
P2P Grid
Network
LMS server

Fig. 7. Web based learning grid system structure

with the P2P grid middleware so that, a learner is not aware of the P2P grid. Fur-
thermore, the architecture is designed in such a way that a learner only needs an
ActiveX-enabled Internet browser to use both the presentation broadcasting and the
P2P grid. The architecture shown in Fig. 7, is explained as well as the gridcast oper-
ations which is listed in this figure in more detail in the following sections.
Gridcast’s technology utilizes the redundant upstream capacity of the end user.
While the viewer watches the video on the downstream, the video is relayed to a peer
on the upstream.
(1) Encoded content is sent to the Media Server.
(2) The Media Server stores and streams content to the end user.
(3) Each user receives a unique stream (unicasting). This results in cumulative
bandwidth consumption equal to the number of users multiplied by the connection
speed.
(4) When a user clicks to get the content, a connection is made to a media server.
ActiveX manages the peering network.
(5) Gridcast server pulls a single stream from the Media Server.
(6) Gridcast server streams the content to the user 1.
(7) The second user joins the broadcast and receives the stream on user’s 1 redundant
upstream. Multiple users can jointly contribute bandwidth so that every spare
kilobit of bandwidth is used most efficiently.
404 R. Al-Khannak

(8) When there is insufficient bandwidth available on the network to support the
broadcast, GS instructs the Media Server to deliver a direct stream to the grid
network.
An E-learning provider will benefit from this integration system for the following
reasons:
(1) Deliver live simultaneous streams to an almost unlimited number of viewers
without network congestion or server duress.
(2) Bandwidth peaks are removed making it possible to manage large live events
without incurring additional cost.
(3) Broadcast quality can be guaranteed to users because the P2P grid uses multiple
sources to always deliver sufficient bandwidth.
(4) As bandwidth consumption falls, the cost savings can be rolled on to the end user
making the provider has more opportunity to double or triple media quality for the
same cost.
The e-learning PC is used by learners to access the presentation gridcasting. At the
same time, such a PC can also be used as a resource for the grid. This has been
modelled by the P2P relationship in Fig. 7 which, also illustrates that at the same time
every resource of the grid needs to have access to the learning management system
(LMS).
The LMS login service makes it possible for the user PC to become a resource in
the grid. When the learner authenticates their PC on the web page that is connected to
the login service of the LMS, the fabric layer ActiveX of the grid can be transferred as
mobile code and be started locally on the user PC. This enables communication with
the grid. The LMS login service transparently calls the grid login operation of the grid
login service with the data and the PC of the user as parameters. This completes the
authentication process of the e-learning PC user in the grid. If a user of the LMS is not
registered at the grid login service, the LMS login service could be given the authority
to create new users in the grid login service, based on a trust relationship between the
two services. These steps keep the login procedure of the grid completely in the
background, so that the learner is not aware of it.

4 Towards an Implementation of a Presentation Gridcasting

To show the feasibility of the concepts developed in the previous section, and to
illustrate the interaction between P2P gridcast, presentation slide and web interface, we
will now outline a possible implementation of a P2P grid application for Presentation
broadcasting that can be used for e-learning events like courses, trainings and con-
ferences. To this end, it is assumed that learners are to be provided with a high quality
video of a live visual presenter and synchronies slides. By opening with the internet
browser, learners can view or joint the live events. Furthermore, learners can give
feedback through different ways of back channel (e.g. poll, chat, e-mail, VOIP and
voice mail).
 E-Learning Web Based Grid by Utilising P2P Technology 405

Video
stream User

User Interface (ActiveX)

Data Streams

Gridcast Service

Live stream Streaming Media Server

Data Streams

Broker

Video
stream

P2P Grid

Fig. 8. Layers of a gridcast application layer

The relevant layers for an implementation of this scenario are shown in Fig. 8. The
user interface is implemented as an ActiveX control and runs inside an Internet browser
of a learner. The gridcast component is implemented as a P2P GS, runs on a grid
application server and represents the grid application layer. The broker provides access
to the grid, which consists of main stream and helpers. The user interface gets several
data streams, constructs back the video stream and display it. The video stream includes
script changing slide.
The GS implements a P2P streaming algorithm. Server provides the initial contact
point for users in the network. User begins receiving the stream, maintaining a buffer of
stream data in memory at all times and simultaneously initiating connections negoti-
ation with other users and receiving/serving the streams. When streaming the content,
the information about who is also joining the grid is sent as metadata with the stream.
In this way, each single user has an address book of other users in the grid. The media
player plays the stream from buffer preventing interruption to the playback. If it has
available resources, the user also continually probes the users from his standby list to
be sure that they are ready to take over if one of the current senders stops sending or
gets congested. The conventional streaming server is still available as a last resort for
users that cannot find sufficient resources in the grid available to use. This guarantees a
constant seed of the stream.
Since users participate with their idle bandwidth by sending data streams of a size
smaller than the live stream, a very flexible system is achieved. A user with an upload
capacity smaller than the size of the stream will just send fewer data streams, and a user
with a higher upstream capacity may send more data streams. If sending from a source
not located within this grid, and doing it by using unicasting, then all the traffic would
have to go through the few lines connecting the group to the rest of the Internet. When
using gridcasting, the signal has to reach one end user in the group only, which then
406 R. Al-Khannak

can spread it to the rest. All users, may contribute to the broadcast with any fraction or
multiple of the stream size without local and global congestion and bottleneck prob-
lems. Therefore, a scalable user and a low cost solution by bandwidth-saving is being
achieved.
Finally, it should be mentioned that presentation gridcasting can be obtained, by
using the presentation broadcasting and gridcasting technique. The functionality of the
P2P GS can even be extended to support collaboration tools, which would be in line
with recent developments in the intersection of tele-presentation and Web services. To
give learners the possibility to save, share files, share application, share white board
and a simulation component or equipment have to be added.

5 Conclusions

In this paper, on the one hand e-learning and web based learning have been argued and
P2P grid on the other, which considered and developed separately in the past, can
fruitfully be brought together, in particular for applications or learning scenarios where
either high bandwidth is needed or the tool sets on which learning should be done are
too expensive to be given out to each and every learner. Beyond making a case for its
feasibility and more importantly, architecture detail for an e-learning grid have outlined
which integrates core grid middleware and presentation broadcasting functionality
appropriately. Finally, we have indicated how a web based learning grid could be
realised on the basis of suitably designed P2P grid learning objects.
Clearly, what has been described in this paper is intended mainly as an attempt to
draw up a research agenda for an exploitation of grid computing in e-learning, and
many details remain to be filled in. Yet, it appears feasible to pursue the area, as there is
considerable hope for being able to extend the achievements of electronic learning
beyond the limits of individual computers.

References
1. Adelsberger, H.H., Collis, B., Pawlowski, J.M. (eds.): Handbook on Information
Technologies for Education and Training. Springer, Berlin (2002)
2. Vossen, G., Westerkamp, P.: E-learning as a web service (Extended Abstract). In:
Proceedings 7th International Conference on Database Engineering and Applications
(IDEAS). IEEE Computer Society Press, Hong Kong, China (2003)
3. Austerberry, D.: The Technology of Video and Audio Streaming, 2nd edn. Focal Press,
Waltham (2005)
4. Gemmell, D.J., Bell, C.G.: Non-collabotative telepresentations come of age. Commun. ACM
40(4), 79–89 (1997)
5. Microsoft Corporation, PowerPoint 2003 Online help (2005)
6. Microsoft Corporation, Windows Media Services help (2005)
7. Demetriades, G.C.: Streaming Media: Building and Implemented a Complete Streaming
System. Wiley, Hoboken (2003)
 E-Learning Web Based Grid by Utilising P2P Technology 407

8. Christensen, E., Curbera, F., Meredith, G., Weerawarana, S.: Web services description
language (WSDL) 1.1. W3C Note, 15 March 2001. http://www.w3.org/TR/2001/NOTE-
wsdl-20010315
9. Box, D., Ehnebuske, D., Kakivaya, G., et al.: Simple object access protocol (SOAP) 1.1,
W3C Note, 08 May 2000. http://www.w3.org/TR/2000/NOTE-SOAP-20000508
10. Bellwood, T., Clément, L., Ehnebuske, D., et al.: UDDI Version 3.0. UDDI Spec Technical
Committee Specification, 19 July 2002. http://uddi.org/pubs/uddi-v3.00-published-
20020719.htm
11. Andrews, T., Curbera, F., Dholakia, H., et al.: Specification: business process execution
language for web services version 1.1. IBM developerWorks, 05 May 2003. http://www.
ibm.com/developerworks/library/ws-bpel/
12. Li, M., Baker, M.: The Grid Core Technologies. Wiley, Hoboken (2005)
13. Subramanian, R., Goodman, B.D.: Peer-To-Peer Computing: The Evolution of a Disruptive
Technology. Idea Group Publishing, Hershey (2005)
14. ESM. http://esm.cs.cmu.edu/ (2006)
15. Peercast. http://www.peercast.org/ (2006)
16. Alstrup, S., Rauhe, T.: Introducing Octoshape, EBU Technical Review No. 303, July 2005
Soft Computing and Fuzzy Logic in
Biometrics (SCFLB)
 A Circular Eccentric Iris Segmentation
Procedure for LG2200 Eye Images

Cristina-Madalina Noaica1,2(&)
1
Department of Computer Science, University of Bucharest, Bucharest,
Romania
[email protected]
2
Applied Computer Science Testing Laboratory (ACSTL), Bucharest, Romania

Abstract. This paper proposes an eccentric circular ring iris extractor as an

instrument for future comparative iris recognition studies and, also, as an
intrinsic result pointing, among other things, to a new kind of segmentation error
documented and exemplified here for the first time, as far as we know, namely
the eccentricity detection error. There are iris code matching errors for which the
eccentricity detection error is the only cause, matching errors which unfortu-
nately are insurmountable. Otherwise, the proposed segmentation procedure
performs reasonably well on such a difficult database (LG2200 subset of
ND-CrossSenssor-Iris-2013 dataset, consisting in 116,564 eye images), proving
a failure rate of 3.26%.

Keywords: Iris segmentation
LG2200

1 Introduction

There is a long discussion in the literature on iris recognition if the best segmentation
procedure should consider the pupillary and iris-sclera boundaries as being ellipses or
circles [3–5, 7, 8, 10, 20, 30]. In the second case, the iris segment can be further treated
as a concentric or eccentric circular ring. All our previous works on iris recognition [15,
17, 21] and the works closely related to them [22–26] assume the concentric circular
ring for the geometry of the iris segment. On the contrary, this paper proposes an
eccentric circular ring iris extractor as an instrument for future comparative studies on
iris recognition and also as an intrinsic result pointing, among other things, to a new
kind of segmentation error, documented and exemplified here for the first time (as far
as we know), namely the eccentricity detection error. There are iris code matching
errors for which the eccentricity detection error is the only cause, matching errors
which unfortunately are insurmountable. Otherwise, the proposed segmentation pro-
cedure performs reasonably well on such a difficult database (LG2200 subset of
ND-CrossSenssor-Iris-2013 dataset [16], consisting in 116,564 eye images), proving a
failure error rate of 3.26%.
The outline of this paper is as follows: the proposed segmentation procedure is
described in the 2nd section in the context given in the 1st section (related works). All
the processing steps of the proposed segmentation procedure are classified with respect
to their purpose and explained in separate subsections: processing steps for edge

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_31
412 C.-M. Noaica

detection, for cropping, for voting and selecting circles of interest. The paper continues
by presenting our experimental results (the 3rd section) and ends with few conclusions
and possible future works.

2 Related Works

There are many iris segmentation techniques that took into consideration the eccen-
tricity of the inner and outer boundaries of the iris. For instance, Guo et al. [6], while
working on CASIA [2], used the eccentricity information to decide if the inner iris
boundary is better approximated by a circle or by an ellipse. On the contrary, others
have chosen to work with the circular concentric model of the iris [26] or with the
circular eccentric model of the iris [8, 13, 14], without expec or documenting a negative
effect eventually affecting the iris matching scores and overall recognition performance
in the iris recognition tests that they have undertaken.
This situation, in which both circular eccentric and concentric approaches to iris
segmentation appear equally well suited for the job, motivated us to define and test our
own first eccentric circular iris segmentation approach, in order to open the possibility
of comparing iris recognition results obtained in the same lab with both concentric and
eccentric circular iris segmentation procedures.
Two of the segmentation techniques designed for the same iris sensor as the one
used in this paper are the following ones. Liu et al. [11] developed an improved version
of Masek’s iris segmentation procedure, called IrisBEE (Iris Biometric Evaluation
Environment) [19]. For IrisBEE, the authors report that the rank-one recognition rate
has increased from 90.92% (Masek’s iris segmentation procedure) to 96.61% when
tested on LG2200 database. The improvement is reported to originate in four major
changes:
– The inversion of the order in which the two boundaries of the iris are detected.
Masek’s segmentation procedure detects firstly the outer boundary of the iris and
then the inner one. Liu et el. decided that, since the pupil boundary is the most
visible one, it should be the one that is firstly detected;
– The elimination of those edges that caused the Hough transform to vote false iris
edges;
– Restricting Hough transform to vote only inside a 30 degrees area on both left and
right sides of the center candidate, unlike Masek’s segmentation version where
Hough transform votes for circles in all directions around the center candidates;
– Selecting the circles corresponding to the inner and outer boundary of the iris
according to a “hypothesize and verify” approach instead of only determining the
maxima in Hough accumulator. Shortly, this approach verifies if the current pupil
circle candidate voted in Hough accumulator actually encloses the pupil and, also, if
the current outer boundary circle candidate is surrounded by the sclera.
However, given the fact that Masek designed his original segmentation procedure
[12] in 2003 while using CASIA-1 database (an ideal database), the rank one recog-
nition rate of 90.92% on LG2200 database (a difficult database) is still a notable
 A Circular Eccentric Iris Segmentation Procedure 413

performance, indicating that this ‘old’ segmentation tool keeps its potential to evolve
surprisingly due to its initial robust design.
The second segmentation method is called VASIR (Video-based Automatic System
for Iris Recognition) and was released in 2005–2006 by Lee et al. [9]. This is an open
source iris segmentation and recognition system, adapted for both video sequences and
still-images. VASIR firstly identifies the location of the pupil inside a binary image
obtained after thresholding the initial eye image with a value that is automatically
generated. A Gaussian filter and some morphological opening and closing operations
are applied on the binary image in order to eliminate a part of the noise. On the resulted
edge image, the pupillary boundary is searched as being a circle with half the radius of
the major axis of the ellipse that approximates the pupil. For finding the limbic
boundary of the iris, VASIR uses the same modified Hough transform as IrisBEE. In
order to avoid false positives for the outer boundary of the iris, VASIR selects the
candidates for the limbic circle inside a disk around the candidate center of the pupil.
Even though these two segmentation techniques (IrisBEE and VASIR) are popular
[18, 20, 27], since the LG2200 database is a very difficult one, there is a lot of room for
improvements. Therefore, the differences between the two iris segmentation techniques
and the one proposed in this paper are as follows. The proposed segmentation tech-
nique starts, in the image processing phase, with the enhancement of the edges present
in the eye image, then, allows the Hough transform to look for iris boundaries in all
directions around the coordinates of the center candidates, and, finally, selects pairs of
circles that have the highest number of votes according to Hough accumulator and that
satisfy several conditions, as expressed in Sect. 3.
The fact that the iris and pupil are eccentric to each other (their centers are distinct,
but the iris always fully includes the pupil) is known ever since Daugman proposed the
rubber sheet method [5] for iris unwrapping. Up to date, many other authors [28–31]
described segmentation procedures built to cope with instances of eccentric iris-pupil
boundaries.

3 Proposed Segmentation Procedure

Our segmentation procedure is described below and illustrated in Fig. 1. It consists in

two processing threads (phases I and II) executed before Hough voting and circle
selection (phase IV), in order to eliminate (phase III) much of the false radius and
center candidates.
ACSTL LG2200 Segmentation Procedure (M. Noaica, 2016):
– Processing steps for edge detection and enhancement (Fig. 1, left side column):
image intensity adjustment; edge detection; noise removal; edge dilation;
– Processing steps for cropping (Fig. 1, right side column): image normalization;
thresholding; morphological operation for filling the gaps caused by the specular
lights;
I. Decimation in the space of radius-centers candidates;
II. Hough voting and circles selection;
414 C.-M. Noaica

Initial image

(a) image intensity ad- (e) image

justment normalization

(b) edge detection (f) center area

approximation

(c) noiseremoval (g) filling the gaps

(d) edge
dilation

pupillary and limbic boundary localization

Fig. 1. ACSTL LG2200 Segmentation Procedure: The contrast of the original image is adjusted
(a) in order to obtain clearer iris boundaries (b)–(c) further dilated (d) and decimated by
thresholding the distance to the area of center candidates (g) which is the region where the centers
of the inner and outer boundaries of the iris region are more likely to be found.

3.1 Processing Steps for Edge Enhancement and Detection

For the large majority of eye images, increasing the contrast results in enhancing the
boundaries of interest, in our case the inner and the outer boundary of the iris. Since an
important part of the unnecessary edges belongs in regions with high chromatic (in-
tensity values above a threshold), as a first step of edges decimation, in the image
resulted after increasing the contrast, all the almost white regions are flatted at 255, and,
therefore, the edges found within these regions lose their chance to interfere in Hough
voting mechanism. Canny edge detector is used in order to obtain the edges in the
resulted image and then a second edge-decimation step takes place such that more
 A Circular Eccentric Iris Segmentation Procedure 415

unnecessary edges vanish (any chain formed with less than 120 pixels is eliminated
from the edges’ map). The final step consists in dilating the edges, which increases the
chances for Hough transform to identify correctly the inner and outer boundaries of the
iris.

3.2 Processing Steps for Cropping

The important contribution of cropping to the segmentation procedure is that it helps
reducing the area where the boundaries of interest may be searched, which results in
eliminating cases of false positives. The cropped area around the pupil, C, is obtained
by thresholding the eye image at 20% intensity:

I
C¼ð Þ  0:2; ð1Þ
255

The result, C, indicates the center for the cropping. In Fig. 1f, the region corre-
sponding to the candidate circle centers is represented with white. The specular lights
may often introduce gaps overlapping with the actual pupil/iris centre coordinates. To
overcome this effect, all the gaps found within the cropped area are filled and taken into
consideration when the circular Hough transform is applied (Fig. 1g).

3.3 Hough Voting and Circle Selection

Since the proposed segmentation procedure considers the iris to be an eccentric circular
ring, it returns two circles with the property that their centers are situated at a certain
distance from each other, distance determined in concordance with the radius of the
pupillary circle. The circular Hough transform evaluates any circle in the [20, 160]
radius interval. Each candidate circle receives a percentage as its final vote. For the kth
candidate circle, the percentage indicates the degree of coverage of the circle, given the
number of pixels that are voting:

ANPðkÞ  100
vðkÞ ¼ ; ð2Þ
ENPðkÞ

where ANP(k) is the actual number of pixels voting for the kth circle and ENP(k) is the
expected number of pixels that should vote (ideally) for the same circle.
The pupillary circle (pC) is selected from the sorted set of pupillary Circles (sspC).
sspC is obtained by sorting in descending order, based on the received percentage of
votes, the circles (spC, set of pupillary Circles) found in the [20, 35] radius interval. For
each selected pC, is determined a set (slC–set of limbic Circles) of candidate limbic
circles, with the property that their radius is at most one and a half bigger than the one of
pC. From slC are selected only those circles whose center coordinates are found in an
interval. The interval starts at [pCx, pCy], the coordinates of pC, and increases until it
reaches one third of pC radius. The limbic circle (lC) is determined from the newly
established interval and is the one with the highest percentage of votes. The pseudo-code
416 C.-M. Noaica

Table 1. Pseudo-code for circle selection procedure

Algorithm 1: Circle selection

INPUT: all_candidate_circles, being all the circles that are found by circular Hough transform
OUTPUT: The circles corresponding to the inner and outer boundaries of the iris
SET spC = {all_candidate_circles | radius(all_candidate_circles)<35};
sspC = SORT(spC);
FOR each pC in sspC
slC = {all_candidate_circles | (radius(all_candidate_circles) > radius(pC) * 1.5)};
SET lC = Matrix(0);
WHILE empty(lC)
SET interval = [pCx,pC y] +/- C (where C is a constant, and pC x and pC y,
respectively, are the cartesian coordinates for the pupil center);
lC = {slC | ([lC x,lC y] interval)} (where [lCx,lC y] are center coordinates for lC)
IF empty(lC)
Increase the value of C
IF C ≥ radius(pC) / 3,
BREAK
END IF
ELSE
SET innerBoundaryCircle = pC;
SET outerBoundaryCircle = max(votes(lC)) % select the
% circle with the most votes from the slC set of candidate
% circles for the limbic boundary
BREAK
END IF
END WHILE
IF (is found a pair of inner and outer boundaries)
BREAK
END IF
END FOR

for selecting the circles that represent the inner and outer boundaries of the iris is found
in Table 1.

3.4 Discussion on the Segmentation Improvements

Figure 2 contains two examples illustrating that performing the image intensity
adjustment leads to improved results. The figure consists in four eye-edge pairs, the eye
image having marked with white circles the estimated inner and outer edges of the iris.
The pairs from the left column (Fig. 2a, c) are obtained adjusting image intensity,
while their right-side pairs (Fig. 2b, d) are the corresponding examples in the case
when the initial images have not been subject to this preprocessing step.
 A Circular Eccentric Iris Segmentation Procedure 417

(a) (b)

(c) (d)

Fig. 2. Examples of segmentation results: (a) and (c) – results obtained after applying
enhancements; (b, d) - results obtained without applying enhancements.

(a)

(b)

(c)

(d)

Fig. 3. The influence of preprocessing: (b, d) edge maps obtained with (d) and without
(b) image enhancements and their corresponding Hough accumulators (a, c).

As it can be seen in Fig. 2b, applying the processing steps listed in Fig. 1b–d
(without preliminary intensity adjustment) could exclude all edges of the pupil, but if
the intensity of the image is adjusted first, this problem is remediated. For the eye
images displayed in Fig. 2c–d, the effect of adjusting the intensity of the eye image was
418 C.-M. Noaica

the elimination of some edges that otherwise can give decisive votes for a false circle
(Fig. 2d).
Furthermore, Fig. 3 illustrates how the boundary detection improves when all
preprocessing steps from the left column of Fig. 1 are applied. Figure 3a shows the
values stored in the Hough accumulator for the edge-map presented in Fig. 3b. By
analogy, Fig. 3c shows the values stored in the Hough accumulator for the edge-map
obtained after enhancement (Fig. 3d). It is interesting to observe that, in Fig. 3a, the
maximum coverage of a circle is 80%, whilst after enhancement (Fig. 3c) the pupillary
boundary was detected with 100% circle coverage. Also, comparing Fig. 3a with
Fig. 2a–b, we can see that, for the edge map obtained without image intensity
enhancement, since the circle selector expects a certain ratio between pupil and iris
radii, it decided that there is no plausible pair of circles detected as inner and outer iris
boundaries.

4 Experimental Results

The segmentation procedure described above was tested on LG2200 dataset (116’537
near infra red iris images). The average execution time for an image was one second
without any speed optimization. As may be seen in Table 2, the initial segmentation
failure rate was 3.88%, but after eliminating all the images that cannot be used for iris
recognition, the segmentation failure rate slightly improved, up to 3.26%.

Table 2. Segmentation failure rate

No. images No. errors ERR_all(%) No. ellim ERR_ellim(%)
116537 4528 3.88 722 3.26

After a visual inspection of all segmented iris images, a number of 4528 images
were marked as segmentation failures, out of which 722 images were found as not
having enough quality to serve for iris recognition. Several examples of such images
may be seen in Fig. 4. After a long and intense study and practice of iris recognition
along the last six years (2010–2016), we conclude that pretending automatic iris
recognition on such images is rather fiction that science. Also, the commercial value of
such approach is very limited because, until now, we didn’t find at least one iris
recognition specialist wishing to protect his data or his banking account with such
images, or able to guarantee under his signature a reasonably security-comfort balance
for third parties that hypothetically could wish to use such images to protect their
assets.
When it comes for iris hunting, the situation is slightly different mainly because
such images can be subject to manual, case-by-case based iris recognition, but even in
this approach, the number of false matches may grow rapidly and uncontrollably, in
counterbalance with the degree of certainty for inferring one specific identity.
 A Circular Eccentric Iris Segmentation Procedure 419

Fig. 4. Images that lack the necessary level of quality for allowing automatic iris recognition

4.1 Segmentation Failures

There are four categories of image characteristics that caused errors among the seg-
mented iris images:
– The edge image does not contain the pupillary boundary (Fig. 5a).
– The candidate-centers image does not indicate any approximation of the centers of
the two circles (Fig. 5b).

(a)

(b)

(c)

Fig. 5. Pure segmentation failure examples. Columns from left to right: edges,
center-candidates, and estimated boundaries for the iris.
420 C.-M. Noaica

– Sometimes, a slight occlusion may cause the algorithm to find a circle of a smaller
or bigger radius than the correct one. This happens when the surrounding edges
from the image vote for that false circle (Fig. 5c).
– The strong occlusion present in some of the images prevented the algorithm to
perform a successful segmentation (Fig. 4).

4.2 Enhancement’s Effect on Recognition Performance

Performing image intensity enhancement leaded to enhancing the iris boundaries and
not necessarily the iris patterns. In order to analyze what is the effect on the iris
patterns, we extracted iris segments from both the original (Fig. 6e) and the enhanced
iris images (Fig. 5f). Let us denote the iris segments extracted from the original images
with S1, and the ones obtained from the enhanced images with S2. In both cases, the iris
segments are extracted using Daugman’s rubber sheet model and then encoded with
Log-Gabor filter. Figure 6 illustrates the histograms of an original (Fig. 6a) and the
corresponding enhanced (Fig. 6b) eye image, the histograms of extracted iris segments
(Fig. 6c–d), the iris segments (Fig. 6e–f) and their corresponding iris codes (Fig. 6g–
h).
In order to determine the degree of alteration that intensity adjustment causes to the
iris patterns, on a subset of 1878 iris codes from S1 and on the corresponding subset of
S2, is performed one-to-one, id-to-id, S1-to-S2 Hamming similarity-based comparison.
The minimum and maximum Hamming similarity scores are recorded here in Table 3.
Also, analyzing Fig. 7, we can see that the majority of the similarity scores are situated
within the (0.92, 0.97) interval.
The conclusion is that the enhancement does not have an important negative effect
on the iris patterns; on the contrary, the minimum score obtained by comparing the iris
codes extracted from original and enhanced iris images is still a great recognition score.
More importantly in this context, the score distribution drawn in Fig. 7 illustrates an
upper bound for the maximally achievable recognition performance (when using
Hamming similarity for iris code matching), mainly because recognizing the same iris
in two different eye postures will be always much more difficult than recognizing the
same iris in exactly the same eye posture but with a slightly different chromatic
enhancing the iris boundaries.
A limitation occurs naturally when it comes to use the same segmentation proce-
dure on a different iris image dataset. It is known that, usually, different enhancement
procedures are necessary for different iris image datasets, as (correctly) remarked in [1]
in the context of analyzing iris camera interoperability.
Figure 8 illustrates some examples of segmentation successfully performed on the
LG2200 database. It can be seen that, as long as the quality of the eye images is good
enough to be used for iris recognition, our segmentation procedure succeeds in
retrieving the iris segment correctly. The image collection in Fig. 8 exemplifies that the
proposed segmentation procedure overcomes reasonable illumination differences, dif-
ferent types of iris-eyelid contact and different levels of dark noises around the iris
(including the makeup).
 A Circular Eccentric Iris Segmentation Procedure 421

(a) (b)

(c) (d)

(e) (f)

(g) (h)

Fig. 6. Histograms of the original (a) and enhanced (b) images; Histograms of the iris segments
obtained from the initial (c) and enhanced (d) images; Iris segments extracted from the initial
(e) and enhanced (f) images, and their corresponding iris codes ((g) and (h)).

Table 3. Hamming similarity between iris codes obtained with and without intensity adjustment
No. Iris Codes Minimum Hamming Similarity Maximum Hamming Similarity
1878 0.8569 0.9833

4.3 The Influence of the Eccentricity Detection Error

When practicing automatic iris segmentation based on the proposed segmentation
procedure, it is possible to obtain, for the same iris and for the same pupil dilation,
different pupil-iris eccentricity. We define here the Eccentricity Detection Error for two
irides with the same pupil dilation as being the difference between the vectors deter-
mined by the iris and pupil centers in the two images:
For example, in one image the pupil center Cp1 may be detected as being 8 pixels
above and 8 pixels to the right of the iris center Ci1 , while in a second image having the
422 C.-M. Noaica

Fig. 7. One-to-one, id-to-id, S1-to-S2 Hamming similarity scores obtained by comparing iris
codes extracted from the original and corresponding enhanced eye images.

Fig. 8. Several iris segmentation examples (ACSTL LG2200 Segmentation Procedure – C.M.
Noaica, 2016) with eye images from LG2200 dataset.

same pupil dilation, the pupil center Cp2 may be detected as being 8 pixels lower and 8
pixels to the right of the iris center Ci2 , case in which EDE = 16. If the vector between
two centers is used in recognition, this is a clear False Reject case. If the iris codes are
matched by Hamming similarity, this is also a False Reject case, while only the visual
examination can lead to a True Accept case.

ƒƒ1ƒ!1 ƒƒ2ƒ!2
EDE :¼ Cp Ci  Cp Ci

; ð3Þ
 A Circular Eccentric Iris Segmentation Procedure 423

5 Conclusions

An eccentric circular ring iris extractor was proposed and tested in this paper as an
instrument for future comparative iris recognition studies and also as an intrinsic result
pointing to a new kind of segmentation error documented and exemplified here for the
first time, namely the Eccentricity Detection Error. There are indeed iris code matching
errors for which the Eccentricity Detection Error is the only cause and they are
insurmountable, unfortunately. Otherwise, the proposed segmentation procedure per-
formed reasonably well on such a difficult database and proved a failure error rate of
3.26%.

Acknowledgement. This work was supported by the University of Bucharest (Romania) and
Applied Computer Science Laboratory (Bucharest, Romania).

References
1. Arora, S.S., Vatsa, M., Singh, R., Jain, A.: On iris camera interoperability. In: IEEE Fifth
International Conference on Biometrics: Theory, Applications and Systems (BTAS).
pp. 346–352 (2012)
2. CASIA: http://biometrics.idealtest.org/dbDetailForUser.do?id=4
3. Cheng, G., Yang, W., Zhang, D., Liao, Q.: A Fast and Accurate Iris Segmentation Approach.
In: Image and Graphics, pp. 53–63. Springer International Publishing (2015)
4. Daugman, J.: New methods in iris recognition. IEEE Trans. Syst. Man Cybern., Part B 37(5),
1167–1175 (2007)
5. Daugman, J.G.: High confidence visual recognition of persons by a test of statistical
independence. IEEE Trans. Pattern Anal. Mach. Intell. 15(11), 1148–1161 (1993)
6. Guo, G., Jones, M.J.: Iris extraction based on intensity gradient and texture difference. In:
IEEE Workshop on Applications of Computer Vision, pp. 1–6, January 2008
7. Jan, F., Usman, I., Agha, S.: Reliable iris localization using Hough transform,
histogram-bisection, and eccentricity. Sig. Process. 93(1), 230–241 (2013)
8. Khan, T.M., Khan, M.A., Malik, S.A., Khan, S.A., Bashir, T., Dar, A.H.: Automatic
localization of pupil using eccentricity and iris using gradient based method. Opt. Lasers
Eng. 49(2), 177–187 (2011)
9. Lee, Y., Micheals, R.J., Filliben, J.J., Phillips, P.J.: VASIR: an open-source research
platform for advanced iris recognition technologies. J. Res. Nat. Inst. Stand. Technol. 118,
218 (2013)
10. Li, P., Ma, H.: Iris recognition in non-ideal imaging conditions. Pattern Recogn. Lett. 33(8),
1012–1018 (2012)
11. Liu, X., Bowyer, K.W., Flynn, P.J.: Experiments with an improved iris segmentation
algorithm. In: Fourth IEEE Workshop on Automatic Identification Advanced Technologies,
pp. 118–123 (2005)
12. Masek, L.: Recognition of human iris patterns for biometric identification, The University of
Western Australia (2003)
13. Meenakshi, B.K., Prasad, M.R., Manjunath, T.C.: Segmentation of iris images which are
affected by light illuminations. In: 2014 International Conference on Control, Instrumen-
tation, Communication and Computational Technologies (ICCICCT), pp. 945–948. IEEE,
July 2014
424 C.-M. Noaica

14. Moi, S.H., Asmuni, H., Hassan, R., Othman, R.M.: A unified approach for unconstrained
off-angle iris recognition. In: 2014 International Symposium on Biometrics and Security
Technologies (ISBAST), pp. 39–44. IEEE, August 2014
15. Motoc, I.M., Noaica, C.M., Badea, R., Ghica, C.G.: Noise influence on the fuzzy-linguistic
partitioning of iris code space. In: Proceedings of 5th IEEE International Conference on Soft
Computing and Applications. Advances in Intelligent Systems and Computing, vol. 195,
pp. 71–82. Springer (2013)
16. ND-CrossSenssor-Iris-2013. https://sites.google.com/a/nd.edu/public-cvrl/data-sets
17. Noaica, C.M., Badea, R., Motoc, I.M., Ghica, C.G., Rosoiu, A.C., Popescu-Bodorin, N.:
Examples of artificial perceptions in optical character recognition and iris recognition. In:
Proceedings of 5th IEEE International Conference on Soft Computing and Applications.
Advances in Intelligent Systems and Computing, vol. 195, pp 57–69. Springer (2013)
18. Ortiz, E., Bowyer, K.W., Flynn, P.J.: A linear regression analysis of the effects of age related
pupil dilation change in iris biometrics. In: 2013 IEEE Sixth International Conference on
Biometrics: Theory, Applications and Systems (BTAS), pp. 1–6. IEEE, September 2013
19. Paone, J., Flynn, P.J.: On the consistency of the biometric menagerie for irises and iris
matchers. In: 2011 IEEE International Workshop on Information Forensics and Security
(WIFS), pp. 1–6 (2011)
20. Pillai, J.K., Puertas, M., Chellappa, R.: Cross-sensor iris recognition through kernel learning.
IEEE Trans. Pattern Anal. Mach. Intell. 36(1), 73–85 (2014)
21. Popescu-Bodorin, N., Noaica, C.M., Penariu, P.S.: Iris recognition with 4 or 5 fuzzy sets. In:
Proceedings of IFSA-EUSFLAT 2015 (16th World Congress of the International Fuzzy
Systems Association & 9th Conference of the European Society for Fuzzy Logic and
Technology), June 30–July 3rd, Gijon, Asturias, Spain, pp. 1438–1445. Atlantis Press
(2015). doi:10.2991/ifsa-eusflat-15.2015.204
22. Popescu-Bodorin, N., Balas, V.E.: Best practices in reporting iris recognition results. In: Soft
Computing Applications. Advances in Intelligent Systems and Computing, vol. 357,
pp. 819–828. Springer, New York (2016) doi:10.1007/978-3-319-18416-6
23. Popescu-Bodorin, N., Balas, V.E.: Fuzzy membership, possibility, probability and negation
in biometrics. Acta Polytech. Hungarica 11(4), 79–100 (2014)
24. Popescu-Bodorin, N., Balas, V.E., Motoc, I.M.: Iris codes classification using discriminant
and witness directions. In: Proceedings of 5th IEEE International Symposium on
Computational Intelligence and Intelligent Informatics (Floriana, Malta, September 15–
17), pp. 143–148. IEEE Press (2011). ISBN: 978-1-4577-1861-8 (electronic),
978-1-4577-1860-1 (print). doi:10.1109/ISCIII.2011.6069760
25. Popescu-Bodorin, N., Balas, V.E.: Comparing haar-hilbert and log-gabor based iris encoders
on bath iris image database. In: Proceedings of 4th International Workshop on Soft
Computing Applications, pp. 191–196. IEEE Press, July 2010. ISBN 978-1-4244-7983-2.
doi:10.1109/SOFA.2010.5565599
26. Popescu-Bodorin, N.: Exploring new directions in iris recognition. In: 11th International
Symposium on Symbolic and Numeric Algorithms for Scientific Computing, Conference
Publishing Services - IEEE Computer Society, pp. 384–391, September 2009. doi:10.1109/
SYNASC.2009.45
27. Porro-Muñoz, D., Silva-Mata, F.J., Mendiola-Lau, V., Hernández, N., Talavera, I.: A New
iris recognition approach based on a functional representation. In: Iberoamerican Congress
on Pattern Recognition, pp. 391–398. Springer, Berlin, November 2013
28. Qiaoli, G., Cao, H., Benqing, D., Xiang, Z: The iris normalization method based on line. In:
2013 Fourth International Conference on Intelligent Systems Design and Engineering
Applications, pp. 669–671. IEEE, November 2013
 A Circular Eccentric Iris Segmentation Procedure 425

29. Shen, Z., Macphie, R.H.: Scattering by a thick off-centered circular iris in circular
waveguide. Microwave Theor. Tech IEEE Trans. 43(11), 2639–2642 (1995)
30. Vatsa, M., Singh, R., Noore, A.: Improving iris recognition performance using segmentation,
quality enhancement, match score fusion, and indexing. Syst. Man Cybern. Part B 38(4),
1021–1035 (2008)
31. Yeo, S.P., Teo, S.G.: Thick eccentric circular iris in circular waveguide. IEEE Trans.
Microwave Theor. Tech. 46(8), 1177–1180 (1988)
 Iris Segmentation in the Last Decade – A Survey

C.M. Noaica1,2 ✉ , P.S. Penariu1,2, and V.C. Stroescu2,3

( )

1
Department of Computer Science, University of Bucharest, Bucharest, Romania
[email protected]
2
Applied Computer Science Testing Laboratory (ACSTL), Bucharest, Romania
{victor,penariu}@lmrec.org
3
‘Goethe’ Deusche National College, Bucharest, Romania

Abstract. The goal of this paper is to gather the most important ideas, develop‐
ments and trends undergoing in iris segmentation through the last ten years and
to offer to those researchers that are interested in iris segmentation, as clear as
possible, a perspective both on the most popular or state of the art approaches and
on ‘exotic’ ideas and possible future directions in the field.

Keywords: Survey · Iris segmentation

1 Introduction

According to Google Scholar, 14.800 papers have been published in the field of iris
recognition in the last decade (2006–2016), while 4.230 of them address the topic of iris
segmentation. Among the latter, are found the papers selected by us as being the most
representative. The number of articles addressing matters of iris recognition and
published between the years 2006–2016 is illustrated in Fig. 1. The two bar graphs there
offer a visual comparison between the number of papers that address any topic of iris
recognition and the number of papers that address iris segmentation specifically. As it
can be seen in the figure, problems of iris recognition have kept up the interest of many
researchers, leading to many and very interesting results, some of them being mentioned
in this paper.
Offering more detailed statistic on the number of scientific publications on any topic
of iris recognition versus scientific publications on iris segmentation, Table 1 reveals
that the latter cover a proportion of the former that is not at all negligible. This obser‐
vation is easily explained by the fact that iris segmentation is not yet a closed topic.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_32
 Iris Segmentation in the Last Decade – A Survey 427

(a) (b)

Fig. 1. Bar graphs showing (a) the number of publications on the problem of iris recognition and
(b) the number of publications that approach the problem of iris segmentation, spanning from
2006 to 2016. The information has been obtained using Google Scholar, (a) by searching the
keywords “iris biometrics” OR “iris recognition” and (b) the keywords “iris segmentation” OR
“iris localization” OR “pupil localization”.

Table 1. The number of published papers on the topic of iris image segmentation for each year,
starting with 2006, and ending with the current year (2016), until the month of May. The values
situated on the row “Segm.”, which stands for the number of published papers, are obtained using
Google Scholar, by searching the keywords “iris segmentation” OR “iris localization” OR “pupil
localization”. The values situated on the row “All”, which stands for the total number of papers
published on the topic of iris recognition, are obtained using Google Scholar, by searching the
keywords “iris biometrics” OR “iris recognition”.
Year 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
Segm 184 205 270 371 403 464 513 552 548 554 160
All 718 828 1030 1210 1410 1600 1790 1930 2020 2090 512

In the literature, can be identified several research directions on the topic of iris image
segmentation. For instance, there are segmentation proposals that tackle the iris boun‐
daries detection either by using a circular Hough transform [84, 92] or different circular
filters [48, 49], or techniques for detecting elliptical edges [71, 85, 87], edges that repre‐
sent the pupillary and limbic boundaries. Aside from elliptical/circular boundaries,
another characteristic of the iris taken into consideration when performing iris segmen‐
tation is the concentricity/eccentricity of the iris boundaries [21, 33, 45, 85]. A key role
reported in optimizing the segmentation algorithms was played by the morphological
operations that were applied to the iris images [1, 16, 25, 34, 81, 84] and, also, by
considering cropped versions of the original eye image as the regions where the centers
of the inner and outer iris boundaries are situated [41, 83]. The fact that the iris and pupil
localization is a difficult problem (especially when it comes to work on noisy iris images)
is indicated in several papers as well. For instance, different methods to pinpoint the
ROI (Region of Interest) within an eye image can be found in [4, 6, 13, 77]. Several
papers go further and propose iris segmentation procedures for eye/face images acquired
as video frames [20, 52], the latter implying a preliminary crop also: firstly, the eye
region is localized (as a ROI), and, only then, iris localization is performed. The last of
the research directions on the topic of iris image segmentation that emerged from the
cited papers refers to assessing iris segmentation quality [30, 43].
428 C.M. Noaica et al.

Each of the above mentioned topics that are related to iris segmentation are detailed
in what follows: the 2nd section of this paper discusses the related works, namely the
surveys not only of iris segmentation techniques but of any important developments in
this research field. The 3rd section puts together a short presentation of the most utilized
iris image databases and, starting with the 4th section, the current paper groups and
discusses the selected papers according to the approached segmentation methods. That
is, if the iris boundaries are considered to be circular or elliptical (4th section), if the
centers of the two boundaries are or not concentric (5th section), different methods of
approximating the image area that is most likely to contain the iris (6th section), applying
morphological operations (7th section), extracting iris segments from video frames
(8th section), using Machine Learning methods to obtain iris segments (9th section), a
comparison between visible and near infrared iris segmentation (10th section) and
ending with proposed methods of evaluating the performance of iris segmentation
(11th section).

2 Related Works

According to Google Scholar, 25 iris recognition surveys were published in the past
decade, five of them being mentioned here because, in our view, they give important
insights onto iris recognition phases, starting with image acquisition and ending with
biometric system evaluation methods. For instance, a survey that goes through every‐
thing from iris segmentation, various filtering techniques, feature vectors and iris
patterns matching (either considering multi- or single-enrollment), is written by Bowyer
et al. Their surveying work is divided in two parts: a paper that covers the years 1990–
2007 [8] and a paper [9] that covers three more years (2008–2010).
Different methods applied in iris recognition (such as Independent Component Anal‐
ysis, Continuous Dynamic Programming, or zero-crossing) are covered in Sheela’s [74]
survey, while De Marsico et al. [19] come with a comprehensive re-view on various
Machine Learning techniques that have been applied in iris recognition. These techni‐
ques are used either to perform pattern recognition, eye localization, or periocular
recognition.
Since the mobile devices have been greatly improved, mobile biometrics has
emerged and certainly gained terrain. Addressing this topic, Meng et al., in their paper
[44], wrote about the physiological (iris included) and behavioral biometric features
used to protect the mobile devices. The survey took into consideration all the drawbacks
caused by the limited computational power available on mobile devices at the level of
the year 2013.

3 Available Iris Image Databases

The available databases for iris recognition present different problems, the majority of
them relating to “unconstrained”/“on the move” point of view.
 Iris Segmentation in the Last Decade – A Survey 429

The most popular databases among the researchers are the ones offered by CASIA
(Chinese Academy of Science Institute of Automation) [12]. CASIA databases offer a
large variety of images:
– Images that are captured at a distance (Fig. 2a), for which the problem of segmen‐
tation is not resumed to only finding the iris, but to finding the eyes in the first place,
in order to perform iris segmentation;

(a) (b) (c)

(d) (e)
(f.1)

(g) (h) (f.2)

Fig. 2. Images selected from several publicly available databases. CASIA Iris v4: (a) CASIA-Iris-
Distance; (b) CASIA-Iris-Interval; (c) CASIA-Iris-Lamp; (d) CASIA-Iris-Syn; (e) CASIA-Iris‐
Thousand; (f.1) (f.2) CASIA-Iris-Twins; ND-CrossSensor-Iris-2013 (g) LG4000 and (h) LG2200

– Images that are captured from close up with different sensors (Fig. 2b–e);
– Synthetic iris images (Fig. 2d);
– Iris images that come from twins (Fig. 2f1, 2)
A database that provides images captured with two different sensors (enabling a
cross-sensor approach) is ND-CrissSensor-Iris-2013 [47]. In this database, the eye
images are taken with LG2200 (Fig. 2h) and LG4000 (Fig. 2g) sensors.
Table 2 comprises the names of several of the available iris databases, along with
the papers that mention them. One can observe that the most popular databases are
430 C.M. Noaica et al.

CASIA v3 (it covers many situations in which correct iris segmentation is difficult to
achieve), and UBIRIS (containing iris images captured in visible light).

Table 2. Databases used for segmentation purposes

Database Citation
CASIA v1 [12] [6, 7, 27, 45, 49, 53, 66, 76]
CASIA v2 [12] [45, 49]
CASIA v3 [12] [3, 6, 11, 21, 27, 29, 66, 72, 73, 85, 86, 91]
CASIA v4 [12] [1, 35, 80, 89]
MMU [46] [71–73, 84, 86]
IITD [31] [76, 84]
WVU [51] [24, 71]
UBIRIS [64, 82] [15, 16, 29, 41, 45, 62, 63, 67, 71, 79, 80, 85, 86, 89]
ICE [32] [29, 85]
FRGC [22] [80, 89]

Table 3. Comparison of the results reported in papers

Authors Database Segmentation results
Umer [84] IITD 98.48%
MMU1 98.22%
Shamsi [73] CASIA v3 99.34%
Essam [21] CASIA v3 98.8%
Gu [27] CASIA v1 100%
Proenca [63] UBIRIS v1 97.88% s2;
98.02% s1
Basit [6] CASIA v1 99.06%
CASIA v3 99.21%
Jiang [33] CASIA 99.4%
Chen [15] UBIRIS v2 98.0%
Costa Filho [16] UBIRIS 97.6%
Zhou [91] CASIA v3 lamp 98.88%,
Interval 99.66%
Twins 97.32%
Radman [67] UBIRIS v1 97.36% s1
96.88% s2
Wan [86] CASIA v3 lamp 97.02%
MMU1 97.83%
UBIRIS v1 96.97%
CASIA v3 interval 97.29%
Kumar [35] CASIA-Iris-Thousand v4 99.3%
Shamsi [73] CASIA v3 99.34%
Abdullah [1] CASIA v4 Iris-Lamp 98.9%
Alvarez-Betancourt [3] CASIA v3-Iris-Interval 98%
Bastos [7] CASIA v1 95.10%
Khan [34] CASIA v1 98.10% class II
Labati [36] CASIA v3 97%
 Iris Segmentation in the Last Decade – A Survey 431

Table 3 puts together the segmentation results that have been reported in some of
the iris segmentation publications that are mentioned in this survey. Analyzing the data
from the table, one can form an approximate impression on the current state of the art
in the field of iris segmentation. For instance, for the CASIA 1 database Gu et al. [27]
report 100% boundary detection accuracy. On a subset of CASIA v3 (Interval), Zhou
[91] report 99.66% boundary detection accuracy, while for CASIA v4 (Iris-Thousand)
Kumar et al. [33] report an iris boundary accuracy of 99.3%. For UBIRIS v1, the highest
segmentation results are reported by Proenca et al. [63].

4 Circles or Ellipses?

If the segmentation procedure should consider the boundaries of the iris as being circular
or elliptical/non-circular (see Fig. 3) is highly dependent on the utilized iris image dataset
and on the purpose of the different segmentation techniques. Therefore, there are papers
that report either using both elliptical/non-circular and circular approaches to detecting
the iris boundaries or one of the two.
When it comes to detecting the inner and outer boundaries of the iris as circles, the
most popular method in the papers that are considered in this survey is the circular
Hough transform [7, 16, 24, 50, 53], most of the times modified to obtain a reduction in
terms of computation cost. For example, Umer et al. [84] applied a circular Hough
transform only on several selected points, points that are found at a certain distance, d,
from each other, within the image.

Fig. 3. Examples of irides whose boundaries are better defined by an elliptical/non-circular model
(left) or by a circle (right). The images are captured with the LG4000 sensor.

Other segmentation approaches are designed to cope with iris boundaries that are
not circular. Such an approach belongs to Punyani et al. [66], who use a ‘negative func‐
tion’ to threshold the eye image and the four neighbor method to determine the boundary
of the pupil. Unfortunately, the ‘negative function’ is not explained at all in [66], nor in
the paper that Punyani cites: Almisreb et al. [2]. The only clarification that Almisreb
gives regarding the ‘negative function’ is that is “used in processing X-Ray images for
increasing the contrast in the images and to ensure it is more readable and assist medical
practitioners to make the diagnose process easier” [2].
The four neighbor method is detailed as follows:
432 C.M. Noaica et al.

– Let I an eye image and I’ its binary version, in which with 1 is represented the pupil‐
lary disk and with 0 the area around the pupil.
– A 3 × 3 window slides over the pupil disk and whenever one of the pixels marked in
Fig. 4 with gray falls over a 0 value, in the original eye image, I, at the coordinates
corresponding to the center of the window i.e. (x,y), the pixel is flagged as being
situated on the pupil edge.

Fig. 4. The 3 × 3 window used for the four neighbor algorithm.

Active contours are often used for non-circular iris boundary detection in iris
segmentation approaches: Shah et al. [71], Farouk et al. [23], Abdullah et al. [1], and
Daugman [18]. All these techniques need a starting contour (usually inside the pupil)
whose position is iteratively changed until some stopping criteria is met.
RANSAC (RANdom SAmple Consensus), a popular method to estimate the param‐
eters of a mathematical model, is used to fit both elliptical and circular iris boundaries.
For example, Wang et al. [87] used RANSAC to find elliptical/non-circular edges in iris
images within CASIA Iris v3 and IITD v1 databases, while Ranjad et al. [68] used the
same boundary detection algorithm for iris images captured in visible light (UPOL) at
ophthalmologic quality. Other results of applying RANSAC on iris images from UBIRIS
v2 database are explained by Li and Ma in their paper [40].
An adaptation of the Starburst algorithm [38], an algorithm that uses RANSAC,
firstly designed for gaze tracking, was proposed by Ryan et al. [70] for elliptical iris
segmentation purposes on ICE database.
Ghodrati et al. [25] tackle the problem of non-circular iris boundaries as well. For
that, they firstly locate the pupil, then, by fitting a window around the estimated location
of the pupil, look for the pupillary boundary. The window assures the minimum level
of noise around the pupillary edge, edge obtained by thresholding. The pupillary boun‐
dary is localized with a course to fine strategy, while the outer one is obtained with
Hough transform adapted for arches. Basically, the authors divide the iris in two sides
(left and right) and apply the Hough transform to vote for arches.
An approach that employs an elliptical model to approximate a raw location of the
inner and outer boundaries of the iris belongs to Vatsa et al. [85]. The refinement of the
estimated boundary locations is obtained with the Mumford-Shah functional [5].
A segmentation method that is based on the Radial Symmetry Transform [42], to
locate the pupil, and on a modified circular integro-differential operator to detect the iris
boundaries is proposed by Zhou et al. [91].
 Iris Segmentation in the Last Decade – A Survey 433

While the papers cited in this section choose either using circular or elliptical models
to perform iris segmentation, the work of Miyazawa et al. [45] is slightly different. They
combine the two approaches (circular and elliptical boundaries) by considering that the
inner boundary is best represented by an ellipse while, in order to simplify the compu‐
tation, the outer boundary may be represented with a circle.

5 Concentric or Eccentric Iris Boundaries?

Since the irides are not always best represented by two concentric circles, the majority
of the proposed algorithms cited here are designed to cope with cases of ec-centric iris
boundaries. A segmentation method that starts from considering the irides as having
eccentric boundaries is proposed by Jiang et al. [33], who firstly identify the pupillary
boundary, then consider that the center of the limbic boundary coincides with the pupil‐
lary one and, finally, with a modified version of the integro-differential method, look
around the detected boundary to see if there is another one that fits better the outer
boundary. Therefore, Jiang’s segmentation techniques even though it starts from consid‐
ering two concentric circles it still comes to look for eccentric circular boundaries.
Even though is not very popular, the concentric segmentation model also lead to low
EER values and/or low segmentation failure rates when applied on ideal images (Bath
and LG4000 databases, [55–61]), noisy images (LG2200, [60]), and even on cross-
sensor comparisons (LG2200-to-LG4000, [60]).

6 Iris and Pupil Center Approximation

Approximating the region of interest (iris) in an eye image may help a lot to reduce the
computational time spent by a segmentation algorithm, especially when is used an
exhaustive search. In order to find the ROI (Region Of Interest) in an eye image, there
have been proposed many different methods in the literature, several of them being
gathered and presented shortly in this paper.
Ideally, the result after applying any type of ROI finding methods is the exact location
of the iris (Fig. 5, right), but, depending on the level of noise, there are cases when the
estimated ROI is not limited to the expected region of interest, but includes all the objects
that satisfy the condition adopted by the ROI finding algorithm (let’s say a threshold).
The most popular methods of ROI finding that result from the articles cited here are
the following:
– Usage of the location information offered by the specular lights;
– Applying a threshold that results, for example, from the analysis of the intensity
histogram;
– Employing different elliptical or circular models;
– Applying Radial Symmetry Transform
– Resizing the image to a certain dimension in order to approximate the location of the
pixels that are most likely to belong to the pupil;
434 C.M. Noaica et al.

Fig. 5. Examples of ROI. Since the eyelashes have sometimes intensity values comparable with
the pupil, the ROI algorithm (considering that the pupil is the region of interest) returns erroneous
results (left image); otherwise, the ROI detection results are precisely the searched ones.

6.1 Specular Lights Detection

While some authors choose only to remove the corneal reflections [1, 3], there are others
that thought to use them in locating the iris [41, 45].
Anisotropic diffusion [54] is one of the methods used for specular lights removal.
The result of applying anisotropic diffusion on an eye image is preserving the eye regions
(e.g. eyelids, iris, pupil), while having a smoothing effect inside them. Even though the
specular lights have well defined edges (strong chromatic gradient), as it can be seen in
Fig. 1a from [1], they do not stretch over a large portion of the image and, consequently,
are eliminated (filled). Therefore, in [1] the authors use anisotropic diffusion in the
preprocessing step for their iris segmentation method to obtain both specular lights
removal and a blurring effect of the eyelashes, which helped to reduce the number of
unnecessary edges. Noise removal based on the anisotropic diffusion technique,
combined with morphologic operations was employed by Wan et al. in [86].
In [41], Li et al. motivate that since the specular lights are often located near or inside
the pupil, it also happens that the two (pupil and specular lights) have uniform upper
and lower gray-levels (levels that are previously set from analyzing the histograms of
several manually selected eye-images) and, therefore, can be localized by using square
windows that scan the image. After setting to a set of candidate pupil regions, the authors
apply a Circular Hough transform to get a raw iris boundary approximation.

6.2 Thresholding
Since different available iris image databases have different acquisition settings, it is
obvious that a certain method of identifying the region of interest does not apply on all
them in the same way. This is why, over time, different approaches have been proposed
to obtain the region of interest, in this section being discussed techniques to obtain a
suited threshold value. Miyazawa et al., in [45], report that, since the experiments were
performed on several different databases, the preprocessing phase has to be done differ‐
ently. For CASIA v2 and ICE databases, the images are thresholded by a value resulted
from the analysis of the pixels located around the corneal reflections, revealing the pupil.
For CASIA v1, the authors use a different approach – an analysis of the intensity histo‐
gram of the entire initial image is the only necessary step to determine the threshold
 Iris Segmentation in the Last Decade – A Survey 435

value that distinguishes between the pupil and the rest of the eye image. Usage of
threshold values (in order to eliminate noise or to locate the region of interest) that are
resulted from the analysis of the intensity histogram is reported also in [4, 6, 21].
Another segmentation technique that uses the information provided by the in-tensity
histogram for iris segmentation is offered by Zhang et al. [88]. The authors consider the
first peak in the histogram to be the upper limit of the pupillary intensity values. With
this information, they binarize the eye image using the threshold determined by the value
of that first peak and obtain the pupil area. Further, they consider a square region from
the resulted binary image as being the location of the inner and outer boundaries centers.
In this way, the computational time required to process an image for iris boundary
detection is considerably reduced.
Puhan et al. [65] approach iris segmentation for images captured with minimum level
of cooperation from the subjects. Their segmentation procedure firstly detects the limbic
boundary and then the pupillary one. The segmentation method is based on Fourier
spectral density (formula 1 in [65]) as a thresholding value, offering a binary eye image
that is the basis of the two edge detection steps. Subbarayudu, in [78], approximated the
iris center by computing the average of the edges that resulted from Canny edge
detector. Around the rough approximation of the iris center, a window of 150 × 150 is
selected and binarized with a value derived from the analysis of the intensity histogram.
The iris boundaries are determined with the Circular Hough transform.
A threshold determined with Otsu’s method is employed in several segmentation
techniques. Thus, Abdullah et al. [1], Alvarez-Betancourt et al. [3] use Otsu’s thresh‐
olding to approximate the location of the pupil. In all these cases, the images were subject
to different preprocessing phases meant to smooth the image by eliminating the noise.
In [1], in this purpose, is used anisotropic diffusion, while in [3] is applied a median
filter. Otsu’s method for threshold selection was used by Chai et al. [14] to find the
locations affected by specular lights reflection.
Two different methods of selecting a threshold are proposed by Essam et al. [21]
and by Miyazawa et al. [45]. In [21], the threshold receives one of three available values
(115, 50 and 85), depending to the information provided by the intensity histogram,
while in [45], for CASIA version 2.0 and ICE 2005 databases, the authors analyze the
corneal reflection to obtain an appropriate threshold value, further used to obtain a binary
image containing the pupil.

6.3 Elliptical or Circular Filters

Different elliptical or circular filters applied to the eye-images in the preprocessing phase
are reported in literature to have great results in the search of the region of interest. On
this line, Nguyen et al., in [49], use a circular filter designed by the authors – basically
a square matrix that has positive values in the center, where the pupillary disk should
fall, and negative values in the rest of the square matrix, to reduce the boundary centers
search space within the eye-image to an area that is at least close to the pupil, if not the
pupil itself. Another circular filter, Circular Gabor filter, is used by Radman et al. [67]
to locate the pupillary area.
436 C.M. Noaica et al.

Uhl [83] used a weighted adaptive Hough transform to detect iris center and a polar
transform for the iris outer boundary. An agent-based pupil localization and boundary
detection using multiple view analysis is presented in [36]. Agent-based refers to
choosing a set of N pairs of coordinate points randomly selected within the image (the
coordinate points are called agents) and decide, based on the analysis of the analysis of
the region around the agents, if the agents “jump” to a different location or not. When
the agent-based algorithm stops the image coordinates corresponding to most of the
agents are most likely to indicate the pupil.

6.4 Resizing the Eye-Image

The most common usage of image resizing in the preprocessing phase is to reduce the
computational cost but Shamsi et al. [72] exploit image resizing for ROI finding
purposes. The ROI finding algorithm based on image resizing is named Average Square
Shrinking (ASS). The Average Square Shrinking algorithm repetitively resizes the input
image to a lower width and weight dimensions and, simultaneously, applies averaging
inside square windows that slide the image. The output, for an image of 240 × 320 size,
is a image of 9 × 11 in which the darkest pixel indicates the center of the pupil. The
purpose of this ROI algorithm is to reduce the computational cost of Daugman’s [17]
circular integro-differential operator by applying it only on the ASS-generated iris center
coordinates. Shamsi used the same ASS for elliptic boundary detection as well in [73].
There the integro-differential operator was adapted for elliptical boundaries.

7 Morphological Operations Used for Noise Removal

Morphological operations are very useful when it comes to eliminating some of the noise
present in eye images. The most common noise eliminated by such operations is the one
caused by eyelids and eyelashes. In this section are succinctly presented some of the
papers that report using morphological operations and the manners in which such image
processing operations improved segmentations results. Morphological operations and
fuzzy membership functions are used in [81] in order to enhance the edges of the iris.
This method was applied on the CASIA iris database. Other segmentation techniques
that imply morphological operations are presented in [1,34].
Kumar et al. [35] proposed a segmentation procedure that was tested on images with
severe degradation from CASIA-Iris-Thousand v4 database. Morphological operations
are used here to remove the severe specular light reflections and, for boundary detection,
the authors used a modified version of the integro-differential operator.
Disk operators are reported to be used in several segmentation techniques, as image
processing steps. For instance, Attarchi et al. [4] use morphological opening and closing
to erode the bridges between the eyelashes and pupil and to eliminate the noise. Essam
et al. [21] use a disk operator to obtain a blur effect over the eyelashes and the overall
noise present in the eye-images. The authors apply a second round of morphological
operations right after thresholding.
 Iris Segmentation in the Last Decade – A Survey 437

Gu et al. [27] proposed a segmentation procedure that starts with a major noise
removal step. Shortly, they consider that any intensity value above 200 is noise and then
go further and smooth the intensity values with a 3 × 3 window, the pixel in the middle
receiving the averaged value resulted from the surrounding pixels within the window.
Also, the authors neglect the pixels situated at a certain distance from the margin of the
image.
Roy et al. [69] thought that, since the specular lights from the pupils may lead to
edges that mislead the segmentation algorithm into selecting them as iris boundaries,
firstly remove those reflections and then, using a called Direct Least Square (DLS),
identify the pupil’s elliptic boundary. The outer boundary is found with a shape guided
approach.

8 Extracting Iris Segments from Video Frames

There are several approaches of iris segmentation from video frames as well. On this
topic, in 2009, Zhou et al. [90] proposed a segmentation method that employs several
units: a quality filter unit that detects those frames that are interlacing or those frames
with complete occlusion or with motion blur, and a segmentation evaluation unit that
automatically detects the segmentation. In 2011, Du et al. [20] continue Zhou’s work
and tackle most of the encountered problems, namely identification of the frames without
the presence of an eye, finding the iris boundaries that are defined by an ellipse and noise
removal. The segmentation method adopted is a coarse to fine one, taken together with
a gradient-based method that removes some of the noise present in the eye frames and
a filter based on the specular lights for those frames in which the eye are not present at
all.
The state-of-the-art in matters of video iris recognition is the work of Lee et al. [37],
VASIR (Video-based Automatic System for Iris Recognition), which is an open source
iris segmentation and recognition system, developed for both video frames (acquired
with an Iris On the Move system [8]) and still images.

9 Machine Learning Methods for Iris Segmentation

Segmentation techniques based on AdaBoost are proposed in [29,39]. In [39], Li et al.

used AdaBoost to learn six class-specific boundary detectors for localization of left/right
pupillary boundaries, left/right limbic boundaries and up-per/lower eyelids respectively.
With weighted Hough transform, on CASIA-Iris-Thousand database, the iris boundaries
are further detected.
He et al. [29], on the other hand, used an AdaBoost-cascade to locate an approximate
center of the iris. Then, with a model called pull and push [28] (used by Bastos et al. [7]
as well) were detected the edges of the iris. This iris segmentation algorithm, because
of a smoothing spline-based edge fitting scheme, is able to identify both circular and
non-circular iris boundaries.
A Machine Learning approach for iris segmentation is made by Tan et al. [79] as
well. There, the input is the image of a face and the purpose of Machine Learning (Feed
438 C.M. Noaica et al.

Forward and SVM) is to automatically pinpoint the eye regions in the input face image.
In a more recent paper of the same authors [80], the Machine Learning (AdaBoost) was
used to recognize the face, then the eyes within images acquired at a distance (no matter
if it is about visible light or NIR acquiring).

10 Visible vs. Near Infrared Iris Image Segmentation

In [16], Costa Filho et al. proposed a segmentation technique based on k-means cluste‐
rization algorithm. The k-means algorithm, starting from the RGB or HSI color spaces,
returned an approximated location of the iris boundaries. The final decision on the exact
location of the boundaries is taken after applying a modified version of Hough trans‐
form. The modification consisted in adding customed weights to the given votes
according to the position of the edge pixels.
A method of segmenting the iris in color images that includes a step that de-tects if
the eye is the left or right one is proposed by Shin et al. [75]. This step assures that there
won’t be any matching attempts between iris patterns that come from the left and the
right eye.
The illumination setup causes segmentation difficulties for visible light iris images
as well. Therefore, Zhao [89] used Single Scale Retinex (SSR) [10] to enhance the illu‐
mination of the eye images. With a median filter, the authors extracted only the red
channel. A Circular Hough transform applied on the red channel gives the coordinates
corresponding to the coordinates of the iris boundaries. Chen et al. [15] chose to use the
HSI color space (the saturation component precisely) to establish an appropriate, adap‐
tive, threshold for sclera localization. Since sclera is located near the iris, the information
resulted after thresholding represents a good start in finding the iris. For iris boundary
detection, the authors used a Circular Hough transform.
Tan et al. [79], on face images, firstly located the eyes using Machine Learning
methods, then, on the edge maps corresponding to the eyes, the authors selected three
coordinate pairs and used the circumcenter of the triangle that fit the selected coordinates
as the center of the iris. In the segmentation method proposed by Tan et al., the iris
boundaries are extracted with a polynomial curve of degree 3.

11 Segmentation Performance Evaluation

In our opinion, automatic iris segmentation quality assessment is an open and unsolved
problem in iris recognition, even today. Some papers have been published on this topic,
[30, 43] – for example, but both approaches are based on using so called ‘ground truth
data’ (GTD). For an iris database, GTD is a collection of records describing the correct
segmentation result as they are given (if the segmentation is manual) or accepted (when
segmentation is automatic) by a human agent. There is an important logical flaw in
accepting evaluation methods that are based on a given GDT: since the overall perform‐
ance of an iris recognition system (reported as EER and FAR-to-FRR interdependence)
heavily depends on iris segmentation, accepting GTD for segmentation comes with the
 Iris Segmentation in the Last Decade – A Survey 439

risk of setting up a lower bound for the improvements that could be made on iris
recognition.
Besides, iris segmentation is a field in which any two different iris recognition tech‐
niques should be free to compete with each other. If they don’t, the only improvement
that can differentiate between them can only come from encoding and matching tech‐
niques. Which client or company could be interested in limiting recognition errors with
a lower bound beyond which is impossible to go because a GTD was used? If this kind
of evaluation method would have been accepted in iris recognition years ago, important
iris recognition results (see Figs. 10 and 11, pp. 42–43, [26]) and the progress in the
field could not have been possible.
Anyone should realize that a GTD is nothing else and nothing more than a collection
of beliefs expressed by a human agent and recorded in a computational manner.
Excluding the performance improvements (that otherwise are possible) just by accepting
such beliefs is the worst logical flaw we’ve heard about in iris recognition so far. On the
other hand, accepting a GTD actually hides the level of uncertainty in iris segmentation
behind the uncertainty in GTD definition. Hence, the problem of evaluating iris segmen‐
tation goes to the problem of evaluating GTD. If we accept this, iris recognition tech‐
nologies that were able to define the ground truth for the databases used in [26] for
establishing the results illustrated there in Fig. 10-11, are those of Iritech and Cogent.
Also, Fig. 11 in [26] gives an example of a situation when establishing the best GTD
available at a given moment is depending on enhancing either the user comfort (Iritech)
or the security level (Cogent) of the system. Hence, the order in the set of all possible
GTD is a partial one and the life of any GTD (that is currently accepted as being the best
one) ends naturally when a better overall iris recognition result is obtained.

12 Conclusions

We think we gathered in this article the most important ideas, developments and trends
undergoing in iris segmentation through the last ten years (2006–2016), from the most
popular ones to some exotic and sometimes surprising ideas that could shape future
directions in the field.

Acknowledgement. This work was supported by the University of Bucharest (Romania) and
Applied Computer Science Testing Laboratory (Bucharest, Romania).

References

1. Abdullah, M.A., Dlay, S.S., Woo, W.L.: Fast and accurate method for complete iris
segmentation with active contour and morphology. In: 2014 IEEE International Conference
on Imaging Systems and Techniques, pp. 123–128 (2014)
2. Almisreb, A.A., Tahir, N.M., Samad, M.: Pupil localization using negative function and the
four neighbors. In: 2nd International Conference on Computational Intelligence, Modelling
and Simulation, pp. 360–363. IEEE (2010)
440 C.M. Noaica et al.

3. Alvarez-Betancourt, Y., Garcia-Silvente, M.: A fast Iris Location based on aggregating

gradient approximation using QMA-OWA operator. In: 2010 IEEE International Conference
on Fuzzy Systems, pp. 1–8. IEEE (2010)
4. Attarchi, S., Faez, K., Asghari, A.: A fast and accurate personal identification method based
on human iris analysis. In: IEEE International Conference on Industrial Technology, ICIT
2008, pp. 1–6. IEEE (2008)
5. Bar, L. Chan, T.F., Chung, G., Jung, M., Kiryati, N., Mohieddine, R., Sochen, N., Vese, L.A.:
Mumford and shah model and its applications to image segmentation and image restoration.
In: Handbook of Mathematical Methods in Imaging, pp. 1095–1157. Springer, New York
(2011)
6. Basit, A., Javed, M.Y., Masood, S.: Non-circular pupil localization in iris images. In: 4th
International Conference on Emerging Technologies, ICET 2008, pp. 228–231. IEEE (2008)
7. Bastos, C.A., Ren, T.I., Cavalcanti, G.D.: Analysis of 2D log-Gabor filters to encode iris
patterns. In: 22nd IEEE International Conference on Tools with Artificial Intelligence
(ICTAI), 2010, vol. 2, pp. 377–378. IEEE (2010)
8. Bowyer, K.W., Hollingsworth, K., Flynn, P.J.: Image understanding for iris biometrics: a
survey. Comput. Vis. Image Underst. 110(2), 281–307 (2008)
9. Bowyer, K.W., Hollingsworth, K.P., Flynn, P.J.: A survey of iris biometrics research: 2008–
2010. In: Handbook of Iris Recognition, pp. 15–54. Springer, London (2013)
10. Brainard, D.H., Wandell, B.A.: Analysis of the retinex theory of color vision. J. Optical Soc.
Am. A 3(10), 1651–1661 (1986)
11. Cao, L., Zhou, Y., Yan, F., Tian, Y.: A novel iris segmentation approach based on superpixel
method. In: 2014 Fourth International Conference on Instrumentation and Measurement,
Computer, Communication and Control (IMCCC), pp. 826–831. IEEE (2014)
12. CASIA Iris database. http://biometrics.idealtest.org/findTotalDbByMode.do?mode=Iris
13. Celik, E.T.: Iris recognition—selecting a fuzzy region of interest in standard eye images. In:
Soft Computing Applications, pp. 793–804. Springer (2016)
14. Chai, T.Y., Goi, B.M., Tay, Y.H., Chin, W.K., Lai, Y.L.: Local Chan-Vese segmentation for
non-ideal visible wavelength iris images. In: 2015 Conference on Technologies and
Applications of Artificial Intelligence (TAAI), pp. 506–511. IEEE (2015)
15. Chen, Y., Adjouadi, M., Barreto, A., Rishe, N., Andrian, J.: A computational efficient iris
extraction approach in unconstrained environments. In: IEEE 3rd International Conference
on Biometrics: Theory, Applications, and Systems, BTAS 2009, pp. 1–7. IEEE (2009)
16. Costa Filho, C.F., Costa, M.G.: Iris segmentation exploring color spaces. In: 2010 3rd
International Congress on Image and Signal Processing (CISP), vol. 4, pp. 1878–1882. IEEE
(2010)
17. Daugman, J.G.: High confidence visual recognition of persons by a test of statistical
independence. IEEE Trans. Pattern Anal. Mach. Intell. 15(11), 1148–1161 (1993)
18. Daugman, J.: New methods in iris recognition. IEEE Trans. SMC B 37(5), 1167–1175 (2007)
19. De Marsico, M., Petrosino, A., Ricciardi, S.: Iris recognition through machine learning
techniques: a survey. Pattern Recogn. Lett. (2016)
20. Du, Y., Arslanturk, E., Zhou, Z., Belcher, C.: Video-based noncooperative iris image
segmentation. IEEE Trans. Syst. Man Cybern. Part B 41(1), 64–74 (2011)
21. Essam, M., Elnaby, M.A., Fikri, M., El-Samie, F.A.: A fast accurate algorithm for iris
localization using a coarse-to-fine approach. In: 2012 Japan-Egypt Conference on Electronics,
Communications and Computers (JEC-ECC), pp. 63–67. IEEE, March 2012
22. Face Recognition Grand Challenge, FRGC database. http://www.nist.gov/itl/iad/ig/frgc.cfm
23. Farouk, R.M.: Iris recognition based on elastic graph matching and Gabor wavelets. Comput.
Vis. Image Underst. 115(8), 1239–1244 (2011)
 Iris Segmentation in the Last Decade – A Survey 441

24. Frigerio, E., Marcon, M., Sarti, A., Tubaro, S.: Correction method for nonideal iris recognition.
In: 2012 19th IEEE International Conference on Image Processing (ICIP), pp. 1149–1152.
IEEE (2012)
25. Ghodrati, H., Dehghani, M.J., Helfroush, M.S., Kazemi, K.: Localization of noncircular iris
boundaries using morphology and arched Hough transform. In: 2010 2nd International
Conference on Image Processing Theory Tools and Applications (IPTA), pp. 458–463. IEEE
(2011)
26. Grother, P., Tabassi, E., Quinn, G.W., Salamon, W.: IREX I - Performance of Iris Recognition
Algorithms on Standard Images, NIST Interagency Report 7629, Information Access
Division, National Institute of Standards and Technology, 30 October 2009
27. Gu, H., Qiao, S., Yang, C.: An efficient iris localization algorithm based on standard
deviations. In: 2011 International Workshop on Open-Source Software for Scientific
Computation (OSSC), pp. 126–130. IEEE (2011)
28. He, Z., Tan, T., Sun, Z.: Iris localization via pulling and pushing. In: Proceedings 18th
International Conference Pattern Recognition, vol. 4, pp. 366–369 (2006)
29. He, Z., Tan, T., Sun, Z., Qiu, X.: Toward accurate and fast iris segmentation for iris biometrics.
IEEE Trans. Pattern Anal. Mach. Intell. 31(9), 1670–1684 (2009)
30. Hofbauer, H., Alonso-Fernandez, F., Wild, P., Bigun, J., Uhl, A.: A ground truth for iris
segmentation. In: 22nd International Conference on Pattern Recognition, ICPR, Stockholm,
Sweden, 24–28 August 2014, pp. 527–532. IEEE Computer Society (2014)
31. IIT Delhi Iris Database (Version 1.0). http://www4.comp.polyu.edu.hk/~csajaykr/IITD/
Database_Iris.htm
32. Iris Challenge Evaluation. http://www.nist.gov/itl/iad/ig/ice.cfm
33. Jiang, L., Zhang, Y., Li, W.: A novel iris location method for fast iris recognition. In: 2nd
International Congress on Image and Signal Processing, CISP 2009, pp. 1–5. IEEE (2009)
34. Khan, T.Z., Podder, P., Hossain, M.F.: Fast and efficient iris segmentation approach based on
morphology and geometry operation. In: 2014 8th International Conference on Software,
Knowledge, Information Management and Applications (SKIMA), pp. 1–8. IEEE (2014)
35. Kumar, V., Asati, A., Gupta, A.: Iris localization based on integro-differential operator for
unconstrained infrared iris images. In: 2015 International Conference on Signal Processing,
Computing and Control, pp. 277–281. IEEE (2015)
36. Labati, R.D., Piuri, V., Scotti, F.: Agent-based image iris segmentation and multiple views
boundary refining. In: IEEE 3rd International Conference on Biometrics: Theory,
Applications, and Systems, BTAS 2009, pp. 1–7. IEEE (2009)
37. Lee, Y., Micheals, R.J., Filliben, J.J., Phillips, P.J.: VASIR: an open-source research platform
for advanced iris recognition technologies. J. Res. NIST 118, 218 (2013)
38. Li, D., Parkhurst, D.J.: Starburst: A Robust Algorithm for Video-Based Eye Tracking, vol. 6.
Elselvier (2005)
39. Li, H., Sun, Z., Tan, T.: Robust iris segmentation based on learned boundary detectors. In:
2012 5th IAPR International Conference on Biometrics (ICB), pp. 317–322. IEEE (2012)
40. Li, P., Ma, H.: Iris recognition in non-ideal imaging conditions. Pattern Recogn. Lett. 33(8),
1012–1018 (2012)
41. Li, P., Liu, X.: An incremental method for accurate iris segmentation. In: 19th International
Conference on Pattern Recognition, ICPR 2008, pp. 1–4. IEEE (2008)
42. Loy, G., Zelinsky, A.: Fast radial symmetry for detecting points of interest. IEEE Trans.
Pattern Anal. Mach. Intell. 25(8), 959–972 (2003)
43. Mahadeo, N.K., Haffari, G., Paplinski, A.P.: Predicting segmentation errors in an iris
recognition system. In: 2015 International Conference on Biometrics (ICB), pp. 23–30. IEEE
(2015)
442 C.M. Noaica et al.

44. Meng, W., Wong, D.S., Furnell, S., Zhou, J.: Surveying the development of biometric user
authentication on mobile phones. Commun. Surv. Tutor. IEEE 17(3), 1268–1293 (2013)
45. Miyazawa, K., Ito, K., Aoki, T., Kobayashi, K., Nakajima, H.: An effective approach for iris
recog-nition using phase-based image matching. IEEE Trans. Pattern Anal. Mach. Intell.
30(10), 1741–1756 (2008)
46. Multimedia University, MMU Iris Image Database (2004). http://pesona.mmu.edu.my/ccteo
47. ND-CrossSenssor-Iris-2013. https://sites.google.com/a/nd.edu/public-cvrl/data-sets
48. Ng, R.Y.F., Tay, Y.H., Mok, K.M.: DSP-based implementation and optimization of an iris
verification algorithm using textural feature. In: Sixth International Conference on Fuzzy
Systems and Knowledge Discovery, FSKD 2009, vol. 5, pp. 374–378. IEEE (2009)
49. Van Nguyen, H., Binh, N.T.H., Kim, H.: Location of iris based on circular and linear filters.
In: 10th International Conference on Control, Automation, Robotics and Vision, ICARCV
2008, pp. 412–416. IEEE (2008)
50. Nigam, A., Gupta, P.: Iris recognition using consistent corner optical flow. In: Computer
Vision–ACCV 2012, pp. 358–369. Springer, Berlin (2012)
51. Off Axis/Angle Iris WVU Dataset, Release 1. http://www.clarkson.edu/citer/research/
collections/wvu_offaxis_release1.html
52. Othman, N., Dorizzi, B.: Impact of quality-based fusion techniques for video-based iris
recognition at a distance. IEEE Trans. Inf. Forensics Secur 10(8), 1590–1602 (2015)
53. Patel, H., Modi, C.K., Paunwala, M.C., Patnaik, S.: Human identification by partial iris
segmentation using pupil circle growing based on binary integrated edge intensity curve. In:
2011 International Conference on Communication Systems and Network Technologies
(CSNT), pp. 333–338. IEEE (2011)
54. Perona, P., Malik, J.: Scale-space and edge detection using anisotropic diffusion. IEEE Trans.
Pattern Anal. Mach. Intell. 12(7), 629–639 (1990)
55. Popescu-Bodorin, N.: Exploring new directions in iris recognition. In: 11th International
Symposium on Symbolic and Numeric Algorithms for Scientific Computing, Conference
Publishing Services—IEEE Computer Society, pp. 384–391 (2009). doi:10.1109/SYNASC.
2009.45
56. Popescu-Bodorin, N., Balas, V.E.: Comparing haar-hilbert and log-gabor based iris encoders
on bath iris image database. In: Proceedings of 4th International Workshop on Soft Computing
Applications, pp. 191–196. IEEE Press (2010). ISBN 978-1-4244-7983-2. doi:10.1109/
SOFA.2010.5565599
57. Popescu-Bodorin, N., Balas, V.E., Motoc, I.M.: Iris codes classification using discriminant
and witness directions. In: Proceedings of 5th IEEE International Symposium on
Computational Intelligence and Intelligent Informatics (Floriana, Malta, September 15–17),
pp. 143–148. IEEE Press (2011). doi:10.1109/ISCIII.2011.6069760. ISBN: 978-1-4577-
1861-8 (electronic), 978-1-4577-1860-1 (print)
58. Popescu-Bodorin, N., Balas, V.E., Motoc, I.M.: 8-Valent Fuzzy logic for iris recognition and
biometry. In: Proceedings of 5th IEEE International Symposium on Computational
Intelligence and Intelligent Informatics, pp. 149–154. IEEE Press (2011)
59. Popescu-Bodorin, N., Balas, V.E.: Exploratory simulation of an intelligent iris verifier
distributed system. In: Proceedings of 6th IEEE International Symposium on Applied
Computational Intelligence and Informatics, pp. 259–262. IEEE Press (2011). doi:10.1109/
SACI.2011.5873010. ISBN 978-1-4244-9108-7
60. Popescu-Bodorin, N., Lucian, G., Balas, V.E., Noaica, C.M., Axenie, I., Horasanli, K., Popa,
J., Munteanu, C., Manu, I., Herea, A., Stroescu, V.: Cross-Sensor Comparison-LG4000-to-
LG2200, Technical Report 460/ 24–07-2013, Rev. No. 4/ 30-09-2013, University of South-
East Europe Lumina, Bucharest, Romania (2013)
 Iris Segmentation in the Last Decade – A Survey 443

61. Popescu-Bodorin, N.: Circular fuzzy iris segmentation using isolines, sine law and lookup
tables. In: SOFA 2016, pp. 24–26 August, Arad, Romania
62. Proenca, H.: Iris recognition: on the segmentation of degraded images acquired in the visible
wavelength. IEEE Trans. Pattern Anal. Mach. Intell. 32(8), 1502–1516 (2010)
63. Proença, H., Alexandre, L.A.: Iris segmentation methodology for non-cooperative
recognition. In: Vision, Image and Signal Processing, IEE Proceedings, vol. 153(2), pp. 199–
205. IET (2006)
64. Proença, H., Filipe, S., Santos, R., Oliveira, J., Alexandre, L.A.: The ubiris. v2: a database of
visible wavelength iris images captured on-the-move and at-a-distance. IEEE Trans. Pattern
Anal. Mach. Intell. 32(8), 1529–1535 (2010)
65. Puhan, N.B., Sudha, N., Kaushalram, A.S.: Efficient segmentation technique for noisy frontal
view iris images using Fourier spectral density. Sig. Image Video Process. 5(1), 105–119
(2011)
66. Punyani, P., Gupta, R.: Iris recognition system using morphology and sequential addition
based grouping. In: International Conference on Futuristic Trends on Computational Analysis
and Knowledge Management, pp. 159–164. IEEE (2015)
67. Radman, A., Jumari, K., Zainal, N.: Fast and reliable iris segmentation algorithm. Image
Process. IET 7(1), 42–49 (2013)
68. Ranjzad, H., Ebrahimnezhad, H., Ebrahimi, A.: A novel compound approach for iris
segmentation. In: Third International Conference on Convergence and Hybrid Information
Technology, 2008. ICCIT 2008, vol. 2, pp. 657–661. IEEE (2008)
69. Roy, K., Bhattacharya, P., Suen, C.Y.: Iris recognition using shape-guided approach and game
theory. Pattern Anal. Appl. 14(4), 329–348 (2011)
70. Ryan, W.J., Woodard, D.L., Duchowski, A.T., Birchfield, S.T.: Adapting starburst for
elliptical iris segmentation. In: 2nd IEEE International Conference on Biometrics: Theory,
Applications and Systems, 2008. BTAS 2008, pp. 1–7, IEEE (2008)
71. Shah, S., Ross, A.: Iris segmentation using geodesic active contours. IEEE Trans. Inf.
Forensics Secur. 4(4), 824–836 (2009)
72. Shamsi, M., Saad, P.B., Ibrahim, S.B., Kenari, A.R.: Fast algorithm for iris localization using
Daugman circular integro differential operator. In: International Conference of Soft
Computing and Pattern Recognition, SOCPAR 2009, pp. 393–398. IEEE (2009)
73. Shamsi, M., Kenari, A.R.: Iris boundary detection using an ellipse integro differential method.
In: 2012 2nd International eConference on Computer and Knowledge Engineering (ICCKE),
pp. 1–5. IEEE (2012)
74. Sheela, S.V., Vijaya, P.A.: Iris recognition methods-survey. Int. J. Comput. Appl. 3(5), 19–
25 (2010)
75. Shin, K.Y., Nam, G.P., Jeong, D.S., Cho, D.H., Kang, B.J., Park, K.R., Kim, J.: New iris
recognition method for noisy iris images. Pattern Recogn. Lett. 33(8), 991–999 (2012)
76. Si, Y., Mei, J., Gao, H.: Novel approaches to improve robustness, accuracy and rapidity of
iris recognition systems. IEEE Trans. Ind. Inform. 8(1), 110–117 (2012)
77. Smereka, J.M.: A new method of pupil identification. Potentials IEEE 29(2), 15–20 (2010)
78. Subbarayudu, V.C., Prasad, M.V.: A novel iris recognition system. In: Third International
IEEE Conference on Signal-Image Technologies and Internet-Based System, SITIS 2007, pp.
883–887. IEEE (2007)
79. Tan, C.W., Kumar, A.: Automated segmentation of iris images using visible wavelength face
images. In: 2011 IEEE Conference on Computer Vision and Pattern Recognition Workshops,
pp. 9–14. IEEE (2011)
80. Tan, C.W., Kumar, A.: Unified framework for automated iris segmentation using distantly
acquired face images. IEEE Trans. Image Process. 21(9), 4068–4079 (2012)
444 C.M. Noaica et al.

81. Tian, Q.C.: A new Iris region segmentation method. In: Fifth International Conference on
Fuzzy Systems and Knowledge Discovery, FSKD 2008, vol. 3, pp. 63–67. IEEE (2008)
82. UBIRIS, Noisy Visible Wavelength Iris Image Databases. http://iris.di.ubi.pt/
83. Uhl, A., Wild, P.: Weighted adaptive hough and ellipsopolar transforms for real-time iris
segmentation. In: 5th IAPR International Conference on Biometrics (ICB), pp. 283–290. IEEE
(2012)
84. Umer, S., Dhara, B.C., Chanda, B.: A fast and robust method for iris localization. In: Fourth
International Conference of Emerging Applications of Information Technology (EAIT),
pp. 262–267. IEEE (2014)
85. Vatsa, M., Singh, R., Noore, A.: Improving iris recognition performance using segmentation,
quality enhancement, match score fusion, and indexing. IEEE Trans. Syst. Man Cybern. B
38(4), 1021–1035 (2008)
86. Wan, H.L., Li, Z.C., Qiao, J.P., Li, B.S.: Non-ideal iris segmentation using anisotropic
diffusion. Image Process. IET 7(2), 111–120 (2013)
87. Wang, K., Qian, Y.: Fast and accurate iris segmentation based on linear basis function and
ransac. In: 2011 18th IEEE Intl. Conf. on Image Processing (ICIP), pp. 3205–3208. IEEE
(2011)
88. Zhang, T., Hu, Y., Zhang, C., An, H., Xu, W.: A new method for iris’ location based on
pointwise detection. In: 4th International Conference on Bioinformatics and Biomedical
Engineering, pp. 1–4. IEEE (2010)
89. Zhao, Z., Ajay, K.: An accurate iris segmentation framework under relaxed imaging
constraints using total variation model. In: Proceedings of the IEEE International Conference
on Computer Vision, pp. 3828–3836 (2015)
90. Zhou, Z., Du, Y., Belcher, C.: Transforming traditional iris recognition systems to work in
nonideal situations. IEEE Trans. Ind. Electron. 56(8), 3203–3213 (2009)
91. Zhou, Z., Yao, P., Zhuang, Z., Li, J.: A robust algorithm for iris localization based on radial
symmetry and circular integro differential operator. In: International Conference on Industrial
Electronics and Applications, pp. 1742–1745 (2011)
92. Ziauddin, S., Dailey, M.N.: A robust hybrid iris localization technique. In: 6th International
Conference on Electrical Engineering/Electronics, Computer, Telecommunications and IT,
ECTI-CON 2009, vol. 2, pp. 1058–1061. IEEE (2009)
Circular Fuzzy Iris Segmentation Using Isolines, Sine Law
and Lookup Tables

N. Popescu Bodorin1 ✉ , V.E. Balas2, and P.S. Penariu1,3

( )

1
Applied Computer Science Lab (ACSL), Bucharest, Romania
{penariu,bodorin}@lmrec.org
2
Faculty of Engineering, University ‘Aurel Vlaicu’ of Arad, Arad, Romania
[email protected]
3
Computer Science Department, University of Bucharest, Bucharest, Romania

Abstract. This paper proposes a new and original fuzzy approach to circular iris
segmentation based on isolines, sine law and lookup tables. All isolines found in
the eye image together form the space in which the inner and outer boundaries of
the iris are searched for, while the sine law is used for identifying clusters of
concyclic points within any given and possibly noisy isoline. The new segmen‐
tation procedure proved a failure rate of 5.83% when tested on 116,564 eye images
(LG2200 subset of ND-CrossSenssor-Iris-2013 database) at an average speed of
four images per second on a single Xeon L5420 core.

Keywords: Iris segmentation · Fuzzy captive isoline · Sine law lookup table ·
Hough transform · Circular fuzzy iris segmentation · Fuzzy isochain · Interface

1 Introduction

One of the latest trends in iris segmentation is to use active contours for finding the inner
and outer iris boundaries. Ultimately, Daugman has sparked our interest in this matter,
after he announced he adopted the active contours as an instrument for iris segmenta‐
tion [4].
The generic approach within all active contours techniques is to start with an initial
curve and to evolve iteratively its shape and its position according to some conditions
(under the pressure of a contour model, for example [6]) up to the point where some
stopping criteria are satisfied (while hoping that the object of interest will be correctly
marked by the final position and shape of the curve). The issues concerning us in this
context are the following:
– If the attempt to map a mathematical model on an image has the effect of writing
something to the image, instead of just reading something from it;
– If the eye images (collections of discrete points) actually satisfy the ideal conditions
in which contour stability and convergence [3] are guaranteed in models of contin‐
uous geometrical spaces;

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_33
446 N. Popescu Bodorin et al.

– If the mathematical model mapped on the eye images is something intrinsic or

extrinsic to the image (is discovered as belonging there naturally or, on the contrary,
is imposed);
– If all of these matters mentioned above translate into excessive computational costs
with respect to the result.
Briefly, our approach is to consider each of the inner and the outer boundaries of the
iris as being simultaneously an almost (fuzzy) circular isoline and a (fuzzy) strong
isoline edge, where the notion of isoline edge is defined below. Circularity is tested using
sine law and lookup tables (instead of using Hough Transform, which is much compu‐
tationally expensive).
The sine law allows for identifying clusters of concyclic points within any given
fuzzy (and possibly) noisy isoline, whereas the lookup tables are indexing sine values
that follow to be used for an easy and fast computation of the radii and centers.
This paper continues by commenting some related articles (in the subsection that
follows), by detailing the principles underlying the proposed segmentation approach (in
the 2nd section), by presenting the proposed segmentation procedure, the experimental
results and our conclusions.

2 Related Works

Efficiently finding the iris boundaries using energy/force based active contours (known
as snakes, [4]) strongly depends on initializing the search algorithm with a particular
snake position, found pretty close to the final/desired one. It happens this way because
the theoretical model of these active contours is not very well supported by the eye
images and, therefore, the errors accumulate in direct proportion with the number of
iterations in such a manner that, sometimes, the target boundary can even be traversed
by the iterating snake. Besides, this active contour model does not support merging and
splitting (however, the extended snake model presented in [6] overcomes this issue). In
general, the snakes are considered elastic curves continuously deformable under the
action of some user defined image forces towards minimization of some user defined
energy field. The so called image forces are responsible for snake self-cohesion, smooth‐
ness, expansion and contraction.
The idea of level-set based active contours is to express the current position as a
constant level set of a potential function which slightly changes over time carrying/
pushing the active contour to its next position [6, 7, 16]. By design, this active contour
model actually supports splitting and merging, but at the cost of a greater computational
effort.
Geodesic active contours got their names mainly by analogy with some particular
situations in geometry and mechanics, their purpose being to extract object boundaries
as a minimal length geodesic. These things go back to Maupertuis’ Principle that tells
[2] in which conditions a snake is a geodesic, i.e. when a given solution minimizing an
energy in an Euclidian (flat) space can be expressed as a solution minimizing a distance
in a Riemannian (curved and twisted) space. However, hidden behind this principle there
is a belief (or an assumption) that any physical space is metrizable in the same manner
 Circular Fuzzy Iris Segmentation 447

in all of its points (there is an intrinsic and global operator defining the Riemannian
distance, in the same way within the entire image) - an assumption that is generally false
(an example will be given below when our approach to this problem will be discussed
on eye images). When it comes for estimating computational complexity of such tech‐
nique in relation with the outcome, one could see (for example in Fig. 13g from [14]
and in Fig. 5 from [13]) that the number of iterations necessary for finding a good enough
approximate of the outer iris boundary can vary (1800 vs. 600 iterations). In the same
figures, it can be seen that the intermediate positions the snake is passing through are
quite meaningless in human understanding. By iterating in this manner, one writes
something to the eye image instead reading suitable candidates that are actually there in
the image, until finding the target boundary, which is also there.

3 Principles Underlying the Proposed Segmentation Approach

We would like to say it was very simple to find this new iris segmentation procedure in
a lucky hour, day, week or even month, but this is very far from being true. A lot of
things that we read or we experienced during the last 9 years guided us in the current
point of our research. During this entire period, while revisiting a number of computing
disciplines, we searched for answers on how to define efficient iris segmentation proce‐
dures with less computational costs for the same or better results.
It happens that these answers came to us not specifically for the purpose of iris
segmentation, but for an entire class of searching problems as a collection of notions
and principles useful each time when someone feels or knows that the speed-accuracy
balance is not the right one. Two of them (in our view, the most important after those
of imperfect /consistent experimental frameworks [9] and Consistent Biometry [10]) are
the notions of vital space and the principle of deflation - another two essential bricks
completing the formal apparatus initially developed in [9, 10], mainly around essential
logical and methodological matters within biometrics.
Regardless its title, this entire section is about practicing iris segmentation efficiently,
not about theoretical principles.

4 The Global Vital Space of a Grayscale Image

In order to maintain the state-space of our algorithm inside a minimal volume of intrinsic
image properties, an eye image (as much as any other two dimensional grayscale image)
is viewed here as a punctual distribution of chromatic information (quantized on a
discrete scale) spread in a rectangular domain of discrete pixels. Hence the natural global
frame coordinate system for the eye image is three-dimensional, discrete, and spans a
volume of a discrete parallelepiped (built with cubic pixels) having a discrete rectangle
of pixels as the horizontal base, and the chromatic discrete scale in any point of the base
oriented in the normal upward direction defined such that, together, the column (x), row
(y) and chromatic (z) axes to form a right-handed coordinate system. This volume,
further referred to as the vital space of a grayscale image, is a granule of space having
an extended convexity property: for any two points found inside it, the shortest paths in
448 N. Popescu Bodorin et al.

between them and the shortest paths from the two points to all coordinate axes and planes
(where distance is defined by taxi-cab norm) are all entirely found inside it also. In other
words, any pixel of the image is reachable from origin and from any other point of its
vital space using paths within this space.
Moreover, given a set of three points, A, B, C, inside this space, any two of them are
connected by a lot of paths passing through the third one and having minimal length (in
taxi-cab norm). In a formal view, we ask if it is possible that all intrinsic properties of
the image to be expressed in a closed formal language using constants, variables, oper‐
ators, predicates and functions defined solely inside its vital space. Regardless the answer
to this ambitious question, it is evident that any iterative procedure able to return a
handler to an object of interest within the image has to traverse the vital space boundary
if it starts somewhere outside this space.
Hence, in order to obtain fast processing algorithms, it is preferably to start them
within the vital space or, at most, on vital space boundary instead of starting them outside
and passing through the boundary towards the object of interest.

5 Search Space Deflation

As a graph, the vital space is fully connected. Therefore, the ensemble (R, Vs, P, I)
formed with the right-handed referential - R, the vital space Vs, the set of all paths within
it – P, and the 8-bit eye image I, is an implicit over-determined description of the image
that can be exhausted completely in exponential time 2255*r*c (where r and c are the
number of rows and columns of the image) as upper bound. In this context, deflation
principle is a principle of action, stating that any search space must be decimated first,
as much as possible, by any means, in order to decrease computational and time
complexity of the search process.
When searching for a given object of interest within an image, deflation process can
occur in successive stages, each of them corresponding to a one step genus-to-species
reduction in which the complement of a species is decimated from the search space. If
a formal, logical and hierarchically nested ontological model object-species-genus exists
for the object of interest, deflation should be easy. Unfortunately, the object of interest
is often defined in natural language, not in data and a programming language - this is
the main source of troubles.
When the definition of the target object is wrong or not rigorous enough, more objects
(belonging actually in other species) or fewer objects of interest are found, error situa‐
tions known as False Accept and False Reject cases. This is why our efforts were directed
to finding suitable computational definitions for the inner and outer boundaries of the
iris in the first place, such that, the process of finding them to include abrupt efficient
deflation stages for the initial search and vital spaces.

6 Defining a Local Vital Space for a Grayscale Image Using Isolines

Each time when this is possible (with reasonable computational effort), we would like
to know a local coordinate system that moves across the image. In this view, a kind of
 Circular Fuzzy Iris Segmentation 449

tangent space is mapped onto the image (an analogy with the tangent space across a
manifold).
First important characteristic of this tangent space is that it is not known (defined)
in all pixels of the image (a first deflation step, and a first fuzzification step – the infor‐
mation becomes partial), whereas the second is that it does not preserve its orientation
for all pixels where it is defined, nor the angles in between its axes. We are not bound
to the idea of imposing an orthogonal local coordinate system (or other classical refer‐
ential) just because it would be easier for us to work with. On the contrary, we are curious
to find what is there, what is supported there by the image itself.
A third important characteristic of our approach is that we consider (numerical)
differentiability (implemented by finite differences) a matter of degree (simply because
its geometrical appearance, namely the smoothness, is also a matter o degree). Hence,
an image cannot be modeled here as a differentiable manifold. On the contrary, (numer‐
ical) lateral partial derivatives will be very often far from matching each other.
Another important aspect, concerning the degeneracy of coordinate curves, will be
commented and illustrated by example below, while talking about isolines.
The particular way in which an intrinsic local coordinate system varies across the
eye image is itself a feature that can be used in detecting regions and/or objects of
interest.

6.1 Fuzzy Isolines

Ideally, a crisp isoline within an eye image is a crisp curve whose points share the same
chromatic level. Since the existence of this ideal form of isoline is not guaranteed, we
must weaken this concept up to fuzzy isoline (f-isoline), by specifying a way of inter‐
preting small chromatic variations as being fuzzy nil, so that, a fuzzy isoline is now a
curve along which the chromatic variation is fuzzy nil. In other words, all points within
an f-isoline are fuzzy critical points of the fuzzy chromatic gradient along the f-isoline.
Semantically speaking, the concept of fuzzy isoline is a handler pointing to curves with
minimal (fuzzy nil) chromatic variability. The fuzzy aspect across the f-isoline (incerti‐
tude in exact line-column position) is neglected (prior to circle finding step of our
approach) in favor of the fuzzy aspect along the f-isoline (incertitude in exact chromatic
value). Hence, it is easy to imagine the f-geometry [15] of a circular f-isoline as if we
spray a straight horizontal line on a paper and then use that paper as a lateral surface for
a cylinder. If the incertitude across a circular isoline is not neglected, its f-geometry
becomes a sprayed torus. In both cases, the prototype curve [15] of the both f-isolines
is a circle. In what follows here, in order to avoid complex graphical representation, the
f-isolines will be graphically represented by their crisp prototype counterparts, as in
Fig. 1b. The f-isolines form the first family of local coordinate curves and span local
directions of fuzzy nil (minimal) chromatic variation.
450 N. Popescu Bodorin et al.

(a)

(b)

(c)

Fig. 1. Detecting the pupil: (a) Fuzzy isolines within the lower chromatic of the eye image;
(b) The f-isolines within an eye image represented by their crisp prototype counterparts on 32
chromatic levels; (c) Pupil isoline chain
 Circular Fuzzy Iris Segmentation 451

6.2 Degenerated f-isolines

A closed f-isoline is a cycle within the image vital space regarded as a graph. If the
interior of a closed f-isoline is large enough and is not containing another isoline, that
f-isoline is the border of a region with fuzzy constant chromatic value (a fuzzy isosur‐
face). Hence, a fuzzy isoline can degenerate to a fuzzy isosurface, meaning that the local
passage of fuzzy nil chromatic variation is not narrow, but wide. In such image regions,
it is impossible to define the second family of local coordinate curves along which chro‐
matic variation is maximal, simply because there, the chromatic variation is locally fuzzy
nil in all directions. The possibility of defining directions of maximal chromatic variation
exists only for pixels situated on the border of the degenerated closed f-isoline, cases in
which, the directions of maximal chromatic variation are all pointing out or in, across
the isoline. When the interior of a degenerate closed f-isoline has fuzzy nil or small area,
it is called a fuzzy spot. A fuzzy membership function can be defined for identifying
white, gray and black spots, as well as white, gray and black isosurfaces.

6.3 Essential Local Chromatic Variations

Two kinds of local directions are considered essential here: those of minimal and
maximal chromatic variation, respectively. Let us consider the set of all pixels covered
by isolines (IP - a pretty sparse truncation of the eye image vital space). For all of them,
local directions of fuzzy nil (minimal) chromatic variation are known (along the isolines
that are going through there). For only a part of them, directions of maximal chromatic
variation can be defined also, while using solely the pixels within IP, as follows below:
– The interface, I, between two isolines is defined here as the set of all sliding segments
drawn from one isoline to the other and having the property that they are not inter‐
secting any other third isoline in between the two given isolines, nor the isolines on
which they slide with their end points. Further, we will say that two isolines interface
each other in point P (which is on one of them, or equivalently, that the two isolines
are locally adjacent in P, or locally connected in P) if there is a non-empty interface
in between them containing a segment starting in or ending with P. An isoline is
interfaced on one side in one of its points P, if there is another isoline interfacing it
on that side in P (i.e. degenerated closed isolines are interfaced on at most one side).
– An isoline is captive on a side if it is interfaced on that side in all of its points by other
one and only isoline. The restriction of the captivity relation to the set of closed
isolines is symmetric.
– For given crisp prototype [15] isolines and their fuzzy counterparts [15], the captivity
relation is weakened here up to the point where we say that a fuzzy isoline is f-captive
on a side if it is interfaced on that side in almost all of its points by other one and
only closed fuzzy isoline.
– An isochain (a fuzzy isochain) is defined here as any sequence containing at least
three (fuzzy) isolines that are pair-wise (fuzzy) captive to each other. The isolines
within such a sequence will be further referred to as being (fuzzy) globally chained.
– Two isolines are said to be connected globally if they are captive to each other. An
intensive parameter ‘cs’ expressing the connection strength as a matter of degree is
452 N. Popescu Bodorin et al.

introduced here as the percentile of all points in which the two isolines are interfaced.
For all pairs of successive (pair-wise captive) isolines found within an isochain,
connection strength parameter cs is unitary and maximal.
– Two isolines are said to be connected locally if there is a common neighborhood such
that their restrictions inside there appear as being connected globally.
– An isoline edge is defined now as any isoline ending an isochain or one/those
supporting the maximum local chromatic variation along an isochain. Its fuzzy coun‐
terpart is defined by analogy.
– Given a point P on an isoline IL1 and a second isoline IL2 interfacing IL1 on one side
in P (where the two isolines can share either different or the same chromatic level),
a local direction of maximal variation in P is the vector defined by the shortest path
trough P found within the interface of IL1 and IL2. If the shortest path from P to IL2
is not unique, a local direction of maximal variation in P is the vector PP’ where P’
is situated on IL2 and PP’ is matching the direction of the resultant of all shortest
paths from P to IL2.
– Given a point P on an isoline IL and two other isolines IL1 and IL2 interfacing IL on
the same side in P, then, regardless the chromatic levels of the isolines, in general
there are two different local directions of essential variation in P across IL, one
maximal with respect to IL1 and the other one maximal with respect to IL2.

6.4 A Local, Isoline-Based, Vital Space of the Grayscale Image

If a non-degenerated isoline IL has only two adjacent isolines IL1, and IL2, one on each
of its sides, in a given point P, a local intrinsic coordinate system in P is given by two
local directions of fuzzy nil chromatic variation (from P, back and forth along IL) and
two local directions of maximal variation (across IL in P, one on each of its sides in P,
towards IL1 and IL2, respectively).
Finally, a local intrinsic coordinate system exists for any pixel P found on an isoline
IL and contains 2 local directions of fuzzy nil chromatic variation (back and forth from
P, along the isoline) and N local directions of maximal chromatic variation (where N is
the number of distinct isolines locally adjacent with IL in P). Hence, the first family of
local coordinate curves contains isolines, whereas the second family of local coordinate
curves contains polygonal lines stepping in between locally adjacent isolines while
simultaneously minimizing the distance and maximizing chromatic variation.
The reunion of all vital spaces associated point-wise to all pixels covered by isolines
is the local isoline-based vital space of the grayscale image. Keeping the computation
inside this space (while searching for objects of interest) makes sense from two points
of view: this space is the result of minimizing the volume while maximizing the infor‐
mation about the local chromatic variations, and secondly, this space is assumed prac‐
tically as it is, not ideally – as having theoretic properties in terms of continuity and
smoothness.
Along a closed isoline that is captive on both sides, the local intrinsic coordinate
system reflecting the essential local chromatic variations in the current point P is 4
dimensional, containing two local directions of fuzzy nil chromatic variation (from P
 Circular Fuzzy Iris Segmentation 453

back and forth, along the isoline) and two directions of maximal chromatic variation
across the isoline, towards the two isolines that captured it in between them.
Exhausting the local isoline-based vital space of a grayscale image is a way of
evidencing how to traverse that image by stepping only on and in between (locally or
globally) connected /adjacent isolines.

7 Circularity Test Based on Sine Law and Lookup Tables

Using the terminology defined above, from a practical point of view expressed in terms
of intrinsic properties of the eye image, the pupil boundary (found within a standard eye
image collected for biometrics purposes) is simultaneously: (1) a fuzzy isoline, (2) fuzzy
circular, (3) a member of an isochain and (4) a fuzzy isoline edge. From all these four
features, three were commented above, whereas the circularity is usually tested by
applying Hough Transform on the edge matrix. Firstly, we are no longer considering
edges, but isolines, isochains and edge isolines and their fuzzy counterparts. Secondly,
we searched for an alternative to Hough voting (which is well known as being compu‐
tationally expensive) and we found one way to estimate circularity and circular/concy‐
clic clusters within the isolines using the sine law, as follows. Let us consider:
– an isoline IL,
– R – the set of rectangles defined by points within IL such that they have a vertical
side,
– T – the set of triangles △ABC defined by points within IL, such that they have acute
angles only, and either a horizontal AB or a vertical AB side, as illustrated in Fig. 2.

Fig. 2. The set T and the indexing scheme within the lookup table LT for r = c/2sin(Ĉ ) values
corresponding to unique (c, d, h) triplets - see also formula (1). IL and IL’ are two iris boundaries
candidates.

– the circumcircle of △ABC heaving the radius r;

– c – the length of AB, H∈(AB) such that CH ⊥ AB;
– d – the distance from A to H, h – the length of CH;
454 N. Popescu Bodorin et al.

– LT – a 3D lookup table containing c/2sin(Ĉ ) values, indexed after c, d, and h,

according to sine law:
c
r(c, d, h) = (1)
̂
2 sin(C(c, d, h))

– a fuzzy predictor pr, of low computational complexity (linear with respect to the
cardinal of T), defined on T and selecting the most voted radius over the set T, radius
that corresponds to the maximum point within a histogram of a signal formed by
reading (not by computing) radius values corresponding to unique (c, d, h) indexing
triplets in the lookup table LT;
– a fuzzy predictor pc, of low computational complexity (linear with respect to the
cardinal of R), defined on R and selecting the most voted center over the set R. Voting
mechanism is based on the fact that any rectangle in R votes its center as being the
center of a concyclic cluster found within the isoline. These centers are found with
maximal efficiency, by averaging two abscise values and two ordinate values as in
Fig. 3 (not by computing square roots for some sums of squares).

Fig. 3. For any circular isoline, by sharing the same vertical and the same horizontal symmetry
axes, all concyclic rectangles within the set R vote for the same center, whose coordinates are
computed by averaging two column and two line indices, respectively.

Fig. 4. Iris segmentation based on isolines, sine law and lookup tables
 Circular Fuzzy Iris Segmentation 455

7.1 Advantages
By design, the two fuzzy predictors pc and pr from above, both do their job well, even
when the isoline is not circular on its entire length: a specular light or other occlusion
situated near the inner iris boundary usually modifies locally the isoline path, but this
perturbation cannot happen simultaneously in all corners of the rectangles within R, or
in all corners of the triangles within T.
The lookup table LT is computed only once and used over and over again as a
replacement for its much computationally expensive alternative, namely Hough accu‐
mulator. From the point of view of a system designer, this means that the entire compu‐
tational effort of writing this table on either a ROM (read only memory) or a RAM can
be done once, prior to exploitation (during system calibration) or during system boot,
respectively.

8 Iris Segmentation Based on Isolines, Sine Law and Lookup Tables

With the terminology introduced in the previous sections, pupil boundary is an almost
circular edge isoline whose radius rp belongs in a given range rrp, isoline belonging in
a highly connected chain (almost unitary connection strength between its adjacent
isolines) that makes the longest passage to one of the darkest and almost circular isoline
found outside a given image border. Any isoline chain that candidates for the inner iris
boundary is discarded if no left and right outer boundaries are found around it. Shortly,
the candidate pupil chains are discarded by backtracking after successful outer left/right
iris boundaries detection. Each of the latter also belongs in a chain (possibly a weakly
connected chain) that makes the passage between two adaptively determined chromatic
thresholds: one lower threshold (lt) detected as the chromatic value of the whitest isoline
of the current pupil chain candidate and one upper threshold (ut) identified as the first
regional (i.e. non-spot) maxima found across the iris radius range on the left and right
sides of candidate pupil chain (hopefully on the sclera, but not mandatory).
Despite, theoretically, it could be possible to finalize the search for the iris boundaries
with two or more candidates, this never happened practically on LG2200 database, and
therefore, we haven’t established a procedure for selecting the most suitable candidate
triplet (inner circular boundary, left outer circular boundary, right outer circular boun‐
dary) in this case.
Over all, pupil and iris are considered concentric, whereas the outer circular iris
boundary is defined by the smallest radius detected for the left and right outer circular
boundaries. This strategy was first used in [12] and is motivated by the fact that doing
that is not giving guarantees on the accuracy of segmenting each individual eye image,
but guarantees that the iris regions neglected in any two different standard eye images
taken from the same user are almost the same (bigger the eccentricity of the iris, wider
the area of neglected iris region).
456 N. Popescu Bodorin et al.

The pseudo-code for the proposed segmentation procedure is as follows:

1. INPUTS:
1.1. the current standard infra-red eye image EI;
1.2. the given pupil radius range and its extreme values: rrp; = [mrrp, Mrrp ];
1.3. the given iris radius range and its extreme values: rri; = [mrri, Mrri ];
1.4. the width of the image border ib;
2. Expand the local vital space of the eye image as a collection of local directions of
minimal & maximal chromatic variations (along & across the isolines found in the
lower image chromatic).
3. Identify the set of pupil candidates PC as containing the first four longest and most
connected chains fitted inside 2Mrrp x 2Mrrp squares and formed with almost
circular isolines whose chromatic values are descending to black (Fig. 1.c);
4. For each pupil candidate pc in PC,
4.1. Get the lower chromatic threshold lt detected as the chromatic value of the
whitest outer isoline belonging in the current pupil chain candidate;
4.2. Get the upper chromatic thresholds utl and utr identified as the first regional
(i.e. non-spot) maxima found across the iris radius range on the left and right
sides of candidate pupil chain;
4.3. Infer pupil radius rp (Fig. 2) and pupil center position cp (Fig. 3)
4.4. Get BN be a ‘butterfly’ neighborhood whose wings extend from the left and
from the right of the pupil candidate, for 60° around the horizontal of pupil
center (±30° above/under).
4.5. To the left and to the right of BN, find left and right outer iris boundary candi‐
dates (libc and ribc) as the longest and most connected chains making the
passage from lt to utl and to utr, respectively.
4.6. Memorize the current (pc, libc, ribc) triplet candidate formed with the edge
isolines within the three corresponding chain candidates in the set of possible
solutions PS.
5. If PS is empty,
5.1. report that the current image does not contain a detectable iris.
ELSE,
5.2. for the best candidate in PS:
5.2.1. Infer iris radius ri as the minimum between the radii of libc and ribc;
5.2.2. Mark pupil circle, and both iris circles (corresponding to libc and ribc);
5.2.3. Get BN be a ‘butterfly’ neighborhood whose wings extend from the top
and to the bottom of the pupil candidate, for 60° around the vertical of
the pupil center (±30° right/left), find the edge isolines that define top/
bottom eye occlusions (if any) and mark top/bottom eye occlusions (if
any) by filling the regions found in-between the outer iris boundary and
top/bottom edge isolines previously determined.
6. OUTPUT: either
6.1. An empty vector (no iris found)
OR
6.2. Two occlusion masks and a vector containing two center coordinates, the inner
and outer radii of the iris.
 Circular Fuzzy Iris Segmentation 457

9 Experimental Results

The new segmentation procedure proposed here proved a failure rate of 5.83% when
tested on 116,564 eye images (LG2200 subset of ND-CrossSenssor-Iris-2013 database,
[8]) at an average speed of four images per second on a single Xeon L5420 core, where
a segmentation failure means that no suitable candidates for inner and outer iris boun‐
daries have been found within an image. However, we know from our previous work
on this ‘difficult’ database that almost a half of the situation in which segmentation fails
(around 3000 eye images), it does that because those images are improper with respect
to iris recognition. Such examples are illustrated here in Fig. 5: for some of them there
is no pupil chain candidate to be found because there is no circular pupil present; in other
cases, a circular pupil is present but it is not black enough, or position of the iris is too
close to image border. The cases of successful segmentation are illustrated in Fig. 4
where one can see examples for: horizontal /vertical butterfly neighborhoods of the iris
center (see 4.4 and 5.2.3 in the pseudo-code of the proposed segmentation procedure),
left and right outer iris boundary candidates represented as the two circles (libc/ribc –
see the instructions 5.2.1, 5.2.2 in the pseudo-code above) and the upper occlusion mask
induced by the presence of eyelashes over the iris area.

Fig. 5. Eye images that are not able to deliver at least one suitable triplet candidate (pc, libc,
ribc) for pupil (inner), left and right (outer) iris boundaries - see above, the steps (4.6) and (5) in
the pseudo-code of the proposed segmentation procedure.
458 N. Popescu Bodorin et al.

10 Conclusions

This paper proposed a new and original fuzzy approach to circular iris segmentation
based on isolines, sine law and lookup tables. All isolines found in the eye image together
form the space in which the inner and outer boundaries of the iris are searched for, while
the sine law was used for identifying clusters of concyclic points within any given and
possibly noisy isoline. The main difference between the related works and our approach
is that we only read something from the eye image searching for three isoline chains
matching certain conditions that together qualify them as being inner and left/right outer
boundaries of the iris.

Acknowledgement. This work was supported by the Applied Computer Science Laboratory
(Bucharest, Romania).

References

1. Caselles, V., Catté, F., Coll, T., Dibos, F.: A geometric model for active contours in image
processing. Numer. Math. 66(1), 1–31 (1993). doi:10.1007/BF01385685
2. Caselles, V., Kimmel, R., Sapiro, G.: Geodesic active contours. Int. J. Comput. Vis. 22(1),
61–79 (1994). Kluwer Academic Publishers, ISSN 0920-5691, 1997
3. Chaudhury, K.N., Ramakrishnan, K.R.: Stability and convergence of the level set method in
computer vision. Pattern Recogn. Lett. 28, 884–893 (2007). Elsevier
4. Daugman, J.: New methods in iris recognition. IEEE Trans. SMC - PART B Cybern. 37(5),
1167–1175 (2007). doi:10.1109/TSMCB.2007.903540. IEEE Press. ISSN 1083-4419
5. Daugman, J.: How iris recognition works. IEEE Trans. Circ. Syst. Video Technol. 14(1), 21–
30 (2004)
6. Nguyen, K., Fookes, C., Sridharan, S.: Fusing shrinking and expanding active contour models
for robust iris segementation. In: Proceedings of 10th International Conference on Information
Science, Signal Processing and their Applications (ISSPA 2010), pp. 185–188. IEEE Press
(2010)
7. Osher, S., Fedkiw, R.: Level Set Methods: An Overview and Some Recent Results, Technical
Report CAM-00-08, Mathematics Department, UCLA, February 2000. https://
www.math.ucla.edu/pub/camreport/cam00-08.ps.gz. Accessed 5 Dec 2016
8. ND-CrossSenssor-Iris-2013 Database. https://sites.google.com/a/nd.edu/public-cvrl/data-
sets
9. Popescu-Bodorin, N., Balas, V.E.: Fuzzy membership, possibility, probability and negation
in biometrics. Acta Polytech. Hung. 11(4), 79–100 (2014)
10. Popescu-Bodorin, N., Balas, V.E.: Learning iris biometric digital identities for secure
authentication: a neural-evolutionary perspective pioneering intelligent iris identification. In:
Recent Advances in Intelligent Engineering Systems. Studies in Computational Intelligence,
vol. 378, pp. 409–434. Springer (2011)
11. Popescu-Bodorin, N., Balas, V.E.: Comparing haar-hilbert and log-gabor based iris encoders
on bath iris image database. In: Proceedings 4th International Workshop on Soft Computing
Applications, pp. 191–196. IEEE Press, July 2010. doi:10.1109/SOFA.2010.5565599. ISBN
978-1-4244-7983-2
 Circular Fuzzy Iris Segmentation 459

12. Popescu-Bodorin, N.: Exploring new directions in iris recognition. In: 11th International
Symposium on Symbolic and Numeric Algorithms for Scientific Computing (SYNASC
2009), pp. 384–391. IEEE Computer Society, September 2009. doi:10.1109/SYNASC.
2009.45
13. Ross, A., Shah, S.: Segmenting non-ideal irises using geodesic active contours. In:
Proceedings of Biometrics Symposium (Baltimore, USA), pp. 1–6. IEEE (2006). doi:10.1109/
BCC.2006.4341625
14. Shah, S., Ross, A.: Iris segmentation using geodesic active contours. IEEE Trans. Inf.
Forensics Secur. 4(4), 824–836 (2009)
15. Zadeh, L.A.: Toward extended fuzzy logic - a first step. Fuzzy Sets Syst. 160, 3175–3181
(2009). doi:10.1016/j.fss.2009.04.009. Elsevier
16. Zhang, X., Sun, Z., Tan, T.: Texture removal for adaptive level set based iris segmentation.
In: Processing of the 17th IEEE International Conference on Image Processing, pp. 1729–
1732. IEEE Press (2010)
 A Parallel Approach on Airport Runways Detection
Using MPI and CImg

P.S. Penariu1,3 ✉ , N. Popescu Bodorin1, and V.C. Stroescu1,2

( )

1
Applied Computer Science Laboratory (ACSL), Bucharest, Romania
{penariu,bodorin,victor}@lmrec.org
2
‘Goethe’ Deusche National College, Bucharest, Romania
3
Computer Science Department, University of Bucharest, Bucharest, Romania

Abstract. This paper proposes a parallel and distributed approach to achieve the
recognition of airport runways in aerial images, by combining Message Passing
Interface (MPI) standard communication protocol with Canny Edge detector and
Radon Transform. The reason for adopting a parallel architecture is obtaining
speed and efficiency when running on multiple processors from a multi-core and
hyper-threaded platform. The airport runways recognition application is imple‐
mented in C++ and uses MPI and CImg library for image processing. The latter
provides instruments for noise removal (Gaussian filtering), edge finding (Canny
Edge detector) and line extraction (Radon Transform), whereas the former allows
for distributing independent parallel computation over a given number of cores
while avoiding any intercommunication overhead. The presented approach
proved a good performance on the recognition of the airport runways through the
experimental results, with an average of 0.4 s for the overall computational time
per image (including reading from and writing to files).

Keywords: Airport runways detection · Message Passing Interface (MPI) · CImg

library · Parallel Architecture

1 Introduction

Nowadays, detecting airport runways can have important advantages in the military and
public field. Various approaches and methods were tested over the years with the purpose
of recognizing airport runways from aerial images, all of them being grouped in the
following three categories:
– Approaches and methods based on straight-line support rage [1], line segments
extraction models [4, 7, 10, 21, 23], Radon Transform [2, 6, 11, 26], segmentation
[13, 17], feature extraction and feature classification techniques [12];
– Approaches and methods based on Neural Networks (Multi-Channel Pulse Coupled
Neural Network [28]), Improved Chain Codes Based Edge Tracking [5] and improve‐
ments of Hough Transform [5, 18];
– Approaches and methods that incorporate different variants of the remote sensing
data concept [14–16, 21, 25, 27].

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_34
 A Parallel Approach on Airport Runways Detection 461

The application presented in this paper uses techniques from the first two categories,
namely methods designed to extract the straight lines from edge images as the basis for
recognizing airport runways. Since the quality of the results is highly dependent on the
level of noise, applying a noise removal technique in the preprocessing step may be
helpful. The goal is to accomplish a good recognition of the airport runways, excluding
taxi-lines, buildings, mobile objects or other features. To concretize this, in the prepro‐
cessing phase is applied a Gaussian filter to re-move additional noise from the aerial
images. The edges are further extracted with Canny Edge detector and the straight lines
are found using Radon Transform.
The outline of our paper is as follows: improvements of existing techniques for edge and
line detection algorithms are given in the Sect. 2. A short description of the methodology,
along with the description of the proposed application (including the main steps of the
parallel algorithm) is presented in the Sect. 3. The experimental results are discussed in the
Sect. 4. Conclusions and ideas for continuing this work follow in the last section.

2 Related Works

Over the time, significant contributions to improve line detection algorithms have been
brought [1, 4, 6, 7, 12, 13, 15–21, 23, 25–27]. According to Duan et al. [8], Hough
Transform represents a voting process in which each point from a pattern gives a vote
for all the possible patterns that are passing through it. A straight line may be defined
by the Eq. (1), where x and y correspond to a coordinate pair, P(x, y) (Fig. 1), from the
image coordinate space.

𝜌 = x cos 𝜃 + y sin 𝜃 (1)

d y

θ
ρ P(x,y)

θ x

Fig. 1. Hough Transform – Relationship between the (x, y) coordinates and (θ, ρ) parameters

The Hough Transform accumulator array receives votes for the pair of parameters
(θ, ρ). θ is the angle between the line d (Fig. 1) and the y-axis (0 ≤ θ ≤ π), while ρ stands
for the distance between the origin and the line. The most accurate lines are indicated
in the Hough accumulator by the maximum values.
The Discrete Radon Transform (unlike the continuous one which is an integral) is a sum
of pixels from the image matrix that is computed in the directions specified by the angles
θ (0 ≤ θ < π) [20]. There are several Discrete Radon Transform algorithms proposed [2,
11, 20], one of them, for periodic vectors, being introduced by Beylkin, in [2].
462 P. S. Penariu et al.

The Hough Transform was subject to many improvements over time. For instance, Di
et al. [5] propose a method that combines Hough Transform with an Improved Chain
Codes Based Edge Tracking (ICCBET) algorithm. ICCBET erases all the short and curving
lines and confirms the approximate orientation between neighbor pixels reducing in this
way the angle range that Hough Transform has to process (the approximate orientation is
chosen as the dominating orientation in the set of the reserved chain lists). Ma et al. [18]
propose another improvement for Hough Transform. They propose a method that auto‐
matically adjusts the Hough Transform threshold to accomplish noise reduction by dividing
the straight lines into two categories (parallel lines and intersection lines), the threshold
being transformed according to different characteristics of the two categories.
Just as in the case of Hough Transform, different contributions and improvements
were proposed for Radon Transform as well. For instance, Dong et al. [6] reduced Radon
Transform’s computational time by applying it only on the edge image, where the left
bottom of the edge image represents the origin point for Radon Transform. In addition,
the incline angle of the straight line, the length of the straight line and the distance
between the origin point and the straight line are all acquired at the same time.
Xiong et al. in [26] propose an algorithm which selects approximately ten of the most
voted objects by Radon Transform, while ignoring the rest (which are set to 0). The
airport runways were recovered according to several conditions (conditions that have
the purpose of avoiding the confusion of the runways edges with the ones of the other
elements present in the images, such as rivers) formulated in the form of a structure
creditable membership and an intensity creditable membership.
Other related works that are worth considering are [13, 21], both are using techniques
of segmentation to extract instances of long rectangular shapes or line segments from
aerial imagery. For instance, Huertas et al. [13] propose a method based on object shape
recognition to detect airport runways, with a procedure that uses line segments and “anti
parallel” pairs of line segments (line segments characterized by parallel lines of opposite
contrast, characterized by their length, orientation, width, end points and color)
computed from aerial imagery.
Another approach to runways detection is presented in [23], where Tripathi et al.
propose an algorithm that is based on the shape and color of the runways. The algorithm
consists of a binary classification (or coarse segmentation) of airport regions, using
anisotropic diffusion (iterative process that provides the noise immunity, smoothing the
image and preserving the edges) and Frangi filter (vessel enhancement filter that compute
the likeliness of an image region to represent structures), followed by a detailed analysis
of these regions using shape and color attributes.
Parallelizing an algorithm can lead to drastically reduction of the computation-al
time, this being shown, for example, by Zhuang et al. in [28] where they propose a Multi-
Channel Pulse Coupled Neural Network (MPCNN) implemented in parallel on a Field-
Programmable Gate Array (FPGA) chip, obtaining an average of 15.6 ms for the
MPCNN approach to recognize fifty to one hundred straight line segments.
 A Parallel Approach on Airport Runways Detection 463

3 Runways Detection Algorithm

Implementing a parallel application efficiently requires an environment that supports a

framework for parallel programming [22]. Therefore, we use a C++ pro-gramming
language in Visual Studio 2012 on a Windows 7/64bit platform with a multi-core (2)
hyper-threaded (4) Intel i5 processor, with Message Passing Inter-face (MPI) [19]
standard communication protocol for inter-communication between processes. As a
toolbox for image processing (Edge Detection, Radon Transform, straight lines detec‐
tion from Radon image), we use CImg library (C++ Template Image Processing
Toolkit) [3]. The library is an adaptable open-source with applicability on multiple
operating systems, compatible with numerous C compilers and libraries. For a better
understanding, in Table 1 is exposed the pseudo-code of our algorithm, followed by the
main scheme of the application presented in Fig. 2.

Table 1. Pseudo-code for parallel approach on airport detection algorithm

464 P. S. Penariu et al.

Run_MPI_Flag is set?
Call Again
YES NO

Run_MPI_Flag is 0? Set Run_MPI_Flag to 0

YES NO

MPI Init Are the MPI config vars (ncores

& hosts) set?
Call
Again
YES NO
Is
Set MPI_Flag to Set MPI config
MASTER? 1 vars

Call MPI
(mpiexec…..)
YES NO

Master: FIND IMAGES

Master: SPLIT IMAGE Slave: Receiving Buffer Name

COLLECTION EQUALLY FOR
SLAVES
Slave: Parses Buffer File

Master: WRITE FILE BUFFERS

FOR SLAVES SEND Slave: Process Current Image
RECEIVE

Master: SEND BUFFER NAMES

Slave: Write Results
TO SLAVES

SEND/RECEIVE
Slave: Acknowledge Master for finish-
ing its job
Master: WAIT FOR SLAVES TO
FINISH THEIR JOBS
MPI Finalize Exit

Fig. 2. The flow chart of the proposed parallel approach on the recognition of complex scenes
from airport imagery

Since it was designed to run in parallel, the algorithm was decomposed into several
individual parts (see Fig. 2). By implementing master–slave architecture, the master
process passes to each slave process, through a send-receive blocking routine, the name
 A Parallel Approach on Airport Runways Detection 465

of the file where the paths of the images are stored (Table 1, steps 7–9). Therefore, every
slave process has the duty to process its own set of image paths (Table 1, step 12), while
the master process waits until all slaves complete their part of image processing
(Table 1, step 10).

(a)

(b)

(c)

Fig. 3. Results obtained after applying the airport runways detector with preliminary Gaussian
filtering. The images from the row (a) show the edges resulted from applying Canny Edge
detector; the images from the row (b) are the image representation of the Radon accumulator,
while on the row (c) are found the airport runways marked with red lines, according to the votes
information accumulated by Radon Transform.

Blocking send–receive routines use message buffers when passing messages

between slaves or between slaves and master, a reason why the system may go into
deadlock. For an efficient and accurate communication between processes, we send data
from master to slaves processes with an individual tag for each process, all MPI Requests
finishing sequentially their part of processing, with no inter-communication between
slaves. The communication (send and receive) is allowed only between master and
slaves, meaning that the slaves do not communicate with each other at all.
The steps performed by each of the slaves to retrieve the airport runways are as
follows. Firstly are applied some preprocessing techniques to the airport images:
– all images from the database are resized to 640 × 480;
– then is applied a Gaussian low pass filter on the resized images with the purpose of
reducing the level of noise in the aerial images, and therefore, to obtain more mean‐
ingful edges (edges that truly represent the runways);
466 P. S. Penariu et al.

– with Canny Edge detector are retrieved the edges from the airport runways images.
Examples of resulted edges can be seen in Fig. 3a.
The algorithm continues with the application of Radon Transform on the edge images
and ends with the runways extraction phase from the resulted Radon images. We choose
to extract the first ten most voted straight lines, by creating the Radon image histogram
where the highest pixels intensity values are representing the most voted straight lines.
The pixels corresponding to those lines are marked with the color red in the output
images.

4 Results

The database used to test the proposed runways detection method is a collection of aerial
images obtained with Earth Explorer Satellite Searcher [9], source owned by United
States Geological Survey [24], a scientific agency of the United States government. The
database includes images with increased difficulty as well, such as multiple runways,
intersection of runways, or buildings, airplanes or taxi-lines. In total, in our database,
there are 492 aerial images taken from approximately the same altitude from a list of 63
airports. There is an average of eight images per airport.
The results obtained at each of the steps performed by the proposed algorithm are
illustrated in Fig. 3 as follows:
– In Fig. 3a are shown the edge maps resulted after applying a Gaussian filter and
Canny Edge detector in the preprocessing phase. The Gaussian low pass filter
smoothed the image, resulting an edge image with less noise (see Fig. 3a—first and
second columns and Fig. 4a—first and second columns), which leaded to correctly
identifying the runways (see Fig. 3c—first and second columns and Fig. 4c—first
and second columns). Comparing the results from Fig. 3a–c, right column, with the
ones from Fig. 4a–c, right column, can be seen that there are images for which
applying a Gaussian filter did not influence the runways detection algorithm to give
wrong results.
– Figure 3b illustrates the Radon accumulator corresponding to the images from
Fig. 3a (column correspondence). Maximum values in the Radon accumulator corre‐
spond (ideally) to runways lines. Since the values from the Radon accumulator were
normalized to [0255] interval, white pixels in the image of the Radon accumulator
indicate the line coordinates/parameters that received the most votes.
– In Fig. 3c is shown the output of the runways detection algorithm, namely the aerial
images with the detected runways marked with red lines.
Applying a preliminary Gaussian filter improves the runways detection accuracy,
indeed, by smoothing the image and removing in this way some of the unnecessary
edges. It also allows further improvements, as can be seen in Fig. 5, which shows the
results obtained by applying (right column) or not (left column) the Gaussian low pass
filter. In both cases, the airport runways detection application has poor results. For
instance, in the left column, where the edge image resulted only from applying Canny
Edge detector, without any other image processing techniques, the proposed algorithm
 A Parallel Approach on Airport Runways Detection 467

(a)

(b)

(c)

Fig. 4. Results obtained after applying the airport runways detector without preliminary Gaus‐
sian filtering. The images from the row (a) show the edges resulted from applying Canny Edge
detector; the images from the row (b) are the image representation of the Radon accumulator,
while on the row (c) are found the airport runways marked with red lines, according to the votes
information accumulated by Radon Transform.

indicates correctly (with one of the ten most voted lines) one of the runways present in
the image, while the rest of them pass through the middle of the second runway. Applying
Gaussian filter (right column) fixed the tendency of the Radon Transform to give many
votes for the edges formed in the middle of the second runway, but it also made almost
invisible the second (perpendicular) runway.
In Fig. 6 are put together several examples where the proposed airport runways
detector gives the correct indication of the runways coordinates. All these images are
complex since the corresponding edge image, due to all the elements comprised in the
aerial images is rich in unnecessary edges induced, for instance, either by the taxi-lines
or by the buildings situated nearby the airport runways.
The most often encountered errors—resulted after applying the proposed algorithm
—are illustrated in Fig. 7, namely the most voted lines indicate an edge that does not
belong to the searched runway (Fig. 7a), or the lines do indicate the runway but with
slight angle deviations (Fig. 7b).
468 P. S. Penariu et al.

(a)

(b)

(c)

Fig. 5. Examples of erroneous results obtained no matter if the Gaussian filter is applied or not.
The left column shows the (a) edge map of the airport aerial image, (b) Radon accumulator and
(c) the lines with the maximum number of votes, marked with red. The right column is homologous
to the left column, but on the original aerial image was applied a Gaussian filter in the image
processing phase.

Fig. 6. Samples of airport scenes after extracting first ten most voted lines from Radon
accumulator. The airport runways are correctly identified.
 A Parallel Approach on Airport Runways Detection 469

(a) (b)

Fig. 7. Experimental results that, after a visual inspection, were marked as being erroneous
because either the most voted lines are outside the runway from the image (a) or the most voted
lines are indeed on the runway’s boundaries but with slight angle deviations (b).

Overall, the presented parallel airport runways detection algorithm proved, during
the experimental results on the aerial images, that is robust and that is able to recognize
the airport runways within different image scenes. In addition, the average time spent
by the proposed algorithm on the processing of one aerial image is about 0.4 s.

5 Conclusions

In this paper, we combined Message Passing Interface communication protocol with

Canny Edge detector and Radon Transform for airport runways detection. The experi‐
mental results show that the proposed method for airport runways de-tection is robust
and able to detect the runways in different scenarios: intersection of runways, presence
of multiple runways, buildings, airplanes or taxi-lines.
We think the presented application leaves enough room for improvements. In this
regard, future works will include improvements of Radon Transform to accomplish a
better recognition of the straight lines. By applying a minimum threshold for the distance
between every two points defining the straight lines and by explicitly calculating θ values
through the arctangent, we expect to increase line detection accuracy.

Acknowledgements. We would like to thank United States Geological Survey for granting us
the necessary rights for using their aerial imagery acquisition tool.
This work was supported by the Applied Computer Science Laboratory (Bucharest, Romania).

References

1. Bai, X., Han, Y., Wu, D., Zhang, F.: The automatic detection method for airport runways. In:
2015 8th International Congress on Image and Signal Processing (CISP), pp. 1015–1019.
IEEE, October 2015
2. Beylkin, G.: Discrete radon transform. IEEE Trans. Acoust. Speech Signal Process. 35(2),
162–172 (1987)
470 P. S. Penariu et al.

3. C++ template image processing toolkit - http://cimg.eu/, licensed under CeCILL v2.0. (http://
www.cecill.info, http://www.cecill.info/licences/Licence_CeCILL_V2-en.html)
4. Cao, S., Jiang, J., Zhang, G., Yuan, Y.: Runway detection using line segment statistical model.
In: 2012 Third International Conference on Intelligent Control and Information Processing
(ICICIP), pp. 601–604. IEEE, July 2012
5. Di, N., Zhu, M., Wang, Y.: Real time method for airport runway detection in aerial images.
In: International Conference on Audio, Language and Image Processing (ICALIP), pp. 563–
567. IEEE, July 2008
6. Dong, Y., Yuan, B., Wang, H., Shi, Z., Liu, Y.: An algorithm for recognizing runway based
on improved Radon Transformation. In: 2011 International Conference on Information
Technology, Computer Engineering and Management Sciences (ICM), vol. 2, pp. 275–279.
IEEE (2011)
7. Dong, Y., Yuan, B., Wang, H., Shi, Z.: A runway recognition algorithm based on heuristic
line extraction. In: 2011 International Conference on Image Analysis and Signal Processing
(IASP), pp. 292–296. IEEE, October 2011
8. Duan, D., Xie, M., Mo, Q., Han, Z., Wan, Y.: An improved Hough transform for line detection.
In: 2010 International Conference on Computer Application and System Modeling (ICCASM
2010), vol. 2, p. V2-354, IEEE, October 2010
9. Earth explorer satellite searcher. http://earthexplorer.usgs.gov/
10. Fengjing, Z., Wenbang, S., Yongsheng, G.: Airport runway extraction method in aviation
reconnaissance image. In: 2013 2nd International Symposium on Instrumentation and
Measurement, Sensor Network and Automation (IMSNA), pp. 152–154. IEEE, December
2013
11. Fishburn, P., Schwander, P., Shepp, L., Vanderbei, R.J.: The discrete Radon transform and
its approximate inversion via linear programming. Discret. Appl. Math. 75(1), 39–61 (1997)
12. Han, P., Chang, L., Shi, Q., Qu, J.: Runways detection based on scattering similarity and
structural characteristics. In: Integrated Communication, Navigation, and Surveillance
Conference (ICNS), p. H2-1. IEEE, April 2015
13. Huertas, A., Cole, W., Nevatia, R.: Detecting runways in complex airport scenes. Comput.
Vision Graph. Image Process. 51(2), 107–145 (1990)
14. Jackson, P.T., et al.: Runway detection in high - resolution remote sensing data. In: 2015 9th
International Symposium on Image and Signal Processing and Analysis (ISPA), pp. 170–175.
IEEE (2015)
15. Li, Z., Liu, Z., Shi, W.: Semiautomatic airport runway extraction using a line-finder-aided
level set evolution. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 7(12), 4738–4749 (2014)
16. Linbo, T., Jinglin, Y., Baojun, Z., Chenwei, D., Baoxian, W.: Airport target detection
algorithm in remote sensing images based on JPEG2000 compressed domain. In: Radar
Conference 2013, IET International, pp. 1–4. IET, April 2013
17. Liu, D., He, L., Carin, L.: Airport detection in large aerial optical imagery. In: Proceedings
IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2004),
vol. 5, p. V-761, May 2004
18. Ma, Y., Wang, J., He, D., Pang, J.: Accurate line detection by adjusting hough transform
threshold adaptively. In: 2010 6th International Conference on Wireless Communications
Networking and Mobile Computing (WiCOM), pp. 1–6. IEEE, September 2010
19. Microsoft Message Passing Interface (Microsoft MPI). https://msdn.microsoft.com/en-us/
library/bb524831(v=vs.85).aspx
20. Press, W.H.: Discrete Radon Transform has an exact, fast inverse and generalizes to operations
other than sums along lines. Proc. Natl. Acad. Sci. 103(51), 19249–19254 (2006)
 A Parallel Approach on Airport Runways Detection 471

21. Tang, G., Xiao, Z., Liu, Q., Liu, H.: A novel airport detection method via line segment
classification and texture classification. IEEE Geosci. Remote Sens. Lett. 12(12), 2408–2412
(2015)
22. Träff, J.L., Gropp, W.D., Thakur, R.: Self-consistent MPI performance guidelines. IEEE
Trans. Parallel Distrib. Syst. 21(5), 698–709 (2010)
23. Tripathi, A.K., Swarup, S.: Shape and color features based airport runway detection. In: 2013
IEEE 3rd International Advance Computing Conference (IACC), pp. 836–841. IEEE,
February 2013
24. United States Geological Survey. https://www.usgs.gov/
25. Wang, X., Lv, Q., Wang, B., Zhang, L.: Airport detection in remote sensing images: a method
based on saliency map. Cogn. Neurodyn. 7(2), 143–154 (2013)
26. Xiong, W., Zhong, J., Zhou, Y.: Automatic recognition of airfield runways based on Radon
Transform and hypothesis testing in SAR images. In: 2012 5th Global Symposium on
Millimeter Waves (GSMM), pp. 462–465. IEEE, May 2012
27. Zhu, D., Wang, B., Zhang, L.: Airport target detection in remote sensing images: a new method
based on two-way saliency. Geosci. Remote Sens. Lett. 12(5), 1096–1110 (2015). IEEE
28. Zhuang, H., Low, K.S.: Real - time runway detection in satellite images using multi-channel
PCNN. In: 2014 9th Conference on Industrial Electronics and Applications (ICIEA), pp. 253–
257. IEEE (2014)
 Identifying Down Syndrome Cases by Combined
Use of Face Recognition Methods

Vlad Dima1, Anca Ignat1 ✉ , and Cristina Rusu2

( )

1
Faculty of Computer Science, “Alexandru Ioan Cuza” University of Iași,
str. General Berthelot nr. 16, 700483 Iasi, Romania
{vlad.dima,ancai}@info.uaic.ro
2
“Grigore T. Popa” University of Medicine and Pharmacy,
str. Universității nr. 16, 700115 Iasi, Romania
[email protected]

Abstract. Down syndrome is a chromosomal disorder, people affected by this

disease having very specific facial characteristics. In this paper we adapt some
well-known face recognition methods (Local Binary Patterns, Eigenfaces) and
test their ability to distinguish between a Down syndrome face and a normal one
in digital images. For classification kNN (k Nearest Neighbor) and Support Vector
Machines (SVM) with polynomial kernel of third degree and Radial Basis Func‐
tion (RBF) are employed. We test the methods using FERET, CAS-PEAL, LFW,
AT&T for normal face images and a collection of Down faces gathered from the
Internet. Accuracy, precision, recall and specificity are computed in order to
evaluate the classification results.

Keywords: Down syndrome · PCA · Eigenfaces · Local Binary Patterns ·

Discrete Wavelet Transform · kNN · SVM

1 Introduction

Down syndrome is the most common chromosomal abnormality defined by typical facial
appearance, short fingers/toes and moderate/severe intellectual disability. These features
are frequently associated with cardio-vascular and gastro-intestinal defects. The most
characteristic facial traits include: round and flat face, upslanting palpebral features
(mongoloid eyes), epichantal fold (skin fold in the inner eyelid angle), Brushfield spots
(spotted iris), small nose with flattened root, small mouth (usually open, with protruding
tongue) and small, low set ears [1, 2].
Down syndrome identification in digital images of faces can be useful in computer-
aided diagnosis, human–computer interaction systems or demographic studies.
Down syndrome persons have very distinctive facial traits, that is why it is easy for
humans to distinguish between a Down syndrome person and a healthy person just by
looking at their faces. In this paper we investigate how effective are some computer
vision techniques for face recognition, to make the same decision: Down/non Down
face.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_35
 Identifying Down Syndrome Cases by Combined Use 473

The subject of face recognition is one of the major research direction in image
processing and computer vision. Some of the main directions of research in this field
are described in [3, 4]. The problem of Down syndrome detection was also approached.
In [5] the authors studied the problem of automatic identification of other (than Down)
genetically disordered patients. In [6, 7] they use for Down syndrome identification a
Hierarchical Constrained Local Model with Independent Component Analysis first for
detecting facial landmarks and then to extract features around these points Local Binary
Patterns (LBP) are employed. They used SVMs (both linear and with Radial Basis
Function kernel), kNN, random forest and Linear Discriminant Analysis as classifiers.
The tests were made using images of young children (50 Down/50 healthy) of different
ethnicity, taken under variable illumination, but in a controlled environment. The best
result was an accuracy of 95.6% with the SVM-RBF. In [8] they also extract information
about fiducial face points and then use contourlets and LBP to extract also textural
features around the landmarks. The tests were made using children images, making two
type of evaluations: one with 24 normal patients and 24 with mixed genetic syndromes
(81.3% accuracy in detecting Down syndrome) and the second with 24 normal faces and
24 Down faces (97.9% accuracy). In [9] the authors worked with 1–12 years old children
images (15 healthy, 15 with Down syndrome), performing a standardization step before
extracting features using Gabor wavelets and employed kNN and linear SVM’s for
classification, the best result being an accuracy of 97.34%.
In our approach we used “real-life” images of persons with Down syndrome, of
different ages and tested them using normal faces of mixed ethnicity, or using only
Chinese. We tested LBP, a mixture of Discrete Wavelet Transform (DWT) with LBP,
and Eigenfaces with Principal Components projection features using kNN, polynomial
and RBF SVMs. For Down syndrome we selected 50 images from the Internet and for
normal faces we used classical face recognition datasets (FERET [10], CAS-PEAL [11],
LFW [12]). The influence of noise (inherent to real-life images) was also studied.

2 Face Characterization Features

There are many ways to extract features for face recognition. Among them are the
Eigenfaces method [13] and Local Binary Patterns (LBP) [14, 15]. In the Eigenfaces
algorithm, first the face images are vectorized and the mean value of these images is
subtracted. A matrix is formed with these vectors as columns. Then the eigenvalues and
eigenvectors of the covariance matrix are computed. The eigenvectors are arranged by
sorting the corresponding eigenvalues in descending order. The eigenvectors are then
employed to project new images on the linear subspace generated by them, thus forming
a feature vector.
Local Binary Patterns [14] is a technique initially developed for texture characteri‐
zation but can also be applied for face recognition [15]. A binary pattern for each pixel
in a grayscale image is computed by comparing the intensities of its 8 neighbors with
the intensity of the considered pixel. If the neighbor’s intensity is bigger, a 1 is put in
the binary pattern associated with the pixel, otherwise a 0. Thus, for each pixel an 8-bit
pattern is computed and usually is converted to a 0–255 decimal value. One obtains a
474 Vlad Dima et al.

matrix of binary patterns associated with the original image. In order to extract LBP
features, the image is divided in blocks and histograms of the binary patterns are
computed for each subimage. The feature vector is formed by concatenating these histo‐
grams. It was observed that among the 256 binary patterns that can be associated to a
pixel, those that have only at most two 0–1 or 1–0 transitions are more important. These
patterns are called uniform and there are 58 such patterns. If one computes histograms
counting only the uniform patterns and the non-uniform ones are all counted in one bin,
the size of histograms reduces from 256 to 59 without losing precisions in results, and
thus the feature vectors are 4 times shorter.
The Wavelet Transform [16] was a solution to the shortcoming of the fact that the
Fourier Transform coefficients provided no local information on frequencies. In 1988
Mallat and Meyer introduced the wavelet multiresolution concept, then Daubechies gave
an algorithm for computing orthogonal wavelets with compact support, and Mallat
deduced the fast wavelet transform, all these steps made the wavelet transform a very
useful tool for signal processing. At one level of Discrete Wavelet (DW) decomposition,
4 filtered images are computed, half the size of the processed image: HH (high-high
filtering both rows and columns of the image), HL (highpass filter is applied to rows and
a lowpass one to columns), LH (low-high filtering rows/columns) and LL (lowpass
filtering both rows and columns).
We use these methods to build feature vectors for Down syndrome identification in
the following ways:
1. The uniform binary patterns are computed for an image, then the image is divided
in 4 × 4 or 8 × 8 blocks, for each block the histogram of the uniform pattern is
computed. The feature vector is formed by concatenating these 16 or 64 histograms.
In Fig. 1 are some examples of images and their uniform binary pattern corre‐
spondent.

Fig. 1. Examples of face images from FERET, CAS-PEAL, LFW, and with Down syndrome
and their corresponding uniform LBP images

2. First apply a 2 levels DW decomposition, for the LL subimage from the second level
of wavelet decomposition we build the feature vector as described in 1 (using 8 × 8
blocks).
 Identifying Down Syndrome Cases by Combined Use 475

3. We first computed the eigenfaces for the AT&T ([17], Fig. 2) face database, then
we projected the image on the most significant of these eigenfaces. The weights
obtained from the projection step form the feature vector.

Fig. 2. Eigenfaces from AT&T dataset

4. For the AT&T faces, the 8 × 8 uniform LBP feature vectors are computed; for these
feature vectors the principal components are calculated; for an image to be classified
the final feature vector is obtained by projecting the 8 × 8 uniform LBP feature vector
of the image on some of the most important principal components.
All the feature vectors were normalized: for the LBP ones, the histogram for each
block was divided by the city-block norm, and for the projection weights we used the
Euclidean norm for normalization.
The LPB method is known to be very sensitive to noise, that is the reason of consid‐
ering applying first a DWT before extracting features with LBP method or projecting
the face images on a set of eigenfaces or principal components.
As classifiers, the Support Vector Machines (SVM) with 3-rd degree polynomial
kernel, or Radial Basis Function (RBF) kernel and kNN with k = 3 and Euclidean norm,
were employed.
For testing the feature vectors, face images form FERET [10, 18], CAS-PEAL [11],
and LFW [12] were employed. We selected randomly, from the Internet 50 images of
faces for persons with Down syndrome.
In Fig. 3 are some examples of images we used in the tests.
476 Vlad Dima et al.

(a) – FERET

(b) - AT&T

(c) – CAS-PEAL

(d) - LFW

(e) - Down Syndrome Faces

Fig. 3. Samples of face images from FERET, AT&T, CAS- PEAL, LFW, Down syndrome
datasets

We evaluated the classification results using the following measures:

TP + TN
Accuracy = (1)
TP + FP + TN + FN
TP
Precision = (2)
TP + FP
TP
Recall = (3)
TP + FN
TN
Specificity = (4)
TN + FP
 Identifying Down Syndrome Cases by Combined Use 477

where TP, TN, FP, FN stands for “true positive”, “true negative”, “false positive” “false
negative”. For example, “true positive” is the number of Down faces identified as having
Down syndrome, “false positive” the number of non-Down faces identified by the clas‐
sifier as having Down syndrome. The accuracy is a measure of the overall success of
the classification process, it is the proportion of the correctly identified images from all
images. The recall (also called sensitivity) and specificity measure the proportion of the
correctly identified Down/non-Down images from all Down/non-Down images, respec‐
tively. The precision is the ratio of the correctly identified Down images from all those
images identified as Down.

3 Results

Our numerical tests were performed using MATLAB and OpenCV in Java. The face
databases that we worked with have the following characteristics. The AT&T [17] data‐
base has 400 images of size 92 × 112, for 40 distinct persons, for each person 10 images
with different facial expressions, under different light conditions and with or without
glasses. The FERET (Face REcognition Technology, [10, 18]) have 14126 images for
1199 persons, with different facial expressions and positions, mixed ethnicity, taken in
controlled environment. The size of images in FERET is 256 × 384. The CAS-PEAL
database is a dataset of Chinese persons consisting of 30900 images for 1040 subjects,
taken under varying pose, expression, accessory and lighting (PEAL) conditions, the
size of images is 360 × 480. The LFW (Labeled Faces in the Wild) is a collection of
13000 face images collected from the Internet (such that Viola and Jones [19] face
detector worked on them) for 5479 individuals, the size of the images is 250 × 250. Our
criterion for selecting the images (Down or non-Down) was also the Viola and Jones
face detector working on the images.
All the images from the AT&T dataset were employed for building 10304 (92*112)
eigenfaces or 3776 principal components. In the projection computations we only used
the most relevant 1500 of these eigenfaces and 125 principal components. From FERET
database 900 frontal images were selected, from CAS-PEAL 300 (frontal and normal)
and from LFW 70. The Viola and Jones face detector was applied to all the selected
images and then feature vectors were computed. Because of the eigenfaces projection
method 3. described in the previous section, all the faces were resized to 92 × 112, the
common size of the AT&T images.
We performed several tests on these databases. The tests were carried out by using
60 (for FERET or CAS-PEAL) or 70 (for LFW) non-Down images and the 50 Down
faces. Thus, for FERET 15 tests were made, for CAS-PEAL, 5 tests and 1 test for LFW
images. As cross validation we used “leave-one-out”. In the tables with results, LBP
(4 × 4 or 8 × 8) means that the uniform binary patterns were used to extract features for
blocks of size 4 × 4 or 8 × 8, respectively. EP stands for Eigenfaces Projection method.
LBP(PCA) is the 8 × 8 uniform LBP feature extractor followed by the projection on the
principal components on AT&T as described in Sect. 2, reducing the dimension of the
feature vectors to 125. For the DWT we used 22 tap biorthogonal Daubechies filters.
478 Vlad Dima et al.

The feature vectors have the following dimensions: 944 for LBP (4 × 4), 3776 for
LBP (8 × 8) and DWT + LBP, 1500 for EP, 125 for PCA (Principal Component Anal‐
ysis) reduced dimension LBP (8 × 8).
In Tables 1, 2 and 3 are the (mean) results for the (1)–(4) measures of classification.
We expected low values for the Recall parameter due to the fact that the selected Down
faces images are low quality and are not standardized in any way. The good classification
results we get for FERET and CAS-PEAL are due to this difference in image quality.
The images in FERET and CAS-PEAL were taken in controlled conditions, with the
same camera. The 50 Down selected images are rather noisy, and had different sizes,
the re-dimensioning process enhanced the ‘defects’ in these images. We suspect that in
the case of CAS-PEAL classification, in fact a Chinese/non-Chinese identification
process was performed rather than Down/non-Down one. The results we obtain for the
LFW database give better information on the capacity of distinguishing between Down
and non-Down face images of the considered features, because these images were
selected in the same way.

Table 1. Mean values for Accuracy (A), Precision (P), Recall (R), Specificity (S) for 15 tests
with 60 FERET faces and 50 Down faces
SVM (polynomial, 3-rd degree) 3NN
A P R S A P R S
LBP 98.18% 99.3%1 96.66% 99.44% 83.45% 99.24% 64.13% 99.55%
(4 × 4)
DWT + 98.72% 99.33% 97.86% 99.44% 71.33% 100% 36.93% 100%
LBP
EP 96.84% 98.22% 94.8% 98.55% 94.66% 98.27% 89.86% 98.66%
LBP 95.65% 99.49% 90.94% 99.58% 91.51% 99.56% 81.76% 99.56%
(8 × 8)
LBP 98.54% 99.19% 97.6% 99.33% 77.34% 100% 50.47% 100%
(PCA)

Table 2. Mean values for Accuracy (A), Precision (P), Recall (R), Specificity (S) for 5 tests with
60 CAS-PEAL faces and 50 Down faces
SVM (polynomial, 3-rd degree) 3NN
A P R S A P R S
LBP 100% 100% 100% 100% 85.81% 100% 68.8% 100%
(4 × 4)
DWT + 99.63% 100% 99.2% 100% 90.72% 100% 79.6% 100%
LBP
EP 97.63% 98.74% 96% 99% 93.27% 97.77% 87.2% 98.33%
LBP 98.72% 99.60% 97.6% 99.66% 89.45% 100% 76.8% 100%
(8 × 8)
LBP 99.63% 99.60% 99.60% 99.66% 84.90% 100% 66.8% 100%
(PCA)
 Identifying Down Syndrome Cases by Combined Use 479

Table 3. Accuracy (A), Precision (P), Recall (R), Specificity (S) for 70 LFW faces and 50 Down
faces
SVM (polynomial, 3-rd degree) 3NN
A P R S A P R S
LBP 90.83% 95.34% 82% 97.14% 76.66% 92.3% 48% 97.14%
(4 × 4)
DWT + 93.33% 97.72% 86% 98.57% 72.5% 100% 34% 100%
LBP
EP 90.83% 91.48% 86% 94.28% 80.83% 75.47% 80% 81.42%
LBP 89.16% 97.43% 76% 98.57% 82.5% 93.93% 62% 97.14%
(8 × 8)
LBP 91.66% 97.61% 82% 98.57% 77.5% 100% 46% 100%
(PCA)

We studied the effect of noise on the considered features by performing the following
test: we trained the classifier using the original images and the same images where
Gaussian noise with zero mean and 0.0005 variance was added. We tested using images
with added Gaussian noise with zero mean and 0.003 variance. The tests were made in
a “leave-one-out” manner, by training the classifier without the corresponding image
we test. We only tested the FERET and CAS-PEAL images using the SVM (for 3NN
the results are unusable). The results are in Tables 4 and 5.

Table 4. Mean values for Accuracy (A), Precision (P), Recall (R), Specificity (S) for 15 tests
with 60 FERET faces and 50 Down faces—Gaussian noise with zero mean and 0.0005 variance
for training and 0.003 variance for test
SVM (polynomial, 3-rd degree)
A P R S
LBP (4 × 4) 84.06% 97.62% 67.06% 98.22%
DWT + LBP 89.81% 99.3% 78.26% 99.44%
EP 96.84% 98.22% 94.8% 98.55%
LBP (PCA) 98.54% 99.19% 97.6% 99.33%

Table 5. Mean values for Accuracy (A), Precision (P), Recall (R), Specificity (S) for 5 tests with
60 CAS-PEAL faces and 50 Down faces—Gaussian noise with zero mean and 0.0005 variance
for training and 0.003 variance for test
SVM (polynomial, 3-rd degree)
A P R S
LBP (4 × 4) 70.36% 60.98% 100% 45.66%
DWT + LBP 82.18% 72.16% 100% 67.33%
EP 97.63% 98.74% 96% 99%
LBP (PCA) 99.09% 99.6% 98.4% 99.66%
480 Vlad Dima et al.

We can observe from these tables that the LBP does not perform well in noisy situa‐
tions (which is well known), the DWT manages to deal better with the noise, but by far,
the best results are obtained by using the projection methods on the principal components
or eigenfaces of the AT&T images (the results are almost unchanged from the normal,
no noise case). The recall becomes lower for the FERET database, the noise enhancing
the ‘defects’ of the Down images and decreasing their identification rate. For the CAS-
PEAL situation, the specificity decreases showing that, in fact, the features are classi‐
fying Chinese/non-Chinese.
We used DWT or eigenfaces projection in order to deal with noise. We made some
tests by first applying median filters of different sizes (3, 5, 7, 9) in order to reduce the
noise and then computing LBP (4 × 4) feature vectors. Tests were made for Down and
LFW images. The results can be seen in Fig. 4. The results tend to decrease when the
size of the filter increases, as expected, with the exception of the 7 sized median filter
and all the results are worse than those obtained by using the DWT and eigenfaces
projection methods.

Fig. 4. Accuracy and Recall for LBP median filtered 70 LFW, 50 Down images

We used a 3-rd degree polynomial kernel for SVM because it needed no parameter
tuning as the RBF kernel needs and the results are good. Of course, in some cases we
obtain better results with RBF SVMs. For example, for the LFW test without noise we
get for the LBP case a better accuracy result, 91.66%, with lower precision (93.47%)
and higher recall (86%). For the DWT + LBP and SVM-RBF case we obtain overall
improved results with accuracy of 95.83%, precision 97.87%, 92% recall and 98.57%
specificity. For the EP case we didn’t manage to find the appropriate parameter for the
RBF kernel in order to improve the polynomial SVM result.
 Identifying Down Syndrome Cases by Combined Use 481

4 Conclusions

In the present paper we tested four feature extraction methods for Down syndrome
identification in digital images. The first is the classical Local Binary Patterns method,
the second is a combination of Discrete Wavelet Transform and LBP, the third method
is an eigenfaces projection method using AT&T as eigenfaces provider and the fourth
is a projection method on principal components of the LBP features of AT&T. All these
methods give good results in distinguishing between faces with Down syndrome and
healthy ones, but in the presence of noise projection methods are the best (especially the
projection on the principal components of the LBP of AT&T faces) followed by the
combination DWT and LBP. The feature vectors were tested using only frontal face
images from the well-known face recognition datasets FERET, CAS-PEAL, LFW and
some Down images collected from Internet, with SVM and 3NN as classifiers.
For the projection methods, we need to further test the influence of the dataset from
which the eigenfaces or principal components are built. Other visually visible genetic
disorders will be approached in a future work.

References

1. Jones, K.L., Jones, M.C., del Campo, M.: Smith’s Recognizable Patterns of Human
Malformation, 7th edn, pp. 7–13. Elsevier Saunders, Philadelphia (2013)
2. Cassidy, S., Allanson, J.: Management of Genetic Syndromes, 3rd edn, pp. 309–335. Wiley
Blackwell, New York (2010)
3. Zhao, W., Chellappa, R.P., Phillips, P.J., Rosenfeld, A.: Face recognition: a literature survey.
ACM Comput. Surv. 35(4), 399–458 (2003)
4. Jafri, R., Arabnia, H.R.: A survey of face recognition techniques. JIPS 5(2), 41–68 (2009)
5. Boehringer, S., Guenther, M., Sinigerova, S., Wurtz, R.P., Horsthemke, B., Wieczorek, D.:
Automated syndrome detection in a set of clinical facial photographs. Am. J. Med. Genet. A.
155A(9), 2161–2169 (2011)
6. Zhao, Q., Okada, K., Rosenbaum, K., Zand, D.J., Sze, R., Summar, M., Linguraru, M.G.,
Mori, K., Sakuma, I., Sato, Y., Barillot, C., Navab, N.: Hierarchical constrained local model
using ICA and its application to down syndrome detection. In: Proceedings of MICCAI 2013,
Part II: 16th Int. Conf. Nagoya, Japan, September 22–26, pp. 222–229 (2013)
7. Zhao, Q., Rosenbaum, K., Sze, R., Zand, D.J., Summar, M., Linguraru, M.G.: Automated
Down syndrome detection using facial photographs. In: 2013 35th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Osaka, pp.
3670–3673 (2013)
8. Zhao, Q., Rosenbaum, K., Okada, K., Zand, D.J., Sze, R., Summar, M., Linguraru, M.G.:
Down syndrome detection from facial photographs using machine learning techniques. In:
SPIE Medical Imaging, Medical Imaging 2013: Computer-Aided Diagnosis, pp. 867003–
867003, 28 February 2013
9. Saraydemir, Ş., Taşpınar, N., Eroğul, O., Kayserili, H., Dinçkan, N.: Down syndrome
diagnosis based on gabor wavelet transform. J. Med. Syst. 36(5), 3205–3213 (2012)
10. Phillips, P.J., Wechsler, H., Huang, J., Rauss, P.: The FERET database and evaluation
procedure for face recognition algorithms. Image Vis. Comput. J. 16(5), 295–306 (1998)
482 Vlad Dima et al.

11. Gao, W., Cao, B., Shan, S., Chen, X., Zhou, D., Zhang, X., Zhao, D.: The CAS-PEAL large-
scale chinese face database and baseline evaluations. IEEE Trans. Syst. Man. Cybern. (Part
A) 38(1), 149–161 (2008)
12. Huang, G.B., Ramesh, M., Berg, T., Learned-Miller, E.: Labeled faces in the wild: a database
for studying face recognition in unconstrained environments. University of Massachusetts,
Amherst, Technical Report 07–49, October 2007
13. Turk, M.A., Pentland, A.P.: Face recognition using eigenfaces. Computer vision and pattern
recognition. In: Proceedings CVPR 1991, pp. 586–591 (1991)
14. Ojala, T., Pietikainen, M., Maenpaa, T.: Multiresolution gray scale and rotation invariant
texture classification with local binary patterns. IEEE Trans. Pattern Anal. Mach. Intell. 24(7),
971–987 (2002)
15. Ahonen, T., Hadid, A., Pietikainen, M.: Face description with local binary patterns:
application to face recognition. IEEE Trans. Pattern Anal. Mach. Intell. 28(12), 2037–2041
(2006)
16. Mallat, S.: A Wavelet Tour of Signal Processing. Academic Press, Cambridge (2008)
17. Samaria, F.S., Harter, A.C.: Parameterisation of a stochastic model for human face
identification. In: Proceedings of the Second IEEE Workshop on Applications of Computer
Vision, 1994, Sarasota, FL, pp. 138–142 (1994)
18. http://www.nist.gov/itl/iad/ig/feret.cfm
19. Viola, P., Jones, M.J.: Rapid object detection using a boosted cascade of simple features. In:
Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern
Recognition, vol. 1, pp. 511–518 (2001)
 Iris Recognition Using Color and Texture
Features

Ioan Păvăloi1 and Anca Ignat2(&)

1
Institute of Computer Science, Romanian Academy,
str. Codrescu nr. 2, 700481 Iaşi, Romania
[email protected]
2
Faculty of Computer Science, “Alexandru Ioan Cuza” University of Iași,
str. General Berthelot nr. 16, 700483 Iaşi, Romania
[email protected]

Abstract. We approach the problem of iris recognition by combining both

color and texture information. For color features, a well-known global color
criterion was extended and for texture we adapted three classical methods:
Gray-Level Co-occurrence Matrix (GLCM), Local Binary Patterns (LBP), and
Gabor filters. As classifiers we employed k-NN (with Euclidean, Manhattan,
Canberra and some variations of these distances), Support Vector Machines
(SVM) and a two steps recognition process based on k-NN. We tested the
methods using RGB, HSV, and LAB color spaces on two irises datasets
(UBIRIS and UPOL). We get better results by combining color and texture
features than by only considering color and texture separately.

Keywords: Color indexing
Color spaces
Texture
Gray-Level

Co-occurrence Matrix
Local Binary Patterns
Gabor filters
K-NN

SVM
Iris recognition

1 Introduction

Because of the increased need for safer systems, the researchers took into consideration
for security applications the iris recognition. This is also stimulated by the decreasing
cost of the eye-scanning technology. Sure, there are other forms of biometric (from
biological measurement) identification, as fingerprints, but iris recognition is becoming
a popular one, considering the trend to incorporate scanners on smartphones and
wearables in the near future.
Iris recognition can facilitate access for persons unable to engage in the standard
mechanical transactions of access because it allows hands-free, automatic, rapid and
reliable identification of persons [1, 2]. We have to consider also that iris patterns have,
in their structure, a high degree of randomness, and this is what makes them unique.
The identification process in iris recognition, consist in two steps. Firstly, one or
more detailed images of the eye are acquired (using sophisticated high resolution
digital camera using visible or infrared wavelengths). Secondly, using a specialized
computer program, the subject’s iris pattern is compared with images already stored in
a database (these comparisons contain computations that involve a great number of

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_36
484 I. Păvăloi and A. Ignat

images per second with a level of precision comparable to digital fingerprint).

Nowadays, Iris recognition applications are also available for the iPhone and other
smartphones.
The iris is the part of the eye between the sclera and the pupil. By using its fine
muscular structure, the iris adjusts the amount of light entering the eye. The pigments
from the iris control its color. Sometimes the color of the iris changes, for example this
fact can appear because of changes in the environment (lighting condition, nearby
colors), but in this study we are interested in testing the proposed methods for color
image recognition.
Color is an important feature which provides useful information to be used in image
recognition. A simple image indexing method (usually named color indexing) that has
some interesting properties (being insensitive to geometric transformation) based on
color histograms was proposed by Swain [3]. This method is robust against variations
such as shifting in viewing position or partial occlusion. Detailed information on the
regularity patterns which can appear in different areas of an image, can be obtained by
analyzing the textural content of the image.
In this paper, we’ll use in iris identification an extension of the known Swain color
indexing criterion. Using features related to the color average of the color pixels in the
image, the performance of image recognition is improved comparing with original
Swain criterion. More, we’ll use also three methods for texture feature extraction. In
our experiments we are using two public and well-known iris databases UBIRIS [4, 5]
and UPOL [6–8].
In Sect. 2 we present the two iris databases (UBIRIS and UPOL) on which we were
worked on. Section 3 introduces the three color spaces (RGB, HSV and LAB) on
which tests were made. Section 4 mainly focuses on feature extraction, using both
color and texture to compute feature vectors and introduces the classifiers and measures
used to test the efficiency of the feature extraction process. Experimental results and
comparisons between the two criteria and discussion are presented in Sect. 5,
respectively. Section 6 summarizes this study and emphasizes some possible contin-
uation directions of this work.

2 Iris Databases

UBIRIS.v1 database [4, 5] is a collection of 1877 images that are collected from 241
persons (see Fig. 1). This database incorporates images with several noise factors, in
contrast with the existing databases, that are noise free, to enable the evaluation of the
robustness of iris recognition methods. The dataset we are using in this study is part of
the UBIRISv1 database. The original images were segmented automatically according
to the details in [9]. Each segmented iris image is RGB coded and it is rectangular with
the same dimension of 360  100 pixels. In this dataset, there are 5 iris images for
 Iris Recognition Using Color and Texture Features 485

Fig. 1. Examples of iris images from UBIRIS database

each of the 241 persons, so a total of 1205 images. We are considering two images as
being similar if they belong to the same person.
The UPOL iris database [6–8], contains 384 iris images belonging to 64 persons,
three images for the left eye and three for the right eye (see Fig. 2). Unlike the images
in UBIRIS, this dataset is of high quality. The images are unsegmented, coded in PNG
format, with 24 bits of RGB color for each pixel are of good resolution (576  768).
The images in this database are of very good quality, few images having defects like
eyelids, eyelashes, shadows or specular reflections. The images were made at the
Palacký University, Olomouc (UPOL), the Czech Republic. The irises were scanned by
TOPCON TRC50IA optical device connected with SONY DXC-950P 3CCD camera.
486 I. Păvăloi and A. Ignat

Fig. 2. Examples of iris images from UPOL database

3 Color Spaces

This study is made for three color spaces [10], RGB, HSV (which is also called HSB)
and LAB. RGB is the most commonly employed color space in computer technology.
RGB color space is difficult to understand intuitively, being device-dependent and the
representation of the colors in this space is adapted for monitors and cameras. RGB is a
linear color space, based on a set of primary colors, and describe colors in terms of a
weighed sum of the intensities of these primary colors (red, green and blue). Other
color spaces are usually preferred for color representation in user interfaces.
In contrast with the RGB color space, LAB (derived from the XYZ standard color
system) is a uniform color space. The international lighting commission CIE recom-
mended the L * a * b * color space, (and also the L * u * v * color space) as an
approximation of uniform color spaces (usually abbreviated as CIELAB or LAB and
respectively CIELUV). LAB colors are absolute and the LAB color space is the only
way to communicate different colors across different devices. When any image is
converted to LAB color space it will look exactly the same, because the LAB color
space contains all possible colors.
HSV and HSZ (named perception-based color spaces), based intuitively on human
color perception, are of interest for the fields of computer vision and computer
graphics. In the HSV color space, Hue (H), saturation (S), and brightness value (V or
B) are used as coordinate axes. Unlike the additive components in the RGB color
 Iris Recognition Using Color and Texture Features 487

space, a color can be more easily described intuitively by values for hue, color satu-
ration, and intensity.

4 Feature Extraction, Measures, Classifiers

4.1 Color Feature Extraction

In the Swain and Ballard color indexing scheme [3], the range of the data for the color
information of an image is split (divided) into “n” equally sized bins. The feature vector
X = (x1, x2,…, xm), m = n  n  n is obtained counting for each bin the number of
points from the data set (here, the colors of the pixels in an image) that fall into it. For
example, in the RGB color space, if we divide each channel R, G, and B into 3 equal
intervals with a length of 64 (0–63, 64–127, …, 192–255), we’ll get a vector with
3  3  3 = 27 features. To compare two images, they should have the same number
of pixels, otherwise normalization will be required. Moreover, they should be parti-
tioned with the same number of bins. The feature vectors for a division is formed from
the set of all vectors generated for all images in the dataset.
In the extension criterion we proposed in [11], we also took into consideration the
color average for all pixels in each bin. As a result, we get other three vectors, in an
“m” dimensional space, m = n  n  n:

ðC1 ; C2 ; C3 . . .; Cm Þ
ðL1 ; L2 ; L3 ; . . .; Lm Þ
ðR1 ; R2 ; R3 ; . . .; Rm Þ

where (Ci, Li, Ri) is the color average for pixels in bin number “i”.
For the extension criterion, a feature vector for an image is obtained concatenating
the features obtained as for Swain and Ballard color indexing scheme with the three 3n-
dimensional vectors described above:

X ¼ ðx1 ; x2 ; . . .; xm ; C1 ; C2 ; C3 ; . . .; Cm ; L1 ; L2 ; L3 ; . . .; Lm ; R1 ; R2 ; R3 ; . . .; Rm Þ:

4.2 Texture Feature Extraction

In image processing texture is a concept that has no precise definition. It encompasses
information about some pattern of spatial regularity. The methods dealing with texture
typically use three approaches: structural, spectral, and statistical. Three of the most
utilized techniques for textural information extraction are the Grey Level
Co-occurrence Matrices (GLCM), Local Binary Patterns (LBP) and Gabor filters.
A GLCM [12] computed for a greyscale image with 256 possible intensities is a
256  256 matrix G ¼ ðgij Þ where:
488 I. Păvăloi and A. Ignat

gij ¼ the number of pairs of pixels that have intensities ði; jÞ:

The manner in which the pixels are paired can lead to different results. In this paper
we considered pairing each pixel with its right neighbor. In features computations one
usually employ instead of gij the normalized values:

pij ¼ gij =total number of pairs of pixels:

As texture features the following measures of a GLCM are computed:

XX
Contrast ¼ ði  jÞ2 pij ;
XX
Energy ¼ p2ij ;
XX pij
Homogeneity ¼ ;
1 þ ji  jj
XX
Entropy ¼  pij log2 pij :

Local Binary Patterns (LBP) is a method developed by Ojala et al. in [13] for
texture characterization. The LBP method has many other applications beside texture,
such as face analysis, biomedical image processing, remote sensing, and so on. For
each pixel in a greyscale image a binary pattern is computed by comparing the pixel’s
intensity with the intensities of its 8 neighbors. The logical result of these comparisons
is stored in a 8-bit array, and then transformed into a base 10 number. An image of
local binary patterns is thus computed from the original one. For feature extraction the
LBP image is divided in cells, for each cell a histogram of binary patterns is calculated.
Concatenating these histograms, one obtains the features vector for texture
characterization.
Gabor wavelets were introduced by Hungarian mathematician Dennis Gabor in
1946 and then adapted for digital images by Daugman [14] who used them to develop
his famous iris code. A Gabor filter is a combination of a complex sinusoid and a
Gaussian kernel and can be expressed in the following way:
 
2pur
Gðu; v; r; k; h; cÞ ¼ gðu; v; r; cÞexp i
k

where
 
u2r þ c2 v2r
gðu; v; r; cÞ ¼ exp  ;
2r2
      
ur cos h sin h u u cos h þ v sin h
¼ ¼ :
vr  sin h cos h v u sin h þ v cos h
 Iris Recognition Using Color and Texture Features 489

We used c ¼ 0:5; and k ¼ 0:5622 r: The result of applying a Gabor filter is a

matrix with complex values. In our computation, the amplitude of the complex filtered
image was employed. In order to build a feature vector, we applied Gabor filters with
different values for the orientation h and frequency r. From each filtered image we kept
only the energy, standard deviation and entropy.
We applied these texture feature extraction methods in the following way. In all the
cases we applied the texture algorithm to each color component and concatenated the
feature vectors. For the GLCM, we divided the image in 2  2 blocks, for each block
we computed the GLC matrix and extract the four above mentioned measures. Thus we
obtain o feature vector with 3*2*2*4 = 48 components. For the second type of fea-
tures, we computed the LBP image and then divided it in 2  2 subimages, con-
catenating the block histograms with bins of size 4. The size of the final feature vector
was 3*2*2*64 = 768. For Gabor wavelets we divided the image in 2  2, then for
each block we used h 2 f45 ; 90 ; 135 g; r 2 f5; 10; 15g, thus obtaining 9 filtered
subimages, for which we computed the three values mentioned above. The feature
vector was 3*2*2*9*3 = 324-dimensional.

4.3 Classifiers and Measures

In this study we have chosen to use two well-known discriminative classifiers, k-NN
(k Nearest Neighbor) and Support Vector Machines (SVM) [15, 16]. We also choose to
use a validation technique named leave-one-out cross validation (LOO-CV) to estimate
the error rate of the k-NN classifier. This validation technique is a special case of
cross-validation, in which the test subset is made of a single vector, all the other vectors
being used as a training set [16]. This step is repeated m times (m being the size of the
dataset), each vector of the dataset being used for testing exactly once. The total error is
obtained by averaging the errors for all the runs. In our study we have selected three
measures (used as distances in k-NN) to compare feature vectors, Manhattan (1),
Euclidean (2) and Canberra (3) distances:

X
m
dM ðx; yÞ ¼ j xi  yi j ð1Þ
i¼1
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X m
dE ðx; yÞ ¼ ð yi  xi Þ 2 ð2Þ
i¼1

X
m
j xi  yi j
dC ðx; yÞ ¼ ð3Þ
i¼1
j xi þ yi j

For feature vectors generated using the extended criterion we defined also other two
measures, a “Manhattan” like measure dPM and an “Euclidean” like measure dPE. We
also tested some Canberra like measures, but the results were not very good.
Considering two vectors:
490 I. Păvăloi and A. Ignat

X ¼ ðx1 ; x2 ; . . .; xm ; C11 ; C12 ; C13 ; . . .; C1m ; L11 ; L12 ; L13 ; . . .; L1m ; R11 ; R12 ; R13 ; . . .; R1m Þ

and

Y ¼ ðy1 ; y2 ; . . .; ym ; C21 ; C22 ; C23 ; . . .; C2m ; L21 ; L22 ; L23 ; . . .; L2m ; R21 ; R22 ; R23 ; . . .; R2m Þ

we defined:

X
m
dPM ðx; yÞ ¼ jxi  yi j  ðjC1i  C2i j þ jL1i  L2i j þ jR1i  R2i jÞ ð4Þ
i¼1
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X m  
dPE ðx; yÞ ¼ ðyi  xi Þ2  jC1i  C2i j2 þ jL1i  L2i j2 þ jR1i  R2i j2 ð5Þ
i¼1

For SVM, we generate randomly sets of tests containing exactly one iris for each
person, the remaining ones form the dataset being used for training. The total error is
obtained by averaging the errors for a number of runs.

5 Results

We considered in this study three kinds of sets of feature vectors. The first one is
obtained using the Swain extension color criterion we proposed in [11], for different
divisions, from 3 to 11. The second one contains texture feature vectors, obtained using
GCLM, Gabor, and LBP. The third one is obtained concatenating color features with
texture features.
We’ll analyze the results obtained on UBIRIS and UPOL databases using two
discriminative classifiers, k-NN and SVM. The tests were made for three color spaces,
LAB, RGB and HSV. In the next tables L stand for LAB, H for HSV, R for RGB color
space and, C for Canberra, E for Euclidian and M for Manhattan distance. So, for
example HE expresses HSV color space and Euclidian distance. In the tests done we
used k = 1 and k = 3 for the k-NN classifier, the best results being obtained for k = 1,
because there are only 5 or 6 iris images for each person.

5.1 Color Features Results

The recognition results (in percents) on color feature vectors sets obtained using k-NN
for k = 1 and using the two defined distances (4) and (5) on the three color spaces we
have considered, are presented in Table 1 for the UBIRIS dataset and in Table 2 for the
UPOL dataset.
In next tables first letter indicates the color space and the other two the distance. For
example, L-PE stands for LAB color space and PE for dPE distance, H-PM for HSV
color space and dPM distance.
The recognition results for color feature vectors sets obtained using SVM are
presented in Table 3 for the UBIRIS dataset and in Table 4 for the UPOL dataset.
 Iris Recognition Using Color and Texture Features 491

Table 1. 1-NN results on UBIRIS database using extended Swain criterion

L-PE L-PM H-PE H-PM R-PE R-PM
333 55.68% 56.60% 93.44% 94.69% 87.39% 89.21%
444 86.31% 88.05% 94.36% 95.27% 92.61% 93.69%
555 88.30% 90.04% 94.61% 95.60% 93.53% 94.44%
666 89.29% 89.63% 94.36% 95.77% 93.53% 94.69%
777 93.11% 94.27% 94.02% 95.44% 94.36% 95.10%
888 91.87% 92.61% 94.02% 95.02% 95.27% 95.60%
999 93.94% 94.85% 93.53% 94.85% 94.52% 95.44%
10  10  10 93.28% 94.52% 93.61% 94.94% 93.94% 95.52%
11  11  11 93.44% 94.85% 92.53% 94.44% 94.44% 95.77%

Table 2. 1-NN results on UPOL database using extended Swain criterion

L-PE L-PM H-PE H-PM R-PE R-PM
333 71.61% 75.52% 89.84% 94.53% 89.32% 91.15%
444 85.94% 90.36% 95.05% 96.61% 90.89% 94.01%
555 91.15% 94.79% 93.75% 96.35% 92.45% 95.05%
666 91.41% 93.23% 95.57% 97.40% 92.97% 95.57%
777 94.53% 96.09% 94.79% 97.40% 95.05% 96.61%
888 95.31% 96.09% 96.61% 97.40% 95.31% 97.66%
999 95.31% 96.88% 95.31% 97.92% 95.83% 98.18%
10  10  10 95.83% 96.61% 95.83% 97.66% 96.09% 97.92%
11  11  11 93.75% 96.35% 95.31% 98.18% 94.79% 97.40%

Table 3. SVM results on UBIRIS database using color features (in percents)
333 444 555 666 777 888 999 10  10  10 11  11  11
LAB 57.68 69.12 71.95 76.35 90.62 83.65 94.02 88.46 94.27
HSV 88.13 92.12 92.78 94.61 94.27 95.02 94.69 93.28 95.35
RGB 72.28 84.48 73.03 91.87 93.94 94.19 94.43 95.18 95.85

Table 4. SVM results on UPOL database using color features

333 444 555 666 777 888 999 10  10  10 11  11  11
LAB 54.17 65.63 93.23 75.00 94.53 92.19 95.05 93.75 96.61
HSV 85.94 93.75 93.75 95.83 96.35 96.09 96.35 96.61 97.40
RGB 83.85 89.06 93.48 95.05 95.57 95.57 95.85 96.09 97.39

The best results obtained using 1-NN on color features is 95.77% for UBIRIS and
98.18% for UPOL and 95.85% for UBIRIS and 97.40% for UPOL using SVM.

5.2 Results Obtained Using Texture Feature Vectors

The 1-NN recognition results (in percents) on texture feature vectors sets obtained on
LAB, HSV and RGB color spaces using three distances, Manhattan, Euclidian and
492 I. Păvăloi and A. Ignat

Fig. 3. 1-NN results on texture feature vectors for UBIRIS

Fig. 4. 1-NN results on texture feature vectors for UPOL

Canberra are presented in Fig. 3 for the UBIRIS dataset and in Fig. 4 for the UPOL
dataset.
The 1-NN results obtained on UPOL using texture features are inferior to those
obtained using color features. For UBIRIS, the best result, 96.18%, is obtained using
Gabor filters on HSV color space with Manhattan distance. More, the results on HSV
color space are superior to those obtained on LAB and RGB color spaces.
 Iris Recognition Using Color and Texture Features 493

Table 5. SVM results on texture feature vectors for UBIRIS and UPOL
UBIRIS UPOL
LAB HSV RGB LAB HSV RGB
GCLM 64.23% 84.32% 83.73% 67.71% 78.39% 65.36%
Gabor 75.85% 87.39% 80.17% 72.14% 84.11% 73.96%
LBP 62.16% 67.14% 70.04% 79.43% 79.69% 80.47%

The recognition results (in percents) on texture feature vectors obtained for UBIRIS
and UPOL using SVM are presented in Table 5 and can be observed that are very poor.
These results are inferior to those obtained using 1-NN and we can conclude that SVM
on texture feature is not a practical solution for iris recognition.

5.3 Results Obtained with Color and Texture Feature Vectors

5.3.1 Concatenating Color and Texture Features
Applying 1-NN on the sets of feature vectors obtained by concatenating color with
GCLM texture features we get the results in Table 6.

Table 6. 1-NN results on UBIRIS database using color and GCLM texture features
LC LE LM HC HE HM RC RE RM
333 75.60 81.66 83.24 93.61 95.35 96.51 92.70 93.61 94.52
444 74.69 83.32 80.33 95.19 95.52 96.68 93.86 94.02 95.02
555 83.24 87.05 90.54 94.69 96.10 96.76 94.52 93.44 95.44
666 77.43 82.66 85.81 95.27 94.36 96.85 94.77 93.44 96.18
777 90.12 91.78 93.44 94.94 93.20 96.43 94.94 92.03 96.18
888 81.74 84.90 88.13 94.44 92.28 95.77 95.10 91.45 95.77
999 89.29 88.22 92.28 94.69 91.87 95.27 96.02 91.37 96.10
10  10  10 88.13 88.63 91.78 94.52 90.54 94.77 95.85 90.46 95.68
11  11  11 88.63 86.64 91.62 94.44 89.46 94.19 95.68 89.29 95.68

1-NN results obtained using color and texture feature generated using Gabor and
LBP are inferior to those obtained using GCLM.
The best 1-NN results for UPOL database are obtained using color and GCLM
texture feature, for HSV color space and using Manhattan distance (Fig. 5), which is
similar with the results for UBIRIS database.
Using SVM classifier we get the results from Tables 7 and 8 for UBIRIS and
UPOL databases, respectively.
We note that the results obtained with SVM on the two databases using color and
texture features are very poor.

5.3.2 Two Steps Recognition

In the first step, using 1-NN classifier on texture features vectors, a set of seven best
matching irises is selected. In the second step, 1-NN is applied on color features for the
494 I. Păvăloi and A. Ignat

Fig. 5. 1-NN results using Manhattan distance on UPOL database using color and GCLM
texture features

Table 7. SVM results on UBIRIS database using color and texture features
HSV RGB
Gabor GLCM LBP Gabor GLCM LBP
3  3  3 86.39 84.07 52.28 80.50 83.49 61.99
4  4  4 85.23 79.50 39.59 79.92 81.99 53.86

Table 8. SVM results on UPOL database using color and texture features
HSV RGB
Gabor GLCM LBP Gabor GLCM LBP
3  3  3 86.20 86.20 86.20 76.56 77.34 80.47
4  4  4 88.80 86.98 88.80 78.65 74.74 80.99

Table 9. 2-Steps results on UBIRIS database using color and texture features and Manhattan
distance
HSV RGB LAB
Gabor GLCM LBP Gabor GLCM LBP Gabor GLCM LBP
3  3  3 95.60 95.85 96.93 92.28 92.95 96.68 93.11 74.19 95.60
4  4  4 95.77 96.35 96.93 92.61 94.02 96.35 93.20 74.85 94.52
 Iris Recognition Using Color and Texture Features 495

Table 10. 2-Steps results on UPOL database using color and GCLM texture features
Manhattan Euclidean
HSV RGB LAB HSV RGB LAB
3  3  3 97.92 97.92 97.92 97.40 97.40 96.61
4  4  4 98.28 97.14 97.40 96.88 96.88 96.61

irises previously selected. The recognition results obtained (in percents) for UBIRIS
and UPOL are presented in Tables 9 and 10.
As can be observed from these tables, in 2-step recognition we get the best results
on UBIRIS and UPOL datasets, 96.93 and 98.28%, respectively. These are the best
results obtained in this study.

6 Conclusions

In this study that we’ll further use to develop a CBIR system, we have considered
combining an extension of the well-known Swain and Ballard simple global color
criterion and three classical methods for texture features. We generated three kinds of
sets of feature vectors and we tested them for image recognition on two iris databases.
The tests were made on three color spaces using two discriminative classifiers k-NN
and SVM. We also defined two new distances and we are using them with three other
well-known distances (Canberra, Euclidean and Manhattan) in the k-NN classifier.
Studying and comparing the results obtained using these three types of feature
vectors we notice that the best results were obtained using k-NN and the HSV color
space. In many of the studied cases the results obtained using HSV and RGB color
spaces were very close, but the best results are obtained in the HSV color space. Sure,
the LAB color space can be also a choice but with 2-steps recognition method. The
results obtained using the SVM classifier are good for color features but are poor on
texture features for all three considered color spaces.
For almost all the tests made using 1-NN, the selection of Manhattan or
Manhattan-like distance is a better choice that than Euclidean or Canberra distances.
Moreover, considering the calculus complexity, Manhattan distance is the better
choice. We have done tests using other distances like Chebyshev and Minkowski, but
we didn’t get better results.
The best results obtained on UBIRIS database using the 2-step recognition method
reached a maximum percentage of recognition of 96.93% superior to the results
obtained on the same dataset using optimal bin boundaries that reached a maximum
percentage of recognition of 96.76% [17, 18]. The UBIRIS image dataset we used in
our tests was obtained using an automatic segmentation technique, hence the impos-
sibility to attain 100% recognition.
For UPOL, working on segmented images and with optimal bin boundaries for the
LAB color space using a 5  5  5 division the maximum percentage reached was of
98.70% in [19].
496 I. Păvăloi and A. Ignat

In this study, using the 2-step recognition we’ve get a maximum of 98.28% but it’s
important to notice that this result was obtained on unsegmented images (the previous
results in [19] was obtained on a manually segmented version of the dataset).

References
1. Proenca, H.: Iris recognition in the visible wavelength. In: Burge, M.J., Bowyer, K.W. (eds.)
Handbook of Iris Recognition, pp. 151–171. Springer, New York (2013)
2. Tankasala, S., Gottemukkula, V., Saripalle, S., Nalamati, V., Derakhshani, R., Pasula, R.,
Ross, A.: A video based hyper focal imaging method for iris recognition in the visible
spectrum. In: IEEE International Conference on Technologies for Homeland Security (HST),
Waltham, MA, pp. 214–219, November 2012
3. Swain, M.J., Ballard, D.H.: Color indexing. Int. J. Comput. Vision 7(1), 11–32 (1991)
4. Proença, H., Alexandre, L.: UBIRIS: a noisy iris image database. In: Roli, F., Vitulano, S.
(eds.) Image Analysis and Processing—ICIAP 2005, vol. 3617, pp. 970–977. Springer,
Berlin (2005)
5. Proença, H., Alexandre, L.A.: UBIRIS - noisy visible wavelength iris image databases
(2004). http://iris.di.ubi.pt/ubiris1.html
6. Dobeš, M., Martinek, J., Skoupil, D., Dobešová, Z., Pospíšil, J.: Human eye localization
using the modified Hough transform. Optik 117(10), 468–473 (2006)
7. Dobeš, M., Machala, L., Tichavský, P., Pospíšil, J.: Human eye iris recognition using the
mutual information. Optik 115(9), 399–405 (2004)
8. Dobeš, M., Machala, L.: Iris Database. http://www.inf.upol.cz/iris/
9. Radu, P., Sirlantzis, K., Howells, W.G.J., Hoque, S., Deravi, F.: A versatile iris segmentation
algorithm. In: Presented at the BIOSIG 2011, Darmstadt, Germany (2011)
10. Fairchild, M.D.: Color and Image Appearance Models Color Appearance Models, pp. 340–
385. Wiley, London (2005)
11. Păvăloi, I., Ignat, A.: A simple global color criterion for image retrieval. In: Accepted at the
13-th International Conference on development and application Systems, Suceava, Romania,
May 19–21 (2016)
12. Gonzales, R.C., Woods, R.E.: Digital Image Processing, 3rd edn., pp. 852–858. Prentice
Hall, Upper Saddle River (2008)
13. Ojala, T., Pietikainen, M., Maenpaa, T.: Multiresolution gray scale and rotation invariant
texture classification with local binary patterns. IEEE Trans. Pattern Anal. Mach. Intell. 24
(7), 971–987 (2002)
14. Daugman, J.G.: Uncertainty relation for resolution in space, spatial frequency, and
orientation optimized by two-dimensional visual cortical filters. J. Opt. Soc. Am. 2(7), 1160–
1169 (1985)
15. Chang, C.-C., Lin, C.-J.: Libsvm: a library for support vector machines, version 2.3 (2001).
http://www.csie.ntu.edu.tw/*cjlin/libsvm
16. Dougherty, G.: Pattern Recognition and Classification - An Introduction. Springer Science
and Business Media, New York (2013)
17. Ciobanu, A., Costin, M., Barbu, T.: Image categorization based on computationally
economic LAB colour features. In: Balaș, V.E., et al. (eds.) Soft Computing Applications.
Advances in Intelligent Systems and Computing, vol. 195, pp. 585–593. Springer, Berlin
(2013)
 Iris Recognition Using Color and Texture Features 497

18. Păvăloi, I., Ciobanu, A., Luca, M.: Iris classification using WinICC and LAB color features.
In: 4th IEEE International Conference on e-Health and Bioengineering - EHB 2013, Iași,
România, pp. 21–23, November 2013
19. Ciobanu, A., Luca, M., Păvăloi, I., Barbu, T.: Iris identification based on optimized lab
histograms applied to iris partitions. Bul. Inst. Politeh. Iaşi, Tomul LX (LXIV),
Fasc. 1 (2014)
 Retinal Vessel Classification Technique

Florin Rotaru(&), Silviu-Ioan Bejinariu, Cristina Diana Niţă,

Ramona Luca, Mihaela Luca, and Adrian Ciobanu

Institute of Computer Science, Romanian Academy Iaşi Branch, Iaşi, Romania

{florin.rotaru,silviu.bejinariu,cristina.nita,ramona.
luca,mihaela.luca,adrian.ciobanu}@iit.
academiaromana-is.ro

Abstract. A retinal vessel classification procedure is proposed. From the image

of thinned vessel network, landmarks are extracted and classified as branching,
crossover and end points. Then a vascular graph is generated. Using a stratified
graph edge labeling procedure the artery/vein map is built. In a first step the
graph branches near the optic disc are localized and classified. Each label is
propagated along the most significant segments linked to initial vessels. The
next labeling phase aims the not processed branches starting from already
classified vessels. Only branches and edges at crossings are labeled. Finally,
using the current labels set, the uncertain cases are solved.

Keywords: Retinal images
Vessel graph
Graph edge labelling

1 Introduction

The retinal vessel morphology can be modified by systemic diseases as diabetic

retinopathy, glaucoma, atherosclerosis or macular degeneration. The arteries and veins
are affected in different ways in such cases. For instance calibers of arteries decrease
relatively to the calibers of veins. In other words the ratio between arteries and veins
diameters (AVR) decreases. Another feature differently modified for arteries and veins
is vessel tortuosity. To compute indicators as AVR or vessel tortuosity a classification
of retinal vessels is mandatory. The vessel labeling phase is preceded by a complex task
of retinal image processing and analysis: optic disc detection and modeling; vessel
network segmentation; network landmarks (branching, crossover, end points) extrac-
tion and analysis; vessel graph generation and filtering. The main rules used by final
labeling process are: a) the 3D vascular structure is cycle free, so each binary tree
included in the 2D vessel graph represents a vessel; b) in the vicinity of the optic disc
area vessels at crossings are of different type (an artery never crosses an artery, the
same for veins).
Most of the proposed methods on the matter of artery/vein (A/V) classification
localize in the vessel skeletons image branching, cross-over and end vessel points.
Using these landmarks a graph reproducing the vessel network is synthesized. To
eliminate the vessel network thinning errors and to simplify the graph different graph
processing techniques have been proposed. In the final phase the A/V classification is
performed by various graph analysis techniques. Subsequently some proposed methods

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_37
 Retinal Vessel Classification Technique 499

following the above strategy are reviewed. In [1, 2] the graph is filtered to solve the
errors eventually generated by vessel network thinning: splitting a real node in two
(appears at the original large vessel intersections); missing links and false links. The
graph analysis decides the type of each intersection point represented by the pertaining
graph node. The 2nd degree nodes might be connecting or meeting points (where two
vessels touch each other at their ends). Vessel crossings or tangential touches or
splitting a vessel in two branches produce 3rd degree nodes. Also, two vessels crossing
each other or tangential touching each other (meeting point) or vessel splitting in three
branches (bifurcation point) generate 4th degree nodes. The 5th degree nodes appear
where more than two vessels are crossing. The nodes are classified as meeting or
branch or cross point depending on the orientation and width of the incident vessel
edges. According with the node type the incident edges are labeled as follows: (a) the
same label as the current one at branching; (b) different label at crossings for two vessel
segments and the same label for the fourth. In the last step, based on a combination of
structural information and vessel intensity information, the domain of initial vessel
labels is restricted to a binary domain (artery/vein). A comparison of results obtained
using a supervised classification technique versus an unsupervised alternative is pro-
vided in [2]. Another artery/vein labeling approach is proposed in [3]. By Dijkstra’s
shortest-path algorithm in the vascular graph vessel subtrees are identified and different
sub-graphs are extracted. Each subgraph is labeled as artery or vein using a fuzzy
C-mean clustering algorithm. In [4] the arteries and veins are classified employing a
technique to extract the optimal forest in the graph given a set of constraints.
A heuristically AC-3 algorithm to differentiate in the graph the two vessel classes is
presented in [5]. Using some manually-labeled starting vessel segments the algorithm
propagates the vessel labels as either artery or vein throughout the vascular tree. The
A/V classification method proposed in [6] relies on estimating the overall vascular
topology in order to finally label each individual vessel.
Some other proposed methods do not use the whole vessel graph to perform the
A/V classification. For instance in [7], first the vessel network is extracted by a tracking
technique. Then the optic disc is localized and its diameter approximated. A concentric
zone around the optic disc is divided into four regions according to the position of the
vessel network two main arcs. In every quadrant for each vessel a feature vector is
synthesized. Finally, using a fuzzy clustering procedure each feature vector is allocated
to one of the two classes. The same technique to separate the two vessel classes is
employed by the author of [8] for AVR estimation. In [9] for each main vessel in a ring
area around optic disc a six features vector has been constructed. The feature vector
includes color intensities of the pixels representing the thinned vessel, the contrast
differences in a 5  5 area of the vessel and in a 10  10 window outside the vessel.
Finally a linear discriminant labels each feature vector as vein or artery. As in [7], the
authors of [10] identify the optic disc and its approximate diameter. Then they divide
the area around optic disc into four quadrants and for each quadrant vessel synthesize a
feature vector. The feature vectors are labeled using a Gaussian Mixture Model and
Expectation-Maximization unsupervised classifier.
Our first attempt to label the retinal vessels was presented in [11]. We proposed a
stratified strategy based on main vessels identification located in the optic disc vicinity.
Each main starting vessel is classified as vein or artery using a supervised classification
500 F. Rotaru et al.

technique. The established label is then propagated gradually along the most significant
descendant vessels connected to the main branch. This way the eventually uncertain
labeling cases appear in the last stages of the classification process for less significant
vessels. To our knowledge this method leads to more accurate results than other
proposed approaches on the A/V classification matter. In the next sections a new A/V
classification approach with better performances versus the first version of the algo-
rithm is presented. As in [11] to test the new method the manually segmented images of
public DRIVE database were used as input. We resume from [11] all the image analysis
steps: (a) the optic disc detection and modeling - the optic disc circle center and the
circle radium r are used to crop on the segmented blood network image a circular ring
bounded by two concentric circles centered in the optic disc center; the radii of the two
circles are of 2r and 12r sizes [4, 12]; (b) in the selected area of segmented image:
network landmark points detection and processing [13]; (c) new vessel’s graph analysis
and classification.
The remainder of the paper is organized as follows. The preliminary image pro-
cessing analysis to extract information used by the final classification is resumed from
[11] in Sect. 2. The new vessel graph filtering approach is presented in Sect. 3. Sec-
tion 4 details the new A/V classification procedure. The work is concluded in Sect. 5.

2 Preliminary Retinal Image Processing Steps

The working image synthesis is illustrated in Fig. 1:

1. Using one of the techniques proposed in [12] the optic disc is detected and modeled.
The radius r of the circle model is used as a parameter to extract the working area in
the thinned image of the vessel network, Fig. 1 - upper and bottom rows;
2. From the DRIVE vessel images, figured by the second row of Fig. 1, a centerline
image, denoted TV, is generated by thinning, Fig. 1 - third row;
3. A working image, noted RV, is built by deleting in TV image any vessel pixel
outside the outer circle or inside the inner circle, Fig. 1 - bottom row.
Additionally two feature images are built, to be used in landmark analysis and final
A/V classification:
• A vessel width image, denoted WV, is computed using the same technique
described in [11];
• For each vessel pixel in TV image the intensity variance in the green channel of the
original color image is computed, using a 14  14 window. The result is stored in a
floating point array, noted VV.

3 Vessel Graph Processing

All nodes in the RV image are labeled as multiple and eventually essential, cross points
(with more than three neighbors vessel points) according to the procedure presented in
[4]. The multiple points set is ordered function of the point distance to optic disc center.
 Retinal Vessel Classification Technique 501

Fig. 1. Working area on four fundus images of drive database (area outside of inner circle and
inside of outer circle) - upper row; manually segmented images - second row; thinning results -
third row (TV); working images for landmarks detection (RV) - bottom row.

Then the method proposed in [13] is employed to process multiple points in order to
generate the significant crossover and bifurcation structure set. This procedure consists
of the following main steps: (a) for each essential point the multiple neighbors are
replaced by the nearest 2-neighbor vessel points on the branch; (b) if two nearby
essential points share a starting point the two essential pixels become equivalent and
only one would be validated, inheriting the starting points of the canceled essential
pixel. It may happen that the valid essential point changes its status from branching to
crossing pixel; (c) the essential points sharing all neighbors (starting points) with other
essential pixels are canceled; (d) the branching points with a very short branch are
canceled. For each significant point structure there are filled multiple point coordinates
and the coordinates of the nearest 2-neighbor vessel pixels considered as starting points
for vessel tracking in the RV image. For each starting point a tracking procedure is
initiated. The tracking ends when another starting point or an end point (1 vessel
neighbor) is met. Every tracking process generates a line structure which includes: line
end points; essential points associated to end points; line length; tracked vessel pixel
coordinates set; a binary attribute to mark the line as allocated (processed) during the
502 F. Rotaru et al.

graph structures generation process, presented in the next paragraph. Also, the line
structure includes a graph index, to indicate the graph membership. Along the tracking
for each vessel pixel the following parameters are grabbed from images generated in
the process described in Sect. 2: vessel width (WV image); intensity variance (VV
image); gray level of the correspondent pixel in green channel of the original RGB
image. Subsequently for each line end the average of the above parameters are com-
puted: (a) if the vessel length is greater than a threshold experimentally established, at
each end a fraction of the stored vessel points are used to compute the parameter
average; (b) for short segments the same procedure is employed using for both ends the
whole set of vessel pixels. Then at each vessel end a slope is computed in the same
manner as for width, intensity and variance intensity averages but this time using the
coordinates of a fraction or of the whole set of stored vessel points. After the whole line
structure set initialization each line structure is completed with pointers to the incident
lines at its ends. For each incident line at crossings the associated line index structure
field is completed if the two lines have almost the same width and local orientation.
Then the line structures are filtered according with the following rules:
1. At crossings of four or more lines delete weak lines (not aligned and very thin). For
every deleted line update also the essential end points structure in order to rebuild
the line structures;
2. At four lines crossings with two non-associated lines detect the weakest line from
the two non-aligned lines. If one of the two aligned lines is short and the other end
is a branch essential point it is possible to split the crossing in two branches. The
closest new branch to the old branch of the short line might hide a crossing,
(Fig. 2);
3. Find out eventually hidden crossings generated by short lines bounded by
3-connectivity essential points, (Fig. 3);
4. Solve tangency for two cases:
4:1. The tangency part is a short line with branching ends, [11];

Fig. 2. Illustration of step 2 of the line structure filtering scheme. The vessel intersection is
marked on the original image area – left figure. The crossing lines are figured on the right in the
corresponding zoomed area of working image RV.
 Retinal Vessel Classification Technique 503

Fig. 3. Case 3 illustration of the line structure filtering scheme. The vessel intersection is
marked on the original image area – left figure. The crossing lines are figured on the right in the
intersection zoomed area of the working image RV.

4:2. The tangency part is a short line with a branch at one end and four or more lines
at the other end. At the end with more than four lines the largest two are
selected and the problem is reduced to the precedent one, (Fig. 4).
5. Using a weaker condition find out the hidden crossings given by short lines
bounded by 3-connectivity essential points, (Fig. 5).
The filtering procedure leads to completion of the association indices at line
crossings, the same indices mentioned for direct line association at line structure
synthesis - steps 1, 2, 3. In the same manner other two association line indices are filled:
for tangency cases (4.1 and 4.2), and an index for association at crossings but for a
more relaxed condition on line slopes - step 5. Except steps 1, 3 and 4.1, presented in
[11], all the other procedures are new filtering approaches. The steps 2, 3 (resumed
from [11]), 4.2 and 5 are illustrated in Figs. 2, 3, 4 and 5. The case 2, depicted by
Fig. 2, is generated by a fake crossing of lines l1, l2, lt and lc. As shown in the left
image of the original crossing area line lt is a very thin one and the real crossing, hidden
by line lc, is of vessel of lines l1 and l3 with the vessel of lines l2 and l4.
Consequently the initial line association at the fake crossing is cancelled. The new
one is set properly in the structures of lines l1 and l3 respectively l2 and l4, given the fact
that the association conditions are fulfilled.
As mentioned in [11] for each line incident at crossings the associated line index
structure field is completed if the two lines have almost the same width and local
orientation. For the case depicted by Fig. 3 vessel of lines l1 and l3 is crossed by vessel
of lines l2 and l4. In [11] there were treated tangency cases for short lines bounded by
3rd degree nodes with the width significantly greater than the width of the incident
lines, case 4.1 from above filtering scheme. In this paper we consider also the case of
short lines thicker than each incident line bounded by a 3rd degree node and a node with
a degree greater than 3. This situation is illustrated in the right image of Fig. 4 where
the line lc is incident with lines l12 and l22 at one end and with lx1, lx2, lx3, l11 and l21
lines at the other end. As mentioned above from the end with more than four lines the
largest two are selected and the problem is reduced to the precedent one. In our case,
from the node with lx1, lx2, lx3, l11 and l21 lines the two widest lines are selected,
respectively l11 and l21. If the association conditions for tangency of vessel l11 and l12
504 F. Rotaru et al.

Fig. 4. Case 4.2 illustration of the line structure filtering scheme. The vessel intersection is
marked on the original image area – left figure. The crossing lines are figured on the right in the
intersection zoomed area of the working image RV.

Fig. 5. Case 5 illustration of the line structure filtering scheme. The vessel intersection is
marked on the original image area – left figure. The crossing lines are figured on the right in the
intersection zoomed area of the working image RV.

with vessel l21 and l22 there are fulfilled the rest of incident lines respectively lx1, lx2, lx3
are placed in hold to be labeled in the last step of the classification process.
The last case of the line structures filtering is illustrated by Fig. 5. Using a weaker
orientation line condition, there are identified the hidden crossings given by short lines
bounded by 3-connectivity essential points. This time the lines l1 and l3, respectively l2
and l4 would finally be associated. At the second step of the labeling procedure the
pairs of weak associated lines are placed in hold to be processed in the last phase of the
process.
The final vessel segment graphs of the four working images set are illustrated on
the bottom row of Fig. 6.

4 Stratified Vessel Classification Procedure

Starting from the line structures sets, illustrated by the bottom row of Fig. 6, graph
structure sets are synthesized. Using the pointers to the incident lines at current line end
by a tracking procedure the set of connected lines is marked with the same graph index.
 Retinal Vessel Classification Technique 505

Fig. 6. Vessel segment graphs (bottom row) of the four working image set (first row).

Also to avoid tracking cycles each line is marked as processed, once tracked. Then, for
each line set with the same graph index there are identified the main branches starting
from the inner circle of the working area. Each starting main branch is tracked in order
to extract the rest of the vessel and to fill the graph structure. The following tracking
rules are applied:
1. If at the current end of the current line only two child lines are linked choose the
most significant vessel, with the largest width;
2. If at the current end are linked two branches having almost the same width choose
the one with the closest slope to the mother line slope;
3. If at the current end are linked more than two branches choose as next branch the
associated or weak associated line if there is an association at crossing;
4. If at crossings the current line is associated and the other pair is not, place the
not-associated line in the holding list, to be labeled in the last step of the classifi-
cation process;
5. If at crossings the current line has the tangency attribute set on, the next branch
would be the associated segment;
6. If at crossings the current line has the complex tangency attribute set on, the next
branch would be the associated segment and the other non-associated line indices
would be placed in the holding list;
7. If all the lines at crossings are not associated choose the most significant vessel if
there is one. Put the rest of lines in the holding list of unsolved pairs: each not
chosen line makes pair with the current line;
8. If all the lines at crossings are not associated and the vessel width range is not large
choose to associate to the current line the line with the most closed slope to the
current one. Put the rest of incident lines in the list of unsolved pairs.
506 F. Rotaru et al.

The graph structure and its completion are depicted below.

 Retinal Vessel Classification Technique 507

The graph structure generation pseudocode:

In order to initiate the vessel labeling process only the branches with a free end
located in the optic disc vicinity are considered. For each starting main branch as many
vessel segments are considered as their length sum is above a predefined threshold. For
each main starting vessel of the graph a feature vector has been synthesized, with the
following components: width, variance intensity, mean gray level. Previously more
groups of statistical classifiers, differentiated by the vessel width range, have been
trained using a feature vector set synthesized from images of public DRIVE database.
508 F. Rotaru et al.

Fig. 7. The result of main branch selection (bottom row) for the four images set (first row).

The current feature vector is assigned by the designated supervised classifier to one of
the two classes (“vein”, “artery”). In cases of uncertainty, when the vessel classification
probability is low, the labeling is interactively accomplished. Once the starting vessel
of a main branch has been labeled the established label is propagated along all seg-
ments of the branch. The result of main branch selection and labeling is illustrated by
the bottom row of Fig. 7 (blue vessels are arteries, red vessels are veins).
The next labeling phase aims the branches starting from already classified main
branches. This time the tracking rules are:
1. If at the current end of the current line are linked only two child lines choose the
most significant vessel, with the largest width. The other child is placed in the stack
to be labeled later;
2. If at the current end are linked two branches having almost the same width choose
the one with the closest slope to the mother line slope. The other child is placed in
the stack to be labeled later;
3. If at the current end are linked more than two branches choose as next branch the
associated line if there is an association at crossing.
4. If at crossings the current line is associated and the other pair is not, place the
not-associated line in the holding list, to be labeled in the last step of the classifi-
cation process.
5. If at crossings the current line is weak associated place the pair in the holding list, to
be labeled in the last step of the classification process. If the stack is not empty go
on labeling the first line in the stack. Otherwise get the next graph labeled branch.
6. If at crossings the current line has the simple tangency attribute set on the next
branch would be the associated segment.
7. If at crossings the current line has the complex tangency attribute set on place the pair in
holding list together with the other non-associated lines indices. If the stack is not empty
go on labeling the first line in the stack. Otherwise get the next graph labeled branch.
8. If all the lines at crossings are not associated place all children in the holding list of
unsolved pairs: each child makes pair with the current line. If the stack is not empty
go on labeling the first line in the stack. Otherwise get the next graph labeled branch.
 Retinal Vessel Classification Technique 509

The new labeling step pseudocode:

510 F. Rotaru et al.

The result of labeling the vessels branching from main branches is depicted in
Fig. 8.
In the last labeling phase the list of unlabeled vessels completed during the first two
steps would be processed. This time due to the fact that most part of vessel network is
classified, so more information is available, the probability for a correct classification of
the initially uncertain cases increases.
The following routine was implemented to complete the graph starting from a
specific vessel and label. The completion might fail if during labelling already clas-
sified branches are met (Fig. 9).

Fig. 8. The result of the second step of labeling process - bottom row; the result of main branch
labeling is illustrated again on the upper row.
Retinal Vessel Classification Technique 511
512 F. Rotaru et al.

As we already mentioned to finalise the classification process all kind of uncertain

cases are considered: non-associated lines at crossings collected in the first two steps
and weak associated lines or complex tangency cases, grabbed in the second step. The
algorithm pseudocode:
 Retinal Vessel Classification Technique 513

Fig. 9. The final result of the labeling process - bottom row; the result of previous labeling step
is illustrated again on the upper row.

5 Conclusions
P cls
wi li
To assess the A/V classification the indicator R ¼ P init has been used, where lcls
i is
w j lj
the length of the correctly classified ith vessel; linit
j is the length of jth vessel from initial
working image, depicted by the bottom row of Fig. 1. The vessel weight wi is defined
widthi widthmin ., where width is the ith vessel width, width
as: wi ¼ 0:1 þ 0:9 width i min is
max widthmin
the minimum vessel width of the network and widthmax is the maximum one.
Tests have been done on 20 manually segmented images from DRIVE database.
The step of generating Line structures was flawless for the entire set. The classification
rate was R = 0.95, better than the labeling rate R = 0.93 of the first version of the
algorithm presented in [11].
Assuming a correct initial labeling of main vessels near the optic disc the stratified
classification method leads to more accurate results than other proposed approaches on
the A/V classification matter. As mentioned in [11] the eventually unsolved cases
appear in the last stages of the labeling process for less significant vessels. However,
the new algorithm performs a more reliable classification in certain cases and a more
detailed selection of uncertain cases in its first two phases. After that more information
is available so the probability for a correct classification of uncertain cases increases.
The reported work is a part of a larger project that will be completed later with the
vessel network segmentation, to fulfill an artery/vein classification process.
The classification procedure was implemented in C++ and OpenCV library was
used for image manipulation and some processing functions.

Acknowledgments. The work was done as part of research collaboration with University of
Medicine and Pharmacy “Gr. T. Popa” Iaşi to analyse retinal images for early prevention of
ophthalmic diseases.
514 F. Rotaru et al.

References
1. Dashtbozorg, B., Mendonça, A.M., Campilho, A.: An automatic graph-based approach for
artery/vein classification in retinal images. IEEE Trans. Image Process. 23(3), 1073–1083
(2014)
2. Dashtbozorg, B.: Advanced image analysis for the assessment of retinal vascular changes.
Ph.D. thesis, University of Porto (2015)
3. Joshi, V.S., Reinhardt, J.M., Garvin, M.K., Abramoff, M.D.: Automated method for
identification and artery-venous classification of vessel trees in retinal vessel networks.
PLoS ONE 9(2), e88061 (2014)
4. Lau, Q.P., Lee, M.L., Hsu, W., Wong, T.Y.: Simultaneously identifying all true vessels from
segmented retinal images. IEEE Trans. Biomed. Eng. 60(7), 1851–1858 (2013)
5. Rothaus, K., Jiang, X., Rhiem, P.: Separation of the retinal vascular graph in arteries and
veins based upon structural knowledge. Image Vis. Comput. 27, 864–875 (2009)
6. Estrada, R., Allingham, M.J., Mettu, P.S., Cousins, S.W., Tomasi, C., Farsiu, S.: Retinal
artery-vein classification via topology Estimation. IEEE Trans. Med. Imaging 34(12), 2518–
2534 (2015)
7. Grisan, E., Ruggeri, A.: A divide et impera strategy for automatic classification of retinal
vessels into arteries and veins. In: Proceedings of the 25th Annual International Conference
of the IEEE Engineering in Medicine and Biology Society, 2003, vol. 1, pp. 890–893 (2003)
8. De Luca, M.: New techniques for the processing and analysis of retinal images in diagnostic
ophthalmology. Ph.D. thesis, University of Padova (2008)
9. Muramatsu, C., Hatanaka, Y., Iwase, T., Hara, T., Fujita, H.: Automated selection of major
arteries and veins for measurement of arteriolar-to-venular diameter ratio on retinal fundus
images. Comput. Med. Imag. Graph. 35(6), 472–480 (2011)
10. Relan, D., MacGillivray, T., Ballerini, L., Trucco, E.: retinal vessel classification: sorting
arteries and veins. In: 2013 35th Annual International Conference of the IEEE Engineering
in Medicine and Biology Society (EMBC), pp. 7396–7399 (2013)
11. Rotaru, F., Bejinariu, S.I., Luca, R., Niţă, C.: Retinal vessel labeling method. In: Proceedings
5th IEEE International Conference on E-Health Bioengineering, Iaşi, pp. 19–21 (2015)
12. Rotaru, F., Bejinariu, S.I., Nita, C.D., Luca, R., Luca, M., Ignat, A.: Optic disc recognition
method for retinal images, soft computing applications volume 357 of the series Advances in
Intelligent Systems and Computing, pp. 875–889. Springer, Heidelberg (2015)
13. Rotaru, F., Bejinariu, S., Bulea, M., Nita, C., Luca, R.: Vascular bifurcations and crossovers
detection and processing in retinal images. In: Proceedings 12th International Symposium on
Signals, Circuits and Systems (ISSCS 2015), Iasi (2015)
Soft Computing Algorithms Applied in
Economy, Industry and Communication
Technology
 Reliable Metrics for Image LSB
Steganography on Mobile Platforms

Dominic Bucerzan and Crina Raţiu(&)

“Aurel Vlaicu” University of Arad, Arad, Romania

[email protected], [email protected]

Abstract. Mobile devices remain the primary means of communication using

standard calls or specialized applications. Technological development in this
field has an accelerated rate of growth. However it is neglected the security of
data used and processed by applications running on mobile devices. With all the
risks involved, mobility remains a top priority in the coming years. Protecting
private information from unauthorized access on smart mobile platforms is
essential. Also it is extremely important to find a way to secretly communicate
using these devices. A solution in achieving this goal is digital steganography.
In this paper we propose a scheme with most used metrics for image LSB
steganography solution that run Android operating system. To analyze and
establish the performances of this type of security solutions, we believe that it is
mandatory to define and to classify some characteristics and metrics that are
wide present in the literature. Therefore this is the main objective of this work.

Keywords: Android
LSB steganography
Steganography metrics

1 Introduction

Internet of Things (IoT) leaves behind M2 M communication (mobile - to - mobile).

Today any device can be connected to the Internet. IoT, evolved from collect, generate
and exchange information between interconnected devices to connecting in real time
people with accurate information, becoming an advanced forecasting tool. Figure 1
shows a prediction of mobile data traffic until 2020.

Prediction of Mobile Data Traffic

40
30.6
30
21.7
20 14.9
9.9
10 3.7
6.2

0
2015 2016 2017 2018 2019 2020

Exabytes per Month

Fig. 1. Global mobile data traffic growth (2015–2020) [7]

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_38
518 D. Bucerzan and C. Raţiu

Mobile devices remain the primary means of communication using standard calls or
specialized applications [1]. They support a wide range of complex software systems,
games, text editors, spreadsheets, web browser, electronic payments, reaching today to
perform tasks which until recently were specific for personal computers [2].
According to the studies presented in [3, 4, 5, 6], mobile devices are more vul-
nerable than desktop platforms:
• due to their mobility, mobile devices can be lost, stolen or exposed to the phe-
nomenon of browsing over the shoulder;
• most times the information stored on these devices is more valuable than their
hardware components;
• through Global Positioning System (GPS), the holder of the devices can be con-
stantly tracked, limiting the right to privacy; if the GPS module is disabled, this risk
is maintained because the operating system knows at all times at which antenna of
the provider was the device connected at last call;
• mobile devices have incorporated cameras and microphones which are controllable
via remote software by unauthorized users, even if apparently the device seem to be
in standby;
• diversity of manufacturers, of mobile devices and of dedicated operating systems
make difficult their efficiency and interoperability of security solutions.
Technological developments in this field has an accelerated rate of growth, but
neglected the security of data used and processed by applications running on mobile
devices. Figure 2 shows the evolution of malware for mobile devices.

10000000
8000000
6000000
4000000
2000000
0
2013 2014 2015 Q1 2015 Q2

Fig. 2. Evolution of malware for mobile devices [3]

Although the risks involved, mobility remains a top priority in the coming years for
social and business environment [3]. From this perspective we consider pertinent
studding and improving the performances of information security solutions for mobile
platforms.
Protecting private information from unauthorized access on smart mobile platforms
is essential. Also it is extremely important to find a way to secretly communicate using
these devices [8]. A solution in achieving this goal is offered by digital steganography.
 Reliable Metrics for Image LSB Steganography 519

2 Related Works

Not long ago in [9], the authors tried to state the fact that steganography can be
successfully implemented and used into next generation of mobile phones.
Today there are studies of performance analysis of steganography solution that runs
Android operating system. The tests performed states that steganography on mobile
platforms is a real fact.
Regarding Steganography solutions for traditional computers, there are well rep-
resented in the literature by a wide and diversified area of steganographic systems with
remarkable performances like in [10, 11].
Table 1 presents the steganography systems for mobile platforms that use images as
cover file, available today. In [12, 11] are described and analyzed some steganography
solutions for mobile platform that are present in the literature.

Table 1. Steganography applications available online [12, 11].

Operating Applications
system
Android CryptApp, PixelKnot, Secret Letter, Shadow, Da Vinci Secret, Camopic
Lite, Steganography Application, Crypsis, MobiStego, Stegais, Stego
Message, Limage, StegoLite, Secret Tidings, PtBox, Steganografia,
Incognito, Secure Message, VipSecret, Data Guardian
iOS iSteg, PicSecret, StegoSec, SpyPix, Hide It In, Steganographia, Pixogram,
Steganomatic, Stega, Concealment, Invisiletter, iClic Reader

There are some solutions for steganography on mobile platforms that run Android,
available on Google Play. These are discussed in [12, 11, 13, 14]. The disadvantages of
these applications include:
• low capacity of embedding data;
• their functionality on different devices;
• the imperceptibility of the embedded data;
• the fact that it is not recorded any scientific performance test regarding hardware
requirements and computational complexity;
• time or speed measurements are not available to establish computational perfor-
mances on mobile platforms.
In the literature there is little discussion about some wide recognized criteria of the
performances of a steganography system. An interesting work about the quality metrics
of a steganography solution is presented in [15].
In [16] the authors introduce another metric that measures image quality entitled
Universal Image Quality Index (UIQI). UIQI considers the human visual system
(HVS) characteristics to incorporate perceptual quality measures
Most authors whom propose a steganographic solution for image LSB steganog-
raphy are using benchmarking parameters like MSE (Mean Square Error), PSNR (Peak
Signal to Noise Ratio), SNR (Signal to Noise Ratio), image evaluation parameters and
520 D. Bucerzan and C. Raţiu

statistical steganalysis to evaluate the efficiency of their solution. This is the reason why
the main target of this work is to study the parameters that analyze and define the
efficiency of a steganographic system.

3 Proposed Solution

In this work we consider the images presented in Fig. 3 for testing and for establishing
the criteria that defines the efficiency of a steganography system. With red are marked
the pixels that are different from the two files. In gray are the pixels that are equal from
the two analyzed images.

a b c

d e f

Fig. 3. Images used for tests

Image 3a represents the original bmp file set as cover. The rest of the images from
Fig. 3 represent the differences from the original cover file and the cover file with secret
information embedded: 3b using 1 LSB, 3c using 2 LSB’s, 3d using 3 LSB’s, 3e using
4 LSB’s and 3f using an improved LSB solution.
The technique used for image 3f [9] is also using the classical LSB principle
optimized with a random selection function based on a secret key, that chooses the bits
from a byte of cover file that will carry the secret information. Also the applied
technique processes the entire cover file and introduces some random data to obfuscate
the length of the secret information.
In this work we used some open access specialized tools for testing the images
involved. To generate the images needed for visual analysis (see Figs. 3 and 4) we used
a free version of the tool called Guiffy Image Diff 11.4 [17]. For statistical analysis
(benchmarking and steganalysis) we used Steganography Studio [18].
 Reliable Metrics for Image LSB Steganography 521

a b c

d e f

Fig. 4. Comparison enlarge 200% of the tested images

The images for the analysis presented in this paper were obtained using smart
devices that run Android. Different versions of LSB steganography were applied using
as a cover bitmap images. The characteristics of the original cover used for testing are:
size – 3.81 MB; dimension – 1000  1333 pixels; bit depth – 24; type – bitmap
(.bmp). The resulted images obtained by embedding the secret information into the
cover file keep the same size, type, dimension and bit depth like the original cover.

3.1 Visual Analysis

Figure 4 (with images 4a, 4b, 4c, 4d, 4e, 4f, corresponding for the description of
images 3a, 3b, 3c, 3d, 3e, 3f, from image 3) shows the same part enlarge 200% of the
original cover file and the cover file with information embedded. The human visual
system cannot detect the differences from the original cover file (image 4a) and the rest
of the images.
Using specialized tools to compute the difference between the analyzed images (see
Fig. 3), an observer with naked eyes can easily determine the differences between the
images. Supposing that the original cover image is available, in case of classical LSB
steganography (see image 3b, 3c, 3d and 3e), one can lightly extract the length of the
secret information hidden in the cover file.
522 D. Bucerzan and C. Raţiu

Regarding the case of image 3f, one can extract the differences between the two
files; however it is a difficult process to determine which of this information is random
and which the secret information is.

3.2 Benchmarking Analysis

Benchmarking test assumes to have access to both original and stego-image (the cover
image with secret information embedded). This type of test analyses the quality
characteristics of the stego-image by using a comparison with the original cover file.
Tables 2 and 3 shows the values of the main parameters obtained during the benchmark
test for the different LSB approaches used in this study.

Table 2. Benchmarking values

Image 3 Average absolute difference Mean squared error Laplacian mean squared error
a 0.0 0.0 0.0
b 0.697362 1.395392 3.031763
c 0.881609 4.006714 7.355984
d 1.253006 11.870066 0.001980
e 1.919045 36.937597 0.006119
f−1 1.620848 4.030456 9.599944
f−2 1.755384 5.182150 0.001287

Table 3. Benchmarking values

Image 3 Peak signal to noise ratio Normalised cross-correlation
a 0.0 1.0000
b 56.226265 1.0000
c 51.645345 1.0000
d 46.928697 1.0000
e 41.998542 1.0000
f−1 51.619685 0.9999
f−2 50.528128 0.9999

Table 4. Example of how to determine the average absolute difference

Original cover (block) Stego-image (block) Absolute difference
Left Center Right
269 32695 1 4 3 0 0 0 4 3 4
387 65879 3 3 1 2 1 0 5 1 2
058 30582 3 5 3 0 0 0 5 3 6
¼10 ¼ −1 ¼ −4
 Reliable Metrics for Image LSB Steganography 523

The results obtained during these tests include a comparison of the original cover
image with itself (image 3a) where the parameters analyzed reach optimal values for
image quality metrics.
Average Absolute Difference (AAD)
AAD metric is used to enhance the similarities between image blocks. In Table 5 is
described the process of computing this parameter. In this example the block from the
center of the stego–image fits better with the searched pattern. We can conclude that a
lower value of this parameter indicates that the analyzed image is the most similar to
the original cover file.
The results obtained for this parameter (see Table 4) show that classical LSB
technique replacement keeps better the pixel arrangement pattern of the original cover
file. The heist value of this parameter is obtained by the LSB replacement using the last
4 bits of a byte of the original image. This means that stego-image 3e, deviates the
most from the pattern of the original cover.
Mean Square Error (MSE)
MSE is used to enhance the difference of two digital images. Equation (1) shows the
way it is computed, having [14]: MxN –image dimension, p –original image, q –image
with hidden data;

1 XM XN 2
MSE ¼ ðpij  qij Þ : ð1Þ
MxN i¼1 j¼1

A low value of this parameter shows that the two analyzed images are similar; a
high value of this parameter shows that there are considerable differences between the
original cover image and the stego-image [14].
The results obtained for the MSE (see Table 2) indicates that image 3b (classical
LSB embedding) is the most similar with the original cover file. Also, it states that the
quality of image 3e was decreased the most comparing with the other analyzed images.
Laplacian Mean Squared Error (LMSE)
LMSE is a metric that quantifies the quality of an image [19]. A high value of LMSE
means that the analyzed image has poor quality.
From the results obtained one can notice that the lowest value of LMSE is reached
by image 3f. The image is processed using LSB embedding optimized with a random
selection function of the bits (0, 1 or 2) that will carry the secret information. The
applied embedding algorithm processes the entire surface of the cover file and it also
hides besides the secret information some random data to obfuscate the size of secret
information.
Normalized Cross-Correlation (NCC)
The NCC metric is used to determine the degree of similarity between two compared
images. The NCC values are between −1 and 1. If NCC value is [20]:
524 D. Bucerzan and C. Raţiu

• 1, it means that the two compared patterns (images in this work) have the same
shape;
• −1, it indicates that the two compared patterns (images in this work) have the same
shape and opposite signs;
• 0, it shows the compared patterns are completely uncorrelated.
• The results obtained in all tested cases equals with 1 or are very close to 1. This
means that the analyzed images are very close; they do not present naked eye visible
distortions.

Peak Signal to Noise Ratio (PSNR)

PSNR metric is used to represent the similarity between the two images. PSNR is
measured in decibels (dB). A high PSNR value means unnoticeable differences
between two analyzed images. PSNR is described in Eq. (2) and it is represented like a
logarithmic function of the ratio between the maximum pixel value in the image and its
MSE value [14]:
2
Cmax
PSNR ¼ 10x log10 : ð2Þ
MSE
The results obtained in this work exceed 40 dB, meaning that the stego-image with
secret information embedded is very similar with the original cover file.

3.3 Statistical Steganalysis

For steganalysis process we applied two techniques present in the literature: RS
Analysis and SP Analysis. Both techniques try to estimate the size of secret information
embedded in a digital image. In this process it is used only the stego-image. Each
technique uses an original algorithm to divide the image in to several groups of pixels.
Based on a comparison of the pixels values with their neighborhood pixel values both
methods approximate the size of the secret information.

Fig. 5. The embedded secret image

Figure 5 represents the image that we used in this work for secret information that
was embedded in the cover file. It is a jpg image of 226 kb (231424 bytes) dimension.
Tables 5 and 6 show the results obtained when applying RS respectively SP
Analysis on the stego-images proposed in this paper. The results express in bytes and
 Reliable Metrics for Image LSB Steganography 525

Table 5. RS analysis results

Image 3 RS analysis (bytes)
Non-overlapping groups Overlapping groups
Red Green Blue Red Green Blue
b 48,466 29,211 190,015 52,804 38,634 202,708
c 28,841 15,516 199,573 32,914 22,727 212,653
d 28,040 8826 212,594 31,284 18,528 225,320
e 21,159 5678 193,024 24,668 14,319 206,073
f−1 416,630 418,332 630,763 421,092 427,738 637,466
f−2 372,103 367,037 376,051 369,234 368,344 377,904

Table 6. SP analysis results

SP analysis (bytes)
Image 3 Red Green Blue
b 47,000 25,744 188,065
c 27,493 18,262 199,642
d 25,387 11,822 209,475
e 18,878 4829 190,159
f-1 413,242 421,275 434,152
f-2 368,869 359,908 371,361

they present an approximation of the size of the secret information hidden in every
RGB color (red, green, blue). Summing the values obtained for every color, we get an
approximation of the size of the secret information embedded.
From the results obtained we can conclude that the techniques of embedding that
used pseudorandom LSB succeed to hide the size of the secret information in com-
parison with classical LSB technique which has a low security from this point of view.

4 Conclusions

Table 7 represents a classification of the steganographic techniques by their quality

metrics analyzed in this paper. We scored from 1 (the best) to 6 (the worst), the values
of the metrics achieved by every technique analyzed in this paper. To complete the
evaluation besides the metrics used in this study we add another two important criteria
namely the speed and the interoperability.
In this work we considered steganographic systems for mobile platforms with
similar performances regarding speed (see [9]). Even though the systems recorded
speed differences, they are unnoticeable to users. This is why we evaluate the analyzed
system for a score of 1 point regarding speed performances.
526 D. Bucerzan and C. Raţiu

Table 7. Quality of the analyzed steganographic methods

Image 3 criteria b c d e f–1 f–2
Visual analysis 6 6 6 6 1 1
Average absolute difference 1 2 3 6 4 5
Mean squared error 1 2 5 6 3 4
Laplacian mean squared error 4 5 2 3 6 1
Peak signal to noise ratio 1 2 5 6 3 4
Normalised cross-correlation 1 1 1 1 1 1
Statistical steganalysis 6 6 6 6 1 1
Speed 1 1 1 1 1 1
Interoperability 3 3 3 3 3 3
Quality 24 28 32 38 23 21

The methods analyzed in this paper work both on Android and Windows platforms
(mobile and traditional computers) however not on iOS. This is the reason why we
evaluate them for a score of 3 points.
From Table 7 we can conclude that from the analyzed steganographic methods
used in this work, the one that sets itself apart with high performances is the one
represented by image 3f (1 and 2). It uses LSB steganography techniques optimized
with a random selection function of the bits that will carry the secret information.
The criteria which made a difference in this research is the security of the method.
Statistical Steganalysis and Visual Analysis are the main metrics that confirms the
security of the steganographic methods.
The benchmark metrics (AAD, MSE, LMSE, PSNR and NCC) can be applied if a
third party has the access to both original cover file and the cover file whit secret
information embedded. This case has a very little probability to happen because the
original cover file is known only by the sender. The receiver doesn’t need the original
cover file to extract the secret information. The best values obtained by image 3b, 3c,
3d and 3e generated by the benchmark metrics are influenced the most by the fact that
the resulted image has an area identical with original cover file. Although the applied
methods for hiding process obtained good values in benchmarking they did not pass the
security test.
In the literature there are other researchers [21, 22] which propose different metrics
for the evaluation of applications on mobile systems. In this paper we propose some
reliable metrics that can measure the quality of stenographic systems running on mobile
platforms.

Acknowledgement. Thanks to the researchers whom by their works and their studies in the field
of information security, steganography and mobile platforms contributed to the results obtained
in this paper.
 Reliable Metrics for Image LSB Steganography 527

References
1. Appthority enterprise mobile threat report. Appthority Inc, San Francisco, CA 94105 (2015)
2. Lupşe, O.S., Căprioru, A., Stoicu-Tivadar, L.: Mobile applications supporting healthy
lifestyle. Soft Computing Applications Soft Computing Applications, Volume 356 of the
series Advances in Intelligent Systems and Computing pp 241–248, Proceedings of the 6th
International Workshop Soft Computing Applications, Volume 1. Springer International
Publishing (2015). ISBN:978-3-319-18295-7
3. David, E., et al.: It Threat Evolution in Q2 2015, vol. 125212. Kaspersky Lab HQ, Moscow
(2015)
4. Financial cyberthreats in 2014. Kaspersky Lab, Securelist (2015)
5. McAfee Labs Threats Report. McAfee Labs, Santa Clara, CA 95054 (2015)
6. Hosmer, C.: Steganography and Smart Phones. Forensicmag. Advantage Business Media
© Copyright 2016 Advantage Business Media, 2 August 2012. http://www.forensicmag.
com/articles/2012/02/steganography-and-smart-phones. Accessed 13 Apr 2016
7. Gartner: Forecast: PCs, ultramobiles and mobile phones, worldwide, 2012–2019, 1Q15
Update, Gartner Inc., Egham, UK (2015)
8. Stănescu, D., Stângaciu, V., Stratulat, M.: Steganography on new generation of mobile
phones with image and video processing abilities. In: Proceedings of IEEE International
Joint Conferences on Computational Cybernetics and Technical Informatics (2010)
9. Bucerzan, D., Raţiu, C., Manolescu, M.J.: SmartSteg: a new android based steganography
application. Int. J. Comput. Commun. Control 8(5) (2013). ISSN:1841-9836
10. Singh, K.M., Chanu, Y.J., Tuithung, T.: Steganalysis of ± k steganography based on
noncausal linear predictor. Int. J. Comput. Commun. Control 9(5), 623–632 (2014).
ISSN:1841-9836
11. Karamchandani, S.H., Gandhi, K.J., Gosalia, S.R., Madan, V.K., Merchant, S.N., Desai, U.B.:
PCA encrypted short acoustic data inculcated in digital color images. Int. J. Comput. Commun.
Control. 10(5), 678–685 (2015). ISSN:1841-9836
12. Mazurczyk, W., Caviglione, L.: Steganography in modern smartphones and mitigation
technique. IEEE Commun. Surv. Tutor. 17 (2015)
13. Bucerzan, D., Raţiu, C.: Steganography and Cryptography on Mobile Platform. Constanţa
Maritime University Annals 14(20) (2013). ISSN:1582 3601
14. Bucerzan, D., Raţiu, C.: Contributions to steganographic techniques on mobile devices. In:
Proceedings of the Security for Information Technology and Communications, LNCS,
vol. 9522. Springer (2015)
15. Dhall, S., Bhushan, B., Gupta, S.: An in-depth analysis of various steganography techniques.
Int. J. Secur. Appl. 9(8) (2015). doi:10.14257/ijsia.2015.9.8.07
16. Wang, Z., Bovik, A.C.: A universal image quality index. IEEE Signal Process. Lett. 9(3)
(2002)
17. Guiffy software. The advanced cross-platform image diff, Copyright © 2016 Guiffy
Software, Inc. All rights reserved. http://www.guiffy.com/Image-Diff-Tool.html. Accessed
24 June 2016
18. Steganography studio. http://stegstudio.sourceforge.net/. Accessed 16 June 2016
19. Gulame, M., Joshi, K.R., Kamthe, R.S.: A full reference based objective image quality
assessment. Int. J. Adv. Electr. Electron. Eng. 2(6) (2013). ISSN (Print):2278–8948
20. Raghavender Rao, Y., Prathapani, N., Nagabhooshanam, E.: Application of normalized
cross correlation to image registration. IJRET 3(5). eISSN:2319-1163, pISSN:2321-7308
(2014)
528 D. Bucerzan and C. Raţiu

21. Boja, C., Batagan, L.: Analysis of m-learning applications quality. WSEAS Trans. Comput.
8(5). ISSN:1109-2750, World Scientific and Engineering Academy and Society (WSEAS)
(2009)
22. Boja, C., Doinea, M.: Usability vs. Security in mobile applications. In: Proceedings of the
12th International Conference on INFORMATICS in ECONOMY (IE 2013), Education,
Research & Business Technologies, Bucharest, Romania (2013)
 Information System Security in the Context
of Administrative Cooperation Through Internal Market
Information System in Romania

Camelia Daciana Stoian1 and Dominic Bucerzan2 ✉

( )

1
“Vasile Goldis” Western University of Arad, Arad, Romania
[email protected]
2
“Aurel Vlaicu” University of Arad, Arad, Romania
[email protected]

Abstract. The importance of understanding the security of information in close

correlation with privacy must be reflected in the curricula of undergraduate of
Informatics field. The syllabus includes the management of the IT security
systems and also the knowledge of electronic platforms used and developed
strictly in administrative cooperation. The last one is achieved through the trans‐
position of directives of European Parliament and of the Council regarding the
services in the internal market in the sectoral legislation in this area. It should be
developed a strategy to facilitate the process of raising awareness of the Romanian
citizen about the risks, threats and legally rights they have when using latest IT
development in everyday economic and social activities. In this paper we
approach these aspects from legal point of view.

Keywords: Electronic government · Electronic administration · Internal market

information system · Alert system · Alert authority · Interactive Policy Making
(IPM)

1 Introduction

The understandings of several terms from Information System Security are hardly
perceived by the vast majority of graduates who wish to become professionals in this
field. Here you have some important items:
• “Alert System” (mechanism for notification by Member States and the European
Commission which shall inform each other, in the shortest time, the situation in which
the facts of a service provider or certain circumstances associated with service activ‐
ities covered by Government emergency Ordinance no. 49/2009 [10], approved with
amendments by Law no. 68/2010, can cause serious health problems or the safety of
persons or the environment);
• “Authority alert” (the competent authority, which receives access to the alert system
and initiate alerts in IMI system);
• “Electronic single point of contact” (information system of public utility, part of the
e-government system of National Electronic System);

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_39
530 C.D. Stoian and D. Bucerzan

• “Digital Romania National Centre” (Authority responsible for the implementation,

operationalization and administration of electronic PCU, a body under the Ministry
of Communications and Information);
• “Interactive Policy Making (IPM)” (Authority responsible for the implementation,
operationalization and administration electronic PCU, a body under the Ministry of
Communications and Information);
• “Electronic Government”;
• “Electronic Administration”;
• “Internal Market Information System”.

2 Internal Market Information System (IMI) versus Incident

Legislative Areas

Government Decision no. 931/2010 regarding the freedom of the establishment of the
coordinators of the Internal Market Information System – IMI and for the approval of
rules and procedures of administrative cooperation through IMI1 provides the legal
framework for the application of:
• Government Emergency Ordinance no. 49/2009 regarding freedom of the establish‐
ment of the service providers and the freedom to provide services in Romania [10],
approved with amendments and completions of the Law no. 68/2010 Chapter VI –
Administrative cooperation;
• Law no. 200/2004 regarding the recognition of diplomas and professional qualifica‐
tions for regulated professions in Romania, with subsequent amendments and
completions, art. 37 paragraph (2)2- “Art. 37 (2). Romanian authorities provided in
Annex 3 are part of the Internal Market Information System – IMI – and it carries
out the obligations laid down in this regard. Romanian competent authorities shall
exchange information with the competent authorities of other Member States relating
to disciplinary or criminal penalties imposed or any other specific circumstances
including serious consequences on the conduct of activities under this law, ensuring
compliance with the rules on the protection of personal data and confidentiality (To
the date of 03- Sep- 2015, Art. 37, paragraph (2) of Chapter VI amended by the Art.
I, Sect. 31 from the Ordinance 43/2015)”;
• Commission Decision 2008/49/EC of 12th of December 2007 regarding the imple‐
mentation of the Internal Market Information System (IMI) as regards the protection
system of personal data;
• Commission Decision 2009/739/EC of 2nd of October 2009 setting out the practical
arrangements for the exchange of information by electronic means between Member
States under Chapter VI from the Directive 2006/123/EC of the European Parliament
and of the Council regarding the services in the Internal Market [1].

1
Published in the Official Monitor number 651 dated from 20th of September 2010.
2
Law no. 200 of 25th of May 2004 regarding the recognition of diplomas and professional
qualifications for regulated professions in Romania, published in the Official Monitor number
500 dated from 3rd of June 2004.
 Information System Security in the Context of Administrative Cooperation 531

Internal Market Information System (IMI) is used as an administrative3 linking

between the competent authorities of the Member States and between the competent
authorities of the Member States and the Commission, useful cooperation for the
implementation of Union acts on the field of Internal Market, by folding on the provi‐
sions of the Article 26, paragraph (2) of the Treaty regarding the functioning of the
European Union (TFUE) [7] which predict the administrative cooperation, including
the exchange of personal data between Member States or between Member States and
the Commission.
Exchange of information with similar institutions in other countries for legislative
achieved following areas:
• Mutual assistance provided by the Directive of the European Parliament and EU
Council no. 2005/36/EC of 7th of September 2005 regarding the recognition of
professional qualifications [6];
• Administrative cooperation provided by the Directive of the European Parliament
and EU Council no. 2006/123/CE regarding the services in the Internal Market [1];
• Patients’ rights, as covered by Directive 2011/24/EU on the application of patients’
rights in cross-border healthcare [3];
• Cross-border transport of euro cash, as governed by Regulation (EU) 1214/2011 on
the professional cross-border road transport of euro cash between euro area Member
States [4];
• Detaching of workers as regulated by Directive 96/71/EC on the posting of workers
to provide services [5];
• E-commerce, as governed by Directive 2000/31/ EC of the European Parliament and
of the Council of 8 June 2000 on certain legal aspects of information society services,
in particular electronic commerce, in the Internal Market (Directive on electronic
commerce) [9];
The legal obligation of the authority of member states to give mutual assistance it is
done including the exchange of information realized through IMI, which requires
however a high level of data protection.
The authority, European of Data Protection Supervisor, is the body entitled to
monitor and follow the liaising with national data protection authorities, including the
relevant provisions on data security. The Ministry of Foreign Affairs (MFA) is the
national coordinator of the IMI, the coordinator of legislative area module dedicated to
the Directive 2006/123/EC regarding services in the Internal Market and performs the
role of a reception center for alerts, ensuring the proper functioning and the proper use
of IMI on national level.

3
“Administrative cooperation” means work in cooperation of the competent authorities of the
Member States or the competent authorities of the Member States and the Commission,
through the exchange and processing of information, including through notifications and
alerts, or by providing mutual assistance, including troubleshooting, in order to improve
implementation of EU law.
532 C.D. Stoian and D. Bucerzan

3 Internal Market Information System (IMI) Versus Alumni,

the Stage of Introducing Them the Skills that Reflect the Current
Needs of Society

In a communication of the Commission to the European Parliament, with reference to

the Strategy Europe 2020, it is stated that “The economy is based on the knowledge that
is required by people who own mix of skills: transversal skills, computer skills for the
digital era, creativity and flexibility, and a thorough understanding of the chosen area”
with the indication that the “public and private employers, including in the intensive
research sectors, increasingly report further imbalances and difficulties in finding the
right people for their evolving needs”4.
Going through professional and transversal competences of graduates which are
insured by a program of undergraduate studies related for example to areas such as,
Systems Engineering, Computers and Information Technology or Informatics, we do
not identify a complex sphere of skills specific to the categories of IMI existing users,
namely: users with local data rights management, user with management rights of the
data of other competent authorities, users with rights of management of applications,
users with rights of viewing the applications.
Basically, we must not forget that IMI is an electronic platform which requires the
use of modern methods of analysis and synthesis of computer science, operation and
management of electronic systems of processing and transmission of information,
familiarization with the latest achievements hardware and software, but in close corre‐
lation with the assuming responsibilities and limits on line access rights, secrecy or other
obligations of confidentiality in accordance with national law or with the Union.
All the categories of users listed, practically the users – the persons appointed in
order to fulfill the obligations of administrative cooperation through IMI, must know
and, consequently, to comply with the dispositions of the national law in force and EU
rules directly applicable, regarding the processing of personal data.
It can thus to be said that it is of great importance the acquisition of the knowledge
in the related areas during the faculty, to help ensure an activity of high professionalism,
both in terms of new areas that require knowledge and also with reference to the scope
of possible occupations for the holder of the diploma, practically all professional activ‐
ities regulated under Romanian law, that make this profession in Romania. The training
sessions organized by the Ministry of European Affairs regarding the operation of IMI
target, has, in a first stage of discussions, the presentation of the relevant legislation and
underlines the obligation of administrative cooperation across borders by IMI.
It can thus be said that it is of great importance in acquiring the knowledge of related
areas during the faculty, to help ensure an activity of high professionalism, both in terms
of look new areas that require knowledge and with reference to the scope of occupations
possible the diploma, virtually all professional activities regulated under Romanian law,
that make this profession in Romania. The training sessions organized by the Ministry
of European Affairs regarding the operation of IMI target in a first stage of discussions

4
Document available online on the address: http://eur-lex.europa.eu/legal-content/RO/TXT/?
uri=CELEX%3A52011DC0567.
 Information System Security in the Context of Administrative Cooperation 533

presenting the relevant legislation and underlines the obligation of administrative coop‐
eration across borders by IMI. Later, however, an important determinant is given to
presentation of how effectively that access and use IMI is organized training activities
assisted creating your password and security code and authentication in IMI by entering
the username, exercises in real time within base training related application, which is a
replica of IMI, insertion and data management, eliminating the old users, and new users,
creating and sending a request for information, management of inquiries received and
the reply, use mechanism alert (only for coordinators), a registry search, records
management data on authority owned or managed by their practical aspects of using the
system by IMI users the advantages of its citizens and service providers.

4 Internal Market Information System (IMI) Versus Authorized

Persons

The Commission provides and manages the software and IT infrastructure for IMI,
assures IMI security, monitors network of IMI national coordinators5 and has a deter‐
minant role in training IMI users (persons designated by a competent authority or an
IMI coordinator) and assistance their technique. Like “IMI actors”, we have identified
competent authorities, IMI coordinators and the Commission, following that the “IMI
users” to be identified as individuals who work under the authority of IMI actor and who
are registered in the system.
Another active category of operable and involved in this system, which is defined in
Article 5, Letter i) of the Regulation (EU) no. 1024/2012 of 25th of October 2012
regarding the administrative cooperation through the Internal Market Information
System and repealing the Decision 2008/49/EC Regulation IMI, being defined as
“external actors”, namely physical or legal persons, others than IMI users, who may
interact with IMI only through a separate technical means and in accordance with a
specific pre – defined workflow provided for this purpose.
The right to grant and withdraw in a reasoned way the access of their own IMI users
in the field of Internal Market for which hold competences, belongs to the competent
authorities, IMI coordinators and the Commission. Member States are those who bear
the burden nomination of IMI coordinators and competent authorities, and also the
internal market areas in which they have competence, in this procedure, the Commission
having only an advisory role.
Regarding the training of IMI users within competent authorities, this aspect is
achieved by IMI coordinators responsible or by the Ministry of Foreign Affairs, when‐
ever it is necessary, either due to their demands or after determining training needs within
competent authorities, being organized, in this sense, training sessions.

5
“IMI coordinator” means a body designated by a Member State to perform support tasks
necessary for the efficient functioning of IMI in accordance with this Regulation.
534 C.D. Stoian and D. Bucerzan

5 Case Study: IMI User Skills

I approached this subject with the idea that, “IMI users” - in fact, the individual appointed
by a competent authority, works on and with an electronic platform, with certain char‐
acteristics, that, in our opinion, implies thorough knowledge of modern methods of
analysis and synthesis in informatics, in the exploitation and management of electronic
processing and transmission of information and in hardware and software achievements.
However, in national legislation which regulates the nomination of these users we
have not identified the specific conditions, reflected in studies, that must be fulfilled by
nominees and that could serve as a guidance for the competent authorities in the process
of nomination. We give as example the ordinance of the Ministry of Education and
Science, issued under no. 5800 of November 16, 2015 concerning the registration of
accredited higher education institutions in Romania in the Internal Market Information
System – IMI, where Articles 1 to 4 refer only to the number of users designated by each
accredited higher education institution in Romania to use the internal market information
system - IMI (between minimum two and maximum five people) and to their attributions
to respond electronically to requests for information through the System for internal
market information – IMI, respectively to use the notification module on the recognition
of professional qualifications (IMI -PQ notifications). The only reference is to the
nominee having a certain function (chief- secretary of the respective HEI) and not to
certain specific conditions. Article 4, following the same guideline, only indicates that
training IMI registered users is performed permanently by the IMI Coordinator of the
legislative area for the “Professional Qualifications Acknowledgement” domain -
Ministry of Education and Scientific Research - National Centre Recognition and Equiv‐
alence of Diplomas.
Analyzing transversal competences of higher education graduates, in relation to how
they are defined in methodology-NQFHE, we notice the highlighting of those skills that
transcend a certain field, respectively any studies with a focus on transdisciplinarity.
Specifically, “they consist teamwork skills, oral and written communication in the
mother tongue/foreign tongue, use of information and communication technology - ICT,
problem solving and decision making, recognition and respect for diversity and multi‐
culturalism, learning autonomy, initiative and entrepreneurship, openness to lifelong
learning and development of professional values and ethics, etc.” Drawing a parallelism
between transversal and professional competencies regarded as a unitary and congealed
range of specific knowledge, properly used and skills learned (critical and constructive
reflection, creativity and innovation), we conclude that the basis of the competence
regarded as a whole can only be set by attending specialized studies.
Considering the above, we believe that experts in information and communications
technology are IMI ideal users, the only ones able to provide advice and to manage all
aspects of information technology systems, meaning hardware (computer components)
as well as software implementation and other aspects related to specific IT applications.
At the same time, a work must be carried out to draft the appropriate documentation
including principles, policies and procedures in order to maintain and support databases
and other information systems to the main purpose of ensuring optimal performance and
data integrity and security, all these lying within IT expert’s competence.
 Information System Security in the Context of Administrative Cooperation 535

We note in the October 2015 edition of an IMI Newsletter, the results of a survey
sent to all users at the beginning of the year. The study was conducted to assess the main
functions of the system, its advantages and ease of use, and needs for training and
support. Hence, according to the publication, it is outlined that “a third of respondents
said they have not received information or training in operation of IMI”. That is why
the European Commission IMI team, establishes: “We will discuss with IMI national
coordinators about these results and we will integrate them in future IT projects. Several
measures have already been defined:
– it includes a number of improvements of the search function for competent authorities
– sets of questions to be reviewed and improved, according to comments received from
users, in the following versions of the software.
– improving the way questions and answers are presented, to help users find relevant
questions.
– Member States and the Commission will further promote the IMI and the related
website.”

6 The Confidentiality of Communications and the Security

of Information in the Processing of Personal Data

“Personal data that can be made public do not constitute a homogeneous category to be
dealt with uniformly from a data protection point of view. Every individual, regardless
of name, age, sex, nationality, marital status, is protected in relation to any institution
that has personal information about him and has rights in relation to that person. He has,
for example, the access to data held by the operator on their person, has the right to know
exactly what they use, to intervene or oppose the processing of their data, and has the
right to complain to the supervisory authority or directly in court” [2].
Nationally, the Law 506/2004 concerning the processing of personal data and
privacy in the electronic communications sector6 transposes the Directive 2002/58/EC
of the European Parliament and of the Council [8] (regulating any operation performed
upon these data, including collection, storage, use or disclosure by transmission or
otherwise. According to the contents arranged in Article 15 of IMI Regulation, IMI team
members of the Commission have the right to access personal data of IMI users. Instead,
it is remarkable that Commission staff has access to personal data “of citizens that
exchanged information (“relevant persons”), unless the Commission fulfills a role in the
exchange of information, for one of the legislative acts covered by IMI. However, for
technical reasons, it is possible that Commission staff to be forced to process the personal
data of a such person to extract, at the request of a competent authority, the personal
data that have been blocked and regarding the person subject who has exercised the right
to access, correct or delete, or delete the personal data of IMI before the expiry of normal

6
Published in the Official Monitor number 1101 from 25th of November 2004, and promulgated
by the Decree 946/2004.
536 C.D. Stoian and D. Bucerzan

storage at the express request of competent authorities involved in administrative coop‐

eration procedure and provided the person to express his consent”7.
This brings us to put the emphasis on the “agreement of the person concerned” that
bind us to the side of the confidentiality of communications made using public electronic
communications networks. Of the exposure which was rendered above, gleaned from
the doctrine, dedicated otherwise in the wording of several laws, it can be outlined the
ensuring of confidentiality of information that can be the subject of any act of listening,
recording, technical storage and any other form of interception or surveillance of
communications and related traffic data. Definitely, this side of privacy becomes inop‐
erable in case of a written consent or a requirement under the law manifested by compe‐
tent authorities. And notice here as an aspect acute, those retained in an article published
in Emergency Medicine Search and Rescue Journal, which targets the right to privacy,
the theme strictly related to the topic always topical, namely, patients’ rights: “We
appreciate the timely the future existence of instructions that target and medical system,
specifically in terms of measures of organizational and technical to protect both
networking between services and administrative structures within a healthcare provider
corroboration extend patient acceptance the first representative/employee in front of
whom proceeded to manifestation processing agreement until the last, which almost
never has direct contact. Only such a rule with its own methodology can cover otherwise
and scope of patient information regarding the knowledge of the right of access to data
held by operator on their person a right to know exactly what they use, to intervene or
oppose the processing of their data” [16].
Understanding by a student, future graduate, the content of definition of personal
data, the implications of the term “any information” that targets an identified physical
person or particular who can be identified, regardless of way, directly or indirectly, in
particular by reference to several factors specific to his physical, physiological, mental,
economic, cultural or social identity, involves an approach consistent with national and
international practices in the field. Only after this time, once it has been well ingrained
the sense, contours and content of the concept can be understood the consequences of
violation of legal rules governing personal data and the implications that may result from
non-compliance with procedures for management of data on the registers held or
managed, and also the use of their benefits to citizens and service providers. And all this
in close conjunction with knowledge of the regulated segment processing of personal
data by Directives 95/46/EC and 2002/58/EC [8], respectively the Regulation (EC) no.
45/2001, in accordance with the principle of the need to know and only in area or areas
of the internal market, for which they were granted access, according to the EU law.

7
The document is available online at the address: http://ec.europa.eu/internal_market/iminet/
_docs/data_protection/privacy_statement_ro.pdf.
 Information System Security in the Context of Administrative Cooperation 537

7 European Professional Card in the Context of the Differences

Between the Activities of a Profession Regulated Differently
in Another Member State

The activation of the European Professional Card is closely related with Information
System (IMI), the two having the well-defined purpose to cement the trust between the
ones interested and the involved authorities, also contributing to ensuring conditions of
transparency and reliability that promote free movement of professionals.
European professional card, sketched as an electronic mechanism, in fact as a web
app, aims to facilitate the recognition of professional qualifications in another EU
country. The main interest is to help professionals that wish to settle in a certain EU
country, to exercise their profession there, permanently or temporarily.
In the document entitled “User’s Guide to the competent authorities’ version of 2
February 2016 published on the website http://ec.europa.eu/internal_market/imi-net/
_docs/library/epc_user_guide_ro.pdf, our European Professional Card described as an
“added value”, thus: “For professionals, EPC adds value because it offers:
– general information about documents, additional procedures and possible costs in the
destination country;
– quick and easy procedure to complete all administrative steps required at any time
and from anywhere in the EU;
– a channel for a clear and transparent communication in the applicant`s native tongue,
with the authorities responsible for managing the applications;
– upgrading and security of personal;
– facilitating applying for Statements of Comparability and the possibility of reusing
documents submitted in earlier applications;
For competent authorities, EPC adds value because it offers:
– structured and standardized procedures common to all EU countries;
– a concise and easy to use tool for processing applications, which facilitates decision
making”.
The regulation of this procedure definitely has a unexplored side allowing, besides
the issues that caused the very establishment of the institution itself, awareness of other
two segments that need to be placed within a legal framework, as proper training of IMI
users, necessarily regarded as professionals as long as they access and manage confi‐
dential data and are responsible for the security, management, processing and updating,
by overlapping with the inclusion of the activities in question in a specific profession
also encompassing a wider area of activities.
If differences between the activities of a profession, regulated differently in the two
states as activities/competencies assigned/skills identified or acquired are determinant,
either a training program would be imposed for the graduate-professional to help
compensate deficiencies or prevent this by changing curricula. “Occupation” is a group
of professional activities that compose it and that are described in a document that
certifies the training. “Occupation” is a complex of knowledge obtained both through
learning and practice. By “professional competence” as defined in methodology-CNCIS
538 C.D. Stoian and D. Bucerzan

means proven ability to select, combine and use the appropriate knowledge, skills and
other acquisitions (values and attitudes) in order to resolve successfully a certain cate‐
gory of situations work or learning circumscribed to that profession, in terms of effec‐
tiveness and efficiency.” Certification of skills and recognition of qualifications involves
strictly regulated steps by each Member State’s own methodology in part, skill certifi‐
cation processes and recognition of qualifications in the European Union member states
being analyzed by specialists in studies shown on the websites http://cnred.edu.ro/ site.

8 Conclusions

The effective takeover by the authorities at central or local level by professional bodies
with regulatory role, authority or control, with specific expertise in one or more legis‐
lative areas, of applications based on information technology, the management of elec‐
tronic platform through which it is realized the exchange of information and commu‐
nication between competent authorities, we must recognize that is still under improve‐
ment.
An electronic platform, a system to get public utility, designed by the legislature to
facilitate both monitoring proper obligations of administrative cooperation between
states and especially access and ensure the provision of legal information by legal and
physical persons concerned, impose a regime appropriate to training and a continuity in
application development line along with various promotion activities and knowledge of
the services provided.
Security procedures which confer a high degree of safety operations must be consis‐
tent with the international practices in the field, and this condition is definitely an imper‐
ative character and respect Internal Market Information System (IMI)8, in fact the elec‐
tronic platform through which administrative cooperation is achieved and which
requires an exchange of information between authorities of a Member State with control
or regulatory role regarding service activities. We aim, in particular, the positions within
the administrative authorities that can be employed, and legal persons or associations
which, in the exercise of self-regulation, creates the legal framework for access to the
activities of services, covering the area of security of information mainly in relation to
confidential information respectively to the insurance of accessibility of the information
only by authorized persons in a legal framework.
From the technical point of view the main issues raised by information security have
reliable and high performance solutions like cryptography, steganography and water‐
marking. These solutions are well present in the literature and worldwide implemented
and used by social, economic and governmental sector. An interesting overview of
cryptography is presented in [11, 12]. Regarding steganography and mobile information
security a comprehensive approach is presented in [13–15].
We appreciate that the Romanian authorities and not only, that take part to the
Internal Market Information System and which carry out the obligations provided in this

8
Emergency Ordinance no. 49/2009 on freedom of establishment for service providers and
freedom to provide services in Romania, in the Official Gazette number 366 of 1 June 2009.
 Information System Security in the Context of Administrative Cooperation 539

regard and who needs professionals specialized in access and management of IMI in
compliance with respecting the rules on the protection of personal data and knowing
how ensure of confidentiality, aspect which may be translated as a proposal for updating
and developing the skills in the specialization of the future graduates.

Acknowledgement. Manny thanks to our colleagues whom offer their support to accomplish
this work.

References

1. Directive of the European Parliament and EU Council no. 2006/123/CE regarding the services
in the Internal Market
2. Răscol, R.: The internet and the protection of personal data, in annals of “Constantin Brancusi”
of Târgu Jiu. Ser. Jurid. Sci. 2, 186–191 (2008)
3. Directive no. 2011/24/ EU on the application of patients’ rights in cross-border healthcare
4. Regulation (EU) no. 1214/2011 on the professional cross-border road transport of euro cash
between euro area Member States
5. Directive no. 96/71/EC on the posting of workers to provide services
6. Directive of the European Parliament and EU Council no. 2005/36/EC of 7th of September
2005 regarding the recognition of professional qualifications; (on the date 10 – Nov – 2014,
the act was related to Regulation from 2014)
7. The TREATY from 25th of March 1957 regarding the functioning of the EUROPEAN
UNION – Published in the Official Journal number 83C dated from 30th of March 2010
8. Directive 2002/58/EC of the European Parliament and of the Council concerning the
processing of personal data and privacy of private life in the electronic communications sector,
published in the Official Journal of the European Communities no. L. 201 of 31st of July 2002
9. Directive no. 2000/31/ EC of the European Parliament and of the Council of 8 June 2000 on
certain legal aspects of information society services, in particular electronic commerce, in the
Internal Market (Directive on electronic commerce)
10. The Emergency Ordinance no. 49/2009 regarding the freedom of establishment of the service
providers and the freedom to provide services in Romania, in the Official Monitor number
366 of the 1st of June 2009
11. Ivan, I., Toma, C.: Informatics Security Handbook, 1st edn. Editura ASE, Bucharest (2006)
12. Schneier, B.: Applied Cryptography Protocols, Algorithms, and Source Code in C. Wiley,
New York (1996)
13. Bucerzan, D., Raţiu, C., Manolescu, M.: SmartSteg: a new android based steganography
application. Int. J. Comput. Commun. Cont. 8(5), 1841–9836 (2013)
14. Boja, C., Doinea, M., Pocatilu, P.: Impact of the Security Requirements on Mobile
Applications Usability, pp. 1582–7941. Economy Informatics, București (2013)
15. Bucerzan, D., Raţiu, C.: Contributions to Steganographic Techniques on Mobile Devices,
Innovative Security Solutions for Information Technology and Communications, Lecture
Notes in Computer Science. Springer International Publishing, Print ISBN:978–
3-319-27178-1, Online ISBN: 978-3-319-27179-8 (2015)
16. Puticiu, M., Stoian, R., Stoian, C.: Privacy of patient through the right to confidentiality of
information, Article published in volume of National Conference of Emergency Medicine and
trousers in special situations, Search and Rescue Emergency Medicine Journal, ISSN 2343–
7030
 An Application of the Multi Period Decision Trees
in the Sustainable Medical Waste Investments

Luiela Magdalena Csorba1 ✉ and Mihaela Crăciun2

( )

1
Department of Economics, Faculty of Economics, “Aurel Vlaicu” University of Arad,
Arad, Romania
[email protected]
2
Electrical and Power Engineering Faculty,
Politehnica University Timisoara, Timisoara, Romania
[email protected]

Abstract. The decision making process built up on the decision tree technique
is one of the most common methods used in almost all fields, in order to generate
predictions and classify previously unseen data. In the present paper we will apply
the multi period decision trees method in a sustainable medical waste investment
process. The purpose of this study is to decide the best investment strategy which
may be taken into consideration by a manager or entrepreneur who are interested
to extend a sustainable medical waste business.

Keywords: Decision making · Decision tree · Multi period decision trees ·

Probabilities · Sustainable investment · Medical waste · Incineration

1 Introduction

One of the most controversial domains regarding sustainability is focused on the field
of waste management, which aims to reduce the adverse effects of waste on people’s
health and on the environment. The waste handling, collection, incineration, recycling
and re-use methods vary widely among different countries and regions. In our study we
are focusing on the incineration of medical waste. The aim of our study is to create a
decision making process using the multi period decision trees by a manager who must
take an investment decision in the field of medical waste incinerators (we have to under‐
line that the term “multi period decision trees” is, in our acceptation, synonym with the
“complex decision trees” or the “multi-sequential decision trees”, being in opposition
with the “mono period decision trees”, the “simple decision trees” or the “mono-sequen‐
tial decision trees”).
Incineration is a disposal method in which solid organic waste are subjected to
combustion. Incinerators convert waste materials into heat, gas, steam and ash [1].
Incineration is a controversial method, due to issues such as emission of gaseous
pollutants.
Combustion in an incinerator is not always perfect and there have been concerns
about pollutants in gaseous emissions from incinerator stacks, which may have serious
environmental consequences. Incineration is also a practical method of disposing of

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_40
 An Application of the Multi Period Decision Trees 541

certain hazardous waste materials (such as medical waste, on which we intend to focus
in this study) [2].
Medical waste is a subset of wastes generated by health care facilities, such as
hospitals, physicians’ offices, dental practices, blood banks, veterinary clinics, as well
as medical research laboratories. Generally, medical waste is a healthcare waste that
may be contaminated by blood, body fluids or other potentially infectious materials.
The Medical Waste Tracking Act of 1988 defines medical waste as “any solid waste
that is generated in the diagnosis, treatment or immunization of human beings or animals,
or in the production or testing of biologicals.” Medical waste can be identified by one
of four different categories: infectious, hazardous, radioactive and general [3].
Infectious waste, labeled as pathological waste, requires specific treatment methods.
Pathological waste is either known or suspected to contain pathogens.
Hazardous waste affects humans in non-infectious ways. Hazardous waste can
include chemicals, both medical and industrial. Old drugs, including chemotherapy
agents, are sometimes hazardous. Around 20–25% of the total waste generated by
healthcare establishments is regarded as hazardous and may create a variety of health
and environmental risks, if not managed and destroyed in an appropriate manner [4].
Radioactive waste can be generated from nuclear medicine treatments, cancer thera‐
pies and medical equipment that use radioactive isotopes.
General waste makes up at least 85% of all waste generated at medical facilities, and
is no different from general household or office waste. It includes paper, plastics, liquids
and any other materials that do not fit into the previous three categories.
Hospital or clinical waste varies and the biggest challenge is to dispose of this wide
range of waste streams using only one solution. Because it has been roughly estimated
that of the 4 kg of waste generated in a hospital, at least 1 kg would be infected, incin‐
eration is an important method for the treatment and decontamination of biomedical and
health care waste, without endangering the health of patients, staff or anyone else coming
into contact with them. Being a high-temperature dry oxidation process (850°C to
1100°C), incineration reduces organic and combustible waste to an inorganic, incom‐
bustible matter [4].
There are two possibilities spread worldwide to incinerate medical waste: the incin‐
eration in the hospital or medical institution or the incineration through a private
company specialized in medical waste incineration. For those who choose the first
method, the market offers a large variety of medical waste incinerators. Generally, they
are three categories of medical waste incinerators: large medical incinerators for hospital
use, small medical incinerators for health clinics/labs, and finally, mobile medical incin‐
erators for multi-site operation. Practically, the best medical, lab or mobile incinerator
must process all types of medical and pathological waste.
For the second choice, the international market offers a large variety of incinerators.
For example, “Inciner 8” is the name of the firm which nowadays is one of the interna‐
tional leaders in offering a full range of medical waste incinerators, designed to provide
efficient waste destruction, using the best available technologies and environmentally
friendly processes. So, the market incinerators offered by this firm starts with I8–20A,
542 L.M. Csorba and M. Crăciun

I8–40A small scale incinerators, continues with I8–55A, I8–75A middle range inciner‐
ators, finally ends with types I8–140, I8–200, model I8–250 for large scale incinerators
and with the largest incinerator, model I8–700.
Not only the size, type or quality of an incinerator are important to assure an efficient
waste management, but also, in order to ensure a sustainable environment in that field
in Europe, the 7th Environment Action Programme established the following priority
objectives for the waste policy:
• to reduce the amount of waste generated;
• to maximize the amount of recycled and re-used wastes;
• to limit incineration of non-recyclable materials;
• to phase out landfilling to non-recyclable and non-recoverable waste;
• to ensure full implementation of the waste policy targets in all Member States.
We have to state that the 7th Environment Action Programme (EAP) guides the
European environment policy until 2020. In order to give more long-term direction, it
sets out a vision beyond that, of where it wants the Union to be by 2050. The programme
entered into force in January 2014. It is up to the EU institutions and the Member States
to ensure it is implemented, and that priority objectives set out are met by 2020.

2 Related Works

2.1 Waste Management in Europe - An Overview

Globalization, the growth of population (at 10 July 2016 they were 7.436 billion peoples
worldwide), consumerism and industrial development have led to a dangerous increase
of waste.
In our everyday life, waste comes in many different forms and may be categorized
in a variety of ways. Generally, waste could be liquid or solid. Both of them could be
hazardous. Liquid and solid waste types can also be grouped into organic, re-usable or
recyclable waste [5].
In a detailed accept, waste can take the form of agricultural and industrial, biode‐
gradable and organic, biomedical, chemical, clinical and toxic waste, and finally, but
not at least, domestic, electronic, human and mixed waste.
The more the number of consumers worldwide, the higher are the quantity of waste
produced. That’s why sustainable waste management is so important in meeting the
environmental challenges of the contemporary society of consume. In order to do that,
governments must spend money for environmental protection and for a sustainable
waste management.
The Classification of the Functions of the Government (COFOG) divide the govern‐
ment’s expenditures into ten main categories: general public services; defense; public order
and safety; economic affairs; environmental protection; housing and community affairs;
health; recreation, culture and religion; education; social protection. These divisions are
further broken down into “groups”. For “environmental protection” the groups are: waste
 An Application of the Multi Period Decision Trees 543

management, waste water management, pollution abatement, protection of biodiversity and

landscape, R&D environmental protection, environmental protection1.
The latest statistical data’s shows that in 2014, the EU-28 total expenditure of general
government on “environmental protection” amounted to 0.8% of GDP. The expenditure
for “waste management” amounted to 0.4% of GDP, while 0.1% of GDP was devoted
to expenditure in waste water management, pollution abatement, protection of biodi‐
versity/landscape and other expenditures related to environmental protection (see
Fig. 1).

Fig. 1. The EU-28 total expenditure of general government on “environmental protection” [6]

The highest ratios to GDP for waste water management were reported by Slovenia
and Luxembourg (0.6% of GDP in 2014), while Greece reported the highest ratio in the
EU for “pollution abatement” (0.6% of GDP) and 0.3% was devoted to “protection of
biodiversity and landscape” in the Czech Republic.

1,8%
1,6%
1,4%
1,2%
1,0%
0,8%
0,6%
0,4%
0,2%
0,0%

Fig. 2. Total expenditure on “environmental protection” in different EU-28 countries (2014) [6]

1
In the framework of the European System of National Accounts (ESA2010), Eurostat collects
data on general government expenditure by economic function according to the international
Classification of the Functions of Government (COFOG).
544 L.M. Csorba and M. Crăciun

In the same year, Greece and Malta devoted 1.6% of GDP to “environmental protec‐
tion”, the Netherlands 1.5% of GDP, while Cyprus, Finland and Sweden reported 0.3%
of GDP (see Fig. 2).
Waste management accounted for 0.8% of GDP in Malta and 0.7% in Greece, the
United Kingdom and Bulgaria, but below 0.05% of GDP in Austria, Ireland, Finland,
Denmark and Croatia (see Fig. 3).

0,9%
0,8%
0,7%
0,6%
0,5%
0,4%
0,3%
0,2%
0,1%
0,0%

Fig. 3. Waste management total expenditure in different EU-28 countries (2014) [6]

Romania amounted in the same year only 0.8% of GDP on environmental protection
and less than 0.4% for waste management. While in Europe people currently use 16
tones of materials per person each year, of which 6 tones become waste, the percentage
above related to Romania are low comparative with other EU countries.
The management of waste continues to improve in the EU, while the European
economy currently still loses a significant amount of potential “secondary raw materials”
such as metals, wood, glass, paper and plastics. In 2012, the total waste production in
the EU amounted to more than 2,514 billion tons. From this, only a limited share (36%)
was recycled. The rest was landfilled or burned.2
Just in terms of household waste alone, each person in Europe is currently producing
per year, on average, half of tone of such waste. Only 40% of it is reused or recycled
and in some countries more than 80% still goes to landfill [6].
An improved waste management helps to reduce health and environmental problems,
greenhouse gas emissions (directly by cutting emissions from landfills and indirectly by
recycling materials which would otherwise be extracted and processed), and avoid the
negative impacts at local level such as landscape deterioration due to landfilling, local
water and air pollution, as well as littering.
The European Union’s approach to waste management is based on the “waste hier‐
archy” which sets the following priority order when shaping waste policy and managing
waste at operational level: prevention, reuse, recycling, recovery and, as the least

2
Conference Report “Separate Waste collection in the context of a circular economy in Europe”,
Brussels, January 2016.
 An Application of the Multi Period Decision Trees 545

preferred option, disposal (which includes land filling and incineration without energy
recovery). In this study we will focus only on the incineration of medical waste.

2.2 The Multiple Uses of the Decision Trees

The decision tree technique is a commonly used method with the aim to predict the value
of a target variable based on several input variables. This method is a modeling approach
used in statistics and data mining. In the decision analysis, a decision tree can be used
to visually and explicitly represent decisions and decision making processes. In data
mining, a decision tree describes data, but not decisions [7]. The present paper deals
with decision trees used in the decision making process.
Decision tree is an inductive method of gathering information [8], which was devel‐
oped during the early 1950s as a device to represent a game in an extensive form. The
method’s purpose is the analysis of a decision problem in order to eventually find a
rational solution. Luce and Raiffa’s famous book (1957) make economic application of
decisions and game theories clearer [9].
A decision tree used in decision making is both, an algorithm and a decision support
tool that uses a tree like graph or model of decisions and their possible consequences,
including chance event outcomes, resource costs and utilities. Decision trees are
commonly used in decision analysis, to help to identify a strategy, in order to reach a
goal. Decision tree analysis involves forecasting future outcomes and assigning proba‐
bilities to some events. As the list of potential outcomes which are contingent on prior
events, become more dynamic with complex decisions, the well-known Bayesian prob‐
ability models have to be implemented to determine the priori probabilities (the revised
probabilities) [10].
All over the world, during the time, different authors used this model in many
domains, from finance to engineering, from philosophy to health care, from economic
forecasting to particular life decisions.
For example, Diallo, Jindal and Lin used the decision tree in a decision modeling
process in the field of software engineering projects. The authors’ recognizes that tradi‐
tionally, decision trees have been the fundamental tools for modeling investment oppor‐
tunities [11].
Gatta uses the decisional tree model in major oil field projects [12].
In some projects, risk is not only discrete, but it is sequential. So, for an investment
to succeed, it has to past through a series of tests, with failure and any point potentially
translating into a complete loss of value. This is the case, for example, with a pharma‐
ceutical drug that is just being tested on human beings. Decision tree allow us to not
only consider the risk in stages, but also to devise the right response to outcomes at each
stage. In that way, Damodaran recognizes the possibility to use decision trees in taking
decision in health care too, and also, he underlined the advantages of this algorithm [13].
In medical decision making (classification, diagnosing), there are many situations
where decision must be made effectively and reliable. That’s why Podgorelec proposed
the use of the decision tree in that field too [15]. The same idea was continued by Bae,
546 L.M. Csorba and M. Crăciun

who proposed to take clinical decision analysis by using the decision tree [14]. Vall‐
verdu, Riano, Bohada proposed a combination of criteria’s necessary to use in medicine
to improve the medical decision tree process [16].
Magee proposed some decision tree model for investments, but also for decisions
which might be taken in our every day life. For example, he proposed a tree for a cocktail
party problem. So, we can see the multiple ways a decision tree may be used [17].
In elaborating our sustainable waste management investment scenario, we will focus
on the general purpose that decision trees serves in all domains, including sustainable
management, which presents interest for our study. At the same time, we have to under‐
line the benefits that accrue from using decision trees, like [13]:
• dynamic response to risk, because decision trees encourage firms to consider how
they should act under different circumstances;
• value of information, because we can measure the expected value of the improved
information in the decision tree;
• risk management; because decision trees provides a picture of how cash flows unfold
over time, they are useful in deciding what risks should be protected against, and the
benefits of doing so.

3 Medical Waste Incineration in Romania

In Romania, the incineration of medical waste is a huge market and a real business
opportunity. With the adoption of the EU directives in terms of waste in the early 2000s
and the closing of the greatest part of the hospitals crematoriums, private companies
having authorized incinerators needed for the disposal of hazardous waste, have
emerged.
Because of the rapid changing’s in the environmental legislation, nowadays, in
Romania are firms which only collect the medical waste, and firms which collect and
burn the waste, in the same time. Because a business which only collects the waste must
have the same authorizations like the one which collect and burn, it is more efficient and
profitable to eliminate the waste.
At this moment, because the incinerator market is at the beginning, in our country
they are only a few firms which possess an eco-incinerator. For such a business, the
minimal investment is around 130,000–150,000 € and the income can reach the value
of 60,000 euro in a year. An eco-incinerator can generate a maximum of 2 Gcal per hour
(a Gcal is somewhere around 27 euro). Within 10 h per day the process brings 270 euro
from heat recovery, plus the money collected from waste incineration. So, the profit is
not insignificant in that field. For example, one of the largest health units in Romania is
the Cluj county hospital, which produces monthly an average of eight tons of medical
waste. The firm which collects and burns this waste is the biggest one in that region, “IF
Technologies”, with foreign shareholders. The company collects every month hundreds
of thousands of lei from health units, printing houses, garages or any other economic
agents who generates hazardous waste in that region. The incinerator from Cluj has a
maximum capacity of 550 kg of waste per hour, but this capacity varies depending on
 An Application of the Multi Period Decision Trees 547

the type of waste. After cremation, results 3–5% of sterile ashes from the amount of the
waste burned [19].
An incinerator consumes thousand of cubic meters of gas, but this are general good
operating costs, imposed by the equipment required, the humans, the related personnel
and by the packaging. Only the costs with packaging represent 30–60% from the cost
charged per kilogram. The most important is the collection network, because the hardest
part of the business is to have a sustained activity over the years. That’s why, almost
half of the employees are drivers, which do thousands of kilometers per month only to
collect waste. The profit of “IF Technologies” increased each year. For example, it was
500,000 lei in 2005, nearly 2 million in 2011 and nearly 3 million in 2014 [18].
Other Romanian companies dealing with hazardous waste management have also a
good financial trend. Iridex Group Import Export SRL Bucureşti had in 2011 a profit of
3,7 million lei; in 2014, nearly 10 million, but in 2015 only 4,2 million lei [20]. Eco Fire
Systems Constanţa had 383,000 lei profit in 2011 and 813,680 lei in 2014 [21], while
SC Mondeco Suceava 168,000 lei profit in 2011 and 1,099,500 lei in 2014 [22].
For the future, in Romania it is desired to create Regional Centers on European
Standards including two - three counties, and to close gradually all the remained old
hospital crematoriums. Thus, it would be easier to monitor and track the elimination of
medical waste. Furthermore, the costs per kilogram of waste incinerated will fall, as the
amount of waste incinerated will be higher and the operating costs lower.

3.1 Objectives of the Study

We underlined before that, as shown in the related works, the decision tree is often used
in many fields. We consider that the main objective of our study is to extend the use of
the decision tree in the field of sustainable waste investments.
One of the priorities of all EU countries is to reduce the environmental impact of
waste. Romania is one of the EU countries with huge waste and landfill management
problems. The Eurostat statistical data shows that Romania and Bulgaria are on the last
place in recycling and use of waste (99% are not recycled). In Germany, Austria,
Holland, for example, around 1% of waste are stored, while the rest are transformed in
money (recycling) or energy (incineration) [23].
A sustainable landfill management can assure an efficient elimination of municipal
solid waste and special toxic waste. But, from 41 Romanian counties, more then a half
doesn’t have a sustainable landfill, while 15 counties have one sustainable landfill and
only a few counties own two sustainable landfills. Big counties like Cluj or Brașov,
which produces huge quantities of waste, did not have an eco- landfill, which determines
the municipality to transport the waste in neighbor counties which own both, an eco-
landfill and an incinerator. Because of that, the issue of a sustainable incinerator does
not call into question. Starting from the above mentioned problems, the present paper
may be a support for an entrepreneur who decided to open a new business in the field
of sustainable waste incineration.
Applying the multi period decision trees method in the field of sustainable medical
waste investments, we intend to predict which alternative will be the best to be chosen
by a manager, taking into account that in this business, the expected profit depends
548 L.M. Csorba and M. Crăciun

directly on two factors or situations which may occur: a large or a moderate amount of
medical waste collected.
Taking into consideration the proposed situations (state of nature) which may appear
regarding the medical waste investments trends, our study try to predict in a proper way
which of the chosen investments alternatives will be the best, taking into account the
experience of a consultant needed by our manager, in the decision making process.

3.2 Research Methodology

The decision tree analysis is a way to frame and solve complex situations that require a
decision. The key in obtaining a useful solution is to define the problem at the very
beginning and to determine what decisions are need to be made. The problem-identifi‐
cation stage includes all the known information and factors that may influence the final
outcome. Once the problem is framed, decision trees help to find a route to an advan‐
tageous solution.
In our example, a company whose sphere of activity is the medical waste incineration
must make decisions regarding how to develop its business in the future. Due to the fact
that in Romania, like in other countries, the waste management legislation is stricter and
the pollution standards are more stringent, the company cannot operate further with a
traditional waste incinerator. The volume of medical waste is increasing, especially in
a big city. Given the national and international trends regarding the medical and toxic
wastes disposal, the company’s manager have to choose one from three possible alter‐
natives:
• A1 = to replace the old unsustainable incinerator, with a new large scale incinerator,
which is also an ecological one;
• A2 = to expand the business, opening a new business unit in the neighboring county,
equipping this workstation with a large scale environmentally friendly incinerator,
able to cover the medical waste incineration needs for more than two counties;
• A3 = to purchase two new sustainable incinerators and to under contract one of this
capacity to other entrepreneurs from related areas, in order to cover a wider market
and to collect the medical waste from the surrounding counties who do not have a
green incinerator.
Based on his experience, the manager knows that, the business income and therefore
its profit depend on the amount of medical waste collected in the time frame considered.
Basically, the following situations (state of nature) may occur:
• S1 = a large amount of waste collected;
• S2 = a moderate amount of waste collected.
The payoffs or consequences which may be obtained are summarized in the payoff
matrix and their size depends on the condition from the medical devices market, the
amount of health care services provided and the amount of waste collected from health
facilities (See Table 1).
 An Application of the Multi Period Decision Trees 549

Table 1. The Payoff Matrix

Annual profits (euro/year)
S1 S2
A1 500,000 300,000
A2 700,000 500,000
A3 400,000 200,000

The manager appreciates that the probability to occur state of nature S1 are 70%,
while the probability to occur state of nature S2 are 30%.
We have to mention that to purchase an environmentally friendly incinerator costs
the manager around 30,000 euro, the cost of expanding the business is around 60,000
euro and the under contracting process needs 40,000 euro.
The manager of the company does not know which may be the right decision? He
has two variants. He may take own decisions without requiring the help of a specialized
consultant. But, because of the market uncertainty, he decides to hire a consultant who
may provide him a report on the market conditions and on the success rate of each
alternative. The consultant may provide a favorable report (F), or an unfavorable one
(N), indicating a good condition of the market or a weak one, for the future. From the
experience of the consultant we turns out that when he predicted good market conditions,
the report was favorable in 70% of the cases, and when he forecasted weak market
conditions, the report was negative in 80% of the cases.
What will be the best decision for the entrepreneur?

Decision tree analysis

Decision trees are diagrams that portray the flow of a decision-making process as a
sequence of events and possible outcomes. Events are represented as points, or nodes,
and out-comes are depicted as branches issuing from each node. Nodes are either deci‐
sion nodes - at which the decision-maker determines what branch to take - or uncertainty
nodes/chance nodes, where chance rules the outcome.
Associated with each branch is the expected monetary value of the outcome. Addi‐
tionally, branches emanating from uncertainty nodes are weighted by the probability of
that outcome occurring. In common notation, decision nodes plot as squares and uncer‐
tainty nodes as circles.
The details of the decision-making process by using the decision tree are given
below. At the first stage (decision node 1) we have the situation when the company has
to choose between two alternatives: the manager should make its own decisions or, to
hire a consulting firm able to give him a proper investment advice. The first option leads
to another decisional node (decision node 3), where the manager has three alternatives:
to replace the old unsustainable incinerator, with a new, ecological one, to expand the
business, opening a new business unit in the neighboring county, equipping this work‐
station with an environmentally friendly incinerator or to purchase two new green incin‐
erators and to under contract one of this capacities to other entrepreneurs from related
areas, in order to cover a wider market and to collect the medical waste from the
surrounding counties who do not have an incinerator. Each of these alternatives leads
550 L.M. Csorba and M. Crăciun

in a chance node, in which two states of nature can arise: the company can find a large
amount of waste or, a moderate amount of waste (chance nodes 6, 7 and 8).
The second decision leads to another chance node (2), in which the following states
of nature are considered: a favorable or an unfavorable report from the investment
consultant. The decisions considered in nodes 4 and 5 are the same alternatives coming
out from the decisional node 3, with the same costs, but with revised probabilities. Each
of this alternatives leads to another chance node (9, 10, 11, 12, 13, 14), which supports
the same two state of nature alternatives like nodes 6, 7, 8, but with revised probabilities.
The branches of the uncertainty node try to capture the most important aspects of
this continuous distribution. The expected value of an uncertainty node is the sum of all
probability-weighted expected values of the outcomes stemming from that node. In this
way, working back from the end or right-hand side of the tree, expected values can be
calculated for every outcome. Once all the expected values have been calculated, the
optimal decision route - the one along maximum expected value - can be taken.
At the beginning we will try to solve the problem using the information owned by
the manager, without requiring the advice of a consultant. Furthermore, we have all the
needed probabilities given in our hypothesis.
The above mentioned situation can be designed in the simple mono period or mono-
sequential decision tree (see Fig. 4), which will be represented again at the end of the
article in the superior part of the extended multi period decision tree (see Fig. 6). The
simple tree is a mono period or mono-sequential tree, which, in our case will show that
the best decision to be chosen by the manager is alternative A2.

p(S1)= 0.7
500,000 €
A1
6
EV1= 440,000 € p(S2)= 0.3
cost= 30,000 € 300,000 €

p(S1)= 0.7
700,000 €
A2
7
EV2= 640,000 € p(S2)= 0.3
3 cost= 60,000 € 500,000 €

p(S1)= 0.7
400,000 €
A3
8
EV3= 340,000 € p(S2)= 0.3
cost= 40,000 € 200,000 €

Fig. 4. The simple mono-sequential decision tree

 An Application of the Multi Period Decision Trees 551

We must underline that the expected values are calculated as follows:

Expected value = success probability ⋅ successful payoff + failure probability ⋅ failure payoff
That means:
• EV1 = 0.7 × 500,000 + 0.3 × 300,000 = 440,000 euro
• EV2 = 0.7 × 700,000 + 0.3 × 500,000 = 640,000 euro
• EV3 = 0.7 × 400,000 + 0.3 × 200,000 = 340,000 euro
The cost of choosing the first alternative is 30,000, so the real expected value will
be: 440,000–30,000 = 410,000 euro. In the case of the second alternative, the implied
costs are 60,000 euro, so, the expected value will remain 640,000–60,000 = 580,000
euro. For the last alternative, the costs are 40,000 euro, so the expected value will remain
340,000–40,000 = 300,000 euro.
The optimal decision is: Do = max {EVi − cos ti}, where i = 1, m.
Because the highest expected value is 580,000 euro, the manager will opt for the
second alternative, when he takes the decisions without the support of a consultant.

The revised probabilities and the Bayes’ Theorem in decision trees

We suppose that a research took place and it predicts the possibility to collect a large
amount of waste.
We symbolized p (F/S1) as being the probability of a favorable report, when state of
nature S1 occur; than, the probability of an unfavorable report, given the state of nature
S1 is p (N/S1). The values of the probabilities are:
( )
p F∕ S1 = 0.7
( )
p N∕ S1 = 0.3

If p (N/S2) is the probability of an unfavorable report, given the state of nature S2,
the probability of a favorable report will be p (F/S2):
( )
p N∕S2 = 0.8
( )
p F∕ S2 = 0.2

Equation (1) is proposed in [24] as a way to change the previous probability p(F/S1)
in a revised probability, symbolized p (S1/F), by using the Bayes’ formula for two vari‐
ables.
If p (S1/F) is the probability to occur the state of nature S1, if the report is favorable,
than p (S2/F) is the probability to occur the state of nature S2, if the report is favorable:
( ) ( )
( ) p S1 × p F∕ S1
p S1 ∕ F = ( ) ( ) ( ) ( ) (1)
p S1 ⋅ p F∕ S1 + p S2 ⋅ p F∕ S2
552 L.M. Csorba and M. Crăciun

( ) 0.7 × 0.7
p S1 ∕ F = = 0.891
0.7 × 0.7 + 0.3 × 0.2
( )
p S2 ∕ F = 1 − 0.891 = 0.109

If p (S1/N) is the probability to occur the state of nature S1, if the report is unfavorable
means that p (S2/N) is the probability to occur the state of nature S2, if the report is
unfavorable, like in Eqs. (2) and (3) proposed by [24].
( ) ( )
( ) p S1 × p N∕ S1
p S1 ∕ N = ( ) ( ) ( ) ( ) (2)
p S1 × p N∕ S1 + p S2 × p N∕ S2

( ) 0.7 × 0.3
p S1 ∕ N = = 0.467
0.7 × 0.3 + 0.3 × 0.8
( ) ( )
( ) p S2 × p N∕ S2
p S2 ∕ N = ( ) ( ) ( ) ( ) (3)
p S2 × p N∕ S2 + p S1 × p N∕ S1

( ) 0.3 × 0.8
p S2 ∕ N = = 0.533
0.3 × 0.8 + 0.7 × 0.3
We note that the probability to obtain a favorable report p (F) is 0.55 and the prob‐
ability for an unfavorable report p (N) is 0.45 (Eqs. 4 and 5) [24]
( ) ( ) ( ) ( )
p(F) = p S1 ⋅ p F∕S1 + p S2 ⋅ p F∕S2 (4)

p(F) = 0.55
( ) ( ) ( ) ( )
p(N) = p S1 ⋅ p N∕S1 + p S2 ⋅ p N∕S2 (5)

p(N) = 0.45
Based on these data we can draw the multi period decision tree (see Fig. 5).
To find the expected value, we have to combine the payoffs and probabilities for
each node. The higher the expected value are (EV), the better will be the particular
decision alternative on average when compared to other alternatives in the decision tree.
To calculate the EV for each chance node, we have to sum all the EVs for each branch
that is connected to the node. For example, we have take again a look at the above
calculated chance nodes 6, 7 and 8, and we calculated further the other EVs.
Furthermore, we calculate the expected value in chance node 2:
VE = 618,200 × 0.55 + 533,400 × 0.45 = 580,040 euro.
So, concluding the obtained levels of the expected values using the revised proba‐
bilities calculated with the Bayes’ Theorem, the final expected payoff obtained when
taking into account de advice of a consultant is 580,040 euro. This value is higher than
the expected payoff obtained when our manager will take his own decision (580,000
euro). At the end, when the manager follows the advice of the specialist he will expand
 An Application of the Multi Period Decision Trees 553

p(S1|F)= 0.891
500,000 €
A1
9
EV1= 478,200 € p(S2|F)= 0.109
cost= 30,000 € 300,000 €

p(S1|F)= 0.891
700,000 €
p(F)= 0.55 Favorable Report A2
4 10
D0= 678,200 € EV2= 678,200 € p(S2|F)= 0.109
D0-cost= 618,200 € cost= 60,000 € 500,000 €

p(S1|F)= 0.891
400,000 €
A3
11
EV3= 378,200 € p(S2|F)= 0.109
cost= 40,000 € 200,000 €

2
p(S1|N)= 0.467
500,000 €
A1
12
EV1= 393,400 € p(S2|N)= 0.533
cost= 30,000 € 300,000 €

p(S1|N)= 0.467
700,000 €
A2
p(N)= 0.45 Unfavorable Report 13
5 EV2= 593,400 € p(S2|N)= 0.533
D0= 593,400 € cost= 60,000 € 500,000 €
D0-cost= 533,400 €

p(S1|N)= 0.467
400,000 €
A3
14
EV3= 293,400 € p(S2|N)= 0.533
cost= 40,000 € 200,000 €

Fig. 5. The proposed multi period decision tree

the business, opening a new business unit in the neighboring county, equipping this
workstation with an environmentally friendly incinerator, able to cover the medical
waste incineration needs for more than two counties.

4 Results and Interpretation

The decision tree allows the assessment and comparison of the potential results of a
strategic decision, by predicting the future events in a complex diagram. If the tree
represented in Fig. 4 is a simply mono-sequential one, it must be shown that a multi
period tree, a complex one, adapted to the strategy development process can be inter‐
preted as a reunion of simple trees, each of them representing one period of time. The
complex multi-sequential (multi period) tree may be extended to the limit of the capaci‐
ties of prediction (see Figs. 5 and 6).
554 L.M. Csorba and M. Crăciun

p(S1)= 0.7
500,000 €
A1 6
EV1= 440,000 € p(S2)= 0.3
cost= 30,000 € 300,000 €

p(S1)= 0.7
700,000 €
A2
7
Manager's own decision EV2= 640,000 € p(S2)= 0.3
3 cost= 60,000 € 500,000 €
EV2= 640,000 €
D0= 580,000 € p(S1)= 0.7
400,000 €
A3 8
EV3= 340,000 € p(S2)= 0.3
1 D0= 580,040 € cost= 40,000 € 200,000 €
p(S1|F)= 0.891
500,000 €
A1
9
EV1= 478,200 € p(S2|F)= 0.109
cost= 30,000 € 300,000 €
p(S1|F)= 0.891
Decision of the consultant p(F)= 0.55 Favorable Report 700,000 €
580,040 € 4 A2 10
D0= 678,200 € EV2= 678,200 € p(S2|F)= 0.109
D0-cost= 618,200 € cost= 60,000 € 500,000 €
p(S1|F)= 0.891
2
400,000 €
A3 11
EV3= 378,200 € p(S2|F)= 0.109
cost= 40,000 € 200,000 €
p(S1|N)=0.467
500,000 €
A1
12
EV1= 393,400 € p(S2|N)=0.533
cost= 30,000 € 300,000 €
p(S1|N)=0.467
700,000 €
p(N)= 0.45 Unfavorable Report A2 13
5 EV2= 593,400 € p(S2|N)=0.533
D0= 593,400 € cost= 60,000 € 500,000 €
D0-cost= 533,400 € p(S1|N)=0.467
400,000 €
A3 14
EV3= 293,400 € p(S2|N)=0.533
cost= 40,000 € 200,000 €

Fig. 6. The proposed final multi period decision tree

Analyzing the simple mono-sequential tree, we compare the expected value obtained
in each chance node, more exactly 440,000 euro (A1), 640,000 euro (A2) and 340,000
euro (A3). The cost of choosing the first alternative is 30,000, so the final expected value
will be 410,000 euro. In the case of the second alternative, the implied costs are 60,000
euro, so, the final expected value will remain 580,000 euro. For the last alternative, the
costs are 40,000 euro, so the final expected value will be 300,000 euro. Because the
highest expected value is 580,000 euro, the manager will opt for the second alternative,
when he takes his own decisions without the support of a consultant.
 An Application of the Multi Period Decision Trees 555

In the case the manager will ask the consultant for a piece of advice, he may obtain
a favorable report which will confirm that when he decided based on his own experience
that alternative A2 will be the best choice, he had right. We can see in Fig. 5 the final
expected value obtained after we diminished the expectance values with each implied
costs, as we have also done in the case of the mono period tree.
Because the highest final expected value is 580.040 euro, the manager must choose
again the second alternative, when he takes decisions with the support of a consultant.

5 Conclusions

Romania is increasingly buried in all kinds of waste: municipal, medical, toxic, and so on.
This happens not only because the Romanian people produce a huge quantity of waste
impossible to be processed by the landfills and incinerators they dispose, but also because
important quantities of legal or illegal waste are imported in Romania from abroad. As
stated in [25], this situation has a negative impact on human health, because, at first, it
affects the natural environment where human beings live. To solve this problem - as a
part of the future waste management - it is needed to build a great number of sustainable
landfills, to improve the recycling and reuse of waste and not a least, to encourage entre‐
preneurs to invest in sustainable incineration businesses. Regarding the last issue, our
article aims to be a support for an investor who needs to take the right decision in the field
of eco-incineration of medical waste. Of course, our prediction model can be extended at
all types of waste. Even if our prediction gave us in a particular manner the same results,
in the case when our manager chose to make his own decision or to ask a consultant for a
proper advice, it can be used furthermore to create a software easy to use by each manager
interested in, a soft which needs as inputs only the alternatives, the state of natures, de
payoffs and the probabilities for each chance.

Acknowledgements. The authors are grateful for the support offered by the colleagues involved
in the present research paper.

References

1. http://www.wrfound.org.uk/articles/incineration.html. Accessed 30 May 2016

2. http://www.energyjustice.net/incineration. Accessed 25 May 2016
3. http://www.wastemed.com/types.htm. Accessed 23 May 2016
4. http://www.inciner8.com/medical-incinerators.php. Accessed 25 May 2016
5. http://www.eschooltoday.com/waste-recycling/types-of-waste.html. Accessed 23 May 2016
6. Environmental Data Centre on Waste, Eurostat. http://ec.europa.eu/eurostat/web/
environment/waste
7. Rokach, L., Maimon, O.: Data mining with decision trees: theory and applications. World
Scientific Pub Co Inc, Singapore (2008)
8. Kast, R., Lapied, A.: Economics and finance of risk and of the future. Wiley, New York (2006)
9. Luce, R.D., Raiffa, H.: Games and decisions. Courier Corporation, North Chelmsford (1957)
10. Quinlan, J.R.: Simplifying decision trees. Int. J. Man Mach. Stud. 27(3), 221 (1987). doi:
10.1016/S0020-7373(87)80053-6
556 L.M. Csorba and M. Crăciun

11. Diallo, M., Jindal, D., Lin, H.: Decision Modeling in Software Engineering Projects (2000)
12. Gatta, S.R.: Decision tree analysis and risk modeling to appraise investments on major oil
field projects. Presented at the Middle East Oil Show and Conference, Bahrain, 20–23
February 1999. http://dx.doi.org/10.2118/53163-MS
13. Damodaran, A.: Applied Corporate Finance. Wiley, New York (2011)
14. Bae, J.M.: The clinical decision analysis using decision tree, Epidemiol. Health Rev. (2014). https://
doi.org/10.4178/epih/e2014025, http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4251295/
15. Podgorelec, V., Kokol, P., Stiglic, B., Rozman, I.: Decision trees: an overview and their use
in medicine. J. Med. Syst. 26(5), 445–463 (2002)
16. Vallverdu, J.A., Riano, D., Bohada, J.: Improving medical decision trees by combining
relevant health-care criteria. J. Expert Syst. Appl. 39(14), 11782–11791 (2012)
17. Magee, J.F.: Decision trees for decision making. Harvard Bus. Rev. 42(4), 126–138 (1964)
18. http://www.firme.info/i-f-tehnologii-srl-cui13501763/. Accessed 21 May 2016
19. http://www.ziardecluj.ro/spitalul-clujana-sparge-monopolul-isi-construieste-incinerator-
pentru-deseurile-medicale. Accessed 20 June 2016
20. http://www.risco.ro/verifica-firma/iridex-group-import-export-cui-398284. Accessed 20
June 2016
21. http://firme.info/eco-fire-sistems-srl-cui14356289/. Accessed 20 June 2016
22. http://www.firme.info/mondeco-srl-cui13800021/. Accessed 20 June 2016
23. http://ec.europa.eu/eurostat/statistics-explained/index.php/. Accessed 22 May 2016
24. Ionescu, G., Cazan, E., Negruţa, A.L.: Modelarea şi optimizarea deciziilor manageriale.
Editura Dacia, Cluj (1999)
25. Geman, O.A: Markov model for analyzing the health status of the Romanian population until
2020, E-Health and Bioengineering Conference (2013)
Containerizing Solution for a Highly Distributed Systems

Vlad Iovanov1, Mihai Buzgau2, and Dominic Bucerzan3 ✉

( )

1
Hewlett Packard Enterprise, Seattle, WA, USA
[email protected]
2
Hewlett Packard Enterprise, Arad, Romania
[email protected]
3
“Aurel Vlaicu” University of Arad, Arad, Romania
[email protected]

Abstract. Containerization technologies draw attention in modern IT industry.

Docker (the open source project), is beginning to be the base for the development
projects of most major IT vendors, on both Linux and Windows operating
systems. With the growing popularity of operating system isolation, in this paper
we propose a solution for containerizing highly distributed systems. The solution
is meant to be automated, using data gathered from deployment systems aimed
at classically virtualized environments. We also borrow from the 12-factor app
pattern, which has its own benefits.

Keywords: Operating system security · Containerization · Docker · Fissile ·

BOSH · Cloud foundry

1 Introduction

Operating-system-level virtualization is a server virtualization method in which the

kernel of an operating system allows the existence of multiple isolated user-space
instances, instead of just one. Such instances, which are sometimes called containers,
software containers, virtualization engines (VEs) or jails (FreeBSD jail or chroot jail),
and may look and feel like a real server from the point of view of its owners and users.

1.1 What Is a Container?

First things first: What is a container, and how is it different than a virtual machine?
We’re all familiar with hypervisors at this point – the hypervisor provides full hardware
virtualization, emulating compute, network, and storage resources for multiple Guest
OSes running on a physical machine. The Guest OSes running in a VM are essentially
the same as the equivalent OS running on a physical machine: they expect to be managing
physical hardware, and they run all of the services necessary to do so – including an OS
kernel, drivers, memory and process management.
Figure 1 shows a comparison between containers and virtual machines.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_41
558 V. Iovanov et al.

Fig. 1. Containers vs. VMs [1]

While this enables strong isolation between VMs, allows customers to increase the
utilization of their hardware, and provides an environment that is compatible with most
legacy workloads – there is a ton of effort duplication involved. Because each VM runs
a full OS, the size of Virtual Machines is fairly large: typically tens of gigabytes.
Each VM is also doing a lot of duplicate work, since each thinks it is exclusively
responsible for managing physical hardware. Incidentally, this is the reason that storage
deduplication works so well in VM environments: there’s a lot of duplication to remove.
On the other hand, containers share a single OS and kernel, as depicted in the right
side of the diagram. Isolation between containers is achieved within the host kernel,
which provides namespace segregation for compute, network, and filesystem resources,
along with control groups (cgroups) to limit resource consumption. These are technol‐
ogies that have been in the Linux kernel for years, but have been difficult for mere mortals
exploit until now, with the advent of container management tools like Docker and
Rocket.
Since containers share the host OS and kernel, they are much more lightweight than
VMs. Applications running in containers run at bare metal speeds, and their footprints
are typically measured in megabytes versus gigabytes. One way to think of containers
relative to virtual machines is to think of them as providing process/application virtu‐
alization, rather than OS virtualization.
Another benefit of containers over virtualization is startup time. Virtual Machines
tend to take a while to start up, because the application can’t start until the Guest OS
boots. At best, you’re probably looking at double-digit seconds. Containers, on the other
hand, generally start in milliseconds – because the OS/kernel is already booted.
This is one reason why live migration (e.g. Vmotion) is not a critical requirement
for containers: if a container dies, restarting it is trivial.
On the other hand, containers lack some of the flexibility of virtual machines. With
a hypervisor, you can run any x86/x64 OS you’d like within a VM. But since a container
shares a single kernel with its host OS, you’re stuck with that kernel; you can’t run a
Windows container on top of a Linux kernel.
 Containerizing Solution for a Highly Distributed Systems 559

Containers, as a fledgling technology, are also much less mature and polished than
their hypervisor counterparts. Security is another concern in the container world –
because the kernel is shared, the attack surface is broader. If an attacker gains access to
the shared kernel, he effectively has access to all host containers.
Over the past two years, Docker’s Linux container technology has revolutionized
the way developers build, ship, and run distributed software applications.
In the past year alone, the project has experienced a 183% increase in contributors,
515% growth in projects on GitHub, and an astounding 18,082% spike in container
downloads. Naturally, this massive project growth and enterprise adoption has also led
to a 1,720% jump in Docker job openings. And it’s not hard to see why the demand is
so high after attending any of the enterprise use case sessions from companies like
Capital One, the New York Times Newsroom, GrubHub, and Orbitz. Not too shabby
for having just celebrated its two-year anniversary [2].
Operating-system-level virtualization is a server virtualization method in which the
kernel of an operating system allows the existence of multiple isolated user-space
instances, instead of just one. Such instances, which are sometimes called containers,
software containers, virtualization engines (VEs) or jails (FreeBSD jail or chroot jail),
and may look and feel like a real server from the point of view of its owners and users.
Docker’s ease of use is due to its high-level API and documentation, which enabled
the DevOps community to dive in full force and create tutorials, official containerized
applications, and many additional technologies. By lowering the barrier to entry for
container technology, Docker has changed the way developers share, test, and deploy
applications [3].
How can Docker help? Well, developers can now package up all the runtimes and
libraries necessary to develop, test, and execute an application in an efficient, standar‐
dized way and are assured that it will deploy successfully in any environment that
supports Docker [3].

1.2 Portability
Containers allow applications, along with their dependencies (e.g. libraries and envi‐
ronmental state), to securely share a single OS in a lightweight, portable deployment
model. This allows applications to be written on, say, a developer’s laptop, and then
easily deployed onto a physical server, VM, or a private/public/hybrid cloud, without
having to worry that the application will behave differently in production than in dev/
test.
Docker for Mac and Docker for Windows are an integrated, easy-to-deploy envi‐
ronment for building, assembling, and shipping applications from Mac or Windows.
Docker for Mac and Windows contain many improvements over Docker Toolbox [4].
560 V. Iovanov et al.

2 Proposed Solution

BOSH has good primitives and contracts, with some exceptions.

Having the correct primitives - packages that get compiled and versioned based on
their signature, jobs that dictate configuration can be immensely helpful in making
automation easy.
The proposal is to use the formalism BOSH requires when implementing a project
using its tools to generate Docker images automatically.

2.1 Fissile and BOSH

During the Cloud Foundry Summit we were very excited to talk about Fissile. Fissile is
a tool that allows us to take BOSH releases and build containers out of them. Being able
to change a BOSH release into a simple Docker container has a number of advantages.
First off, we are no longer coupled to BOSH and that means that suddenly anything
that was built as a BOSH release can now be deployed without it. Which leads us to the
next great benefit is that we can use anything that can deploy and manage a container.
This unlocks Kubernetes, Docker Swarm, CoreOS and other container management
systems for us to deploy our BOSH releases on.
Suddenly anything that previously was only deployable using BOSH can take
advantage of the tremendous momentum behind containers in the cloud ecosystem
today.
We had to dig through the internals of stem cell creation to build base layers. These
may change over time, which means humans will need to re-compile fissile in order to
maintain compatibility with the latest BOSH Releases.
Because of the good contracts, software packaged as BOSH release can now enjoy
a multitude of new deployment targets.

2.2 Dealing with Tight Coupling

One thing that’s not apparent when you start using Cloud Foundry is that it’s actually
tightly coupled to BOSH. Cloud Foundry components are composable and have well
defined boundaries and APIs but ironically the lines between BOSH and Cloud Foundry
are blurred (see Fig. 2).
An example of this is BOSH templates which are written in ERB meaning that only
Ruby-like tools can parse them. ERB templates contain control scripts which are equiv‐
alent to Cloud Foundry templatized code which are based on BOSH primitives.
Additionally, these templates include logic, functions, and even Ruby class defini‐
tions which have slipped into them over time. All of this makes porting this to another
system infeasible.
Furthermore, we believe that having choice is important and the fact that Monit is
the only service daemon that you can use is not ideal.
 Containerizing Solution for a Highly Distributed Systems 561

Fig. 2. BOSH and cloud foundry

2.3 Image Layers, Compilation and the Entry Point

Fissile on the other hand consumes the BOSH release contract and produces images.
It’s interesting to see what makes up one of these Docker images so let’s take a dive into
one. All of our containers start from an Ubuntu 14.04 Docker image. On top of this
image we add a stem cell layer (runtime dependencies) and on top of that we add the
BOSH packages and jobs. We also provide an entry point as well as a way to provide
configuration (see Fig. 3).

Fig. 3. Fissile architecture

One interesting difference from BOSH is that all the compilation is done before you
deploy. In BOSH this compilation is managed by the director as you deploy the different
components. Doing this at build time saves time during deployment. Fissile leverages
Docker and Go to compile all the packages on your machine in parallel which can help
speed up the process as well.
In addition to being highly parallel Fissile also does smart detection of dependencies
and will build only the artifacts that are required by what’s mentioned in the Role Mani‐
fest. Another difference from BOSH is that we separate runtime and compilation
dependencies, so the stem cell layer does not contain build-time libraries and the compi‐
lation layer doesn’t contain any runtime dependencies.
In BOSH there is the manifest that it needs in order to deploy but in Fissile there are
three key components to building the container images. First, we have the role manifest
which contains a list of all the Docker Images we want to build (and what they’re made
of) as well as some configuration setup.
562 V. Iovanov et al.

3 Using the Solution - Cloud Foundry Containerized

Then come the opinion files which are essentially BOSH deployment manifests with
only a properties section. The “Light” opinions file contains defaults for properties the
end-user will never touch. “Dark” opinions are a failsafe – we use them to erase any
defaults that may exist in the system for secrets (e.g. user passwords).
The last part that we need to include is all of our BOSH releases that we’ve referenced
in our Role Manifest. Once we have all of these inputs we can use fissile to create Docker
images!
Finally, a note on using Fissile in order to run Cloud Foundry; first you have to
configure it. However, this is more daunting than you might think. There are over 500
configuration settings in the cf-release spec files alone (see Fig. 4).

Fig. 4. Fissile’s configuration mechanisms

Because Fissile’s configuration mechanisms allows us to hide some and templatize

other configuration values we’re able to whittle this down to about 100 configuration
values in a deployment that we think users will actually care enough to set (see Table 1).

Table 1. Description
Name # Description
Specs 500 Each job has spec files
Opinions 200 Our curated defaults which started from BOSH manifest generation
User global 90 Values set by a user specific to their environment
User role 20 Values set by a user specific to their environment AND a specific VM

We’ve talked about some of the benefits, as well as what the process looks like to
change a BOSH release into a container and what those containers are made of. We’d
like to share a proof of concept demonstration we’ve created to illustrate a (very nearly)
fully highly available Cloud Foundry instance running on Kubernetes, but again you
could run it on your choice of Docker orchestration tool – even plain old Docker itself!
 Containerizing Solution for a Highly Distributed Systems 563

We think this is an innovative approach to deploying Cloud Foundry and we’re keen
to see it grow into its own and therefore we’re open sourcing Fissile and Configgin. In
the past having multiple implementations to a single spec (the BOSH contract) means
that both implementations stay true to the promises that have been made as well as
provide some needed flexibility.
We’re excited for the community to get their hands on Fissile and we’re planning
on really flushing out some documentation and examples in the coming weeks. Before
that please feel free to play around and see what you can do with it.
This solution uses containers in containers. This is not ideal, because it requires one
of the images (the role that represents the elastic runtime) to run in privileged mode. A
more secure and elegant solution would be to run the Cloud Foundry application
containers on the same scheduler that runs Cloud Foundry itself.
This project proves that in the end, Cloud Foundry can be run on its own without
BOSH. It may take a lot of man hours, but the net result is that there is no technical
impediment.
The solution is not idiomatic Docker - we run multiple processes per container, and
one could say we’re using Docker containers like you would a VM.
The service discovery mechanism we’re using is not pluggable. Environment vari‐
ables are used to point each container to the location of other roles.

4 Conclusions

For a long time, containerization was understood by a small minority of insiders and
unavailable to the vast majority. Docker has changed that irrevocably.
The ready availability of containerization technology massively increases the
economic advantage of microservices-based application architectures.
Virtualization is turning into a technology ghetto where legacy applications go to
live out their remaining days.
Automation of conversion is important for large systems, otherwise it’s difficult to
find the resources to achieve something like this. The Cloud Foundry community would
need to spend a lot of man hours to try and make a project like this a reality from scratch.
Automating the conversion takes us into the right direction and proves that a solution is
possible.
Reasoning further from the previous point, we can conclude that it’s required that
the deployment and configuration mechanism of the software is described in a machine
readable format.
For security reasons in the future develop of the project we will study the opportunity
to integrate up to date cryptographic services [5, 6].
Furthermore, projects that follow best practices for configuration and service
discovery are more likely to result in a successful project.

Acknowledgement. Thanks to all friends and colleagues whom had help to develop this work.
564 V. Iovanov et al.

References

1. http://www.rightscale.com/blog/cloud-management-best-practices/docker-vs-vms-combining-
both-cloud-portability-nirvana. Accessed 18 April 2016
2. http://osdelivers.blackducksoftware.com/2015/06/24/dockercon-2015-stacking-containers-of-
excitement-and-announcements/. Accessed 15 May 2016
3. https://insights.sei.cmu.edu/devops/2015/01/devops-and-docker.html. Accessed 18 April 2016
4. https://blog.docker.com/2016/03/docker-for-mac-windows-beta/. Accessed 21 June 2016
5. Bucerzan, D., Raţiu, C.: Contributions to Steganographic Techniques on Mobile Devices. In:
Proceedings of the Security for Information Technology and Communications, LNCS, vol.
9522. Springer (2015)
6. Bucerzan, D., Raţiu, C., Manolescu, M.J.: SmartSteg: a new android based steganography
application. Int. J. Comput. Commun. Control (IJCCC) 8(5), 681–688 (2013)
 Data Acquisition and Processing with Fusion Sensors,
Used in Algorithms for Increasing Accuracy
of Position Measurement of Mobile Robots

Sandru Florin1, Nanu Sorin1 ✉ , Negrea Romeo2, Popescu Anca1, and Fericel Sergiu1
( )

1
Faculty of Automation and Computers, Politehnica University Timisoara,
Bv. V. Parvan 2, 300223 Timisoara, Romania
[email protected], [email protected],
[email protected], [email protected]
2
Department of Mathematics, Politehnica University Timisoara,
Victoria Sq 2, 300030 Timisoara, Romania
[email protected]

Abstract. For a mobile robot movement, the position measurement using

internal reference sensors are of low accuracy. Encoder measurements are
affected by skid, accelerometer results are affected by noise and final calculation
by additive errors. In this paper is presented a method for acquiring and processing
data in order to be used for the correction to be performed in order to determine
the position of a mobile robot with increased accuracy. The used sensors in the
proposed system are accelerometer and encoders, while the perturbation filtering
is made with Kalman method. The data obtained from sensors are processed and
analyzed, in order to reduce the measurement error.

Keywords: Mobile robot · Accelerometer · Encoder · Fusion sensors · Kalman

filter · Data processing

1 Introduction

The position of a mobile robot is defined by three parameters: x coordinate, y coordinate

and orientation angle relative to one axis. The position can be determined with internal
and external reference sensors. The most used sensors with internal reference are
encoder, accelerometer, compass and gyro. When in move, the position is very difficult
to be calculated due to the particularities of the sensors, the computing methods, etc.
The encoders that are mounted on the wheels or on the actuators have the problem
of skid. If one point of the wheel is sliding on the surface, the wheel is not rotating,
consequently the encoder is not rotating, even if the mobile is changing the position.
The system called ESP (Electronic Stability Program) provides a solution for the
problem of skid, but in a more advanced way, with help from the Antilock Braking
System (ABS) and the Traction Control System (TCS). The ABS prevents the wheels
from locking during braking and the TCS prevents the wheels from spinning when
starting off and accelerating.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_42
566 S. Florin et al.

While ABS and TCS intervene on a vehicle’s longitudinal dynamics, ESP addition‐
ally improves the lateral dynamics, thus ensuring stable driving in all directions. But
this cannot be implemented on a mobile robot.
The accelerometer provides the momentarily radial and tangential acceleration, of
one point of the robot. In order to determine the position, this values have to be integrated
twice, the operation generates huge additive errors, reducing dramatically the accuracy
of the measurement. Moreover, the position cannot be determined from information of
only one accelerometer.
The paper presents a hardware and a software structure built around a mobile robot,
to enable acquisition of information from two encoders and one 2-D accelerometer. The
position in some points of the trajectory are calculated using different methods. The data
from accelerometer was filtered with a Kalman filter. The points obtained are compared
and a mathematical analysis is performed in order to find algorithms to improve accuracy
of the measurement.

2 Calculus of Preliminary Data

The robot is a 3-D physical object that moves in 2-D space. Its position is defined by
three parameters: x, y, θ that represents position of a point P that replaces physical object,
on x axis, y axis, respectively angle θ relative to x axis (Fig. 1). In order to obtain the
position of the robot (x, y, θ), primary data from sensors (number of pulses from
encoders, radial and tangential acceleration from accelerometer) are processed obtaining
preliminary data (rotation angle of robot, speed of P, acceleration of P). In the third
phase, the x, y, θ values are obtained for every moment of calculation [2] (Fig. 2). Values
are analyzed and compared to real values of position and a processing is performed in
order to reduce the differences between calculated data and real data.

Fig. 1. Parameters of robot position

 Data Acquisition and Processing with Fusion Sensors 567

Fig. 2. The diagram of operations

3 Geometry of Robot

The robot has a rectangular 16 cm × 12 cm platform, two DC motors in front, with

integrated encoder and wheels mounted on shaft. The rear mount is a plastic stem that
keeps the robot in horizontal position. The encoders are placed on both wheels. The
geometrical parameters of robot is presented in Fig. 3.

Fig. 3. View of the robot

The accelerometer used has three axis of sensing, with a resolution of 13 bit. It was
placed between the wheels in the middle of the distance (Fig. 3), point P. Further on,
the point P will be considered the material point, replacing the rigid body of the robot.
The parameters x, y of robot position will be considered the x, y values of P.
The geometrical parameters of the robot are:

- D - distance between central points of wheels

- R - wheel radius (1)
- N - number of pulses per rotation of encoder
568 S. Florin et al.

4 Trajectory Parameters Obtained from Encoders

If the trajectory is a straight line, the speed of the wheels and of the point P are equal,
computed with the formula [3]:

2∗𝜋∗R∗n
v= (2)
N ∗ Δt
Where n is the number of pulses from encoder in time Δt.
If the trajectory is a curve line, robot moving from 1 to 2, the linear speed of the
wheels vL and vR are not equal. The wheels are moving on two circles with radius r − D/
2 for the right wheel, and r + D/2 for the left wheel [1] (Fig. 4).

Fig. 4. Trajectory geometry of a turning robot

The angular speed ω of the points P is valid for both wheels:

vL vR
𝜔= , 𝜔=
D D (3)
r+ r−
2 2
The radius r of the circle is obtained from:

vL = 𝜔(r + D∕2)
(4)
vR = 𝜔(r − D∕2)

Eliminating ω in (4), we obtain

vL + vR D
r= ⋅ (5)
vL − vR 2
 Data Acquisition and Processing with Fusion Sensors 569

The speed of point P is

vL + vR
vP = (6)
2
The change angle α of robot from (1) to (2) is calculated:

α = 𝜔 ⋅ Δt (7)
Considering known the start angle θ0, the actual angle θ is calculated iteratively: adding
α for every moment of measurement to the initial angle θ0.
Considering change in x (Δx) and y (Δy) position from a moment of measurement
to the next moment of measurement (Δt)

Δx =vP ⋅ Δt ⋅ sin(𝛼)
(8)
Δy =vP ⋅ Δt ⋅ cos(𝛼)

where vP is obtained from (6) and (2).

Absolute value of x and y are obtained adding cumulatively every Δx and Δy to x0
respectively y0.

5 Trajectory Parameters Obtained from Accelerometer

The outputs of accelerometer are the acceleration of the point P on the direction longi‐
tudinal to robot, named ax, and perpendicular to robot, named ay [4–6]. When moving
on the trajectory, ax becomes tangential acceleration at, while ay becomes radial accel‐
eration ar (Fig. 5).

Fig. 5. Accelerations measured by accelerometer

The speed vP is computed with formula:

ΔvP
at = (9)
Δt
570 S. Florin et al.

v2P
ar = (10)
r
The variable r is the radius of trajectory and cannot be obtained from information
provided by accelerometer. It is calculated in (5) from information of encoders.
In case of constant speed, vP = constant, and the trajectory is linear at = 0, ar = 0 vP
is unchanged. In case of constant speed, vP = constant, and the trajectory is circular at = 0,
ar ≠ 0 and the value of vp is obtained from (9). When vP is not constant and trajectory is
circular at ≠ 0, ar ≠ 0, vP is increasing from vP to vP + ΔvP.
The absolute values of position are calculated as in (7) and (8) using information
from encoders.

6 Kalman Adjustment

The Kalman filter is considered one of the most important tools in the field of statistical
estimation theory [7].
First theorized by Rudolf E. Kálmán in 1960 in the paper “A new approach to linear
filtering and prediction problems” the algorithm was very well received due to the large
spectrum of fields it can be applied but also due to the relative simple mathematical
structure it has [8].
At its basic form it is an algorithm for solving the linear quadratic problem, this
problem consists in estimating the state of linear dynamic system on which white noise
is acted on through a series of measurements (also affected by white noise) related to
the state.
The filtering algorithm is in most interpretations divided into two separate stages:
the prediction phase and the update phase although it may be reformulated into a single
step execution of booth phases.
Given the following linear dynamic system containing white noise.

xt = Ft ∗ Xt−1 + Bt ut + wt (11)

where xt is the state vector containing the terms of interest for the system, ut is the vector
containing any control input, Ft is the state transition matrix which applies the effect of
each system state parameter at time t − 1 on the system state at time t, Bt is the control
input matrix, wt is the vector containing the process noise.
The measurement samples for the system can be written in the following form

z t = H t xt + vt , (12)

where zt is the vector of measurements, Ht is the transformation matrix that maps the
state vector parameters into the measurement domain, vt is the vector containing the
measurement noise terms for each observation in the measurement vector [9].
The prediction phase consists in the computation of the estimated state and the esti‐
mate covariance.
 Data Acquisition and Processing with Fusion Sensors 571

The update phase consists in the computation of the Kalman Gain, update of the
estimate and covariance and the projection onto k + 1.
Figure 6 [10] shows a visual representation of the algorithm.

Fig. 6. Kalman filter algorithm

The filter is recursive in its nature therefore the prediction and the update phase will
be reiterated as necessary.
Although the Kalman filter is not a direct solution to a problem it is a great tool for
estimation and performance analysis of estimators.

7 Experimental Results

The experiments took place using a physical robot with encoders integrated in DC motor,
accelerometer ADXL345, ARM® Cortex™-M4 STM32F407ZG microcontroler board.
Δt = 1 s.
The results were transmitted through wi-fi to another identical board that commu‐
nicated with a PC through USB line. The data was processed on-line. The robot was
programmed to follow a predefined trajectory [7].
The results of the measurements and processed data using algorithms are presented
in Fig. 7. Trajectories are in (x, y) coordinates. The orientation of robot is not presented.
a. Trajectory (xc, yc) is the real trajectory followed by robot, and was recorded with a
pen drawing on the floor
b. Trajectory (xe, ye) is obtained from encoders
c. Trajectory (xa, ya) is obtained from accelerometers
d. Trajectory (xk, yk) is obtained from accelerometers with data filtered with Kalman
filter
572 S. Florin et al.

Fig. 7. The trajectories of the robot

Figures 8, 9 and 10 present the variation in time of the errors between values of
trajectory calculated from encoders, accelerometers respectively accelerometers filtered
and the real position.

Fig. 8. Errors between position calculated and real position (encoder information)
 Data Acquisition and Processing with Fusion Sensors 573

Fig. 9. Errors between position calculated and real position (accelerometer information)

Fig. 10. Errors between position calculated and filtered with Kalman filter and real position

8 Results Analysis

We try to find a (statistical) relation between the real coordinate (xc, yc) and the corre‐
sponding coordinate given by encoder (xe, ye), acceleration (xa, ya) and given by filtered
(Kalman) acceleration (xk, yk).
First, we made an analysis of linear models. We obtain very good linear models for
the abscise coordinate for all three sensors. For all models the P-value of t-test and F-
test are less than 0.00001. Also, the coefficient of determination R^2 has very good
values [10–14].
For the encoder, the coefficient of determination R^2 is 0.9391 and average of resi‐
dues is 1.727014e-16. The graph from Fig. 11 represent the linear model in this case.
574 S. Florin et al.

Fig. 11. Relation between xe and xc

Similar, for the acceleration, we obtain the coefficient of determination R^2 is 0.9611
and average of residues is 9.854193e–17. The graph from Fig. 12 represent the linear
model in this case.

Fig. 12. Relation between xa and xc

Respectively, for filtered acceleration, we obtain the coefficient of determination R^2

is 0.9654 and average of residues is 4.254891e–16. The graph from Fig. 13 represents
the linear model in this case.
 Data Acquisition and Processing with Fusion Sensors 575

Fig. 13. Relation between xk and xc

Unfortunately, this good aspect (i.e., a simple relation between the real abscise coor‐
dinate and the abscise coordinate given by sensors), is not verify for the second (ordinate)
coordinate. The P-values for t-test and F-test are less than 0.001 but he values for
coefficient of determination R^2 are not so good.
For example, for the encoder, we have R^2 = 0.704. The graph of linear models
(Fig. 14) suggest a non-linear regression.
Similar, for the second sensor, we have R^2 = 0.4472 and P-value for F-test is greater
than 0.01, the graph of this linear model (Fig. 15) and respectively, for the last case,
R^2 = 04528 and P-values for T-test and F-test are greater than 0.01. The graph of linear
model in this case (Fig. 16).

Fig. 14. Relation between ye and yc

576 S. Florin et al.

Fig. 15. Relation between ya and yc

Fig. 16. Relation between yk and yc

For the encoder, we obtain some good results for a cubic polynomial model,
R^2 = 0.9004 and the graph from Fig. 17 shows this non-linear models.
 Data Acquisition and Processing with Fusion Sensors 577

Fig. 17. Nonlinear relation between ye and yc

Probably, some polynomial models or some non-sophisticated nonlinear models can

be obtained for the second and third case.

9 Conclusions

Different methods have been proposed for combining various sensor information to
obtain a more accurate positioning in the presence of errors and disturbances. Even if
the use of an accelerometer for precision mobile robot control seems promising in terms
of both performance and cost-effectiveness, the proposed system is willing to give better
results than the solutions proposed before.
Given that the mobile robot trajectory, linear or curve, is influenced by several
factors: the physical properties of the surface, the wheels’ characteristics, the centre of
gravity or the weight distribution on every wheel, a better determination of the trajectory
is needed.
As shown in the proposed system, it concludes that using the sensors fusion: encoders
and accelerometer with error correction based on Kalman filtering will solve the
problem.
This paper submits the different aspects of reading the trajectory only from encoders
or accelerometer or both in the same time and shows the importance of reading the
correct values from accelerometer and correlate them with the encoders’ values for
having in the end the correct position of the mobile robot.
The application presented is a step in a more complex system for mobile robot posi‐
tion measurement with increased accuracy. More sensors will be added, analysing the
efficiency of using them. Moreover, it will be a part of a position control system for a
robot.
578 S. Florin et al.

References

1. Dudek, G., Jenkin, M.: Computational Principles of Mobile Robots, 2nd edn. Cambridge
University Press, New York (2010)
2. Grewal, M.S., Andrews, A.P.: Kalman Filtering: Theory and Practice Using MATLAB, 4th
edn. Wiley-IEEE Press, Hoboken (2015)
3. Henderson, T.: 1-D Kinematics. http://www.physicsclassroom.com/
4. Shim, H., Kochem, M., Tomizuka, M.: Use of Accelerometer for Precision Motion Control
of Linear Motor Driven Positioning System. University of California at Berkeley, Berkeley
(1998)
5. Zhu, W.H., Lamarche, T.: Velocity estimation by using imperfect accelerometer and encoder
for rigid contact modeling and control. In: Proceedings 2006 IEEE International Conference
on Robotics and Automation; ICRA; Orlando, FL, pp. 4234–4239 (2006). doi:10.1109/
ROBOT.2006.1642353
6. Zhu, W.H., Lamarche, T.: Velocity estimation by using position and acceleration sensors.
IEEE Trans. Ind. Electron. 54(5), 2706–2715 (2007). doi:10.1109/TIE.2007.899936
7. Nanu, S.: Educational aspects in control system design technology. Sci. Tech. Bull. 4(1), 584–
589 (2007)
8. Welch, G.F.: History: the use of the Kalman filter for human motion tracking in virtual reality.
Presence Teleoperators Virtual Environ. 18(1), 72–91 (2009)
9. Faragher, R.: Understanding the basis of the Kalman filter via a simple and intuitive derivation
[Lecture Notes]. IEEE Signal Process. Mag. 29(5), 128–132 (2012). doi:10.1109/MSP.
2012.2203621
10. Negrea, R.: Modelare statistica si stocastica. Aplicatii in inginerie si economie; Ed. Politehnica
Timisoara (2006)
11. Schittkowski, K.: Data Fitting in Dynamical Systems. Kluwer, Boston (2002). ISBN
1402010796
12. Bethea, R.M., Duran, B.S., Boullion, T.L.: Statistical Methods for Engineers and Scientists.
Marcel Dekker, New York (1985). ISBN 0-8247-7227-X
13. Ravishankar, N., Dey, D.K.: A First Course in Linear Model Theory. Chapman and Hall/CRC,
Boca Raton (2002)
14. Fotheringham, A.S., Brunsdon, C., Charlton, M.: Geographically weighted regression: the
analysis of spatially varying relationships; (Reprint ed.). Wiley, Chichester (2002). ISBN
978-0-471-49616-8
Modelling and Applications in Textiles
 Specialized Software for Tracking and Management
Machinery Used in Textile Industry

Pavel Farcaș1 ✉ , Ciprian Pop-Moldovan2, and Marius Constantin Popescu1

( )

1
Vasile Goldis Western University Arad, 94-96 Revolutiei Blvd., Arad, Romania
[email protected], [email protected]
2
Top Consulting, 32 Octavian Goga St., Arad, Romania
[email protected]

Abstract. This paper presents the design and implementation of tracking devices
used for textile industry - clothing. The control system is flexible and modular.
One of the most important aims in textile industry is to obtain constant quality.
Currently, in the Romanian textile industry the processes are manual or semi-
automatic, and products made in this way are not always highly qualitative.
Moreover, this industry has both skilled and unskilled workers. Therefore, close
monitoring is necessary. Due to the fact that classical automation solutions for
such industries are not always accessible is necessary to find an economic, reliable
and robust automation solution. This paper refers to the need for automation of
the textile machineries control and describes a system designed, developed and
implemented by the authors.

Keywords: Software · Monitoring · Control · Ethernet · PHP · MySQL

1 Introduction

Textile-garment industry in the European Union is concentrated in Italy, Germany,

France, Spain and Britain, countries that hold 3/4 of EU production. In countries such
as Portugal, Greece, Poland, Slovakia, Romania, Bulgaria and the Czech Republic the
textile sector plays an important role in the national economy realizing over 10% of
exports of manufactures. In 2013, 1774 companies were operating in Romania, not all
automated. From this, according to the Ministry of Finance, 30% had zero turnovers,
and more than half of total failed to make profit. In 2012, only 39 have exceeded a annual
sales of 5 million euros, and on the market there is no player with a turnover of over 100
million. A low-quality product leads to a lack of customer satisfaction and ultimately
loss. This is applicable also to the textile industries.
The latest generation of textile machinery is able to produce transformations of
classic products through effective use and embedding of sensors, processors and devices
in wearable products. Today, the textile product must create and provide the wearer
multifunctional features, sensors, actuators. Application field is very diverse: sport and
leisure, civil security and professional use, industrial textile, transportation, military,
home, construction and geotextiles, agriculture, packaging, fashion, entertainment and
games and “wearable computer”.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_43
582 P. Farcaș et al.

It is estimated that the market of technical textiles will record in 2016 an increase of
about 23%. Therefore it is necessary to automate this sector [1, 2]. Some companies use
a centralized control system SCADA [3] others PLC for driving of frame machines,
widely used in the textile [4, 5]. The product introduced, can be adapted to the specific
production lines and can be used on all process streams that require measurement of the
various processes in the industry. The device designed and produced by the authors is
able to monitor equipment and control them and to perform data recording in the data‐
base for predictive maintenance purposes and monitor production. The software
provided by the same team gathers analog and digital signals transmit them to the server
and offers necessary data for monitoring maintenance and production, improving
production efficiency and quality.

2 The Software

The cost of investment in this system depends on the possibility of measuring the
parameters of machinery/equipment (modular data acquisition existing or provided by
Top Consulting Ltd), configuration of network devices connected (General Packet Radio
Service - GPRS, 3G, 4G Long-Term Evolution LTE, Ethernet or industry standards)
and server/servers used. The software was designed to run on any operating system that
supports Structured Query Language - SQL (Oracle, MySQL, Microsoft SQL Server)
server and a Web Server (Internet Information Services - IIS, Apache) how Php: Hyper‐
textPreprocessor. The web interface is accessible on mobile devices (tablets/smart‐
phones with iOS or Android) via their browser. It is also available a client application
written for Android 4.0 and higher [6]. The tracking machines system is scalable, so it
can run on multiprocessor or mono-processor servers and the required memory is
minimum 16 GB RAM (Random Access Memory). It is recommended that the server
have at least two 1 GB network cards and minimum 1 TB storage capacity.
The monitoring system can be installed on Windows Server 2008 Standard and
higher, Linux Ubuntu 14.04 LTS (other versions of Linux) and Mac OS X, running IIS
or Apache PHP module with standard configuration. The database server can be Micro‐
soft SQL Server 2008 or higher, MySQL 5.5 or higher, Oracle v.10 or higher. The system
uses Open Database Connectivity (ODBC) for data connections. The implementations
made so far are using Windows 2008 Server with IIS, PHP 5.5 and MySQL 5.5. The
network between data collection devices is 1 GB Ethernet [7]. Data Collection devices
are provided with one RS232 port and a barcode reader Motorola Symbol LS2208 [8,
9]. In this case, this system was used for monitoring and control of machines in the textile
industry. The equipment together with the software can be provided with other sensors
types and programmable outputs to perform certain operations such as ventilation,
acoustic and/or visual warning etc.
 Specialized Software for Tracking and Management Machinery 583

3 Description of the Software Application

3.1 Application’s Menu

The application is composed of multiple software modules written in PHP 5.5 and ANSI
- American National Standards Institute - SQL which are interconnected. The applica‐
tion security is managed by a module that prohibits unauthorized access to other modules
and manages access levels for each user. Access to the application menu is based on
username and password (Fig. 1). Depending on the user rights the application will
display the respective menus. Management module options are located in the “User –
Utilizator” menu, these are: “Authentication – Autentificare (in Romanian)” (displays
the login screen), “Change Password - Schimbare parola” (enable logging password
change); “User Groups - Grupuri utilizatori” (for easier management of software users,
security groups are implemented - for example, the group of administrators, technicians,
managers, etc. - who can define, remove, edit by checking and unchecking the appro‐
priate fields).

Fig. 1. The “Users - Utilizatori” menu which allows user creation for the application

Exiting the application is done by logging off option which will close the user’s
session ensuring the deletion of logging information from browser. The menu “Nomen‐
clature - Nomenclatoare (in Romanian)” contains options for information management
from application tables. Selecting information from software’s tables can be done in
three different ways by selecting information from an available list, from a selection list
(which appears when characters are placed in that field) and by entering information
entirely in the particular field. In “Manage compartments - Gestionare secții” module
are defined compartments, respectively (Fig. 2): Production, Warehouse, Workshop and
grinding. For each of these we can define schedules that are needed to generate summa‐
ries and statistics for departments. “Management zones - Gestiune orare” lets you define
schedules used in the application, namely: the days of validity, periods (Fig. 3).
584 P. Farcaș et al.

Fig. 2. Interface for module “Management departments - Gestionare secții”

Fig. 3. Interface for module “Schedule management - Gestiune orare”

The “Days off Management - Gestiune zile libere” is a complement to the schedule

in which can be defined those days in which the unit does not work. In “Maintenance
management - Gestiune mentenante” are defined the maintenance processes and machi‐
nery adjustmens. For each maintenance is defined an unique barcode in application in
order to select the maintenance process from tracking device (Fig. 4).

Fig. 4. Interface of module “Maintenance management - Gestiune mentenante”

 Specialized Software for Tracking and Management Machinery 585

In “Machine types management - Gestiune tipuri masini (in Romanian)” mode are
defined the types of machines, respective the type of SME’s (are spare parts usually very
expensive) used and types of maintenance/adjustments and the period when have to be
executed (Fig. 5).

Fig. 5. Interface for module “Machine types management - Gestiune tipuri mașini”

In “Manage machine-specific parameters - Gestiune parametri specifici mașini” are

defined the modes in which behave (data acquisition device) the mounted machine
interfaces. Additionally defined parameters are: Trot min (minimum value between the two
signals received from sensor to be considered valid signal), vnl (nominal machine speed
whose value that can be learned by software), Dev (maximum deviation from vnl to
consider an abnormality of the machine functionality), Tdf (downtime accepted as
normal; if the machine does not work for a period of time greater than this value is
considered as defect) (Fig. 6).

Fig. 6. Interface for module “Manage machine-specific parameters - Gestiune parametri specifici
mașini”
586 P. Farcaș et al.

In “Devices Management - Gestiune module (in Romanian)” are managed interfaces

which gathers data from machine (Fig. 7). In “Machines Management - Gestiune mașini”
(see Fig. 8) are managed the machines monitored by our software. For each machine is
set: machine type, type specific parameters (for each type of machine we can define
more specific parameters depending on what is running on that machine), history of the
machine in question (the time needed for mounting a particular module on the machine,
in which location was positioned the machine).

Fig. 7. Interface for module “Devices Management - Gestiune module”

Fig. 8. Interface for module “Machines Management - Gestiune mașini”

In “Types of defects - Tipuri defecte (in Romanian)” are defined the types of defects
that may occur during operation of the machine (Fig. 9).
 Specialized Software for Tracking and Management Machinery 587

Fig. 9. Interface from module “Types of defects - Tipuri defecte”

In “Technicians Management - Gestiune tehnicieni” are defined technicians who are

getting a bar code to be registered in the database (Fig. 10). Machine operators can call
these technicians using the same codes. In the module “Personnel Management” are
defined all the machine operators who will work on the monitored machines.

Fig. 10. Interface from module “Technicians Management - Gestiune tehnicieni”

3.2 The Management Module

There are 3 ways of management, each with different characteristics, namely: “Purchase
- Achiziționare” (Fig. 11) involves taking in SME’s when they are purchased: the SME
code is suggested by the application and it’s an internal unique code; SME type (sono‐
troda, Enclume, Mulet); section in which part is taken (section type magazine) and orig‐
inal size (just sonotrode).

Fig. 11. Interface for module “Purchase – Achiziționare”

588 P. Farcaș et al.

Fig. 12. Interface for module “Manage – Gestionare”

Fig. 13. Interface for SME management module “Collect SME rectification - Preluare SME
rectificare”

Fig. 14. Interface for SME management module “Mounting/dismounting SME - Montare/
demontare SME”

After purchasing these SME’s are managed individually by the module “Manage‐
ment - Gestionare” (in Romanian). Depending on where the SME is can be mounted on
the machine/removed/sent to/returned from rectification/scrapped, etc. Also on this
screen you can see the number of rotations made by SME’s and running time (Fig. 12).
The Manage SME module has “Collect SME rectification - Preluare SME rectificare”
(Fig. 13) for taking simultaneous more SME’s from rectification (for sonotrode are
introduced the dimensions before and after rectification) and “Mounting/dismounting
SME - Montare/demontare SME” which allows mounting/dismounting SME’s on
machines in a centralized manner (Fig. 14).
 Specialized Software for Tracking and Management Machinery 589

3.3 Tracking Module

The module “Tracking - Urmărire (in Romanian)” includes options related to tracking
the daily activity of the machines. On “General panel - Panou general” is displayed in
real time the current status of the machine. Each machine is represented through a circle
with a label (machine name). Representation is entirely configurable for each of the
machine state and location (Fig. 15). Both the size and the colors for each state of the
machine can be customized. Status can be: operation (machine works), unused (machine
has not been active lately - parameterized), defective (machine was announced as
defect), intervention (the machine is being repaired), maintenance (machine was in
maintenance), auto-maintenance (machine is in auto-maintenance) settings (to set the
machine) communication problems (the device has lost contact with the server and can
not determine her condition) and learning (machine is running a learning procedure).

Fig. 15. Interface “General panel - Panoul general”

The panel “Technicians panel - Panou tehnicieni” is updated automatically in real

time about machines status which need technician’s intervention. “Shift Report - Raport
tura” shows only to the maintenance leader the machines situation. The display is similar
to the previous one, the technicians display (Fig. 16).

Fig. 16. Module “Report shift - Raport tura”

590 P. Farcaș et al.

In the module “Tracking – Urmarire (in Romanian)”, “Fault Announcement - Anun‐

tare defect” allows the announcement of a defect from the administrator console. The
“Reports – Rapoarte” module centralizes stored data, and contains “Reports technicians
- Rapoarte tehnicieni” (technicians situation), “Maintenance - Mentenante” (mainte‐
nance situation) and “Machine Reports - Rapoarte masini” (machine situation). It can
generate detailed reports (Fig. 17) and graphics reports.

Fig. 17. Module “Reports - Rapoarte”, generation detailed

3.4 Mobile Version of the Application

This app is used by technicians to gain mobility. Getting the data from the machine is
done via a device that has 4 digital inputs, a barcode reader, 4 buttons (+, − OK, Cancel
– Fig. 18a). The machine status is displayed with communication problems on the control
panel (Fig. 18b).

a) b)

Fig. 18. Device screen: (a) data acquisition; (b) display of the machine with communication
problems
 Specialized Software for Tracking and Management Machinery 591

After connecting to the network the module will display the information from the
application (Fig. 19a). The machine will appear online on the application control panel
(Fig. 19b).

a) b)

Fig. 19. Device screen: (a) information from the application; (b) application control panel

The device requires the operator to scan the barcode. After scanning it will display
the name of the operator working on that machine (Fig. 20a). If the operator will start
working on the machine it will switch to functioning state (Fig. 20b).

a) b)

Fig. 20. Device screen: (a) information from the application; (b) application control panel

In the event of a defect, the operator scans the defect code (Fig. 21a) and the machine
switches into status “Defective” (Fig. 21b).
592 P. Farcaș et al.

a) b)

Fig. 21. Device screen: (a) defect code; (b) machine in state “defect”

The technician’s panel will display a new record as well on technician’s mobile
phone a notice will appear. The technician can assign the defect in three ways: by phone
(Fig. 22a) from the desktop application (Fig. 22b) or by scanning own badge on the
device.

a) b)

Fig. 22. Assign the defect from: (a) the phone; (b) the desktop application

After the log on to the machine the technician has several statuses. Completion of
the intervention can be done from the mobile phone or computer (Fig. 23). The machine

a) b)

Fig. 23. Completing the intervention of the phone

 Specialized Software for Tracking and Management Machinery 593

is back in status “Unused – Neutilizata”. Similar operations can be made from the
module: Maintenance/Automaintenance/Settings.

4 Conclusions

This system increases the efficiency of maintenance and preventative maintenance,

tracking production, applying best practices in the operation of equipment (by general‐
izing the positive experiences of their operators) and streamlines the management’s
access to the production lines state from the machine level up to synthesis of the entire
division or plant in real-time. The main advantage of using the system designed by the
authors is its flexibility through the simple adaptation to various types of industrial
equipment and transformation of obsolete equipment/machines in Internet of Things
devices (IoT), thus enabling full integration of maintenance and tracking production
with reasonable investments. Thanks to intuitive programming, this system is a cost-
effective and reliable solution for automation of textile-apparel industries. The system
described can be adapted to any production profile unit in a factory, facilitating manage‐
ment and pursuing of machine tools, electrical, hydro etc. For other types of signals,
different from those presented herein, our devices must be provided with a new software
version. We are working on a more advanced version, which will be customizable
through a web interface and prepared to work on mobile platforms using secured
connections [10].

Acknowledgement. Thanks to all colleagues whom offered their unconditional support to

develop this work.

References

1. Chen, Z., Xing, M.: Upgrading of textile manufacturing based on Industry 4.0. In: 5th
International Conference on Advanced Design and Manufacturing Engineering (ICADME),
pp. 2143–2146. Atlantis Press (2015)
2. David, E., et al., It Threat Evolution in Q2 2015, Kaspersky Lab HQ, Moscow, 125212 (2015).
(Jain, P.: Outsourcing Business Model of Indian Textile Traders from Supply Chain
Management’s perspective, Thesis submitted in partial fulfillment of the requirements for the
award of the degree of master of production engineering in the faculty of Engineering &
Technology, Jadavpur University (2014). http://dspace.jdvu.ac.in/bitstream/123456789/
29488/1/Acc.%20No.%20DC%201832.pdf)
3. Bharatia, N., Thakur, N.D.: Design and implementation of labview based SCADA
for textile mills. Int. J. Res. Eng. Technol. 3(5), 773–779 (2014)
4. Carvalho, H., Silva, L.F., Rocha, A., Monteiro, J.: Automatic presser-foot force control for
industrial sewing machines. Int. J. Cloth. Sci. Technol. 24(1), 36–55 (2012)
5. Linke, M., Greb, C., Klingele, J., Schnabel, A.: Automating Textile Preforming Technology
for Mass Production of Fibre-reinforced Polymer (FRP) Composites. Woodhead Publishing
Limited, Cambridge (2012)
6. Podina, C., Popescu, M., Naaji, A.: Integrator for Android Applications, Recent Advances in
Computer Science, pp. 221–226, 16–19 July 2013
594 P. Farcaș et al.

7. Spatariu, E., Popescu, M.C., Naaji, A.: Intelligent system with remote transmission for
monitoring telecommunication equipment parameters. In: IEEE 21st International
Symposium for Design and Technology in Electronic Packaging (SIITME), Brașov, 22–25
October 2015
8. The Internet of Things: An Overview (2015). http://www.internetsociety.org/doc/iot-
overview
9. Software Innovations (2015). https://www.bosch-si.com/internet-of-things/iot-downloads/
iot-information-material.html
10. Bucerzan, D., Raţiu, C., Manolescu, M.J.: SmartSteg: a new android based steganography
application. Int. J. Comput. Commun. Control (IJCCC), 8(5). ISSN:1841-9836 (2013)
 Treatment Optimization for Bioscouring
in Ultrasound of Cotton/Cottonised Hemp
Blended Fabrics

Pustianu Monica1,2, Dochia Mihaela2, Sîrghie Cecilia1,2(&),

Gavrilaș Simona2,3, and Tomescu Daniel2
1
Faculty of Engineering, “Aurel Vlaicu” University of Arad, Arad, Romania
[email protected], [email protected]
2
Research Development Innovation in Technical and Natural Science Institute,
“Aurel Vlaicu” University of Arad, Arad, Romania
[email protected], [email protected],
[email protected], [email protected]
3
Faculty of Food Engineering, Tourism and Environmental Protection,
“Aurel Vlaicu” University of Arad, Arad, Romania

Abstract. The optimization of the bio-scouring treatment in ultrasounds on

50% cotton +50% cottonised hemp fabrics were studied in the present study.
A commercial product Beisol PRO (which is a mixtures of enzymes used for
pre-treatment of cellulose with a very good hydrophilic effect) were used for
cleaning the materials. In order to establish the optimal treatment parameters a
central compound rotatable program with two independent variables: enzyme
concentration and treatment time, were used. The weight loss and hydrophilic
property of the treated samples fabric were considered as dependent variables in
order to establishing the optimum parameters of bio-scouring treatment in
ultrasounds.

Keywords: Optimization
Bio-scouring treatment
Ultrasound

Cotton/cottonised hemp fabrics
Enzymes
Weight loss

1 Introduction

The bio-scouring procedure tends to be used more often due to its advantages com-
pared with the alkaline one. The reaction conditions are milder: temperature 50–60 °C,
pH: 8–9, 5, compared with those used in the classical alkaline procedure: 90–100 °C,
pH = 13–14. Also, this procedure is in accordance with the global trends of environ-
mental protection, the waste waters being less polluting.
One problem in using the bio-scouring treatment seems to be the fact that it is
difficult to obtain optimum result in the case of the light-colored shades dyed textile
products [1].
In the past years, the good results obtained after the application of the bio-scouring
treatment opened the way for the process optimization. Many researches presented the
results obtained after using different enzymatic products. Enzymes such as pectinases,
cellulases, proteases, cutinases, xylanases and lipases were used for bioscouring
© Springer International Publishing AG 2018
V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_44
596 P. Monica et al.

treatment, but pectinases seems to be the most effective for scouring of cellulosic
fabrics [2, 14].
A xylano-pectinolytic system showed an important synergetic effect [3–5] its
activity presented important influence on the bio-scouring process. It is considered that
an addition of some Ca2+ could improve the thermal stability and the activity of the
enzyme [10].
Good results in the bio-scouring process could also be obtained if the reactions that
are taking place in an ultra sound media [7]. It seams that also the type of ultrasound
power used influences the bio-scouring process [6, 10].
Most of the enzymatic commercial products are designed for a special applications
on a proper surface (cellulosic, ligno-, proteic, etc.) but no so many studies [12] report
their activity on a blended fabrics such: celluloses + lignocelluloses.
In this case, due to their chemical composition (higher percent of lignin) the sub-
strate could have an unexpected behaviour in bio-scouring process especially when the
ultrasound are present.
All processes conducted during the fundamental research must pass through an
optimization step in which the results obtained are analyzed and the best condition to
apply in industry are established. A modeled process is based on a reliable statistical
operation and will lead to an optimum pattern for the reaction conditions to apply. In
this way for the bio-scouring was determined the best model for the time enzyme
quantity dependence [2, 8, 9, 12, 13, 15, 16].

2 Experimental Part

This study presents the optimization of bio-scouring treatment in ultrasound applied on

50% of cotton + 50% of cottonised hemp woven fabrics. 13 samples of fabrics were
used for the experimental model.
The fabric was a plain woven with the following characteristics: width
120 ± 3 cm, weight 220 ± 10 g/m2, warp sett 100 ± 10 fibers/10 cm, weft sett
100 ± 10 fibers/10 cm. For bioscouring process the commercial product Beisol PRO
in buffer solution of 0.1 M sodium phosphate/disodium phosphate (pH 8) were used. In
order to have a synergetic effect between enzymes and ultrasounds the treatments were
performed in an Ultrasonic Bath Elmasonic X-tra Basic 250 with a frequency of
45/25 kHz.
Sodium citrate was used as a complexing agent and a nonionic detergent Denimcol
Wash RGN supplied by CHT Bezema [11]. All the experiments were carried out at
55 °C at a 10:1 ratio liquid to fabrics but with variable enzyme concentration (between
1–3%) and variable treatment time (between 15–55 min).
Beisol PRO is a mixture of enzymes used for pre-treatment of cellulose materials
and their mixtures which gives a very good hydrophilic effect. The hydrophilic effect
was appreciated by the weight loss of the treated samples fabrics and the hydrophilic
property according AATCC/ASTM Test Method TS-018 by using an analytical bal-
ance Kern ABT 220–4 M for measuring the fabric weight after conditioning in an oven
Caloris ECV 500.
 Treatment Optimization for Bioscouring 597

The treatment optimization was made by using a central, rotatable second order
compound program with two independent variables: enzyme concentration and treat-
ment time [11]. The weight loss of the treated samples fabrics (%) and the hydrophilic
property (time in seconds) were considered as a responses variables and used for
establishing the optimum parameters for bio-scouring treatment in ultrasound condi-
tions. Therefore, the variation limits of the independent variables are presented in
Table 1.

Table 1. The variation limits of independent variables

Real value Value. code
−1,414 −1 0 1 +1,414
x - enzyme concentration (% compared to fibre) 1 1,3 2 2,7 3
y - time (minutes) 15 21 35 49 55

The experimental matrix and the measured values of the response functions are
shown in Table 2:

Table 2. Experimental matrix and the measured values of the response function
No. Independent variables Answers
x y Y
x x (real) y y (real) (Y1) (Y2)
(cod.) Enzyme (cod.) Time Weight Hydrophilicity
concentration (minutes) loss (%) (seconds)
(%)
1 −1 1.30 −1 21.00 0.14 0.58
2 1 2.70 −1 21.00 0.15 0.65
3 −1 1.30 1 49.00 0.09 0.33
4 1 2.70 1 49.00 0.04 0.75
5 −1.414 1.00 0 35.00 0.19 0.85
6 1.414 3.00 0 35.00 0.26 0.58
7 0 2.00 −1.414 15.00 0.31 0.64
8 0 2.00 1.414 55.00 0.04 1.05
9 0 2.00 0 35.00 0.14 0.57
10 0 2.00 0 35.00 0.19 0.75
11 0 2.00 0 35.00 0.35 0.66
12 0 2.00 0 35.00 0.49 0.52
13 0 2.00 0 35.00 0.40 0.79
598 P. Monica et al.

2.1 Interpretation of the Mathematical Model Obtained

The two chosen independent variables were:
x - enzyme concentration [%]; y - time (minutes).
The weight loss (denoted by Y1) and the hydrophilicity (denoted by Y2) were used
as goal-functions. The interval of variation of the independent variables was established
based on Beisol producer recommendations (for cotton fabrics) as follow: enzyme
concentration between 1–3% and for treatment time between 15–55 min.

A. The weight loss study (Y1).

The second order central compound rotatable program has the following mathematical
expression:

Y ¼ b0 þ b1 x þ b2 y þ b12 xy þ b11 x2 þ b22 y2 ð1Þ

For analysing the experimental data and to obtaining the regression equations, the
Mathcad Professional and Excel programs were used [2, 6–8]. The significance of the
regression equation coefficients was tested using Student test with critical table value
for the test ta,m = t0,05;6 = 2,132.
The regression equation obtained after eliminating insignificant coefficients is:

Y1ðthe weight lossÞ ¼ 0:314 þ 0:007x þ ð0:067Þy

þ ð0:068Þx2 þ ð0:093Þ y2 þ ð0:015Þxy ð2Þ

The obtained models scans are viewed geometric as hyper-surfaces in three-

dimensional space of independent variables. The hyper-surface represents the response
of the model; because the extreme points (maximum, minimum) of the hyper-surfaces
present technological interest their exact location is searched or at least knowledge
about the shape of the surface in the extreme field neighbouring. The surface can be cut
by planes of type y = ct, resulting the response contours.

Fig. 1. The dependence of the goal-function on the independent variables: (x – enzyme

concentration and y – treatment time)
 Treatment Optimization for Bioscouring 599

Figure 1 presents the plot which shows the dependence of the goal-function (Y1) on
the two independent variables (x and y). The response surface of the regression
equation is an elliptical plan.
The constant level curves obtained by cutting the response surface with constant
level plans (Fig. 2) allows the evaluation of the dependent variable (Y1) for various
values of weight loss between 0.04 to 0.49, according to the conditions imposed by the
independent variables x and y.

Fig. 2. Contour curves for various values of Y (weight loss)

On the elliptical type surface response from Fig. 2 it can be observed that the centre
is elongated in the direction of x axis because the b11 and b22 coefficients are negative
and b11 is lower than b22. The dependence sows a stationary point which is a point of
maximum for the coordinates x = 0.092 and y = −0.368. These coded values are
associated with an enzyme concentration of 2.5% and a treatment time of 29 min.
The value of the goal function at this point is Y1 = 0.327.

f(x,y) for x=constant

0.326

g( x , y)

g( x1 , y)
weight loss [%]

g( x2 , y)

g( x3 , y)

g( x4 , y)

− 0.12
2 1 0 1 2
− 1.414 y 1.406
time [min]
x=0
x1=-1.414
x2=-1
x3=1
x4=1.414

Fig. 3. The dependence of the goal function on all significant values of y parameters for
x = constant
600 P. Monica et al.

Figure 3 shows the dependence of the goal function on one of the two independent
variables for all significant values of the y parameters, given that the second inde-
pendent variable x is constant. It can be observed how, for a constant value of enzyme
concentration, the graphic representing the variation of weight loss versus time indi-
cates the existence of a maximum point nearly to 29 min, which indicates a strong
influence of this parameter on the weight loss.

f(x,y) for y=constant

0.314

0.34
f ( x , y) 0.28

f ( x , y1) 0.22
weight loss [%}

f ( x , y2) 0.16
0.1
f ( x , y3)
0.04
f ( x , y4) 0.02

0.08

− 0.121 0.14
0.2
2 1.6 1.2 0.8 0.4 0 0.4 0.8 1.2 1.6 2
− 1.414 x 1.406
concentration of enzyme [%]
y=0
y=-1.414
y=-1
y=1
y=1.414

Fig. 4. The dependence of the goal function on all significant values of x parameters for
y = constant

Figure 4 shows the dependence of the goal function on one of the two independent
variables for all significant values of the x parameters, given that the second inde-
pendent variable y is constant.
It can be observed how, for a constant value of treatment time - y, the graphic
representing the variation of weight loss versus enzyme concentration indicates the
existence of a maximum point close to 2.5%, which indicates a influence of this
parameter on the weight loss.
All the results confirm the dependence of the weight loss on the enzyme concen-
tration and treatment time, the best results was recorded at 2.5% of enzyme concen-
tration and 29 min of treatment time.
The strong influence of the time in this process could be explained by the presence
of ultrasound which enhances the effect of the enzymes on the substrates. This effect is
more outlined especially when the lignocelluloses fibres are present into the substrate.
 Treatment Optimization for Bioscouring 601

B. Hydrophilic property (Y2) study.

The regression equation obtained after eliminating insignificant coefficients is:

Y2ðHydrophilyÞ ¼ 0:658 þ 0:013x þ 0:053y þ ð0:021Þx2 þ 0:043y2 þ 0:087xy ð3Þ

Figure 5 presents the plot which shows the dependence of the goal-function (Y2) on
the two independent variables (x and y). The response surface of the regression
equation is an elliptical plane.

Fig. 5. The dependence of the goal-function Y2 on the independent variables: (x – enzyme

concentration and y – treatment time)

The constant level curves will be obtained by cutting the response surface with
constant level plans which are presented in Fig. 6 allows the evaluation of the
dependent variable (Y2) for various values of the hydrophilicity between 0.33 to 1.05,
according to the conditions imposed by the independent variables x and y.
The responses surface from Fig. 6 is hyperbolic type which means that are two
distinct regions of maximum and minimum. The coefficients b11 and b22 are different,

Fig. 6. Contour curves for various values of Y2 (hydrophilicity)

602 P. Monica et al.

b11 is negative and b22 is positive which is translating into the fact that interdepen-
dency between the independent variables is present.
In Fig. 7 is shown the dependence of the goal-function on all significant values of y
parameters when the enzymes concentrations is constant. As well, in Fig. 8 is shown
the dependence of the goal-function on all significant values of y parameters when the
time is constant. It can be observed that for a constant value of the enzymes concen-
tration the graphic representing the variation of the hydrophilicity indicates the exis-
tence of a minimum point close to 29 min which means a strong influence of this
parameter on the hydrophilicity.

f(x,y) for x=constant

0.967

g (x y, )

g (x1 y, )
hydrophlic [s]

g (x2 y, )

g (x3 y, )

g (x4 y, )

0.471

2 1 0 1 2
− 1.414 y 1.406
time [min]
x=0
x1=-1.414
x2=-1
x3=1
x4=1.414

Fig. 7. The dependence of the goal-function on all significant values of y parameters for
x = constant

f(x,y) for y=constant

0.969 1
0.94
f ( x , y) 0.88

f ( x , y1) 0.82
hydrophilic [s]

f ( x , y2) 0.76
0.7
f ( x , y3)
0.64
f ( x , y4) 0.58

0.52
0.46
0.473
0.4
2 1.6 1.2 0.8 0.4 0 0.4 0.8 1.2 1.6 2
− 1.414 x 1.406
concentration of enzyme [%]
y=0
y=-1.414
y=-1
y=1
y=1.414

Fig. 8. The dependence of the goal-function on all significant values of x parameters for
y = constant
 Treatment Optimization for Bioscouring 603

The optimal values for the independent variables in order to obtain a better
hidrophilicity are: (a) Enzyme concentrations = 1.7%; (b) Treatment time = 29 min.
For these coordinates the value of the function Y2 is 0.648.

3 Conclusions

The bio-scouring process, of the fabrics with lignocelluloses content, in ultrasound

conditions is strongly influenced by the treatment time and secondly by the enzyme
concentration. The obtained results confirm other studies [12] concerning the removal
of non-celluloses compounds by usage of ultrasound in enzymatic treatments in view
of reducing treatment time and/or enzymes concentrations. The optimal treatments
parameters according to Beisol producer are: 2.5–2.7% enzymes for a time of 35–
40 min on cotton fabrics.
The present study proved once more that, the usage of ultrasounds could reduce the
treatment time as well as the enzyme concentrations even if the percentage of ligno-
celluloses fibres into the mixtures (higher amount of non-celluloses content into the
fabrics) is considerable. The bio-scouring process is proper done when a very good
hydrophilic effect is obtained (minimum time for water absorbtion into the fabric and
optimal weight loss). According with the present research results we conclude that the
optimal recipies for the usage of Beisol PRO for the bioscouring of the blended fabrics
(celluloses + lignocelluloses) în ultrasound conditions is between 1.7–2.5% enzymes
concentrations and the treatment time between 29–30 min.

Acknowledgement. This work was supported by a grant of the Romanian National Authority
for Scientific Research and Innovation, CNCS – UEFISCDI, project number
PN-II-RU-TE-2014-4-1370 and by the CESTN (Centru de Cercetare în Științe Tehnice și Nat-
urale) project, no. 621/2014 POS-CCE, ID/SMIS -1826/48720 - supported by Structural Founds
“Investment in Your Future”.

References
1. Rocky, A.M.K.B.P.: Comparison of effectiveness between conventional scouring &
bio-scouring on cotton fabrics. IJSER 3(8) (2012). http://www.ijser.org
2. Wang, Q., Fan, X.-R., Huab, Z.-Z., Chen, J.: Optimizing bioscouring condition of cotton
knitted fabrics with an alkaline pectinase from Bacillus subtilis WSHB04-02 by using
response surface methodology. Biochem. Eng. J. 34, 107–113 (2007)
3. Singh, A., Kaur, A., Patra, A.K., Mahajan, R.: Xylano-pectinolytic synergism from a
bacterial isolate for ecofriendly processing of jute fabric. J. Clean. Prod. 124, 120–125
(2016)
4. Shi, A.Q., Hu, H., Zheng, F., Long, L.K., Ding, S.J.: Biochemical characteristics of an
alkaline pectate lyase PelA from Volvariella volvacea: roles of the highly conserved
N-glycosylation site in its secret ion and activity. Appl. Microb. Biotechnol. 99(8), 3447–
3458 (2015)
604 P. Monica et al.

5. Wang, H.L., Li, X.M., Ma, Y.H., Song, J.N.: Process optimization of high-level extracellular
production of alkaline pectate lyase in recombinant Escherichia coli BL21 (DE3). Biochem.
Eng. J. 93, 38–46 (2015)
6. Easson, M.W., Wojkowski, S., Condon, B., Yeater, K., Slopek, R., Fortier, C.:
Ultrasound-enhanced bioscouring of greige cotton: regression analysis of process factors.
AATCC J. Res. 2(1), 16–23 (2015)
7. Eren, H.A., Erismis, B.: Ultrasound-assisted bioscouring of cotton. Color. Technol. 129(5),
360–366 (2013)
8. Xu, Y.H., Xu, D.H., Zhang, L.P., Zhou, X.H.: A new inverse problem for the determination
of textile fabrics thickness. Inverse Probl. Sci. Eng. 23(4), 635–650 (2015)
9. Istoc, I.V., Pustianu, M., Bucevschi, A., Dochia, M., Sîrghie, C.: Study regarding the
optimization of the bioscouring treatment on 100% cotton materials. In: Proceeding of the
2nd International Conference on Energy and Environment Technologies and Equipment
(EEETE 2013). Advanced in environmental technologies, agriculture, food and animal
science, vol. 10, Brasov, Romania (2013)
10. Abdel-Halim, E.S., Fahmy, H.M., Moustafa, M., Fouda, G.: Bioscouring of linen fabric in
comparison with conventional chemical treatment. Carbohydr. Polym. 74, 707–711 (2008)
11. Catalog DyStar Textilfarben GmbH (2007)
12. Pustianu, M., Dochia, M., Pernevan, S., Sîrghie, C.: Study regarding the influence of the
bioscouring treatment in ultrasound on 60% cotton +40% hemp materials, Part 1: study
regarding the optimization of the bioscouring treatment. In: Annals of the University of
Oradea, Fascicle of Textiles, Leatherwork, vol. 15(1), pp. 93–98 (2014)
13. Popescu, A., Grigoriu, A., Zaharia, C., Mureşan, R., Mureşan, A.: Mathematical modeling
and the technological process optimization for the bio-scouring of the cotton textile
materials. Indus. Textilă 61, 70–80 (2010)
14. Butnaru, R., Stoichitescu, L.: Special finishing procedures of textiles fabrics. Editura Gh.
Asachi, Iasi (1995)
15. Mihail, R.: Introduction to the Strategy Experimenting with Applications in Chemical
Technology. Scientific and Encyclopedic, Bucharest (1976)
16. Akhnazarova, S., Kafarov, V.: Experiment Optimization in Chemistry and Chemical
Engineering. MIR Publishing, Moscow (1982)
 Fuzzy Method Applied on the Control
of Textiles Biotechnologies

Bălaș Marius1,2 ✉ , Sîrghie Cecilia1,2, Pustianu Monica1,2, Dochia Mihaela2,

( )

Gavrilaș Simona2,3, and Tomescu Daniel2

1
Faculty of Engineering, Aurel Vlaicu University of Arad, Arad, Romania
[email protected], [email protected],
[email protected]
2
Research Development Innovation in Technical and Natural Science Institute,
Aurel Vlaicu University of Arad, Arad, Romania
[email protected], [email protected],
[email protected]
3
Faculty of Food Engineering, Tourism and Environmental Protection,
Aurel Vlaicu University of Arad, Arad, Romania

Abstract. This paper is dealing with the optimization of the bio-scouring process
in ultrasound on cotton & cottonised hemp fabrics. The tests were performed
using an experimental program designed after a central, rotatable second order
compound program with two independent variables CRSCP2IV. The mathemat‐
ical models of the bio-scouring processes are needed for the optimization of the
process parameters. Special knowledge is needed to operate such programs. That
is why we also used a simpler yet effective software tool for modeling and decision
making, including process optimization, the Fuzzy Inference System (FIS) from
Matlab environment. The experimental data processed according to CRSCP2IV
in Mathcad were processed in parallel with FIS in Matlab, which has high
compatibility with all the important data processing and automate control prod‐
ucts and offers numerous dedicated tools.

Keywords: Fuzzy logic · Bio-scouring treatment · Ultrasound · Cotton/

cottonised hemp fabrics · Enzymes · Weight loss · Hidrophily

1 Introduction

In the past 20 years the applied biotechnologies in the textiles industry proved an envi‐
ronmental friendly alternative to classical technologies (chemical) for many practical
applications such as: scouring, desizing, cottonisation, bleaching, etc.
Although the benefits of these technologies have been extensively exhibited in
specialty papers [1, 3, 4, 10, 14] they are not widely applied in industry. One reason
could be the lack of control tools of these processes when they are applied to industrial
scale.

© Springer International Publishing AG 2018

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9_45
606 B. Marius et al.

Regarding only the bio-scouring process, the specialty literature is able to provide
many studies regarding process optimization in terms of optimal parameters [2, 5, 8, 10,
12, 13, 15, 16] for the best results, but no paper for describing a process controller
because all studies were focused on finding optimal value of the process parameters by
mathematical modeling.
On the other side, the fuzzy logic methods are currently applied for designing adap‐
tive industrial controllers; some versions (Takagi-Sugeno and fuzzy-interpolative) are
particularly effective in applications [18].
The usage of fuzzy method instead a laborious CRSCP2IV for optimizing the bio-
scouring process could be a real benefit for increasing the industrial process automation
as well as for the development of commercial enzymatic products designed for many
industrial applications.
While CRSCP2IV has the advantages of better approximation of the optimal treat‐
ment parameters using less experimental points, FIS opens the possibility of designing
process controllers for industrial process automation even when the number of process
parameters increases. In our case we used two parameters enzyme concentration and
time, but working with FIS we intend to take in account as fast as possible a third
parameter: the temperature.

2 Experimental Part

This study presents the possibility of optimizing the bio-scouring treatment in ultra‐
sound, applied on 50% of cotton + 50% of cottonised hemp woven fabrics, using a fuzzy
method (FIS) when the experimental program were designed according with a central,
rotatable second order compound program with two independent variables – CRSCP2IV
based on:
(a) Two independent variables: x1 = enzyme concentration and x2 = treatment time;
(b) Two dependent variables: y1 = the weight loss and y2 = hydrophilic effect.
The range in which the experiments were carried out is:
(1) Enzymes concentrations from 1% to 3% as follows: 1%, 1.3%, 2%, 2,7% and 3%;
(2) Treatment time from 15 min to 55 min as follows: 15 min, 21 min, 35 min, 49 min
and 55 min.
In this study we considered the same process variables that vary in the same steps
(as those requested by CRSCP2IV program) like in Table 1, so we can compare the
results.

Table 1. Experimental program

Variables Values
x1 - Enzyme concentration (% compared to fibre) 1 1.3 2 2.7 3
x2 - Time (minutes) 15 21 35 49 55
 Fuzzy Method Applied on the Control of Textiles Biotechnologies 607

The experimental matrix and the measured values of the dependent variables are
shown in Table 2.

Table 2. Experimental matrix (independent variables to be optimized and the answerers).

x1 (real) enzyme x2 (real) time (y1) weight loss [%] (y2) hidrophily
concentration [%] [minutes] (seconds)
1 1.30 21.00 0.14 0.58
2 2.70 21.00 0.15 0.65
3 1.30 49.00 0.09 0.33
4 2.70 49.00 0.04 0.75
5 1.00 35.00 0.19 0.85
6 3.00 35.00 0.26 0.58
7 2.00 15.00 0.31 0.64
8 2.00 55.00 0.04 1.05
9 2.00 35.00 0.31 0.66

The bio-scouring process was conducted as follow: 9 samples of fabrics (woven

fabrics made from 50% of cotton + 50% of cottonised hemp) with the following char‐
acteristics: width 120 ± 3 cm, weight 220 ± 10 g/m2, warp sett 100 ± 10 fibers/10 cm,
weft sett 100 ± 10 fibers/10 cm was treated in ultrasound conditions (Ultrasonic Bath
Elmasonic X-tra Basic 250) at 45/25 kHz with a commercial product Beisol PRO in
buffer solution of 0.1 M sodium phosphate/disodium phosphate (pH 8).
Some auxiliaries as sodium citrate were used as a complexing agent and a non-ionic
detergent Denimcol Wash RGN supplied by CHT Bezema [11]. All the experiments
were carried out at 55 °C at a 10:1 ratio liquid to fabrics but with variable enzyme
concentration (between 1–3%) and variable treatment time (between 15–55 min).
Beisol PRO is a mixture of enzymes used for pre-treatment of cellulose materials
with a very good hydrophilic effect. The hydrophilic effect was appreciated by the weight
loss of the treated samples fabrics and the hydrophilic property using AATCC/ASTM
Test Method TS-018 by counting the time of water absorbtion into the fabric. For meas‐
uring the weight loss samples were first conditioned in oven Caloris model ECV 500
till constant mass and then weighed using an analytical balance Kern model ABT
220-4 M. When the non-celluloses compounds are properly removed the hydrophilicity
of the fabric is better which means that short time for water absorption into the fabric is
register while the samples weight is decreasing (weight loss increasing).
Therefore, the optimal treatment parameters correspond to the point where the
weight loss is maximal and the time for water absorption into the fabric surface is
minimal.
For optimizing the bio-scouring process the experimental data were processed with
Matlab. We used FIS (Fuzzy Inference System) when plotting the 3D graphs of the
numeric model of the process, and we intend to further use this model in Matlab-Simu‐
link simulations, aiming to test different control algorithms.
608 B. Marius et al.

FIS was chosen because of its capability to merge numeric and linguistic represented
knowledge. Each experimental measuring point is modeled by an inference rule. All
linguistic labels -input and output- are triangular (a Mamdani controller).
In order to obtain a very transparent model, easy to adjust by researchers which are
not specialized in IT, the linguistic labels are named after their corresponding numerical
values.
The fuzzification of the two input variables -enzyme concentration and treatment
time- and of the two output variables -weight loss and the hydrophilicity- is the following
(Figs. 1 and 2).

Fig. 1. The fuzzification

Fig. 2. The inferences

The weight loss and the hydrophilicity inferences are simply reproducing the exper‐
imental data of Table 2.
The resulting surfaces are the following (Fig. 3).
Fuzzy Method Applied on the Control of Textiles Biotechnologies 609

a) The Weight Loss

b) the hydrophilicity

Fig. 3. The surfaces

610 B. Marius et al.

3 Discussions

(a) According with the 3D surface of the weight loss we could conclude that the area
for which the process parameters could lead to obtaining optimal values for the
weight loss has the following coordinates: enzyme concentration between 1.6–2.4%
and treatment time between 27–37 min.
(b) According with the 3D surface of the hydrophilicity we could conclude that the
area for which the process parameters could lead to obtaining a better hydrophilic
effect has the following coordinates: enzyme concentration between 1.4–2.1% and
treatment time between 30–45 min.
Considering the intersection of both optimal surfaces we could estimate that the bio-
scouring process of the blended fabrics in ultrasound conditions is proper done when
the enzymes concentrations varies between 1.6–2.1% and the treatment time between
30–37 min.
These coordinates are very closely to those obtained using the CRSCP2IV program
for optimization of the bio-scouring process applied on blended fabrics in ultrasound
condition (enzymes concentration 1.7–2.5% and treatment time 29–30 min) reported in
reference [17].

4 Conclusions

Fuzzy method using FIS is most handle alternative for optimizing a process comparing
with CRSCP2IV program. Moreover, could be used further for testing different control
algorithms by simulation in Matlab-Simulink for industrial process automation. In the
same time the number of variables (parameters) could be increased taking into account
other needs.

Acknowledgements. This work was supported by a grant of the Romanian National Authority
for Scientific Research and Innovation, CNCS – UEFISCDI, project number PN-II-RU-
TE-2014-4-1370 and by CESTN (Centru de Cercetare în Științe Tehnice și Naturale) project, no.
621/2014 POS-CCE, ID/SMIS -1826/48720- supported by Structural Founds “Investment in Your
Future”.

References

1. Bahrum Prang Rocky, A.M.K.: Comparison of effectiveness between conventional scouring

& bio-scouring on cotton fabrics. IJSER 3(8) (2012). http://www.ijser.org
2. Wang, Q., Fan, X.R., Huab, Z., Chen, J.: Optimizing bioscouring condition of cotton knitted
fabrics with an alkaline pectinase from Bacillus subtilis WSHB04-02 by using response
surface methodology. Biochem. Eng. J. 34, 107–113 (2007)
3. Singh, A., Kaur, A., Patra, A.K., Mahajan, R.: Xylano-pectinolytic synergism from a bacterial
isolate for ecofriendly processing of jute fabric. J. Clean. Prod. 124, 120–125 (2016)
 Fuzzy Method Applied on the Control of Textiles Biotechnologies 611

4. Shi, A.Q., Hu, H., Zheng, F., Long, L.K., Ding, S.J.: Biochemical characteristics of an alkaline
pectate lyase PelA from Volvariella volvacea: roles of the highly conserved N-glycosylation
site in its secret ion and activity. Appl. Microb. Biotech. 99(8), 3447–3458 (2015)
5. Wang, H.L., Li, X.M., Ma, Y.H., Song, J.N.: Process optimization of high-level extracellular
production of alkaline pectate lyase in recombinant Escherichia coli BL21 (DE3). Biochem.
Eng. J. 93, 38–46 (2015)
6. Easson, M.W., Wojkowski, S., Condon, B., Yeater, K., Slopek, R., Fortier, C.: Ultrasound-
enhanced bioscouring of greige cotton: regression analysis of process factors. AATCC J. Res.
2(1), 16–23 (2015)
7. Eren, H.A., Erismis, B.: Ultrasound-assisted bioscouring of cotton. Color. Technol. 129(5),
360–366 (2013)
8. Xu, Y.H., Xu, D.H., Zhang, L.P., Zhou, X.H.: A new inverse problem for the determination
of textile fabrics thickness. Inverse Probl. Sci. En. 23(4), 635–650 (2015)
9. Istoc, I.V., Pustianu, M., Bucevschi, A., Dochia, M., Sîrghie, C.: Study regarding the
optimization of the bioscouring treatment on 100% cotton materials. In: Advanced in
Environmental Technologies, Agriculture, Food and Animal Science. Proceedings of the 2nd
International Conference on Energy and Environment Technologies and Equipment (EEETE
2013), series 10, Brasov, Romania (2013)
10. Abdel-Halim, E.S., Fahmy, H.M., Moustafa, M., Fouda, G.: Bioscouring of linen fabric in
comparison with conventional chemical treatment. Carbohyd. Polym. 74, 707–711 (2008)
11. Catalog DyStar Textilfarben GmbH (2007)
12. Pustianu, M., Dochia, M., Pernevan, S., Sîrghie, C.: Study regarding the influence of the
bioscouring treatment in ultrasound on 60% cotton + 40% hemp materials, Part 1: study
regarding the optimization of the bioscouring treatment. Ann. Univ. Oradea Fasc. Text
Leatherwork XV(1), 93–98 (2014)
13. Popescu, A., Grigoriu, A., Zaharia, C., Mureşan, R., Mureşan, A.: Mathematical modeling
and the technological process optimization for the bio-scouring of the cotton textile materials.
Industria Textilă 61, 70–80 (2010)
14. Butnaru, R., Stoichitescu, L.: Special finishing procedures of textiles fabrics. Editura Gh.
Asachi, Iasi (1995)
15. Mihail, R.: Introduction to the strategy experimenting with applications in chemical
technology, Ed. Scientific and Encyclopedic, Bucharest (1976)
16. Akhnazarova, S., Kafarov, V.: Experiment Optimization in Chemistry and Chemical
Engineering. MIR Publishers, Moscow (1982)
17. Pustianu, M., Dochia, M., Sîrghie, C., Gavrilaș, S., Tomescu, D.: Treatment optimization for
bioscouring in ultrasound of cotton/cottonised hemp blended fabrics. In: Proceedings of SOFA
(2016, in Press)
18. Balas, M.M.: The fuzzy interpolative methodology. In: Balas, V.E., Varkonyi-Koczy, A.,
Fodor, J. (eds.) Soft Computing Based Modeling in Intelligent Systems. Studies in
Computational Intelligence, vol. 196, pp. 145–167. Springer, Berlin (2009)
Author Index

A E
Al-Khannak, Rafid, 395 Ebadi, Hamed, 96
Anca, Popescu, 565
Andone, Diana, 218, 225 F
Andrei, M., 265 Farcaș, Pavel, 581
Asmis, A., 117 Filip, Ioan, 253
Azar, Ahmad Taher, 87 Florin, Sandru, 565

B G
Badea, Andra, 194, 201 Ghiormez, Loredana, 253
Balas, Marius M., 278
Balas, Valentina Emilia, 96, 286, 314, 330, 445 H
Barna, Cornel, 140, 153, 162 Hosseinabadi, Ali Asghar Rahmani, 96
Bejinariu, Silviu-Ioan, 498
Bel, W., 117 I
Bocicor, Iuliana, 54 Ignat, Anca, 472, 483
Bold, Nicolae, 28 Ionescu, Vlad-Sebastian, 71
Bucerzan, Dominic, 517, 529, 557 Iovanov, Vlad, 557
Buzgau, Mihai, 557
K
C Kardgar, Maryam, 96
Cecilia, Sîrghie, 595, 605 Kaul, Deeksha, 153
Ciobanu, Adrian, 498
Ciocarlie, Horia, 177 L
Ciuclea, Camelia, 218, 225 Lile, Ramona, 278
Coroiu, Adriana Mihaela, 3 Lochan, Kshetrimayum, 365
Crăciun, Mihaela, 235, 540 Luca, Mihaela, 498
Csorba, Luiela Magdalena, 235, 540 Luca, Ramona, 498
Czibula, Gabriela, 54, 71
M
D Mansilla, D., 117
Daniel, Tomescu, 595, 605 Marius, Bălaș, 605
Das, Hemjyoti, 297 Matei, Oliviu, 87
De Luise, D. López, 117 Mazandarani, Mehran, 314, 330
Dey, Rajeeb, 286, 297, 354, 365 Mazumdar, Aniket Samrat, 297
Dima, Vlad, 472 Mehta, Suhasini, 153
Draghici, Anca, 225 Mihaela, Dochia, 595, 605

© Springer International Publishing AG 2018 613

V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent
Systems and Computing 634, DOI 10.1007/978-3-319-62524-9
614 Author Index

Mircea, Ioan-Gabriel, 54 R
Misra, Anusha, 153 Ramadan, Rabie A., 87
Monica, Pustianu, 595, 605 Raţiu, Crina, 517
Muller, Emanuela V., 278 Romeo, Negrea, 565
Rotaru, Florin, 498
Roy, Binoy Krishna, 354, 365
N
Roy, Lintu, 297
Naaji, Antoanela, 17
Roy, Reetika, 153, 162
Najariyan, Marzieh, 314, 330
Roy, Sanjiban Sekhar, 140, 153, 162
Nath, Anirudh, 286
Rusu, Cristina, 472
Niazi, Muaz, 87
Nicolau, V., 265
Niţă, Cristina Diana, 498 S
Noaica, Cristina-Madalina, 411, 426 Saemi, Behzad, 96
Samui, Pijush, 162
Sebu, Maria Laura, 177
P Sergiu, Fericel, 565
Panoiu, Manuela, 253 Simona, Gavrilaș, 595, 605
Păvăloi, Ioan, 483 Singh, Jay Prakash, 354
Pavone, Mario, 87 Sinha, Abhishek, 162
Penariu, P.S., 426, 445, 460 Sorin, Nanu, 565
Pintea, Camelia-M., 87 Stoian, Camelia Daciana, 529
Popa, Călin-Adrian, 36, 45 Stroescu, V.C., 426, 460
Popa, Stefan Cosmin, 17 Subudhi, Bidyadhar, 365
Popescu Bodorin, N., 445, 460
Popescu, Alexandru Ion, 28
Popescu, Doru Anastasiu, 28 T
Popescu, Marius Constantin, 17, 581 Teletin, Mihai, 71
Pop-Moldovan, Ciprian, 581
Pratyush, Chitransh, 140 V
Prostean, Gabriela, 194, 201, 218, 225 Vasar, Cristian, 194, 201
Prostean, Octavian, 253 Vossen, Paul Hubert, 381